2023-04-27 12:19:53,976 INFO [train.py:976] (4/8) Training started 2023-04-27 12:19:53,977 INFO [train.py:986] (4/8) Device: cuda:4 2023-04-27 12:19:53,978 INFO [train.py:995] (4/8) {'frame_shift_ms': 10.0, 'allowed_excess_duration_ratio': 0.1, 'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 50, 'reset_interval': 200, 'valid_interval': 3000, 'feature_dim': 80, 'subsampling_factor': 4, 'warm_step': 2000, 'env_info': {'k2-version': '1.23.4', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'a23383c5a381713b51e9014f3f05d096f8aceec3', 'k2-git-date': 'Wed Apr 26 15:33:33 2023', 'lhotse-version': '1.14.0.dev+git.b61b917.dirty', 'torch-version': '1.13.1', 'torch-cuda-available': True, 'torch-cuda-version': '11.6', 'python-version': '3.1', 'icefall-git-branch': 'master', 'icefall-git-sha1': '45c13e9-dirty', 'icefall-git-date': 'Mon Apr 24 15:00:02 2023', 'icefall-path': '/k2-dev/yangyifan/icefall-master', 'k2-path': '/k2-dev/yangyifan/anaconda3/envs/icefall/lib/python3.10/site-packages/k2-1.23.4.dev20230427+cuda11.6.torch1.13.1-py3.10-linux-x86_64.egg/k2/__init__.py', 'lhotse-path': '/k2-dev/yangyifan/anaconda3/envs/icefall/lib/python3.10/site-packages/lhotse-1.14.0.dev0+git.b61b917.dirty-py3.10.egg/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-3-0423201227-84b4557756-8lx4n', 'IP address': '10.177.6.147'}, 'world_size': 8, 'master_port': 12354, 'tensorboard': True, 'num_epochs': 30, 'start_epoch': 1, 'start_batch': 0, 'exp_dir': PosixPath('pruned_transducer_stateless7/exp_multidataset'), 'bpe_model': 'data/lang_bpe_500/bpe.model', 'base_lr': 0.05, 'lr_batches': 5000, 'lr_epochs': 3.5, 'context_size': 2, 'prune_range': 5, 'lm_scale': 0.25, 'am_scale': 0.0, 'simple_loss_scale': 0.5, 'seed': 42, 'print_diagnostics': False, 'inf_check': False, 'save_every_n': 2000, 'keep_last_k': 30, 'average_period': 200, 'use_fp16': True, 'use_multidataset': True, 'num_encoder_layers': '2,4,3,2,4', 'feedforward_dims': '1024,1024,2048,2048,1024', 'nhead': '8,8,8,8,8', 'encoder_dims': '384,384,384,384,384', 'attention_dims': '192,192,192,192,192', 'encoder_unmasked_dims': '256,256,256,256,256', 'zipformer_downsampling_factors': '1,2,4,8,2', 'cnn_module_kernels': '31,31,31,31,31', 'decoder_dim': 512, 'joiner_dim': 512, 'full_libri': True, 'manifest_dir': PosixPath('data/fbank'), 'cv_manifest_dir': PosixPath('data/en/fbank'), 'max_duration': 700, 'bucketing_sampler': True, 'num_buckets': 30, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'drop_last': True, 'return_cuts': True, 'num_workers': 2, 'enable_spec_aug': True, 'spec_aug_time_warp_factor': 80, 'enable_musan': True, 'input_strategy': 'PrecomputedFeatures', 'blank_id': 0, 'vocab_size': 500} 2023-04-27 12:19:53,978 INFO [train.py:997] (4/8) About to create model 2023-04-27 12:19:54,635 INFO [zipformer.py:178] (4/8) At encoder stack 4, which has downsampling_factor=2, we will combine the outputs of layers 1 and 3, with downsampling_factors=2 and 8. 2023-04-27 12:19:54,651 INFO [train.py:1001] (4/8) Number of model parameters: 70369391 2023-04-27 12:19:57,193 INFO [train.py:1016] (4/8) Using DDP 2023-04-27 12:19:58,263 INFO [multidataset.py:46] (4/8) About to get multidataset train cuts 2023-04-27 12:19:58,263 INFO [multidataset.py:49] (4/8) Loading LibriSpeech in lazy mode 2023-04-27 12:19:58,282 INFO [multidataset.py:65] (4/8) Loading GigaSpeech 1998 splits in lazy mode 2023-04-27 12:20:00,752 INFO [multidataset.py:72] (4/8) Loading CommonVoice in lazy mode 2023-04-27 12:20:00,755 INFO [asr_datamodule.py:230] (4/8) Enable MUSAN 2023-04-27 12:20:00,755 INFO [asr_datamodule.py:231] (4/8) About to get Musan cuts 2023-04-27 12:20:02,993 INFO [asr_datamodule.py:255] (4/8) Enable SpecAugment 2023-04-27 12:20:02,993 INFO [asr_datamodule.py:256] (4/8) Time warp factor: 80 2023-04-27 12:20:02,993 INFO [asr_datamodule.py:266] (4/8) Num frame mask: 10 2023-04-27 12:20:02,994 INFO [asr_datamodule.py:279] (4/8) About to create train dataset 2023-04-27 12:20:02,994 INFO [asr_datamodule.py:306] (4/8) Using DynamicBucketingSampler. 2023-04-27 12:20:07,492 INFO [asr_datamodule.py:321] (4/8) About to create train dataloader 2023-04-27 12:20:07,492 INFO [asr_datamodule.py:435] (4/8) About to get dev-clean cuts 2023-04-27 12:20:07,493 INFO [asr_datamodule.py:442] (4/8) About to get dev-other cuts 2023-04-27 12:20:07,494 INFO [asr_datamodule.py:352] (4/8) About to create dev dataset 2023-04-27 12:20:07,733 INFO [asr_datamodule.py:369] (4/8) About to create dev dataloader 2023-04-27 12:20:25,622 INFO [train.py:904] (4/8) Epoch 1, batch 0, loss[loss=7.528, simple_loss=6.826, pruned_loss=7.008, over 16594.00 frames. ], tot_loss[loss=7.528, simple_loss=6.826, pruned_loss=7.008, over 16594.00 frames. ], batch size: 68, lr: 2.50e-02, grad_scale: 2.0 2023-04-27 12:20:25,622 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 12:20:32,880 INFO [train.py:938] (4/8) Epoch 1, validation: loss=6.911, simple_loss=6.238, pruned_loss=6.721, over 944034.00 frames. 2023-04-27 12:20:32,881 INFO [train.py:939] (4/8) Maximum memory allocated so far is 12057MB 2023-04-27 12:20:36,284 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=5.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-27 12:20:52,225 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=23.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:20:55,582 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=91.77 vs. limit=5.0 2023-04-27 12:21:03,471 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=7.56 vs. limit=2.0 2023-04-27 12:21:17,132 INFO [train.py:904] (4/8) Epoch 1, batch 50, loss[loss=1.323, simple_loss=1.175, pruned_loss=1.324, over 16423.00 frames. ], tot_loss[loss=2.15, simple_loss=1.948, pruned_loss=1.936, over 755115.30 frames. ], batch size: 146, lr: 2.75e-02, grad_scale: 2.0 2023-04-27 12:21:21,689 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=15.11 vs. limit=2.0 2023-04-27 12:21:46,457 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:22:02,791 WARNING [train.py:894] (4/8) Grad scale is small: 0.001953125 2023-04-27 12:22:02,791 INFO [train.py:904] (4/8) Epoch 1, batch 100, loss[loss=1.099, simple_loss=0.9313, pruned_loss=1.312, over 17015.00 frames. ], tot_loss[loss=1.633, simple_loss=1.453, pruned_loss=1.613, over 1329908.50 frames. ], batch size: 41, lr: 3.00e-02, grad_scale: 0.00390625 2023-04-27 12:22:13,669 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 5.700e+01 2.326e+02 5.095e+02 1.135e+03 3.099e+06, threshold=1.019e+03, percent-clipped=0.0 2023-04-27 12:22:18,242 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=25.15 vs. limit=2.0 2023-04-27 12:22:20,962 WARNING [optim.py:388] (4/8) Scaling gradients by 0.0112030990421772, model_norm_threshold=1019.0284423828125 2023-04-27 12:22:21,068 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.24, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.974e+09, grad_sumsq = 5.165e+10, orig_rms_sq=3.822e-02 2023-04-27 12:22:43,745 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=144.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-27 12:22:46,126 WARNING [optim.py:388] (4/8) Scaling gradients by 0.0022801109589636326, model_norm_threshold=1019.0284423828125 2023-04-27 12:22:46,229 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.68, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.358e+11, grad_sumsq = 3.286e+12, orig_rms_sq=4.131e-02 2023-04-27 12:22:49,650 WARNING [optim.py:388] (4/8) Scaling gradients by 0.04246773198246956, model_norm_threshold=1019.0284423828125 2023-04-27 12:22:49,755 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.92, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=5.309e+08, grad_sumsq = 1.285e+10, orig_rms_sq=4.131e-02 2023-04-27 12:22:51,350 WARNING [optim.py:388] (4/8) Scaling gradients by 0.000716241542249918, model_norm_threshold=1019.0284423828125 2023-04-27 12:22:51,453 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.93, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.876e+12, grad_sumsq = 4.542e+13, orig_rms_sq=4.131e-02 2023-04-27 12:22:52,995 INFO [train.py:904] (4/8) Epoch 1, batch 150, loss[loss=1.07, simple_loss=0.9079, pruned_loss=1.172, over 16323.00 frames. ], tot_loss[loss=1.394, simple_loss=1.222, pruned_loss=1.443, over 1777410.44 frames. ], batch size: 165, lr: 3.25e-02, grad_scale: 0.00390625 2023-04-27 12:22:53,727 WARNING [optim.py:388] (4/8) Scaling gradients by 0.049951765686273575, model_norm_threshold=1019.0284423828125 2023-04-27 12:22:54,553 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.92, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=3.845e+08, grad_sumsq = 8.968e+09, orig_rms_sq=4.287e-02 2023-04-27 12:22:58,541 WARNING [optim.py:388] (4/8) Scaling gradients by 0.00609818659722805, model_norm_threshold=1019.0284423828125 2023-04-27 12:22:58,643 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.49, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.379e+10, grad_sumsq = 3.140e+11, orig_rms_sq=4.392e-02 2023-04-27 12:23:00,022 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=16.23 vs. limit=2.0 2023-04-27 12:23:16,870 WARNING [optim.py:388] (4/8) Scaling gradients by 0.059935860335826874, model_norm_threshold=1019.0284423828125 2023-04-27 12:23:16,970 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.63, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.808e+08, grad_sumsq = 3.916e+09, orig_rms_sq=4.617e-02 2023-04-27 12:23:28,341 WARNING [optim.py:388] (4/8) Scaling gradients by 0.060559310019016266, model_norm_threshold=1019.0284423828125 2023-04-27 12:23:28,444 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.90, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=2.556e+08, grad_sumsq = 6.221e+09, orig_rms_sq=4.108e-02 2023-04-27 12:23:30,566 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=108.00 vs. limit=5.0 2023-04-27 12:23:42,820 WARNING [train.py:894] (4/8) Grad scale is small: 0.00390625 2023-04-27 12:23:42,820 INFO [train.py:904] (4/8) Epoch 1, batch 200, loss[loss=1.005, simple_loss=0.8436, pruned_loss=1.068, over 16658.00 frames. ], tot_loss[loss=1.255, simple_loss=1.089, pruned_loss=1.315, over 2124537.63 frames. ], batch size: 57, lr: 3.50e-02, grad_scale: 0.0078125 2023-04-27 12:23:50,503 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([6.1196, 6.1168, 6.0974, 6.0825, 6.0945, 6.1080, 6.1198, 6.1034], device='cuda:4'), covar=tensor([0.0057, 0.0028, 0.0123, 0.0072, 0.0186, 0.0116, 0.0025, 0.0164], device='cuda:4'), in_proj_covar=tensor([0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010], device='cuda:4'), out_proj_covar=tensor([1.0064e-05, 1.0005e-05, 9.7931e-06, 9.7327e-06, 1.0010e-05, 9.9739e-06, 9.5224e-06, 1.0035e-05], device='cuda:4') 2023-04-27 12:23:51,004 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 2.076e+02 2.707e+02 4.762e+02 1.423e+06, threshold=5.415e+02, percent-clipped=11.0 2023-04-27 12:23:51,004 WARNING [optim.py:388] (4/8) Scaling gradients by 0.002041660714894533, model_norm_threshold=541.4743041992188 2023-04-27 12:23:51,106 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.86, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=6.060e+10, grad_sumsq = 1.544e+12, orig_rms_sq=3.924e-02 2023-04-27 12:24:00,574 WARNING [optim.py:388] (4/8) Scaling gradients by 0.02974529005587101, model_norm_threshold=541.4743041992188 2023-04-27 12:24:00,679 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.85, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=2.829e+08, grad_sumsq = 7.259e+09, orig_rms_sq=3.897e-02 2023-04-27 12:24:01,477 WARNING [optim.py:388] (4/8) Scaling gradients by 0.01955481991171837, model_norm_threshold=541.4743041992188 2023-04-27 12:24:01,584 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.84, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=6.426e+08, grad_sumsq = 1.649e+10, orig_rms_sq=3.897e-02 2023-04-27 12:24:03,063 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=11.75 vs. limit=2.0 2023-04-27 12:24:06,136 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5588, 5.5560, 5.5139, 5.5785, 5.5202, 5.4620, 5.5735, 5.5475], device='cuda:4'), covar=tensor([0.0502, 0.0261, 0.0336, 0.0398, 0.0525, 0.0699, 0.0151, 0.0293], device='cuda:4'), in_proj_covar=tensor([0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010], device='cuda:4'), out_proj_covar=tensor([1.0074e-05, 1.0011e-05, 9.8028e-06, 9.7544e-06, 1.0031e-05, 9.9907e-06, 9.5290e-06, 1.0054e-05], device='cuda:4') 2023-04-27 12:24:22,963 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=11.44 vs. limit=2.0 2023-04-27 12:24:30,805 INFO [train.py:904] (4/8) Epoch 1, batch 250, loss[loss=0.9232, simple_loss=0.7756, pruned_loss=0.9216, over 16670.00 frames. ], tot_loss[loss=1.159, simple_loss=0.9974, pruned_loss=1.206, over 2398611.50 frames. ], batch size: 76, lr: 3.75e-02, grad_scale: 0.0078125 2023-04-27 12:24:33,593 WARNING [optim.py:388] (4/8) Scaling gradients by 0.057925041764974594, model_norm_threshold=541.4743041992188 2023-04-27 12:24:33,697 INFO [optim.py:450] (4/8) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.59, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=5.145e+07, grad_sumsq = 1.327e+09, orig_rms_sq=3.876e-02 2023-04-27 12:25:05,800 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.7862, 5.7843, 5.7867, 5.7803, 5.7809, 5.7866, 5.7856, 5.7811], device='cuda:4'), covar=tensor([0.0060, 0.0091, 0.0083, 0.0089, 0.0066, 0.0040, 0.0123, 0.0088], device='cuda:4'), in_proj_covar=tensor([0.0010, 0.0009, 0.0009, 0.0010, 0.0009, 0.0009, 0.0009, 0.0009], device='cuda:4'), out_proj_covar=tensor([9.4819e-06, 9.6424e-06, 9.4582e-06, 9.2291e-06, 9.6284e-06, 9.1975e-06, 9.1798e-06, 9.4019e-06], device='cuda:4') 2023-04-27 12:25:14,803 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=5.03 vs. limit=2.0 2023-04-27 12:25:16,014 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:25:20,575 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=300.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-27 12:25:21,119 WARNING [train.py:894] (4/8) Grad scale is small: 0.0078125 2023-04-27 12:25:21,119 INFO [train.py:904] (4/8) Epoch 1, batch 300, loss[loss=0.9243, simple_loss=0.7674, pruned_loss=0.9118, over 16852.00 frames. ], tot_loss[loss=1.089, simple_loss=0.9314, pruned_loss=1.119, over 2609514.86 frames. ], batch size: 42, lr: 4.00e-02, grad_scale: 0.015625 2023-04-27 12:25:30,077 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 9.163e+01 1.282e+02 1.731e+02 2.635e+02 2.769e+04, threshold=3.463e+02, percent-clipped=7.0 2023-04-27 12:25:48,294 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=5.27 vs. limit=2.0 2023-04-27 12:26:12,598 INFO [train.py:904] (4/8) Epoch 1, batch 350, loss[loss=0.8921, simple_loss=0.7427, pruned_loss=0.8319, over 16874.00 frames. ], tot_loss[loss=1.04, simple_loss=0.8837, pruned_loss=1.052, over 2758793.55 frames. ], batch size: 116, lr: 4.25e-02, grad_scale: 0.015625 2023-04-27 12:26:18,683 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=357.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-27 12:26:52,671 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=387.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:27:06,004 INFO [train.py:904] (4/8) Epoch 1, batch 400, loss[loss=0.8844, simple_loss=0.7273, pruned_loss=0.8202, over 16869.00 frames. ], tot_loss[loss=1.007, simple_loss=0.8496, pruned_loss=1.002, over 2876320.10 frames. ], batch size: 109, lr: 4.50e-02, grad_scale: 0.03125 2023-04-27 12:27:10,821 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=15.52 vs. limit=2.0 2023-04-27 12:27:17,878 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 7.937e+01 1.110e+02 1.413e+02 1.837e+02 3.136e+02, threshold=2.826e+02, percent-clipped=0.0 2023-04-27 12:27:46,059 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=439.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-27 12:27:48,663 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=2.37 vs. limit=2.0 2023-04-27 12:27:56,252 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=448.0, num_to_drop=2, layers_to_drop={1, 2} 2023-04-27 12:27:59,545 INFO [train.py:904] (4/8) Epoch 1, batch 450, loss[loss=0.8456, simple_loss=0.6862, pruned_loss=0.7817, over 16440.00 frames. ], tot_loss[loss=0.9842, simple_loss=0.8242, pruned_loss=0.9638, over 2981765.98 frames. ], batch size: 75, lr: 4.75e-02, grad_scale: 0.03125 2023-04-27 12:28:39,658 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=7.52 vs. limit=2.0 2023-04-27 12:28:51,183 INFO [train.py:904] (4/8) Epoch 1, batch 500, loss[loss=0.846, simple_loss=0.6821, pruned_loss=0.7659, over 16525.00 frames. ], tot_loss[loss=0.9655, simple_loss=0.8029, pruned_loss=0.9297, over 3051166.77 frames. ], batch size: 75, lr: 4.99e-02, grad_scale: 0.0625 2023-04-27 12:29:01,368 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 7.996e+01 1.277e+02 1.626e+02 2.121e+02 5.144e+02, threshold=3.251e+02, percent-clipped=17.0 2023-04-27 12:29:44,925 INFO [train.py:904] (4/8) Epoch 1, batch 550, loss[loss=0.8879, simple_loss=0.7137, pruned_loss=0.7832, over 16746.00 frames. ], tot_loss[loss=0.9517, simple_loss=0.7862, pruned_loss=0.9, over 3113158.45 frames. ], batch size: 83, lr: 4.98e-02, grad_scale: 0.0625 2023-04-27 12:29:46,086 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.99 vs. limit=2.0 2023-04-27 12:29:58,112 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=562.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:30:03,836 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=568.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:30:26,867 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=590.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:30:38,230 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=600.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 12:30:38,747 INFO [train.py:904] (4/8) Epoch 1, batch 600, loss[loss=0.9438, simple_loss=0.7564, pruned_loss=0.8126, over 17167.00 frames. ], tot_loss[loss=0.939, simple_loss=0.7712, pruned_loss=0.8716, over 3154072.54 frames. ], batch size: 46, lr: 4.98e-02, grad_scale: 0.125 2023-04-27 12:30:48,387 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 8.862e+01 1.447e+02 1.955e+02 2.870e+02 6.309e+02, threshold=3.911e+02, percent-clipped=20.0 2023-04-27 12:31:02,530 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=623.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 12:31:07,542 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=629.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-27 12:31:27,515 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=3.41 vs. limit=2.0 2023-04-27 12:31:28,004 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=648.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:31:30,996 INFO [train.py:904] (4/8) Epoch 1, batch 650, loss[loss=0.8678, simple_loss=0.6987, pruned_loss=0.7198, over 16930.00 frames. ], tot_loss[loss=0.9254, simple_loss=0.7568, pruned_loss=0.8413, over 3180373.93 frames. ], batch size: 109, lr: 4.98e-02, grad_scale: 0.125 2023-04-27 12:31:31,420 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=651.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-27 12:31:32,142 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=652.0, num_to_drop=2, layers_to_drop={1, 2} 2023-04-27 12:31:50,680 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8731, 4.3379, 4.7821, 3.4600, 5.3119, 5.3185, 3.0428, 3.9412], device='cuda:4'), covar=tensor([2.1888, 3.0490, 1.6541, 4.6794, 0.6504, 0.7049, 2.3213, 3.1684], device='cuda:4'), in_proj_covar=tensor([0.0033, 0.0042, 0.0044, 0.0035, 0.0034, 0.0037, 0.0027, 0.0036], device='cuda:4'), out_proj_covar=tensor([2.8319e-05, 3.2620e-05, 3.2323e-05, 2.9874e-05, 2.7697e-05, 2.9867e-05, 2.4818e-05, 3.1393e-05], device='cuda:4') 2023-04-27 12:32:10,472 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=15.01 vs. limit=5.0 2023-04-27 12:32:18,574 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4507, 4.8180, 5.2839, 4.2293, 5.9580, 5.8970, 3.8242, 4.8948], device='cuda:4'), covar=tensor([2.1701, 1.9163, 0.9716, 1.9871, 0.2264, 0.4530, 2.4359, 2.4015], device='cuda:4'), in_proj_covar=tensor([0.0035, 0.0043, 0.0044, 0.0034, 0.0036, 0.0039, 0.0028, 0.0035], device='cuda:4'), out_proj_covar=tensor([2.9161e-05, 3.3099e-05, 3.2407e-05, 2.9112e-05, 2.8418e-05, 3.1477e-05, 2.5146e-05, 3.0007e-05], device='cuda:4') 2023-04-27 12:32:22,394 INFO [train.py:904] (4/8) Epoch 1, batch 700, loss[loss=0.8355, simple_loss=0.676, pruned_loss=0.6691, over 16439.00 frames. ], tot_loss[loss=0.9118, simple_loss=0.7445, pruned_loss=0.8083, over 3212657.92 frames. ], batch size: 75, lr: 4.98e-02, grad_scale: 0.25 2023-04-27 12:32:25,925 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.60 vs. limit=2.0 2023-04-27 12:32:31,773 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 2.100e+02 3.098e+02 3.988e+02 9.500e+02, threshold=6.196e+02, percent-clipped=26.0 2023-04-27 12:33:00,773 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=739.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-27 12:33:05,887 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=743.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-27 12:33:14,006 INFO [train.py:904] (4/8) Epoch 1, batch 750, loss[loss=0.8386, simple_loss=0.696, pruned_loss=0.6241, over 16497.00 frames. ], tot_loss[loss=0.899, simple_loss=0.7352, pruned_loss=0.7735, over 3234332.91 frames. ], batch size: 68, lr: 4.97e-02, grad_scale: 0.25 2023-04-27 12:33:27,869 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2391, 2.3747, 2.2379, 2.3225, 2.5050, 2.0670, 2.2501, 2.0292], device='cuda:4'), covar=tensor([0.3779, 0.4517, 0.2986, 0.3010, 0.1939, 0.4483, 0.3194, 0.4298], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0031, 0.0027, 0.0028, 0.0027, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([2.5936e-05, 3.3138e-05, 2.6128e-05, 2.7136e-05, 2.5020e-05, 2.6676e-05, 2.5657e-05, 2.6862e-05], device='cuda:4') 2023-04-27 12:33:51,950 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=787.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:34:06,256 INFO [train.py:904] (4/8) Epoch 1, batch 800, loss[loss=0.7249, simple_loss=0.6185, pruned_loss=0.4999, over 17141.00 frames. ], tot_loss[loss=0.8732, simple_loss=0.7174, pruned_loss=0.7272, over 3252439.31 frames. ], batch size: 47, lr: 4.97e-02, grad_scale: 0.5 2023-04-27 12:34:17,532 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.716e+02 2.855e+02 4.131e+02 5.648e+02 8.889e+02, threshold=8.261e+02, percent-clipped=19.0 2023-04-27 12:34:54,103 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=847.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:34:58,583 INFO [train.py:904] (4/8) Epoch 1, batch 850, loss[loss=0.6781, simple_loss=0.5808, pruned_loss=0.4573, over 16844.00 frames. ], tot_loss[loss=0.8395, simple_loss=0.6948, pruned_loss=0.675, over 3267376.72 frames. ], batch size: 109, lr: 4.96e-02, grad_scale: 0.5 2023-04-27 12:35:50,488 INFO [train.py:904] (4/8) Epoch 1, batch 900, loss[loss=0.6897, simple_loss=0.5981, pruned_loss=0.4479, over 17113.00 frames. ], tot_loss[loss=0.8032, simple_loss=0.6707, pruned_loss=0.623, over 3282381.38 frames. ], batch size: 47, lr: 4.96e-02, grad_scale: 1.0 2023-04-27 12:35:57,659 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=908.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 12:36:00,422 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.021e+02 3.478e+02 3.989e+02 5.953e+02 1.563e+03, threshold=7.979e+02, percent-clipped=10.0 2023-04-27 12:36:08,840 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=918.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 12:36:14,424 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=924.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 12:36:34,642 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4207, 2.3658, 2.2806, 2.3378, 2.3609, 2.1755, 2.1774, 2.4228], device='cuda:4'), covar=tensor([0.3125, 0.2268, 0.2647, 0.2605, 0.2634, 0.3329, 0.3530, 0.2513], device='cuda:4'), in_proj_covar=tensor([0.0035, 0.0029, 0.0033, 0.0032, 0.0032, 0.0034, 0.0035, 0.0033], device='cuda:4'), out_proj_covar=tensor([3.1570e-05, 2.8677e-05, 3.4424e-05, 3.1435e-05, 3.1200e-05, 3.4541e-05, 3.3823e-05, 3.2719e-05], device='cuda:4') 2023-04-27 12:36:37,788 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=946.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-27 12:36:42,343 INFO [train.py:904] (4/8) Epoch 1, batch 950, loss[loss=0.67, simple_loss=0.5894, pruned_loss=0.4182, over 17213.00 frames. ], tot_loss[loss=0.7717, simple_loss=0.65, pruned_loss=0.578, over 3293742.22 frames. ], batch size: 44, lr: 4.96e-02, grad_scale: 1.0 2023-04-27 12:36:44,105 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=952.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 12:37:21,927 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.80 vs. limit=2.0 2023-04-27 12:37:35,470 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=1000.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:37:35,994 INFO [train.py:904] (4/8) Epoch 1, batch 1000, loss[loss=0.6205, simple_loss=0.5379, pruned_loss=0.3945, over 16474.00 frames. ], tot_loss[loss=0.737, simple_loss=0.6261, pruned_loss=0.5342, over 3292597.96 frames. ], batch size: 75, lr: 4.95e-02, grad_scale: 1.0 2023-04-27 12:37:46,287 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.400e+02 3.913e+02 4.932e+02 6.154e+02 1.349e+03, threshold=9.864e+02, percent-clipped=6.0 2023-04-27 12:37:53,415 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.33 vs. limit=2.0 2023-04-27 12:38:11,058 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=8.06 vs. limit=5.0 2023-04-27 12:38:21,597 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1043.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-27 12:38:29,481 INFO [train.py:904] (4/8) Epoch 1, batch 1050, loss[loss=0.6175, simple_loss=0.5589, pruned_loss=0.3585, over 17003.00 frames. ], tot_loss[loss=0.7073, simple_loss=0.6066, pruned_loss=0.4959, over 3298135.83 frames. ], batch size: 55, lr: 4.95e-02, grad_scale: 1.0 2023-04-27 12:39:12,051 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=1091.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:39:21,795 INFO [train.py:904] (4/8) Epoch 1, batch 1100, loss[loss=0.5779, simple_loss=0.5402, pruned_loss=0.313, over 17047.00 frames. ], tot_loss[loss=0.6779, simple_loss=0.5873, pruned_loss=0.4598, over 3298387.56 frames. ], batch size: 55, lr: 4.94e-02, grad_scale: 1.0 2023-04-27 12:39:33,202 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.538e+02 4.173e+02 5.189e+02 6.744e+02 1.137e+03, threshold=1.038e+03, percent-clipped=1.0 2023-04-27 12:40:17,535 INFO [train.py:904] (4/8) Epoch 1, batch 1150, loss[loss=0.5189, simple_loss=0.4901, pruned_loss=0.2748, over 16844.00 frames. ], tot_loss[loss=0.6516, simple_loss=0.5701, pruned_loss=0.4284, over 3302218.16 frames. ], batch size: 42, lr: 4.94e-02, grad_scale: 1.0 2023-04-27 12:40:19,494 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1153.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:41:10,398 INFO [train.py:904] (4/8) Epoch 1, batch 1200, loss[loss=0.5658, simple_loss=0.525, pruned_loss=0.3098, over 16646.00 frames. ], tot_loss[loss=0.6268, simple_loss=0.5538, pruned_loss=0.4002, over 3300932.64 frames. ], batch size: 62, lr: 4.93e-02, grad_scale: 2.0 2023-04-27 12:41:12,651 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1203.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-27 12:41:12,766 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2499, 3.3594, 4.0039, 3.9560, 2.7869, 4.1835, 3.9852, 3.9895], device='cuda:4'), covar=tensor([0.1708, 0.1874, 0.1460, 0.2364, 0.5433, 0.1148, 0.1965, 0.1352], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0042, 0.0036, 0.0035, 0.0032, 0.0033, 0.0044, 0.0035], device='cuda:4'), out_proj_covar=tensor([3.5544e-05, 3.2181e-05, 2.9116e-05, 3.0705e-05, 2.9524e-05, 2.8353e-05, 3.7479e-05, 2.9265e-05], device='cuda:4') 2023-04-27 12:41:21,556 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.465e+02 4.824e+02 5.887e+02 7.169e+02 1.569e+03, threshold=1.177e+03, percent-clipped=1.0 2023-04-27 12:41:25,710 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1214.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-27 12:41:29,986 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1218.0, num_to_drop=2, layers_to_drop={1, 2} 2023-04-27 12:41:34,830 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1224.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 12:41:52,790 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9356, 2.9438, 2.7621, 2.8328, 2.9329, 2.6199, 2.6033, 2.8962], device='cuda:4'), covar=tensor([0.3333, 0.2423, 0.3822, 0.2623, 0.3145, 0.3443, 0.3727, 0.2501], device='cuda:4'), in_proj_covar=tensor([0.0061, 0.0056, 0.0058, 0.0057, 0.0062, 0.0061, 0.0063, 0.0058], device='cuda:4'), out_proj_covar=tensor([5.6538e-05, 5.2222e-05, 5.9044e-05, 5.5684e-05, 5.9350e-05, 5.9376e-05, 5.9132e-05, 5.5562e-05], device='cuda:4') 2023-04-27 12:41:57,182 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1246.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:42:03,009 INFO [train.py:904] (4/8) Epoch 1, batch 1250, loss[loss=0.4872, simple_loss=0.4599, pruned_loss=0.2581, over 16396.00 frames. ], tot_loss[loss=0.6075, simple_loss=0.541, pruned_loss=0.3779, over 3296691.72 frames. ], batch size: 68, lr: 4.92e-02, grad_scale: 2.0 2023-04-27 12:42:17,895 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=1266.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:42:25,414 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=1272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:42:49,929 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=1294.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:42:56,579 INFO [train.py:904] (4/8) Epoch 1, batch 1300, loss[loss=0.5309, simple_loss=0.5036, pruned_loss=0.2788, over 17061.00 frames. ], tot_loss[loss=0.5903, simple_loss=0.5311, pruned_loss=0.3572, over 3307158.30 frames. ], batch size: 55, lr: 4.92e-02, grad_scale: 2.0 2023-04-27 12:43:07,776 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.547e+02 4.841e+02 5.972e+02 7.377e+02 1.990e+03, threshold=1.194e+03, percent-clipped=4.0 2023-04-27 12:43:51,772 INFO [train.py:904] (4/8) Epoch 1, batch 1350, loss[loss=0.5382, simple_loss=0.5157, pruned_loss=0.2778, over 16729.00 frames. ], tot_loss[loss=0.5741, simple_loss=0.5216, pruned_loss=0.3388, over 3308620.68 frames. ], batch size: 57, lr: 4.91e-02, grad_scale: 2.0 2023-04-27 12:44:49,662 INFO [train.py:904] (4/8) Epoch 1, batch 1400, loss[loss=0.5167, simple_loss=0.4781, pruned_loss=0.2821, over 16871.00 frames. ], tot_loss[loss=0.5594, simple_loss=0.5125, pruned_loss=0.3229, over 3318141.95 frames. ], batch size: 96, lr: 4.91e-02, grad_scale: 2.0 2023-04-27 12:45:00,101 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.619e+02 5.087e+02 6.130e+02 7.329e+02 1.559e+03, threshold=1.226e+03, percent-clipped=4.0 2023-04-27 12:45:27,715 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1434.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:45:46,439 INFO [train.py:904] (4/8) Epoch 1, batch 1450, loss[loss=0.5474, simple_loss=0.4956, pruned_loss=0.3076, over 16798.00 frames. ], tot_loss[loss=0.5431, simple_loss=0.5017, pruned_loss=0.3074, over 3324546.93 frames. ], batch size: 83, lr: 4.90e-02, grad_scale: 2.0 2023-04-27 12:46:34,504 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1495.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-27 12:46:41,354 INFO [train.py:904] (4/8) Epoch 1, batch 1500, loss[loss=0.4749, simple_loss=0.4572, pruned_loss=0.244, over 16833.00 frames. ], tot_loss[loss=0.5303, simple_loss=0.4933, pruned_loss=0.2951, over 3326461.12 frames. ], batch size: 42, lr: 4.89e-02, grad_scale: 2.0 2023-04-27 12:46:44,188 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1503.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 12:46:51,036 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1509.0, num_to_drop=2, layers_to_drop={2, 3} 2023-04-27 12:46:52,703 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.422e+02 5.181e+02 6.478e+02 8.944e+02 1.260e+03, threshold=1.296e+03, percent-clipped=1.0 2023-04-27 12:47:18,385 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.03 vs. limit=2.0 2023-04-27 12:47:23,046 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.93 vs. limit=2.0 2023-04-27 12:47:38,770 INFO [train.py:904] (4/8) Epoch 1, batch 1550, loss[loss=0.4563, simple_loss=0.4563, pruned_loss=0.2217, over 17178.00 frames. ], tot_loss[loss=0.5237, simple_loss=0.4899, pruned_loss=0.2874, over 3320783.03 frames. ], batch size: 46, lr: 4.89e-02, grad_scale: 2.0 2023-04-27 12:47:39,032 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=1551.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:47:56,074 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1566.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:48:35,335 INFO [train.py:904] (4/8) Epoch 1, batch 1600, loss[loss=0.4899, simple_loss=0.4766, pruned_loss=0.2487, over 17171.00 frames. ], tot_loss[loss=0.5179, simple_loss=0.4875, pruned_loss=0.2806, over 3315653.54 frames. ], batch size: 46, lr: 4.88e-02, grad_scale: 4.0 2023-04-27 12:48:47,163 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.141e+02 5.212e+02 6.871e+02 8.557e+02 2.137e+03, threshold=1.374e+03, percent-clipped=9.0 2023-04-27 12:48:47,459 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1611.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:49:05,612 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1627.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 12:49:08,130 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4399, 5.5540, 5.0511, 4.9768, 5.5892, 5.6094, 5.1205, 5.5538], device='cuda:4'), covar=tensor([0.0263, 0.0238, 0.0263, 0.0334, 0.0116, 0.0208, 0.0236, 0.0224], device='cuda:4'), in_proj_covar=tensor([0.0060, 0.0067, 0.0078, 0.0073, 0.0063, 0.0065, 0.0080, 0.0075], device='cuda:4'), out_proj_covar=tensor([5.1138e-05, 6.2263e-05, 7.7298e-05, 6.6213e-05, 5.5539e-05, 6.0491e-05, 7.7813e-05, 7.0158e-05], device='cuda:4') 2023-04-27 12:49:32,287 INFO [train.py:904] (4/8) Epoch 1, batch 1650, loss[loss=0.4571, simple_loss=0.4431, pruned_loss=0.2336, over 16542.00 frames. ], tot_loss[loss=0.5108, simple_loss=0.4841, pruned_loss=0.2731, over 3321148.42 frames. ], batch size: 75, lr: 4.87e-02, grad_scale: 4.0 2023-04-27 12:49:57,817 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1672.0, num_to_drop=2, layers_to_drop={2, 3} 2023-04-27 12:50:29,818 INFO [train.py:904] (4/8) Epoch 1, batch 1700, loss[loss=0.4286, simple_loss=0.4412, pruned_loss=0.2023, over 17214.00 frames. ], tot_loss[loss=0.5039, simple_loss=0.4812, pruned_loss=0.266, over 3327930.28 frames. ], batch size: 45, lr: 4.86e-02, grad_scale: 4.0 2023-04-27 12:50:32,840 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3798, 5.5706, 5.3876, 5.7150, 5.4578, 5.5271, 5.5372, 5.4575], device='cuda:4'), covar=tensor([0.0362, 0.0501, 0.0406, 0.0208, 0.0394, 0.0284, 0.0281, 0.0249], device='cuda:4'), in_proj_covar=tensor([0.0080, 0.0090, 0.0088, 0.0071, 0.0084, 0.0075, 0.0076, 0.0070], device='cuda:4'), out_proj_covar=tensor([6.0753e-05, 7.9418e-05, 7.2547e-05, 4.7830e-05, 6.1285e-05, 5.5540e-05, 5.8414e-05, 5.4031e-05], device='cuda:4') 2023-04-27 12:50:41,826 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.189e+02 5.218e+02 6.714e+02 8.049e+02 1.427e+03, threshold=1.343e+03, percent-clipped=1.0 2023-04-27 12:50:48,520 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.90 vs. limit=5.0 2023-04-27 12:51:00,277 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-27 12:51:22,888 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4239, 3.7221, 3.2553, 3.3643, 3.3499, 3.7251, 3.7273, 3.4943], device='cuda:4'), covar=tensor([0.0365, 0.0295, 0.0506, 0.0399, 0.0340, 0.0272, 0.0372, 0.0244], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0077, 0.0078, 0.0070, 0.0076, 0.0077, 0.0068], device='cuda:4'), out_proj_covar=tensor([6.6529e-05, 6.6081e-05, 6.8379e-05, 6.9385e-05, 6.1887e-05, 6.6663e-05, 6.5285e-05, 6.1831e-05], device='cuda:4') 2023-04-27 12:51:28,761 INFO [train.py:904] (4/8) Epoch 1, batch 1750, loss[loss=0.4083, simple_loss=0.4237, pruned_loss=0.1915, over 16029.00 frames. ], tot_loss[loss=0.4944, simple_loss=0.4763, pruned_loss=0.2577, over 3330770.43 frames. ], batch size: 35, lr: 4.86e-02, grad_scale: 4.0 2023-04-27 12:52:15,501 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1790.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:52:27,926 INFO [train.py:904] (4/8) Epoch 1, batch 1800, loss[loss=0.4504, simple_loss=0.4619, pruned_loss=0.2158, over 16714.00 frames. ], tot_loss[loss=0.488, simple_loss=0.4742, pruned_loss=0.2513, over 3328028.48 frames. ], batch size: 57, lr: 4.85e-02, grad_scale: 4.0 2023-04-27 12:52:37,428 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1809.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:52:38,890 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.321e+02 5.485e+02 6.556e+02 7.752e+02 2.000e+03, threshold=1.311e+03, percent-clipped=5.0 2023-04-27 12:53:06,025 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.95 vs. limit=2.0 2023-04-27 12:53:10,969 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1838.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:53:25,087 INFO [train.py:904] (4/8) Epoch 1, batch 1850, loss[loss=0.4577, simple_loss=0.4566, pruned_loss=0.2277, over 16754.00 frames. ], tot_loss[loss=0.4826, simple_loss=0.4728, pruned_loss=0.246, over 3308546.36 frames. ], batch size: 134, lr: 4.84e-02, grad_scale: 4.0 2023-04-27 12:53:32,838 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=1857.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:53:33,003 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1857.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:53:48,680 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1870.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:54:22,028 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1899.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-27 12:54:23,725 INFO [train.py:904] (4/8) Epoch 1, batch 1900, loss[loss=0.4305, simple_loss=0.4505, pruned_loss=0.2032, over 17113.00 frames. ], tot_loss[loss=0.4718, simple_loss=0.4668, pruned_loss=0.2378, over 3314453.12 frames. ], batch size: 49, lr: 4.83e-02, grad_scale: 4.0 2023-04-27 12:54:36,241 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.249e+02 5.423e+02 6.685e+02 8.506e+02 1.861e+03, threshold=1.337e+03, percent-clipped=8.0 2023-04-27 12:54:44,502 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1918.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-27 12:54:49,109 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1922.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 12:55:00,069 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1931.0, num_to_drop=2, layers_to_drop={2, 3} 2023-04-27 12:55:22,814 INFO [train.py:904] (4/8) Epoch 1, batch 1950, loss[loss=0.4806, simple_loss=0.4928, pruned_loss=0.2333, over 15370.00 frames. ], tot_loss[loss=0.4631, simple_loss=0.4624, pruned_loss=0.2311, over 3307166.81 frames. ], batch size: 190, lr: 4.83e-02, grad_scale: 4.0 2023-04-27 12:55:42,110 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1967.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:56:23,562 INFO [train.py:904] (4/8) Epoch 1, batch 2000, loss[loss=0.438, simple_loss=0.4682, pruned_loss=0.2039, over 16770.00 frames. ], tot_loss[loss=0.4572, simple_loss=0.4593, pruned_loss=0.2269, over 3311917.95 frames. ], batch size: 57, lr: 4.82e-02, grad_scale: 8.0 2023-04-27 12:56:36,808 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.491e+02 5.646e+02 7.242e+02 8.659e+02 1.834e+03, threshold=1.448e+03, percent-clipped=2.0 2023-04-27 12:57:05,775 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9013, 4.0619, 3.5784, 3.0727, 3.6785, 3.9459, 3.2851, 3.6654], device='cuda:4'), covar=tensor([0.0217, 0.0183, 0.0551, 0.0538, 0.0396, 0.0254, 0.0598, 0.0297], device='cuda:4'), in_proj_covar=tensor([0.0063, 0.0063, 0.0062, 0.0068, 0.0055, 0.0064, 0.0068, 0.0071], device='cuda:4'), out_proj_covar=tensor([5.0569e-05, 4.6282e-05, 4.5751e-05, 5.4352e-05, 4.2120e-05, 4.7572e-05, 5.1903e-05, 5.4740e-05], device='cuda:4') 2023-04-27 12:57:27,604 INFO [train.py:904] (4/8) Epoch 1, batch 2050, loss[loss=0.4461, simple_loss=0.4463, pruned_loss=0.2229, over 16747.00 frames. ], tot_loss[loss=0.4484, simple_loss=0.4541, pruned_loss=0.2208, over 3304140.42 frames. ], batch size: 134, lr: 4.81e-02, grad_scale: 8.0 2023-04-27 12:58:06,079 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1039, 5.2086, 4.8620, 5.0176, 4.9945, 5.1931, 5.2557, 4.8972], device='cuda:4'), covar=tensor([0.0407, 0.0541, 0.0704, 0.0851, 0.0876, 0.0447, 0.0512, 0.0986], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0132, 0.0123, 0.0138, 0.0149, 0.0109, 0.0100, 0.0140], device='cuda:4'), out_proj_covar=tensor([1.0079e-04, 1.2184e-04, 1.0942e-04, 1.2432e-04, 1.4388e-04, 1.0186e-04, 9.2151e-05, 1.3680e-04], device='cuda:4') 2023-04-27 12:58:19,042 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2090.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 12:58:32,223 INFO [train.py:904] (4/8) Epoch 1, batch 2100, loss[loss=0.421, simple_loss=0.4282, pruned_loss=0.2068, over 16736.00 frames. ], tot_loss[loss=0.4415, simple_loss=0.4507, pruned_loss=0.2157, over 3313651.89 frames. ], batch size: 89, lr: 4.80e-02, grad_scale: 16.0 2023-04-27 12:58:41,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6258, 4.8112, 4.6279, 4.8865, 4.5449, 4.8640, 4.6389, 4.6790], device='cuda:4'), covar=tensor([0.0348, 0.0486, 0.0432, 0.0228, 0.0453, 0.0287, 0.0277, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0098, 0.0115, 0.0104, 0.0084, 0.0102, 0.0093, 0.0096, 0.0086], device='cuda:4'), out_proj_covar=tensor([7.9105e-05, 1.0009e-04, 8.6743e-05, 6.0006e-05, 7.8898e-05, 7.2351e-05, 7.4728e-05, 7.0068e-05], device='cuda:4') 2023-04-27 12:58:45,932 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.641e+02 4.108e+02 4.910e+02 5.858e+02 1.001e+03, threshold=9.819e+02, percent-clipped=0.0 2023-04-27 12:59:19,750 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=2138.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:59:35,891 INFO [train.py:904] (4/8) Epoch 1, batch 2150, loss[loss=0.3516, simple_loss=0.3892, pruned_loss=0.157, over 17009.00 frames. ], tot_loss[loss=0.4353, simple_loss=0.4471, pruned_loss=0.2114, over 3317197.64 frames. ], batch size: 41, lr: 4.79e-02, grad_scale: 16.0 2023-04-27 13:00:32,068 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=2194.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:00:40,252 INFO [train.py:904] (4/8) Epoch 1, batch 2200, loss[loss=0.4049, simple_loss=0.421, pruned_loss=0.1944, over 16415.00 frames. ], tot_loss[loss=0.4287, simple_loss=0.4433, pruned_loss=0.2068, over 3315848.33 frames. ], batch size: 146, lr: 4.78e-02, grad_scale: 16.0 2023-04-27 13:00:53,195 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.805e+02 5.075e+02 6.122e+02 8.359e+02 1.658e+03, threshold=1.224e+03, percent-clipped=12.0 2023-04-27 13:00:56,684 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=2213.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:01:07,599 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2222.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-27 13:01:11,873 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=2226.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:01:44,909 INFO [train.py:904] (4/8) Epoch 1, batch 2250, loss[loss=0.3825, simple_loss=0.4272, pruned_loss=0.1689, over 17044.00 frames. ], tot_loss[loss=0.4228, simple_loss=0.4407, pruned_loss=0.2023, over 3315258.00 frames. ], batch size: 50, lr: 4.77e-02, grad_scale: 16.0 2023-04-27 13:02:05,711 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2267.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:02:08,815 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=2270.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:02:49,153 INFO [train.py:904] (4/8) Epoch 1, batch 2300, loss[loss=0.364, simple_loss=0.4132, pruned_loss=0.1574, over 17127.00 frames. ], tot_loss[loss=0.4153, simple_loss=0.4362, pruned_loss=0.197, over 3321449.49 frames. ], batch size: 47, lr: 4.77e-02, grad_scale: 16.0 2023-04-27 13:02:55,973 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.40 vs. limit=5.0 2023-04-27 13:02:59,292 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.36 vs. limit=5.0 2023-04-27 13:03:01,864 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.985e+02 4.330e+02 5.680e+02 7.295e+02 1.284e+03, threshold=1.136e+03, percent-clipped=1.0 2023-04-27 13:03:07,886 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=2315.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:03:39,661 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=2340.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:03:53,136 INFO [train.py:904] (4/8) Epoch 1, batch 2350, loss[loss=0.4147, simple_loss=0.4517, pruned_loss=0.1889, over 17050.00 frames. ], tot_loss[loss=0.4093, simple_loss=0.4324, pruned_loss=0.1929, over 3328749.70 frames. ], batch size: 53, lr: 4.76e-02, grad_scale: 16.0 2023-04-27 13:04:54,412 INFO [train.py:904] (4/8) Epoch 1, batch 2400, loss[loss=0.4064, simple_loss=0.449, pruned_loss=0.1819, over 17001.00 frames. ], tot_loss[loss=0.4049, simple_loss=0.4305, pruned_loss=0.1896, over 3337125.35 frames. ], batch size: 55, lr: 4.75e-02, grad_scale: 16.0 2023-04-27 13:04:55,639 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=2401.0, num_to_drop=2, layers_to_drop={2, 3} 2023-04-27 13:05:07,294 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.161e+02 4.808e+02 6.214e+02 8.409e+02 1.581e+03, threshold=1.243e+03, percent-clipped=3.0 2023-04-27 13:05:55,891 INFO [train.py:904] (4/8) Epoch 1, batch 2450, loss[loss=0.4053, simple_loss=0.4512, pruned_loss=0.1797, over 17064.00 frames. ], tot_loss[loss=0.4034, simple_loss=0.43, pruned_loss=0.1883, over 3324805.37 frames. ], batch size: 53, lr: 4.74e-02, grad_scale: 16.0 2023-04-27 13:06:51,082 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2494.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:06:58,989 INFO [train.py:904] (4/8) Epoch 1, batch 2500, loss[loss=0.375, simple_loss=0.4224, pruned_loss=0.1638, over 16756.00 frames. ], tot_loss[loss=0.3971, simple_loss=0.4264, pruned_loss=0.1838, over 3328928.33 frames. ], batch size: 62, lr: 4.73e-02, grad_scale: 16.0 2023-04-27 13:07:11,499 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.680e+02 4.365e+02 5.111e+02 7.039e+02 1.163e+03, threshold=1.022e+03, percent-clipped=0.0 2023-04-27 13:07:14,807 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2513.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:07:16,001 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.15 vs. limit=5.0 2023-04-27 13:07:29,210 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2526.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:07:50,934 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=2542.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:08:01,505 INFO [train.py:904] (4/8) Epoch 1, batch 2550, loss[loss=0.4053, simple_loss=0.4416, pruned_loss=0.1845, over 16608.00 frames. ], tot_loss[loss=0.392, simple_loss=0.4231, pruned_loss=0.1803, over 3328185.90 frames. ], batch size: 62, lr: 4.72e-02, grad_scale: 16.0 2023-04-27 13:08:15,978 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=2561.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:08:32,316 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=2574.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:08:39,754 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7702, 4.5432, 4.8753, 5.0151, 5.3324, 4.8582, 4.7517, 5.1143], device='cuda:4'), covar=tensor([0.0283, 0.0306, 0.0511, 0.0404, 0.0235, 0.0276, 0.0462, 0.0187], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0101, 0.0124, 0.0122, 0.0115, 0.0110, 0.0115, 0.0096], device='cuda:4'), out_proj_covar=tensor([1.1511e-04, 1.1993e-04, 1.3661e-04, 1.2665e-04, 1.2506e-04, 1.2081e-04, 1.1788e-04, 9.5156e-05], device='cuda:4') 2023-04-27 13:09:07,692 INFO [train.py:904] (4/8) Epoch 1, batch 2600, loss[loss=0.3963, simple_loss=0.425, pruned_loss=0.1838, over 16223.00 frames. ], tot_loss[loss=0.3867, simple_loss=0.4199, pruned_loss=0.1767, over 3315510.56 frames. ], batch size: 165, lr: 4.71e-02, grad_scale: 16.0 2023-04-27 13:09:20,088 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.935e+02 4.794e+02 6.148e+02 7.560e+02 1.171e+03, threshold=1.230e+03, percent-clipped=4.0 2023-04-27 13:10:11,887 INFO [train.py:904] (4/8) Epoch 1, batch 2650, loss[loss=0.3808, simple_loss=0.4353, pruned_loss=0.1631, over 17101.00 frames. ], tot_loss[loss=0.3839, simple_loss=0.4191, pruned_loss=0.1743, over 3316241.70 frames. ], batch size: 55, lr: 4.70e-02, grad_scale: 16.0 2023-04-27 13:10:52,862 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.66 vs. limit=2.0 2023-04-27 13:10:55,267 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=2685.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:11:10,028 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=2696.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:11:15,041 INFO [train.py:904] (4/8) Epoch 1, batch 2700, loss[loss=0.4257, simple_loss=0.4411, pruned_loss=0.2051, over 16382.00 frames. ], tot_loss[loss=0.3809, simple_loss=0.4178, pruned_loss=0.172, over 3323002.22 frames. ], batch size: 145, lr: 4.69e-02, grad_scale: 16.0 2023-04-27 13:11:29,644 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.771e+02 4.333e+02 5.264e+02 6.238e+02 1.242e+03, threshold=1.053e+03, percent-clipped=1.0 2023-04-27 13:11:30,953 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2257, 4.0115, 4.2514, 4.2875, 4.6639, 4.1493, 4.1279, 4.4330], device='cuda:4'), covar=tensor([0.0288, 0.0280, 0.0460, 0.0442, 0.0254, 0.0316, 0.0426, 0.0239], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0101, 0.0124, 0.0123, 0.0117, 0.0110, 0.0113, 0.0096], device='cuda:4'), out_proj_covar=tensor([1.1673e-04, 1.2414e-04, 1.3841e-04, 1.2981e-04, 1.2986e-04, 1.2601e-04, 1.1849e-04, 9.7788e-05], device='cuda:4') 2023-04-27 13:12:14,208 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=2746.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 13:12:20,232 INFO [train.py:904] (4/8) Epoch 1, batch 2750, loss[loss=0.361, simple_loss=0.4197, pruned_loss=0.1511, over 17107.00 frames. ], tot_loss[loss=0.3781, simple_loss=0.4165, pruned_loss=0.1699, over 3323376.74 frames. ], batch size: 49, lr: 4.68e-02, grad_scale: 16.0 2023-04-27 13:13:23,779 INFO [train.py:904] (4/8) Epoch 1, batch 2800, loss[loss=0.4073, simple_loss=0.4236, pruned_loss=0.1955, over 16710.00 frames. ], tot_loss[loss=0.3756, simple_loss=0.4147, pruned_loss=0.1682, over 3322328.38 frames. ], batch size: 134, lr: 4.67e-02, grad_scale: 16.0 2023-04-27 13:13:35,358 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.742e+02 4.329e+02 5.433e+02 6.254e+02 2.106e+03, threshold=1.087e+03, percent-clipped=5.0 2023-04-27 13:13:54,300 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=2824.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:14:25,999 INFO [train.py:904] (4/8) Epoch 1, batch 2850, loss[loss=0.3775, simple_loss=0.3998, pruned_loss=0.1776, over 16899.00 frames. ], tot_loss[loss=0.3714, simple_loss=0.4113, pruned_loss=0.1657, over 3327200.25 frames. ], batch size: 109, lr: 4.66e-02, grad_scale: 16.0 2023-04-27 13:15:09,614 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=2885.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-27 13:15:16,456 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7310, 3.5764, 3.9268, 4.2228, 3.9205, 3.7804, 4.1948, 4.1496], device='cuda:4'), covar=tensor([0.0147, 0.0242, 0.0203, 0.0116, 0.0092, 0.0207, 0.0096, 0.0103], device='cuda:4'), in_proj_covar=tensor([0.0023, 0.0021, 0.0023, 0.0021, 0.0023, 0.0020, 0.0024, 0.0019], device='cuda:4'), out_proj_covar=tensor([2.0510e-05, 1.9630e-05, 2.0430e-05, 1.9059e-05, 2.1070e-05, 1.8720e-05, 2.0343e-05, 1.6231e-05], device='cuda:4') 2023-04-27 13:15:27,285 INFO [train.py:904] (4/8) Epoch 1, batch 2900, loss[loss=0.3647, simple_loss=0.4185, pruned_loss=0.1555, over 17142.00 frames. ], tot_loss[loss=0.3702, simple_loss=0.4089, pruned_loss=0.1657, over 3326362.98 frames. ], batch size: 48, lr: 4.65e-02, grad_scale: 16.0 2023-04-27 13:15:32,173 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9155, 4.6954, 5.0184, 5.0926, 5.4312, 4.9847, 4.8962, 5.1783], device='cuda:4'), covar=tensor([0.0276, 0.0220, 0.0485, 0.0375, 0.0278, 0.0249, 0.0315, 0.0209], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0101, 0.0131, 0.0126, 0.0125, 0.0112, 0.0117, 0.0097], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:15:40,739 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.037e+02 5.015e+02 6.760e+02 8.832e+02 1.641e+03, threshold=1.352e+03, percent-clipped=13.0 2023-04-27 13:15:53,094 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.83 vs. limit=2.0 2023-04-27 13:16:13,921 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=2937.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:16:32,273 INFO [train.py:904] (4/8) Epoch 1, batch 2950, loss[loss=0.3002, simple_loss=0.3494, pruned_loss=0.1255, over 16983.00 frames. ], tot_loss[loss=0.3687, simple_loss=0.4068, pruned_loss=0.1653, over 3321614.54 frames. ], batch size: 41, lr: 4.64e-02, grad_scale: 16.0 2023-04-27 13:17:15,527 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2029, 5.8236, 5.6380, 5.6171, 5.7408, 6.0097, 6.1040, 5.5811], device='cuda:4'), covar=tensor([0.0500, 0.0714, 0.0671, 0.1225, 0.1298, 0.0491, 0.0509, 0.1252], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0174, 0.0145, 0.0159, 0.0190, 0.0127, 0.0123, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 13:17:29,957 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2996.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:17:32,190 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=2998.0, num_to_drop=2, layers_to_drop={1, 2} 2023-04-27 13:17:35,317 INFO [train.py:904] (4/8) Epoch 1, batch 3000, loss[loss=0.3412, simple_loss=0.3948, pruned_loss=0.1438, over 17227.00 frames. ], tot_loss[loss=0.3681, simple_loss=0.4061, pruned_loss=0.1651, over 3318713.76 frames. ], batch size: 44, lr: 4.63e-02, grad_scale: 16.0 2023-04-27 13:17:35,317 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 13:17:45,053 INFO [train.py:938] (4/8) Epoch 1, validation: loss=0.2847, simple_loss=0.3895, pruned_loss=0.08992, over 944034.00 frames. 2023-04-27 13:17:45,053 INFO [train.py:939] (4/8) Maximum memory allocated so far is 15802MB 2023-04-27 13:17:59,838 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.008e+02 4.355e+02 5.212e+02 6.501e+02 1.071e+03, threshold=1.042e+03, percent-clipped=0.0 2023-04-27 13:18:09,273 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3019.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:18:28,025 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3375, 5.0544, 4.2806, 4.6670, 5.0426, 5.3625, 4.6375, 5.1813], device='cuda:4'), covar=tensor([0.0113, 0.0183, 0.0190, 0.0301, 0.0096, 0.0081, 0.0133, 0.0108], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0047, 0.0068, 0.0066, 0.0047, 0.0054, 0.0063, 0.0059], device='cuda:4'), out_proj_covar=tensor([6.4085e-05, 5.6334e-05, 9.6815e-05, 8.3902e-05, 5.1098e-05, 6.1886e-05, 8.2490e-05, 7.6434e-05], device='cuda:4') 2023-04-27 13:18:36,948 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3041.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:18:41,574 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=3044.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:18:42,753 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5423, 4.8702, 4.3443, 3.4125, 3.0039, 2.2047, 4.5619, 4.6187], device='cuda:4'), covar=tensor([0.0300, 0.0193, 0.0170, 0.0885, 0.0787, 0.1159, 0.0092, 0.0191], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0035, 0.0044, 0.0066, 0.0061, 0.0065, 0.0032, 0.0030], device='cuda:4'), out_proj_covar=tensor([4.3051e-05, 3.9787e-05, 4.3313e-05, 6.5802e-05, 5.7956e-05, 6.1905e-05, 3.2911e-05, 3.3926e-05], device='cuda:4') 2023-04-27 13:18:43,742 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9467, 3.9564, 4.1997, 4.0993, 4.4194, 4.1817, 3.9915, 4.3153], device='cuda:4'), covar=tensor([0.0401, 0.0304, 0.0398, 0.0413, 0.0283, 0.0254, 0.0446, 0.0263], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0102, 0.0129, 0.0129, 0.0126, 0.0112, 0.0119, 0.0098], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:18:50,094 INFO [train.py:904] (4/8) Epoch 1, batch 3050, loss[loss=0.3402, simple_loss=0.3947, pruned_loss=0.1429, over 17196.00 frames. ], tot_loss[loss=0.3662, simple_loss=0.4049, pruned_loss=0.1637, over 3313522.47 frames. ], batch size: 46, lr: 4.62e-02, grad_scale: 16.0 2023-04-27 13:19:27,613 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3080.0, num_to_drop=2, layers_to_drop={1, 2} 2023-04-27 13:19:43,618 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3801, 3.4980, 3.5527, 3.1269, 3.3776, 3.5707, 3.6249, 3.4891], device='cuda:4'), covar=tensor([0.0176, 0.0164, 0.0143, 0.0182, 0.0212, 0.0159, 0.0181, 0.0186], device='cuda:4'), in_proj_covar=tensor([0.0045, 0.0036, 0.0036, 0.0041, 0.0039, 0.0041, 0.0045, 0.0040], device='cuda:4'), out_proj_covar=tensor([5.2335e-05, 4.5961e-05, 4.4042e-05, 4.7786e-05, 4.6056e-05, 5.6694e-05, 5.3498e-05, 4.7302e-05], device='cuda:4') 2023-04-27 13:19:53,977 INFO [train.py:904] (4/8) Epoch 1, batch 3100, loss[loss=0.4186, simple_loss=0.4216, pruned_loss=0.2078, over 16768.00 frames. ], tot_loss[loss=0.3643, simple_loss=0.4036, pruned_loss=0.1625, over 3324878.48 frames. ], batch size: 124, lr: 4.61e-02, grad_scale: 16.0 2023-04-27 13:20:07,649 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.721e+02 4.392e+02 5.238e+02 7.570e+02 1.450e+03, threshold=1.048e+03, percent-clipped=8.0 2023-04-27 13:20:44,349 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8631, 3.9453, 3.4657, 2.3240, 3.3877, 3.8595, 3.7065, 3.7754], device='cuda:4'), covar=tensor([0.0147, 0.0158, 0.0193, 0.1233, 0.0329, 0.0156, 0.0123, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0037, 0.0046, 0.0063, 0.0039, 0.0038, 0.0038, 0.0040], device='cuda:4'), out_proj_covar=tensor([3.9295e-05, 4.0496e-05, 4.9842e-05, 6.5789e-05, 4.7623e-05, 3.9638e-05, 4.6777e-05, 4.2081e-05], device='cuda:4') 2023-04-27 13:21:00,201 INFO [train.py:904] (4/8) Epoch 1, batch 3150, loss[loss=0.3328, simple_loss=0.3987, pruned_loss=0.1335, over 17097.00 frames. ], tot_loss[loss=0.3614, simple_loss=0.4013, pruned_loss=0.1608, over 3319324.91 frames. ], batch size: 49, lr: 4.60e-02, grad_scale: 16.0 2023-04-27 13:21:38,600 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3180.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:22:05,296 INFO [train.py:904] (4/8) Epoch 1, batch 3200, loss[loss=0.3297, simple_loss=0.3663, pruned_loss=0.1466, over 16866.00 frames. ], tot_loss[loss=0.3587, simple_loss=0.3996, pruned_loss=0.1589, over 3319617.97 frames. ], batch size: 116, lr: 4.59e-02, grad_scale: 16.0 2023-04-27 13:22:07,843 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.67 vs. limit=2.0 2023-04-27 13:22:17,349 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.273e+02 4.612e+02 6.158e+02 7.733e+02 1.150e+03, threshold=1.232e+03, percent-clipped=3.0 2023-04-27 13:22:40,562 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.03 vs. limit=2.0 2023-04-27 13:23:07,319 INFO [train.py:904] (4/8) Epoch 1, batch 3250, loss[loss=0.3685, simple_loss=0.4003, pruned_loss=0.1684, over 16884.00 frames. ], tot_loss[loss=0.3586, simple_loss=0.3993, pruned_loss=0.1589, over 3315559.34 frames. ], batch size: 96, lr: 4.58e-02, grad_scale: 16.0 2023-04-27 13:23:56,306 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3288.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:24:02,463 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3293.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:24:12,251 INFO [train.py:904] (4/8) Epoch 1, batch 3300, loss[loss=0.3239, simple_loss=0.3888, pruned_loss=0.1294, over 17271.00 frames. ], tot_loss[loss=0.3575, simple_loss=0.3992, pruned_loss=0.1579, over 3320262.68 frames. ], batch size: 52, lr: 4.57e-02, grad_scale: 16.0 2023-04-27 13:24:25,122 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.531e+02 4.292e+02 5.268e+02 6.867e+02 1.392e+03, threshold=1.054e+03, percent-clipped=2.0 2023-04-27 13:24:53,539 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2945, 4.1951, 4.4554, 4.4681, 4.7669, 4.4227, 4.2533, 4.5057], device='cuda:4'), covar=tensor([0.0304, 0.0277, 0.0500, 0.0498, 0.0294, 0.0273, 0.0500, 0.0252], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0107, 0.0139, 0.0138, 0.0143, 0.0120, 0.0131, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:25:03,657 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3341.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:25:16,130 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3349.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-27 13:25:17,759 INFO [train.py:904] (4/8) Epoch 1, batch 3350, loss[loss=0.3976, simple_loss=0.4246, pruned_loss=0.1853, over 16227.00 frames. ], tot_loss[loss=0.3563, simple_loss=0.399, pruned_loss=0.1568, over 3330388.98 frames. ], batch size: 165, lr: 4.56e-02, grad_scale: 16.0 2023-04-27 13:25:36,467 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9053, 5.2451, 4.9010, 5.0441, 5.0381, 5.2838, 5.4097, 5.0134], device='cuda:4'), covar=tensor([0.0497, 0.0679, 0.0674, 0.0889, 0.1139, 0.0592, 0.0461, 0.1037], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0182, 0.0149, 0.0161, 0.0192, 0.0138, 0.0129, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 13:25:50,433 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3375.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:26:09,266 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=3389.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:26:24,812 INFO [train.py:904] (4/8) Epoch 1, batch 3400, loss[loss=0.3535, simple_loss=0.3909, pruned_loss=0.1581, over 16646.00 frames. ], tot_loss[loss=0.3538, simple_loss=0.3972, pruned_loss=0.1552, over 3323577.28 frames. ], batch size: 134, lr: 4.55e-02, grad_scale: 16.0 2023-04-27 13:26:39,219 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.681e+02 4.195e+02 5.225e+02 6.801e+02 1.040e+03, threshold=1.045e+03, percent-clipped=0.0 2023-04-27 13:27:11,907 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-04-27 13:27:32,306 INFO [train.py:904] (4/8) Epoch 1, batch 3450, loss[loss=0.2912, simple_loss=0.3626, pruned_loss=0.1099, over 17157.00 frames. ], tot_loss[loss=0.3507, simple_loss=0.3947, pruned_loss=0.1533, over 3325262.49 frames. ], batch size: 46, lr: 4.54e-02, grad_scale: 16.0 2023-04-27 13:27:36,906 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3948, 3.9883, 3.6664, 3.3744, 3.4252, 3.8843, 3.8231, 3.7537], device='cuda:4'), covar=tensor([0.0247, 0.0127, 0.0174, 0.0197, 0.0232, 0.0158, 0.0233, 0.0167], device='cuda:4'), in_proj_covar=tensor([0.0046, 0.0034, 0.0035, 0.0044, 0.0036, 0.0040, 0.0045, 0.0040], device='cuda:4'), out_proj_covar=tensor([6.0286e-05, 4.7021e-05, 4.6203e-05, 5.6155e-05, 4.8211e-05, 6.2593e-05, 5.7918e-05, 5.2231e-05], device='cuda:4') 2023-04-27 13:28:12,486 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3480.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:28:39,712 INFO [train.py:904] (4/8) Epoch 1, batch 3500, loss[loss=0.3332, simple_loss=0.3809, pruned_loss=0.1427, over 16515.00 frames. ], tot_loss[loss=0.3477, simple_loss=0.3925, pruned_loss=0.1515, over 3326170.60 frames. ], batch size: 75, lr: 4.53e-02, grad_scale: 16.0 2023-04-27 13:28:53,813 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.034e+02 4.599e+02 5.593e+02 7.423e+02 2.273e+03, threshold=1.119e+03, percent-clipped=10.0 2023-04-27 13:29:16,291 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=3528.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:29:40,064 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3546.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:29:46,546 INFO [train.py:904] (4/8) Epoch 1, batch 3550, loss[loss=0.339, simple_loss=0.3761, pruned_loss=0.1509, over 16910.00 frames. ], tot_loss[loss=0.3461, simple_loss=0.3899, pruned_loss=0.1511, over 3313448.60 frames. ], batch size: 96, lr: 4.51e-02, grad_scale: 16.0 2023-04-27 13:29:54,814 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.98 vs. limit=5.0 2023-04-27 13:30:44,944 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3593.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:30:54,496 INFO [train.py:904] (4/8) Epoch 1, batch 3600, loss[loss=0.3262, simple_loss=0.3883, pruned_loss=0.1321, over 17120.00 frames. ], tot_loss[loss=0.342, simple_loss=0.3874, pruned_loss=0.1483, over 3317034.42 frames. ], batch size: 48, lr: 4.50e-02, grad_scale: 16.0 2023-04-27 13:31:04,124 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3607.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-27 13:31:08,702 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.446e+02 5.845e+02 7.592e+02 1.096e+03 2.095e+03, threshold=1.518e+03, percent-clipped=22.0 2023-04-27 13:31:24,588 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-27 13:31:52,479 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=3641.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:31:56,476 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3644.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:32:06,056 INFO [train.py:904] (4/8) Epoch 1, batch 3650, loss[loss=0.3132, simple_loss=0.354, pruned_loss=0.1362, over 16789.00 frames. ], tot_loss[loss=0.3405, simple_loss=0.3847, pruned_loss=0.1481, over 3309776.79 frames. ], batch size: 102, lr: 4.49e-02, grad_scale: 16.0 2023-04-27 13:32:27,711 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.06 vs. limit=5.0 2023-04-27 13:32:42,349 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3675.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:32:43,477 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3676.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:33:03,103 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8796, 2.0081, 2.0790, 2.2094, 2.0166, 2.1432, 2.3354, 2.2115], device='cuda:4'), covar=tensor([0.0288, 0.0383, 0.0216, 0.0251, 0.0209, 0.0134, 0.0203, 0.0246], device='cuda:4'), in_proj_covar=tensor([0.0025, 0.0036, 0.0028, 0.0029, 0.0024, 0.0028, 0.0028, 0.0032], device='cuda:4'), out_proj_covar=tensor([2.5600e-05, 3.3935e-05, 2.7421e-05, 2.8016e-05, 2.2348e-05, 2.2773e-05, 2.4437e-05, 2.8063e-05], device='cuda:4') 2023-04-27 13:33:20,375 INFO [train.py:904] (4/8) Epoch 1, batch 3700, loss[loss=0.3675, simple_loss=0.3906, pruned_loss=0.1722, over 11486.00 frames. ], tot_loss[loss=0.3373, simple_loss=0.3801, pruned_loss=0.1472, over 3290279.99 frames. ], batch size: 247, lr: 4.48e-02, grad_scale: 16.0 2023-04-27 13:33:35,071 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.010e+02 4.933e+02 6.416e+02 8.029e+02 1.327e+03, threshold=1.283e+03, percent-clipped=0.0 2023-04-27 13:33:54,061 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=3723.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:34:15,104 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3737.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-27 13:34:33,709 INFO [train.py:904] (4/8) Epoch 1, batch 3750, loss[loss=0.3178, simple_loss=0.3548, pruned_loss=0.1404, over 16826.00 frames. ], tot_loss[loss=0.3357, simple_loss=0.3777, pruned_loss=0.1469, over 3286041.17 frames. ], batch size: 102, lr: 4.47e-02, grad_scale: 16.0 2023-04-27 13:35:05,166 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1871, 3.6509, 3.4088, 3.2295, 3.6455, 3.4676, 3.4036, 3.5876], device='cuda:4'), covar=tensor([0.0299, 0.0151, 0.0163, 0.0196, 0.0148, 0.0183, 0.0286, 0.0152], device='cuda:4'), in_proj_covar=tensor([0.0046, 0.0030, 0.0031, 0.0040, 0.0033, 0.0034, 0.0040, 0.0036], device='cuda:4'), out_proj_covar=tensor([6.5949e-05, 4.5413e-05, 4.5643e-05, 5.4466e-05, 4.7389e-05, 5.6977e-05, 5.6660e-05, 5.1484e-05], device='cuda:4') 2023-04-27 13:35:35,653 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7861, 4.7194, 4.2881, 4.3178, 4.7750, 4.6551, 4.3279, 4.5979], device='cuda:4'), covar=tensor([0.0123, 0.0101, 0.0105, 0.0250, 0.0054, 0.0153, 0.0089, 0.0112], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0045, 0.0063, 0.0070, 0.0047, 0.0055, 0.0061, 0.0058], device='cuda:4'), out_proj_covar=tensor([8.2026e-05, 6.4211e-05, 1.0741e-04, 1.0605e-04, 6.4036e-05, 8.1066e-05, 9.9351e-05, 9.5279e-05], device='cuda:4') 2023-04-27 13:35:46,283 INFO [train.py:904] (4/8) Epoch 1, batch 3800, loss[loss=0.3994, simple_loss=0.4186, pruned_loss=0.1901, over 16798.00 frames. ], tot_loss[loss=0.3369, simple_loss=0.3781, pruned_loss=0.1479, over 3286684.14 frames. ], batch size: 124, lr: 4.46e-02, grad_scale: 16.0 2023-04-27 13:36:00,501 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.355e+02 5.070e+02 6.208e+02 8.135e+02 1.675e+03, threshold=1.242e+03, percent-clipped=4.0 2023-04-27 13:36:23,985 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1542, 4.3485, 3.8134, 4.3956, 3.7968, 4.4059, 4.1712, 4.3848], device='cuda:4'), covar=tensor([0.1035, 0.1162, 0.1468, 0.0557, 0.1223, 0.0705, 0.0959, 0.0696], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0169, 0.0139, 0.0107, 0.0140, 0.0123, 0.0146, 0.0102], device='cuda:4'), out_proj_covar=tensor([1.3427e-04, 1.6289e-04, 1.2436e-04, 9.4002e-05, 1.2557e-04, 1.1030e-04, 1.4054e-04, 9.9896e-05], device='cuda:4') 2023-04-27 13:36:57,571 INFO [train.py:904] (4/8) Epoch 1, batch 3850, loss[loss=0.3357, simple_loss=0.3721, pruned_loss=0.1497, over 16753.00 frames. ], tot_loss[loss=0.3357, simple_loss=0.3766, pruned_loss=0.1474, over 3275007.74 frames. ], batch size: 83, lr: 4.45e-02, grad_scale: 16.0 2023-04-27 13:37:20,126 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.12 vs. limit=2.0 2023-04-27 13:38:09,799 INFO [train.py:904] (4/8) Epoch 1, batch 3900, loss[loss=0.2528, simple_loss=0.328, pruned_loss=0.08875, over 16704.00 frames. ], tot_loss[loss=0.331, simple_loss=0.3728, pruned_loss=0.1446, over 3274427.44 frames. ], batch size: 57, lr: 4.44e-02, grad_scale: 16.0 2023-04-27 13:38:11,790 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3902.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:38:24,734 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.787e+02 4.914e+02 5.725e+02 7.504e+02 1.784e+03, threshold=1.145e+03, percent-clipped=3.0 2023-04-27 13:38:46,395 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3926.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:39:11,901 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3944.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:39:21,370 INFO [train.py:904] (4/8) Epoch 1, batch 3950, loss[loss=0.3039, simple_loss=0.3497, pruned_loss=0.129, over 16767.00 frames. ], tot_loss[loss=0.3296, simple_loss=0.3713, pruned_loss=0.1439, over 3272780.65 frames. ], batch size: 102, lr: 4.43e-02, grad_scale: 16.0 2023-04-27 13:39:57,545 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0987, 2.9670, 2.5910, 3.8590, 3.2571, 3.7341, 2.5138, 3.5991], device='cuda:4'), covar=tensor([0.1753, 0.0229, 0.0851, 0.0093, 0.0229, 0.0204, 0.0494, 0.0125], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0060, 0.0099, 0.0040, 0.0038, 0.0054, 0.0079, 0.0057], device='cuda:4'), out_proj_covar=tensor([1.5002e-04, 6.5107e-05, 1.1430e-04, 5.1573e-05, 5.2617e-05, 7.1374e-05, 8.9276e-05, 6.3249e-05], device='cuda:4') 2023-04-27 13:40:14,005 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3987.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:40:20,465 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=3992.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:40:20,548 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3748, 4.2923, 3.8736, 4.6638, 4.6174, 4.6882, 4.6131, 4.4199], device='cuda:4'), covar=tensor([0.0376, 0.0291, 0.1504, 0.0420, 0.0508, 0.0283, 0.0413, 0.0369], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0129, 0.0220, 0.0159, 0.0135, 0.0139, 0.0130, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:40:37,508 INFO [train.py:904] (4/8) Epoch 1, batch 4000, loss[loss=0.3208, simple_loss=0.3725, pruned_loss=0.1346, over 17202.00 frames. ], tot_loss[loss=0.3269, simple_loss=0.3689, pruned_loss=0.1424, over 3277524.41 frames. ], batch size: 44, lr: 4.42e-02, grad_scale: 16.0 2023-04-27 13:40:52,171 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.081e+02 4.378e+02 5.733e+02 7.371e+02 1.613e+03, threshold=1.147e+03, percent-clipped=6.0 2023-04-27 13:41:15,408 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-04-27 13:41:24,074 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=4032.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:41:50,490 INFO [train.py:904] (4/8) Epoch 1, batch 4050, loss[loss=0.266, simple_loss=0.3304, pruned_loss=0.1008, over 16666.00 frames. ], tot_loss[loss=0.3186, simple_loss=0.3643, pruned_loss=0.1364, over 3265401.88 frames. ], batch size: 57, lr: 4.41e-02, grad_scale: 16.0 2023-04-27 13:41:57,763 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-27 13:42:21,588 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0704, 3.5882, 3.3312, 3.5863, 2.9202, 2.0711, 3.5931, 3.8824], device='cuda:4'), covar=tensor([0.1851, 0.0752, 0.1085, 0.0336, 0.1556, 0.1832, 0.0368, 0.0107], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0073, 0.0108, 0.0060, 0.0082, 0.0096, 0.0060, 0.0033], device='cuda:4'), out_proj_covar=tensor([1.3455e-04, 8.4052e-05, 1.0584e-04, 5.9500e-05, 9.8139e-05, 9.5232e-05, 6.1628e-05, 3.6992e-05], device='cuda:4') 2023-04-27 13:42:31,901 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.02 vs. limit=5.0 2023-04-27 13:43:04,571 INFO [train.py:904] (4/8) Epoch 1, batch 4100, loss[loss=0.3439, simple_loss=0.3944, pruned_loss=0.1466, over 16693.00 frames. ], tot_loss[loss=0.313, simple_loss=0.362, pruned_loss=0.1321, over 3271783.94 frames. ], batch size: 134, lr: 4.40e-02, grad_scale: 32.0 2023-04-27 13:43:17,320 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-27 13:43:18,961 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.292e+02 4.086e+02 5.427e+02 7.425e+02 1.337e+03, threshold=1.085e+03, percent-clipped=4.0 2023-04-27 13:44:04,990 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=4141.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:44:20,297 INFO [train.py:904] (4/8) Epoch 1, batch 4150, loss[loss=0.3472, simple_loss=0.3973, pruned_loss=0.1485, over 16676.00 frames. ], tot_loss[loss=0.3258, simple_loss=0.3733, pruned_loss=0.1392, over 3250294.93 frames. ], batch size: 57, lr: 4.39e-02, grad_scale: 32.0 2023-04-27 13:44:40,438 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8433, 3.5101, 2.8630, 3.2550, 2.7270, 2.0932, 3.3766, 3.7764], device='cuda:4'), covar=tensor([0.1796, 0.0551, 0.1136, 0.0320, 0.1420, 0.1704, 0.0294, 0.0079], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0075, 0.0108, 0.0062, 0.0082, 0.0097, 0.0060, 0.0033], device='cuda:4'), out_proj_covar=tensor([1.3648e-04, 8.7649e-05, 1.0717e-04, 6.2084e-05, 9.8856e-05, 9.7286e-05, 6.1979e-05, 3.7176e-05], device='cuda:4') 2023-04-27 13:45:37,087 INFO [train.py:904] (4/8) Epoch 1, batch 4200, loss[loss=0.3406, simple_loss=0.4074, pruned_loss=0.1369, over 16902.00 frames. ], tot_loss[loss=0.3333, simple_loss=0.3821, pruned_loss=0.1423, over 3220239.46 frames. ], batch size: 109, lr: 4.38e-02, grad_scale: 16.0 2023-04-27 13:45:39,619 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4202.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:45:39,657 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=4202.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:45:48,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1232, 5.3101, 4.9478, 5.3493, 4.8390, 5.0553, 5.0664, 5.5821], device='cuda:4'), covar=tensor([0.0283, 0.0563, 0.0502, 0.0277, 0.0564, 0.0312, 0.0340, 0.0208], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0158, 0.0137, 0.0104, 0.0134, 0.0117, 0.0141, 0.0097], device='cuda:4'), out_proj_covar=tensor([1.2912e-04, 1.5349e-04, 1.2161e-04, 9.4919e-05, 1.2363e-04, 1.0732e-04, 1.3968e-04, 9.6484e-05], device='cuda:4') 2023-04-27 13:45:52,579 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.534e+02 4.783e+02 6.571e+02 8.274e+02 1.863e+03, threshold=1.314e+03, percent-clipped=9.0 2023-04-27 13:46:50,183 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=4250.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:46:50,856 INFO [train.py:904] (4/8) Epoch 1, batch 4250, loss[loss=0.3113, simple_loss=0.367, pruned_loss=0.1278, over 16300.00 frames. ], tot_loss[loss=0.3342, simple_loss=0.3843, pruned_loss=0.1421, over 3199633.47 frames. ], batch size: 165, lr: 4.36e-02, grad_scale: 16.0 2023-04-27 13:47:09,575 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.97 vs. limit=5.0 2023-04-27 13:47:09,624 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=2.00 vs. limit=2.0 2023-04-27 13:47:37,206 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=4282.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:47:45,262 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=4288.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:47:53,832 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.98 vs. limit=2.0 2023-04-27 13:48:04,724 INFO [train.py:904] (4/8) Epoch 1, batch 4300, loss[loss=0.3278, simple_loss=0.3881, pruned_loss=0.1338, over 16365.00 frames. ], tot_loss[loss=0.3311, simple_loss=0.3837, pruned_loss=0.1392, over 3197279.15 frames. ], batch size: 146, lr: 4.35e-02, grad_scale: 16.0 2023-04-27 13:48:21,408 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.968e+02 4.766e+02 5.895e+02 7.865e+02 1.445e+03, threshold=1.179e+03, percent-clipped=2.0 2023-04-27 13:48:25,674 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3503, 3.5482, 3.3463, 2.7433, 3.4322, 3.4839, 3.4635, 3.0659], device='cuda:4'), covar=tensor([0.0464, 0.0053, 0.0081, 0.0163, 0.0077, 0.0053, 0.0084, 0.0135], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0027, 0.0028, 0.0039, 0.0029, 0.0030, 0.0035, 0.0035], device='cuda:4'), out_proj_covar=tensor([8.2499e-05, 4.4519e-05, 4.5288e-05, 5.8842e-05, 4.7260e-05, 5.1669e-05, 5.4267e-05, 5.3919e-05], device='cuda:4') 2023-04-27 13:48:52,520 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4332.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:48:52,839 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-27 13:49:17,770 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=4349.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:49:20,274 INFO [train.py:904] (4/8) Epoch 1, batch 4350, loss[loss=0.3286, simple_loss=0.3918, pruned_loss=0.1327, over 16766.00 frames. ], tot_loss[loss=0.3352, simple_loss=0.3876, pruned_loss=0.1414, over 3173889.30 frames. ], batch size: 124, lr: 4.34e-02, grad_scale: 16.0 2023-04-27 13:50:04,951 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=4380.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:50:36,405 INFO [train.py:904] (4/8) Epoch 1, batch 4400, loss[loss=0.3294, simple_loss=0.3918, pruned_loss=0.1335, over 17026.00 frames. ], tot_loss[loss=0.3369, simple_loss=0.3887, pruned_loss=0.1425, over 3146985.62 frames. ], batch size: 55, lr: 4.33e-02, grad_scale: 16.0 2023-04-27 13:50:51,959 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.736e+02 5.030e+02 6.587e+02 8.143e+02 1.430e+03, threshold=1.317e+03, percent-clipped=9.0 2023-04-27 13:51:37,256 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.65 vs. limit=2.0 2023-04-27 13:51:48,407 INFO [train.py:904] (4/8) Epoch 1, batch 4450, loss[loss=0.3693, simple_loss=0.4077, pruned_loss=0.1655, over 11774.00 frames. ], tot_loss[loss=0.3363, simple_loss=0.3904, pruned_loss=0.1411, over 3160217.59 frames. ], batch size: 248, lr: 4.32e-02, grad_scale: 16.0 2023-04-27 13:52:57,209 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=4497.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:53:02,592 INFO [train.py:904] (4/8) Epoch 1, batch 4500, loss[loss=0.3099, simple_loss=0.3768, pruned_loss=0.1215, over 16905.00 frames. ], tot_loss[loss=0.3331, simple_loss=0.3885, pruned_loss=0.1389, over 3172102.05 frames. ], batch size: 96, lr: 4.31e-02, grad_scale: 8.0 2023-04-27 13:53:20,026 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.477e+02 4.034e+02 4.961e+02 6.637e+02 1.457e+03, threshold=9.923e+02, percent-clipped=1.0 2023-04-27 13:54:14,092 INFO [train.py:904] (4/8) Epoch 1, batch 4550, loss[loss=0.3636, simple_loss=0.41, pruned_loss=0.1586, over 16695.00 frames. ], tot_loss[loss=0.3317, simple_loss=0.3877, pruned_loss=0.1379, over 3177414.85 frames. ], batch size: 62, lr: 4.30e-02, grad_scale: 8.0 2023-04-27 13:54:57,991 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4582.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:55:25,714 INFO [train.py:904] (4/8) Epoch 1, batch 4600, loss[loss=0.3169, simple_loss=0.3819, pruned_loss=0.126, over 16253.00 frames. ], tot_loss[loss=0.3295, simple_loss=0.3873, pruned_loss=0.1359, over 3197000.68 frames. ], batch size: 165, lr: 4.29e-02, grad_scale: 8.0 2023-04-27 13:55:43,311 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.332e+02 3.980e+02 4.876e+02 6.501e+02 1.584e+03, threshold=9.751e+02, percent-clipped=6.0 2023-04-27 13:56:07,431 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=4630.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:56:28,167 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=4644.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:56:36,565 INFO [train.py:904] (4/8) Epoch 1, batch 4650, loss[loss=0.3065, simple_loss=0.3701, pruned_loss=0.1215, over 16872.00 frames. ], tot_loss[loss=0.3255, simple_loss=0.3838, pruned_loss=0.1336, over 3191629.45 frames. ], batch size: 116, lr: 4.28e-02, grad_scale: 8.0 2023-04-27 13:57:06,608 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3705, 3.8530, 4.0548, 3.0849, 3.6099, 4.0494, 3.8516, 3.6697], device='cuda:4'), covar=tensor([0.0704, 0.0089, 0.0061, 0.0198, 0.0123, 0.0062, 0.0114, 0.0120], device='cuda:4'), in_proj_covar=tensor([0.0059, 0.0027, 0.0027, 0.0040, 0.0028, 0.0028, 0.0035, 0.0036], device='cuda:4'), out_proj_covar=tensor([9.4477e-05, 4.5987e-05, 4.7192e-05, 6.4398e-05, 4.8675e-05, 5.0594e-05, 5.6077e-05, 5.8488e-05], device='cuda:4') 2023-04-27 13:57:33,860 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 13:57:42,705 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-27 13:57:50,180 INFO [train.py:904] (4/8) Epoch 1, batch 4700, loss[loss=0.3088, simple_loss=0.3694, pruned_loss=0.1241, over 16747.00 frames. ], tot_loss[loss=0.3222, simple_loss=0.3807, pruned_loss=0.1318, over 3198498.65 frames. ], batch size: 124, lr: 4.27e-02, grad_scale: 8.0 2023-04-27 13:58:07,984 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.460e+02 4.210e+02 5.632e+02 6.697e+02 1.082e+03, threshold=1.126e+03, percent-clipped=2.0 2023-04-27 13:58:12,824 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6313, 3.3840, 3.2853, 3.9048, 3.1764, 3.6172, 2.7777, 3.1706], device='cuda:4'), covar=tensor([0.0377, 0.0424, 0.0406, 0.0302, 0.1264, 0.0373, 0.0878, 0.0920], device='cuda:4'), in_proj_covar=tensor([0.0063, 0.0069, 0.0060, 0.0062, 0.0123, 0.0067, 0.0085, 0.0072], device='cuda:4'), out_proj_covar=tensor([7.2124e-05, 7.6066e-05, 6.4721e-05, 7.4892e-05, 1.3643e-04, 7.4615e-05, 8.7354e-05, 8.9127e-05], device='cuda:4') 2023-04-27 13:58:27,408 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8100, 1.5582, 1.6666, 2.3368, 1.6240, 2.5759, 2.2590, 2.6489], device='cuda:4'), covar=tensor([0.0071, 0.0546, 0.0185, 0.0155, 0.0149, 0.0110, 0.0136, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0024, 0.0042, 0.0027, 0.0027, 0.0025, 0.0028, 0.0026, 0.0029], device='cuda:4'), out_proj_covar=tensor([2.7012e-05, 5.0232e-05, 2.9081e-05, 3.0752e-05, 2.5648e-05, 2.7969e-05, 2.8431e-05, 3.0669e-05], device='cuda:4') 2023-04-27 13:59:02,499 INFO [train.py:904] (4/8) Epoch 1, batch 4750, loss[loss=0.2587, simple_loss=0.3344, pruned_loss=0.09149, over 16854.00 frames. ], tot_loss[loss=0.3197, simple_loss=0.3777, pruned_loss=0.1308, over 3189117.29 frames. ], batch size: 96, lr: 4.26e-02, grad_scale: 8.0 2023-04-27 13:59:08,332 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0525, 4.8240, 5.3012, 5.4300, 4.4276, 5.2076, 4.7400, 4.8605], device='cuda:4'), covar=tensor([0.0450, 0.0224, 0.0122, 0.0082, 0.0856, 0.0201, 0.0107, 0.0129], device='cuda:4'), in_proj_covar=tensor([0.0079, 0.0060, 0.0104, 0.0078, 0.0131, 0.0085, 0.0073, 0.0081], device='cuda:4'), out_proj_covar=tensor([1.3467e-04, 9.6338e-05, 1.7487e-04, 1.2116e-04, 1.8645e-04, 1.5033e-04, 1.2369e-04, 1.4053e-04], device='cuda:4') 2023-04-27 14:00:11,425 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4797.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:00:16,558 INFO [train.py:904] (4/8) Epoch 1, batch 4800, loss[loss=0.374, simple_loss=0.4241, pruned_loss=0.1619, over 15302.00 frames. ], tot_loss[loss=0.3158, simple_loss=0.3742, pruned_loss=0.1286, over 3189967.64 frames. ], batch size: 190, lr: 4.25e-02, grad_scale: 8.0 2023-04-27 14:00:34,035 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.251e+02 4.322e+02 5.200e+02 6.651e+02 1.076e+03, threshold=1.040e+03, percent-clipped=0.0 2023-04-27 14:00:45,527 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3015, 4.3317, 4.1737, 4.6432, 4.6208, 4.4974, 4.6083, 4.4689], device='cuda:4'), covar=tensor([0.0265, 0.0200, 0.0813, 0.0244, 0.0278, 0.0191, 0.0191, 0.0216], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0127, 0.0216, 0.0152, 0.0128, 0.0132, 0.0116, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:01:22,948 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=4845.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:31,748 INFO [train.py:904] (4/8) Epoch 1, batch 4850, loss[loss=0.2896, simple_loss=0.3638, pruned_loss=0.1078, over 16718.00 frames. ], tot_loss[loss=0.3151, simple_loss=0.3747, pruned_loss=0.1277, over 3192705.90 frames. ], batch size: 76, lr: 4.24e-02, grad_scale: 8.0 2023-04-27 14:01:50,357 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.53 vs. limit=5.0 2023-04-27 14:02:45,444 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9510, 3.8062, 3.9055, 4.1369, 3.1639, 3.7953, 3.1873, 3.2456], device='cuda:4'), covar=tensor([0.0239, 0.0296, 0.0187, 0.0237, 0.1141, 0.0249, 0.0613, 0.0839], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0072, 0.0063, 0.0066, 0.0126, 0.0069, 0.0090, 0.0076], device='cuda:4'), out_proj_covar=tensor([7.9639e-05, 8.0401e-05, 6.9736e-05, 8.3194e-05, 1.4338e-04, 7.9004e-05, 9.3965e-05, 9.4699e-05], device='cuda:4') 2023-04-27 14:02:49,096 INFO [train.py:904] (4/8) Epoch 1, batch 4900, loss[loss=0.3158, simple_loss=0.3688, pruned_loss=0.1313, over 12051.00 frames. ], tot_loss[loss=0.3109, simple_loss=0.3724, pruned_loss=0.1247, over 3189905.60 frames. ], batch size: 246, lr: 4.23e-02, grad_scale: 8.0 2023-04-27 14:03:07,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.068e+02 3.780e+02 4.613e+02 5.959e+02 1.211e+03, threshold=9.227e+02, percent-clipped=3.0 2023-04-27 14:03:54,918 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4944.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:04:04,716 INFO [train.py:904] (4/8) Epoch 1, batch 4950, loss[loss=0.3433, simple_loss=0.3955, pruned_loss=0.1456, over 11849.00 frames. ], tot_loss[loss=0.3125, simple_loss=0.3736, pruned_loss=0.1257, over 3184209.07 frames. ], batch size: 246, lr: 4.21e-02, grad_scale: 8.0 2023-04-27 14:04:37,679 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2355, 1.9116, 2.0059, 2.3435, 2.5250, 2.6867, 2.5761, 2.5990], device='cuda:4'), covar=tensor([0.0111, 0.0844, 0.0398, 0.0168, 0.0149, 0.0174, 0.0122, 0.0161], device='cuda:4'), in_proj_covar=tensor([0.0033, 0.0065, 0.0046, 0.0040, 0.0032, 0.0036, 0.0039, 0.0034], device='cuda:4'), out_proj_covar=tensor([4.8749e-05, 1.1398e-04, 7.7760e-05, 5.7605e-05, 4.8213e-05, 5.5890e-05, 5.3252e-05, 5.0488e-05], device='cuda:4') 2023-04-27 14:04:53,887 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.40 vs. limit=5.0 2023-04-27 14:05:04,811 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=4992.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:05:17,360 INFO [train.py:904] (4/8) Epoch 1, batch 5000, loss[loss=0.2849, simple_loss=0.3607, pruned_loss=0.1046, over 16469.00 frames. ], tot_loss[loss=0.3142, simple_loss=0.3754, pruned_loss=0.1265, over 3181650.85 frames. ], batch size: 68, lr: 4.20e-02, grad_scale: 8.0 2023-04-27 14:05:35,391 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.744e+02 4.820e+02 5.822e+02 7.344e+02 1.526e+03, threshold=1.164e+03, percent-clipped=12.0 2023-04-27 14:06:31,099 INFO [train.py:904] (4/8) Epoch 1, batch 5050, loss[loss=0.3227, simple_loss=0.3865, pruned_loss=0.1294, over 16714.00 frames. ], tot_loss[loss=0.3127, simple_loss=0.375, pruned_loss=0.1252, over 3185244.65 frames. ], batch size: 124, lr: 4.19e-02, grad_scale: 8.0 2023-04-27 14:07:32,890 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3688, 3.2808, 2.9667, 2.9487, 2.4764, 2.0884, 3.3497, 3.8076], device='cuda:4'), covar=tensor([0.1672, 0.0596, 0.0878, 0.0382, 0.1516, 0.1295, 0.0311, 0.0072], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0114, 0.0150, 0.0088, 0.0136, 0.0124, 0.0086, 0.0050], device='cuda:4'), out_proj_covar=tensor([1.8878e-04, 1.3887e-04, 1.5933e-04, 9.8489e-05, 1.6614e-04, 1.3736e-04, 1.0083e-04, 5.9689e-05], device='cuda:4') 2023-04-27 14:07:42,590 INFO [train.py:904] (4/8) Epoch 1, batch 5100, loss[loss=0.2755, simple_loss=0.3475, pruned_loss=0.1018, over 16714.00 frames. ], tot_loss[loss=0.3103, simple_loss=0.3727, pruned_loss=0.1239, over 3187481.05 frames. ], batch size: 124, lr: 4.18e-02, grad_scale: 8.0 2023-04-27 14:07:59,829 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.617e+02 4.246e+02 5.535e+02 6.678e+02 1.197e+03, threshold=1.107e+03, percent-clipped=1.0 2023-04-27 14:08:58,107 INFO [train.py:904] (4/8) Epoch 1, batch 5150, loss[loss=0.2818, simple_loss=0.3623, pruned_loss=0.1007, over 16843.00 frames. ], tot_loss[loss=0.3097, simple_loss=0.3728, pruned_loss=0.1233, over 3173083.34 frames. ], batch size: 102, lr: 4.17e-02, grad_scale: 8.0 2023-04-27 14:09:12,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8981, 4.0825, 3.3826, 1.8000, 2.9266, 2.2461, 3.4892, 3.9799], device='cuda:4'), covar=tensor([0.0182, 0.0208, 0.0357, 0.2008, 0.0833, 0.1239, 0.0695, 0.0108], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0053, 0.0085, 0.0132, 0.0122, 0.0117, 0.0106, 0.0050], device='cuda:4'), out_proj_covar=tensor([1.1424e-04, 8.9656e-05, 1.1590e-04, 1.6398e-04, 1.6212e-04, 1.4898e-04, 1.5534e-04, 8.2151e-05], device='cuda:4') 2023-04-27 14:10:12,903 INFO [train.py:904] (4/8) Epoch 1, batch 5200, loss[loss=0.3152, simple_loss=0.3683, pruned_loss=0.131, over 16721.00 frames. ], tot_loss[loss=0.3102, simple_loss=0.3721, pruned_loss=0.1241, over 3169981.02 frames. ], batch size: 134, lr: 4.16e-02, grad_scale: 8.0 2023-04-27 14:10:30,245 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.582e+02 4.001e+02 4.794e+02 6.093e+02 1.086e+03, threshold=9.588e+02, percent-clipped=0.0 2023-04-27 14:11:26,081 INFO [train.py:904] (4/8) Epoch 1, batch 5250, loss[loss=0.2956, simple_loss=0.357, pruned_loss=0.1171, over 16574.00 frames. ], tot_loss[loss=0.3065, simple_loss=0.3677, pruned_loss=0.1226, over 3184656.14 frames. ], batch size: 62, lr: 4.15e-02, grad_scale: 8.0 2023-04-27 14:11:54,275 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.4635, 2.9132, 2.7751, 2.3804, 2.9712, 2.9258, 2.8801, 2.6387], device='cuda:4'), covar=tensor([0.1143, 0.0120, 0.0139, 0.0234, 0.0109, 0.0111, 0.0202, 0.0226], device='cuda:4'), in_proj_covar=tensor([0.0077, 0.0031, 0.0034, 0.0045, 0.0033, 0.0034, 0.0040, 0.0044], device='cuda:4'), out_proj_covar=tensor([1.3121e-04, 5.8345e-05, 6.2991e-05, 7.8504e-05, 6.1143e-05, 6.8265e-05, 7.3291e-05, 7.7296e-05], device='cuda:4') 2023-04-27 14:12:15,231 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-27 14:12:37,169 INFO [train.py:904] (4/8) Epoch 1, batch 5300, loss[loss=0.2397, simple_loss=0.314, pruned_loss=0.08272, over 16802.00 frames. ], tot_loss[loss=0.3015, simple_loss=0.363, pruned_loss=0.12, over 3199536.19 frames. ], batch size: 83, lr: 4.14e-02, grad_scale: 8.0 2023-04-27 14:12:54,710 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.864e+02 4.537e+02 5.350e+02 6.217e+02 1.130e+03, threshold=1.070e+03, percent-clipped=3.0 2023-04-27 14:13:49,533 INFO [train.py:904] (4/8) Epoch 1, batch 5350, loss[loss=0.319, simple_loss=0.3757, pruned_loss=0.1312, over 16719.00 frames. ], tot_loss[loss=0.2991, simple_loss=0.3607, pruned_loss=0.1188, over 3191691.43 frames. ], batch size: 89, lr: 4.13e-02, grad_scale: 8.0 2023-04-27 14:14:17,883 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8089, 3.5109, 3.3287, 3.1216, 3.6916, 3.3006, 3.3236, 3.5836], device='cuda:4'), covar=tensor([0.0105, 0.0111, 0.0128, 0.0370, 0.0070, 0.0244, 0.0111, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0049, 0.0036, 0.0055, 0.0074, 0.0043, 0.0058, 0.0054, 0.0051], device='cuda:4'), out_proj_covar=tensor([1.0711e-04, 7.5171e-05, 1.2371e-04, 1.4780e-04, 8.1825e-05, 1.2053e-04, 1.1691e-04, 1.1781e-04], device='cuda:4') 2023-04-27 14:15:00,991 INFO [train.py:904] (4/8) Epoch 1, batch 5400, loss[loss=0.3265, simple_loss=0.3858, pruned_loss=0.1336, over 17042.00 frames. ], tot_loss[loss=0.3022, simple_loss=0.3645, pruned_loss=0.12, over 3216413.14 frames. ], batch size: 50, lr: 4.12e-02, grad_scale: 8.0 2023-04-27 14:15:18,319 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.226e+02 4.427e+02 5.350e+02 6.399e+02 9.942e+02, threshold=1.070e+03, percent-clipped=0.0 2023-04-27 14:15:36,074 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5044, 4.1384, 4.0004, 3.6229, 4.4987, 4.1665, 3.9253, 4.3185], device='cuda:4'), covar=tensor([0.0125, 0.0110, 0.0102, 0.0412, 0.0059, 0.0156, 0.0103, 0.0097], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0037, 0.0056, 0.0075, 0.0043, 0.0059, 0.0054, 0.0052], device='cuda:4'), out_proj_covar=tensor([1.1110e-04, 7.7680e-05, 1.2794e-04, 1.5072e-04, 8.3011e-05, 1.2218e-04, 1.1978e-04, 1.2395e-04], device='cuda:4') 2023-04-27 14:15:44,177 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-27 14:16:19,547 INFO [train.py:904] (4/8) Epoch 1, batch 5450, loss[loss=0.3373, simple_loss=0.3832, pruned_loss=0.1457, over 16350.00 frames. ], tot_loss[loss=0.3078, simple_loss=0.3689, pruned_loss=0.1234, over 3196070.07 frames. ], batch size: 35, lr: 4.11e-02, grad_scale: 8.0 2023-04-27 14:17:14,696 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6649, 1.3576, 1.6051, 1.5860, 2.1763, 1.8011, 1.9642, 2.0441], device='cuda:4'), covar=tensor([0.0113, 0.0454, 0.0155, 0.0314, 0.0104, 0.0227, 0.0231, 0.0180], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0054, 0.0034, 0.0033, 0.0032, 0.0037, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([3.3329e-05, 7.4148e-05, 3.9963e-05, 4.1106e-05, 3.4594e-05, 4.1058e-05, 3.6425e-05, 3.7445e-05], device='cuda:4') 2023-04-27 14:17:37,159 INFO [train.py:904] (4/8) Epoch 1, batch 5500, loss[loss=0.3569, simple_loss=0.4106, pruned_loss=0.1516, over 16835.00 frames. ], tot_loss[loss=0.3261, simple_loss=0.3817, pruned_loss=0.1353, over 3147516.75 frames. ], batch size: 102, lr: 4.10e-02, grad_scale: 8.0 2023-04-27 14:17:53,669 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7681, 1.6719, 1.7502, 2.0376, 2.4872, 2.4531, 2.4480, 2.2987], device='cuda:4'), covar=tensor([0.0143, 0.0887, 0.0340, 0.0222, 0.0101, 0.0173, 0.0134, 0.0145], device='cuda:4'), in_proj_covar=tensor([0.0036, 0.0075, 0.0052, 0.0043, 0.0035, 0.0037, 0.0043, 0.0037], device='cuda:4'), out_proj_covar=tensor([5.6134e-05, 1.3342e-04, 9.2815e-05, 6.7141e-05, 5.4570e-05, 5.9256e-05, 6.2770e-05, 5.9799e-05], device='cuda:4') 2023-04-27 14:17:56,197 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.545e+02 5.862e+02 7.575e+02 9.403e+02 2.285e+03, threshold=1.515e+03, percent-clipped=16.0 2023-04-27 14:18:57,385 INFO [train.py:904] (4/8) Epoch 1, batch 5550, loss[loss=0.4569, simple_loss=0.4725, pruned_loss=0.2206, over 15361.00 frames. ], tot_loss[loss=0.3444, simple_loss=0.3942, pruned_loss=0.1473, over 3114901.61 frames. ], batch size: 190, lr: 4.09e-02, grad_scale: 8.0 2023-04-27 14:19:10,304 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4053, 3.3103, 3.6460, 3.7785, 3.9369, 3.5088, 3.5927, 3.7984], device='cuda:4'), covar=tensor([0.0301, 0.0345, 0.0640, 0.0458, 0.0398, 0.0344, 0.0567, 0.0252], device='cuda:4'), in_proj_covar=tensor([0.0108, 0.0100, 0.0129, 0.0124, 0.0138, 0.0108, 0.0135, 0.0099], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:20:17,844 INFO [train.py:904] (4/8) Epoch 1, batch 5600, loss[loss=0.372, simple_loss=0.4183, pruned_loss=0.1629, over 16427.00 frames. ], tot_loss[loss=0.3556, simple_loss=0.4016, pruned_loss=0.1548, over 3075227.99 frames. ], batch size: 146, lr: 4.08e-02, grad_scale: 8.0 2023-04-27 14:20:37,625 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.538e+02 5.637e+02 6.821e+02 8.546e+02 1.933e+03, threshold=1.364e+03, percent-clipped=3.0 2023-04-27 14:20:52,097 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5590, 2.6391, 2.5417, 1.8093, 2.6045, 2.7247, 2.4394, 2.6439], device='cuda:4'), covar=tensor([0.0190, 0.0113, 0.0165, 0.1295, 0.0137, 0.0114, 0.0207, 0.0182], device='cuda:4'), in_proj_covar=tensor([0.0063, 0.0061, 0.0056, 0.0138, 0.0059, 0.0055, 0.0062, 0.0076], device='cuda:4'), out_proj_covar=tensor([8.8597e-05, 8.7622e-05, 8.5529e-05, 1.9529e-04, 9.1248e-05, 8.2976e-05, 1.0297e-04, 1.0755e-04], device='cuda:4') 2023-04-27 14:21:39,869 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=5650.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 14:21:40,596 INFO [train.py:904] (4/8) Epoch 1, batch 5650, loss[loss=0.3886, simple_loss=0.4279, pruned_loss=0.1746, over 16892.00 frames. ], tot_loss[loss=0.3682, simple_loss=0.4096, pruned_loss=0.1633, over 3036041.68 frames. ], batch size: 96, lr: 4.07e-02, grad_scale: 8.0 2023-04-27 14:22:55,616 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0348, 3.4790, 3.3110, 4.1690, 3.1804, 4.1949, 3.2236, 3.1229], device='cuda:4'), covar=tensor([0.0255, 0.0365, 0.0347, 0.0243, 0.1174, 0.0193, 0.0573, 0.1110], device='cuda:4'), in_proj_covar=tensor([0.0077, 0.0081, 0.0070, 0.0081, 0.0149, 0.0080, 0.0100, 0.0092], device='cuda:4'), out_proj_covar=tensor([9.6559e-05, 9.8636e-05, 8.3128e-05, 1.0504e-04, 1.8300e-04, 9.7035e-05, 1.1091e-04, 1.2179e-04], device='cuda:4') 2023-04-27 14:22:59,474 INFO [train.py:904] (4/8) Epoch 1, batch 5700, loss[loss=0.3746, simple_loss=0.4397, pruned_loss=0.1547, over 16752.00 frames. ], tot_loss[loss=0.3707, simple_loss=0.4118, pruned_loss=0.1648, over 3045601.82 frames. ], batch size: 83, lr: 4.06e-02, grad_scale: 8.0 2023-04-27 14:23:15,468 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=5711.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 14:23:17,957 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.964e+02 6.202e+02 7.607e+02 9.700e+02 2.079e+03, threshold=1.521e+03, percent-clipped=5.0 2023-04-27 14:24:21,183 INFO [train.py:904] (4/8) Epoch 1, batch 5750, loss[loss=0.3584, simple_loss=0.4158, pruned_loss=0.1506, over 16630.00 frames. ], tot_loss[loss=0.3728, simple_loss=0.4144, pruned_loss=0.1656, over 3042962.82 frames. ], batch size: 62, lr: 4.05e-02, grad_scale: 8.0 2023-04-27 14:24:28,949 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=5755.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:25:42,935 INFO [train.py:904] (4/8) Epoch 1, batch 5800, loss[loss=0.362, simple_loss=0.3928, pruned_loss=0.1656, over 12187.00 frames. ], tot_loss[loss=0.3693, simple_loss=0.4132, pruned_loss=0.1627, over 3051304.22 frames. ], batch size: 247, lr: 4.04e-02, grad_scale: 8.0 2023-04-27 14:26:01,838 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.312e+02 5.567e+02 6.621e+02 8.933e+02 1.804e+03, threshold=1.324e+03, percent-clipped=2.0 2023-04-27 14:26:06,999 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=5816.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:27:02,341 INFO [train.py:904] (4/8) Epoch 1, batch 5850, loss[loss=0.3956, simple_loss=0.4304, pruned_loss=0.1803, over 15321.00 frames. ], tot_loss[loss=0.3658, simple_loss=0.4106, pruned_loss=0.1605, over 3035718.74 frames. ], batch size: 191, lr: 4.03e-02, grad_scale: 8.0 2023-04-27 14:27:14,307 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=5858.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:28:25,812 INFO [train.py:904] (4/8) Epoch 1, batch 5900, loss[loss=0.3789, simple_loss=0.4157, pruned_loss=0.1711, over 15369.00 frames. ], tot_loss[loss=0.3638, simple_loss=0.4093, pruned_loss=0.1591, over 3040461.91 frames. ], batch size: 190, lr: 4.02e-02, grad_scale: 8.0 2023-04-27 14:28:48,067 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.490e+02 5.425e+02 6.564e+02 8.517e+02 1.690e+03, threshold=1.313e+03, percent-clipped=2.0 2023-04-27 14:28:58,926 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=5919.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:29:49,226 INFO [train.py:904] (4/8) Epoch 1, batch 5950, loss[loss=0.3492, simple_loss=0.403, pruned_loss=0.1477, over 16945.00 frames. ], tot_loss[loss=0.3596, simple_loss=0.4082, pruned_loss=0.1554, over 3056792.77 frames. ], batch size: 109, lr: 4.01e-02, grad_scale: 8.0 2023-04-27 14:29:50,087 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.96 vs. limit=2.0 2023-04-27 14:30:52,711 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=5990.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:31:14,068 INFO [train.py:904] (4/8) Epoch 1, batch 6000, loss[loss=0.3391, simple_loss=0.3873, pruned_loss=0.1455, over 16483.00 frames. ], tot_loss[loss=0.3585, simple_loss=0.4072, pruned_loss=0.1549, over 3082694.67 frames. ], batch size: 146, lr: 4.00e-02, grad_scale: 8.0 2023-04-27 14:31:14,068 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 14:31:23,948 INFO [train.py:938] (4/8) Epoch 1, validation: loss=0.2762, simple_loss=0.3752, pruned_loss=0.08863, over 944034.00 frames. 2023-04-27 14:31:23,949 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17599MB 2023-04-27 14:31:31,586 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6006.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 14:31:41,452 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.294e+02 5.611e+02 7.114e+02 9.006e+02 1.900e+03, threshold=1.423e+03, percent-clipped=2.0 2023-04-27 14:32:08,806 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1634, 3.5261, 3.3621, 1.1499, 3.4921, 3.4786, 3.2020, 3.4731], device='cuda:4'), covar=tensor([0.0214, 0.0174, 0.0215, 0.2498, 0.0178, 0.0143, 0.0191, 0.0199], device='cuda:4'), in_proj_covar=tensor([0.0065, 0.0062, 0.0057, 0.0137, 0.0059, 0.0054, 0.0061, 0.0075], device='cuda:4'), out_proj_covar=tensor([9.5131e-05, 8.9231e-05, 8.9102e-05, 1.9869e-04, 9.4251e-05, 8.5218e-05, 1.0479e-04, 1.1035e-04], device='cuda:4') 2023-04-27 14:32:29,985 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6043.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:32:42,910 INFO [train.py:904] (4/8) Epoch 1, batch 6050, loss[loss=0.3397, simple_loss=0.4005, pruned_loss=0.1394, over 16670.00 frames. ], tot_loss[loss=0.356, simple_loss=0.4055, pruned_loss=0.1533, over 3103280.82 frames. ], batch size: 57, lr: 3.99e-02, grad_scale: 8.0 2023-04-27 14:32:44,231 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6051.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:33:10,676 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.99 vs. limit=2.0 2023-04-27 14:34:03,525 INFO [train.py:904] (4/8) Epoch 1, batch 6100, loss[loss=0.3093, simple_loss=0.3743, pruned_loss=0.1222, over 17071.00 frames. ], tot_loss[loss=0.3522, simple_loss=0.4034, pruned_loss=0.1505, over 3106694.07 frames. ], batch size: 49, lr: 3.98e-02, grad_scale: 8.0 2023-04-27 14:34:09,279 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6104.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:34:21,380 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6111.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:34:24,016 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.820e+02 4.604e+02 6.189e+02 8.357e+02 1.892e+03, threshold=1.238e+03, percent-clipped=2.0 2023-04-27 14:34:59,893 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6135.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:35:09,376 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.93 vs. limit=2.0 2023-04-27 14:35:26,096 INFO [train.py:904] (4/8) Epoch 1, batch 6150, loss[loss=0.3471, simple_loss=0.4007, pruned_loss=0.1468, over 16767.00 frames. ], tot_loss[loss=0.3493, simple_loss=0.4005, pruned_loss=0.1491, over 3103113.90 frames. ], batch size: 124, lr: 3.97e-02, grad_scale: 8.0 2023-04-27 14:35:57,109 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6170.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:36:39,088 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6196.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:36:46,262 INFO [train.py:904] (4/8) Epoch 1, batch 6200, loss[loss=0.3548, simple_loss=0.4014, pruned_loss=0.1541, over 16385.00 frames. ], tot_loss[loss=0.3491, simple_loss=0.3991, pruned_loss=0.1496, over 3081031.98 frames. ], batch size: 35, lr: 3.96e-02, grad_scale: 8.0 2023-04-27 14:36:49,250 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6203.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:37:03,898 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6211.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:37:05,931 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.990e+02 5.160e+02 6.786e+02 8.603e+02 1.824e+03, threshold=1.357e+03, percent-clipped=7.0 2023-04-27 14:37:08,080 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6214.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:37:33,284 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6231.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:38:02,377 INFO [train.py:904] (4/8) Epoch 1, batch 6250, loss[loss=0.4057, simple_loss=0.4367, pruned_loss=0.1874, over 11547.00 frames. ], tot_loss[loss=0.348, simple_loss=0.3983, pruned_loss=0.1488, over 3080371.50 frames. ], batch size: 248, lr: 3.95e-02, grad_scale: 8.0 2023-04-27 14:38:20,958 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6264.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:38:24,197 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6266.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:38:32,071 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:39:17,016 INFO [train.py:904] (4/8) Epoch 1, batch 6300, loss[loss=0.3348, simple_loss=0.3858, pruned_loss=0.1418, over 15487.00 frames. ], tot_loss[loss=0.3482, simple_loss=0.3987, pruned_loss=0.1488, over 3079537.11 frames. ], batch size: 191, lr: 3.94e-02, grad_scale: 8.0 2023-04-27 14:39:25,739 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6306.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 14:39:36,720 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.400e+02 5.487e+02 6.666e+02 8.343e+02 1.856e+03, threshold=1.333e+03, percent-clipped=2.0 2023-04-27 14:39:59,126 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6327.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:40:28,621 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6346.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:40:36,332 INFO [train.py:904] (4/8) Epoch 1, batch 6350, loss[loss=0.4291, simple_loss=0.4311, pruned_loss=0.2135, over 11359.00 frames. ], tot_loss[loss=0.3549, simple_loss=0.4024, pruned_loss=0.1537, over 3048634.28 frames. ], batch size: 246, lr: 3.93e-02, grad_scale: 8.0 2023-04-27 14:40:40,272 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6354.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 14:41:20,010 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9535, 4.2557, 4.0628, 1.7675, 4.1794, 4.2128, 3.9400, 4.0358], device='cuda:4'), covar=tensor([0.0112, 0.0129, 0.0146, 0.1886, 0.0091, 0.0074, 0.0107, 0.0118], device='cuda:4'), in_proj_covar=tensor([0.0072, 0.0064, 0.0063, 0.0143, 0.0060, 0.0054, 0.0066, 0.0080], device='cuda:4'), out_proj_covar=tensor([1.0627e-04, 9.6539e-05, 1.0051e-04, 2.1181e-04, 9.4903e-05, 8.5956e-05, 1.1315e-04, 1.1973e-04], device='cuda:4') 2023-04-27 14:41:38,346 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7796, 4.5199, 4.3971, 2.0336, 3.4822, 2.4473, 4.1886, 4.9551], device='cuda:4'), covar=tensor([0.0226, 0.0200, 0.0251, 0.2151, 0.0857, 0.1338, 0.0670, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0091, 0.0065, 0.0103, 0.0143, 0.0140, 0.0132, 0.0124, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:41:50,566 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6399.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:41:53,485 INFO [train.py:904] (4/8) Epoch 1, batch 6400, loss[loss=0.3004, simple_loss=0.3657, pruned_loss=0.1176, over 16783.00 frames. ], tot_loss[loss=0.3535, simple_loss=0.4008, pruned_loss=0.1531, over 3061414.76 frames. ], batch size: 83, lr: 3.92e-02, grad_scale: 8.0 2023-04-27 14:42:08,609 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6411.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:42:10,410 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 4.039e+02 5.999e+02 7.400e+02 9.043e+02 1.587e+03, threshold=1.480e+03, percent-clipped=2.0 2023-04-27 14:43:09,626 INFO [train.py:904] (4/8) Epoch 1, batch 6450, loss[loss=0.3431, simple_loss=0.4012, pruned_loss=0.1425, over 16511.00 frames. ], tot_loss[loss=0.3472, simple_loss=0.3972, pruned_loss=0.1487, over 3078118.44 frames. ], batch size: 35, lr: 3.91e-02, grad_scale: 8.0 2023-04-27 14:43:21,076 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6770, 4.6193, 4.2696, 1.8735, 3.1412, 2.5470, 3.9607, 4.9348], device='cuda:4'), covar=tensor([0.0247, 0.0327, 0.0282, 0.2234, 0.1095, 0.1363, 0.0863, 0.0133], device='cuda:4'), in_proj_covar=tensor([0.0092, 0.0064, 0.0101, 0.0140, 0.0136, 0.0130, 0.0122, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:43:22,743 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6459.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:43:22,856 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1751, 3.1719, 3.1612, 3.4315, 3.3870, 3.3076, 3.4075, 3.3591], device='cuda:4'), covar=tensor([0.0385, 0.0328, 0.0913, 0.0341, 0.0462, 0.0562, 0.0338, 0.0310], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0153, 0.0239, 0.0169, 0.0145, 0.0149, 0.0127, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:43:26,546 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0488, 4.6413, 4.4084, 4.0362, 4.8518, 3.7678, 4.5734, 4.7819], device='cuda:4'), covar=tensor([0.0086, 0.0071, 0.0087, 0.0362, 0.0047, 0.0350, 0.0070, 0.0096], device='cuda:4'), in_proj_covar=tensor([0.0046, 0.0035, 0.0051, 0.0073, 0.0038, 0.0066, 0.0051, 0.0049], device='cuda:4'), out_proj_covar=tensor([1.1246e-04, 8.5663e-05, 1.2948e-04, 1.6119e-04, 8.7561e-05, 1.5061e-04, 1.2918e-04, 1.3355e-04], device='cuda:4') 2023-04-27 14:43:26,569 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6462.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:43:33,802 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6467.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:44:12,250 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6491.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:44:26,971 INFO [train.py:904] (4/8) Epoch 1, batch 6500, loss[loss=0.3841, simple_loss=0.4096, pruned_loss=0.1793, over 11240.00 frames. ], tot_loss[loss=0.3455, simple_loss=0.3953, pruned_loss=0.1478, over 3064269.03 frames. ], batch size: 248, lr: 3.90e-02, grad_scale: 16.0 2023-04-27 14:44:45,181 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.861e+02 4.801e+02 5.896e+02 8.038e+02 1.836e+03, threshold=1.179e+03, percent-clipped=2.0 2023-04-27 14:44:46,906 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6514.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:45:00,321 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6523.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 14:45:05,060 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6526.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:45:07,595 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6528.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:45:35,798 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.47 vs. limit=5.0 2023-04-27 14:45:42,981 INFO [train.py:904] (4/8) Epoch 1, batch 6550, loss[loss=0.3212, simple_loss=0.3966, pruned_loss=0.1229, over 16318.00 frames. ], tot_loss[loss=0.3481, simple_loss=0.3989, pruned_loss=0.1486, over 3082292.52 frames. ], batch size: 146, lr: 3.89e-02, grad_scale: 16.0 2023-04-27 14:45:55,587 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6559.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:46:00,517 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6562.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:46:08,766 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6567.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:46:44,108 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.02 vs. limit=2.0 2023-04-27 14:46:51,027 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.25 vs. limit=5.0 2023-04-27 14:46:51,979 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2968, 4.2691, 4.2780, 4.6774, 4.6322, 4.5483, 4.6991, 4.4371], device='cuda:4'), covar=tensor([0.0427, 0.0339, 0.0998, 0.0286, 0.0388, 0.0262, 0.0255, 0.0313], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0156, 0.0245, 0.0173, 0.0143, 0.0156, 0.0131, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:46:59,692 INFO [train.py:904] (4/8) Epoch 1, batch 6600, loss[loss=0.3658, simple_loss=0.4157, pruned_loss=0.1579, over 16505.00 frames. ], tot_loss[loss=0.3496, simple_loss=0.4014, pruned_loss=0.149, over 3103544.42 frames. ], batch size: 146, lr: 3.89e-02, grad_scale: 16.0 2023-04-27 14:47:18,180 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.822e+02 5.266e+02 6.465e+02 8.044e+02 1.550e+03, threshold=1.293e+03, percent-clipped=3.0 2023-04-27 14:47:32,784 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6622.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:47:43,277 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6629.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:48:10,727 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6646.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:48:18,078 INFO [train.py:904] (4/8) Epoch 1, batch 6650, loss[loss=0.3841, simple_loss=0.4205, pruned_loss=0.1739, over 16189.00 frames. ], tot_loss[loss=0.3519, simple_loss=0.4023, pruned_loss=0.1508, over 3089970.69 frames. ], batch size: 165, lr: 3.88e-02, grad_scale: 16.0 2023-04-27 14:48:40,818 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7346, 3.1121, 2.9909, 1.5073, 3.1367, 3.1401, 2.8866, 2.8642], device='cuda:4'), covar=tensor([0.0262, 0.0117, 0.0210, 0.2119, 0.0100, 0.0074, 0.0238, 0.0246], device='cuda:4'), in_proj_covar=tensor([0.0075, 0.0064, 0.0060, 0.0142, 0.0059, 0.0053, 0.0065, 0.0080], device='cuda:4'), out_proj_covar=tensor([1.1320e-04, 9.8695e-05, 9.8374e-05, 2.1154e-04, 9.5945e-05, 8.6293e-05, 1.1334e-04, 1.2237e-04], device='cuda:4') 2023-04-27 14:49:18,687 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6690.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:49:24,165 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6694.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:49:27,254 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.69 vs. limit=2.0 2023-04-27 14:49:32,173 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6699.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:49:34,321 INFO [train.py:904] (4/8) Epoch 1, batch 6700, loss[loss=0.3451, simple_loss=0.388, pruned_loss=0.1511, over 16415.00 frames. ], tot_loss[loss=0.3509, simple_loss=0.4009, pruned_loss=0.1505, over 3089076.21 frames. ], batch size: 35, lr: 3.87e-02, grad_scale: 16.0 2023-04-27 14:49:48,764 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.88 vs. limit=5.0 2023-04-27 14:49:52,616 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.128e+02 5.771e+02 7.025e+02 8.722e+02 1.711e+03, threshold=1.405e+03, percent-clipped=7.0 2023-04-27 14:50:27,195 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.69 vs. limit=2.0 2023-04-27 14:50:45,164 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6747.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:50:50,989 INFO [train.py:904] (4/8) Epoch 1, batch 6750, loss[loss=0.3082, simple_loss=0.364, pruned_loss=0.1262, over 16681.00 frames. ], tot_loss[loss=0.352, simple_loss=0.4006, pruned_loss=0.1517, over 3053712.43 frames. ], batch size: 62, lr: 3.86e-02, grad_scale: 16.0 2023-04-27 14:51:50,858 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6791.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:52:05,969 INFO [train.py:904] (4/8) Epoch 1, batch 6800, loss[loss=0.3254, simple_loss=0.3985, pruned_loss=0.1261, over 17184.00 frames. ], tot_loss[loss=0.3486, simple_loss=0.3985, pruned_loss=0.1494, over 3070276.65 frames. ], batch size: 45, lr: 3.85e-02, grad_scale: 16.0 2023-04-27 14:52:24,955 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.031e+02 5.369e+02 6.197e+02 7.365e+02 1.417e+03, threshold=1.239e+03, percent-clipped=1.0 2023-04-27 14:52:33,196 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6818.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 14:52:41,649 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6823.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:52:46,850 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6826.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:53:06,062 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6839.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:53:06,501 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-27 14:53:23,170 INFO [train.py:904] (4/8) Epoch 1, batch 6850, loss[loss=0.3025, simple_loss=0.3843, pruned_loss=0.1104, over 17268.00 frames. ], tot_loss[loss=0.3482, simple_loss=0.399, pruned_loss=0.1487, over 3089089.68 frames. ], batch size: 52, lr: 3.84e-02, grad_scale: 16.0 2023-04-27 14:53:35,280 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6859.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:53:46,383 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6867.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:53:57,788 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6874.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:54:03,471 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6878.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 14:54:37,391 INFO [train.py:904] (4/8) Epoch 1, batch 6900, loss[loss=0.4865, simple_loss=0.4776, pruned_loss=0.2477, over 11295.00 frames. ], tot_loss[loss=0.3475, simple_loss=0.4006, pruned_loss=0.1472, over 3100417.38 frames. ], batch size: 248, lr: 3.83e-02, grad_scale: 16.0 2023-04-27 14:54:47,053 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6907.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:54:49,771 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6909.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:54:55,475 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.157e+02 5.195e+02 6.018e+02 7.768e+02 1.319e+03, threshold=1.204e+03, percent-clipped=1.0 2023-04-27 14:54:59,971 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6915.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:55:09,425 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6922.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:55:22,399 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9399, 1.4307, 1.6412, 1.3608, 1.8561, 1.8656, 2.0465, 2.1650], device='cuda:4'), covar=tensor([0.0107, 0.0557, 0.0207, 0.0304, 0.0209, 0.0259, 0.0195, 0.0152], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0063, 0.0042, 0.0044, 0.0037, 0.0044, 0.0032, 0.0033], device='cuda:4'), out_proj_covar=tensor([3.8109e-05, 9.3000e-05, 5.7915e-05, 6.0536e-05, 4.9988e-05, 5.9179e-05, 4.6600e-05, 4.6785e-05], device='cuda:4') 2023-04-27 14:55:36,575 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6939.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 14:55:54,253 INFO [train.py:904] (4/8) Epoch 1, batch 6950, loss[loss=0.4316, simple_loss=0.4466, pruned_loss=0.2083, over 11132.00 frames. ], tot_loss[loss=0.3529, simple_loss=0.4039, pruned_loss=0.151, over 3080953.92 frames. ], batch size: 246, lr: 3.82e-02, grad_scale: 16.0 2023-04-27 14:56:16,813 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0928, 3.8847, 3.7786, 3.1534, 3.9688, 3.9187, 4.0448, 2.2988], device='cuda:4'), covar=tensor([0.1261, 0.0088, 0.0092, 0.0322, 0.0068, 0.0071, 0.0069, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0104, 0.0042, 0.0042, 0.0068, 0.0039, 0.0039, 0.0044, 0.0070], device='cuda:4'), out_proj_covar=tensor([1.9985e-04, 8.6991e-05, 9.1653e-05, 1.3756e-04, 8.2336e-05, 8.8481e-05, 9.0741e-05, 1.4312e-04], device='cuda:4') 2023-04-27 14:56:25,068 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=6970.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:56:25,213 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6970.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:56:47,680 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6985.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:57:12,232 INFO [train.py:904] (4/8) Epoch 1, batch 7000, loss[loss=0.4243, simple_loss=0.4352, pruned_loss=0.2067, over 11531.00 frames. ], tot_loss[loss=0.3494, simple_loss=0.4024, pruned_loss=0.1482, over 3099673.06 frames. ], batch size: 250, lr: 3.81e-02, grad_scale: 16.0 2023-04-27 14:57:30,860 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.210e+02 5.287e+02 6.266e+02 8.014e+02 1.368e+03, threshold=1.253e+03, percent-clipped=3.0 2023-04-27 14:57:33,121 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8021, 3.4433, 3.7210, 3.8236, 3.2211, 3.6556, 3.5313, 3.4589], device='cuda:4'), covar=tensor([0.0343, 0.0199, 0.0181, 0.0124, 0.0773, 0.0211, 0.0527, 0.0214], device='cuda:4'), in_proj_covar=tensor([0.0086, 0.0063, 0.0115, 0.0089, 0.0139, 0.0091, 0.0081, 0.0095], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:57:37,931 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-27 14:57:41,905 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5069, 3.5926, 3.1267, 1.8030, 2.6614, 1.9690, 2.9291, 3.4472], device='cuda:4'), covar=tensor([0.0306, 0.0276, 0.0298, 0.1929, 0.0923, 0.1259, 0.0993, 0.0249], device='cuda:4'), in_proj_covar=tensor([0.0093, 0.0069, 0.0104, 0.0148, 0.0141, 0.0134, 0.0131, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:58:27,037 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7048.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 14:58:28,868 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1230, 1.2248, 1.6378, 1.1802, 1.7298, 1.7193, 2.0346, 1.8777], device='cuda:4'), covar=tensor([0.0056, 0.0378, 0.0166, 0.0262, 0.0116, 0.0233, 0.0119, 0.0136], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0063, 0.0042, 0.0045, 0.0037, 0.0045, 0.0032, 0.0034], device='cuda:4'), out_proj_covar=tensor([3.5869e-05, 9.2507e-05, 5.7689e-05, 6.2209e-05, 5.0003e-05, 6.0321e-05, 4.7070e-05, 4.7639e-05], device='cuda:4') 2023-04-27 14:58:31,408 INFO [train.py:904] (4/8) Epoch 1, batch 7050, loss[loss=0.3682, simple_loss=0.4182, pruned_loss=0.1591, over 16841.00 frames. ], tot_loss[loss=0.3513, simple_loss=0.4041, pruned_loss=0.1492, over 3099239.21 frames. ], batch size: 116, lr: 3.80e-02, grad_scale: 16.0 2023-04-27 14:58:53,378 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2114, 2.8739, 2.6458, 3.4509, 2.7104, 3.3760, 2.7280, 2.4661], device='cuda:4'), covar=tensor([0.0313, 0.0297, 0.0264, 0.0251, 0.0859, 0.0172, 0.0472, 0.0883], device='cuda:4'), in_proj_covar=tensor([0.0104, 0.0101, 0.0085, 0.0111, 0.0179, 0.0097, 0.0122, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 14:59:02,047 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3062, 3.1194, 3.0284, 2.8312, 3.3010, 2.5691, 3.0303, 3.1666], device='cuda:4'), covar=tensor([0.0113, 0.0091, 0.0118, 0.0315, 0.0065, 0.0492, 0.0094, 0.0106], device='cuda:4'), in_proj_covar=tensor([0.0049, 0.0038, 0.0054, 0.0076, 0.0041, 0.0075, 0.0052, 0.0053], device='cuda:4'), out_proj_covar=tensor([1.2733e-04, 9.9530e-05, 1.4282e-04, 1.7604e-04, 1.0086e-04, 1.7707e-04, 1.3501e-04, 1.5406e-04], device='cuda:4') 2023-04-27 14:59:51,865 INFO [train.py:904] (4/8) Epoch 1, batch 7100, loss[loss=0.4033, simple_loss=0.4178, pruned_loss=0.1945, over 11537.00 frames. ], tot_loss[loss=0.3488, simple_loss=0.4015, pruned_loss=0.1481, over 3094269.73 frames. ], batch size: 248, lr: 3.79e-02, grad_scale: 16.0 2023-04-27 15:00:05,804 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7109.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 15:00:07,388 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.92 vs. limit=2.0 2023-04-27 15:00:11,238 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.812e+02 5.623e+02 6.703e+02 8.193e+02 2.007e+03, threshold=1.341e+03, percent-clipped=3.0 2023-04-27 15:00:20,537 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7118.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 15:00:27,638 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7123.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:01:11,086 INFO [train.py:904] (4/8) Epoch 1, batch 7150, loss[loss=0.4246, simple_loss=0.4335, pruned_loss=0.2078, over 11438.00 frames. ], tot_loss[loss=0.3461, simple_loss=0.3985, pruned_loss=0.1469, over 3095361.01 frames. ], batch size: 246, lr: 3.78e-02, grad_scale: 8.0 2023-04-27 15:01:30,094 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7163.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:01:34,845 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=7166.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:01:41,726 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=7171.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:01:57,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8583, 2.6731, 2.2184, 3.1681, 2.4712, 2.8932, 2.5470, 2.2240], device='cuda:4'), covar=tensor([0.0318, 0.0293, 0.0272, 0.0284, 0.0842, 0.0257, 0.0525, 0.1012], device='cuda:4'), in_proj_covar=tensor([0.0106, 0.0104, 0.0087, 0.0114, 0.0186, 0.0101, 0.0128, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 15:02:27,356 INFO [train.py:904] (4/8) Epoch 1, batch 7200, loss[loss=0.2953, simple_loss=0.3597, pruned_loss=0.1154, over 16365.00 frames. ], tot_loss[loss=0.3429, simple_loss=0.3959, pruned_loss=0.145, over 3082982.90 frames. ], batch size: 146, lr: 3.78e-02, grad_scale: 8.0 2023-04-27 15:02:46,726 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.108e+02 4.516e+02 5.501e+02 7.177e+02 1.508e+03, threshold=1.100e+03, percent-clipped=3.0 2023-04-27 15:02:48,416 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8673, 4.0415, 3.8098, 3.9827, 3.4785, 3.8286, 3.7993, 3.9989], device='cuda:4'), covar=tensor([0.0376, 0.0617, 0.0675, 0.0291, 0.0619, 0.0491, 0.0489, 0.0561], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0196, 0.0188, 0.0124, 0.0159, 0.0126, 0.0172, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 15:02:57,729 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0367, 2.8840, 3.3460, 3.3015, 3.3892, 3.0076, 3.2392, 3.3160], device='cuda:4'), covar=tensor([0.0292, 0.0416, 0.0443, 0.0474, 0.0391, 0.0404, 0.0567, 0.0280], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0116, 0.0134, 0.0133, 0.0145, 0.0119, 0.0166, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:03:02,475 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7224.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 15:03:19,584 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7234.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 15:03:46,512 INFO [train.py:904] (4/8) Epoch 1, batch 7250, loss[loss=0.273, simple_loss=0.3425, pruned_loss=0.1018, over 16783.00 frames. ], tot_loss[loss=0.3384, simple_loss=0.3921, pruned_loss=0.1424, over 3084973.26 frames. ], batch size: 102, lr: 3.77e-02, grad_scale: 8.0 2023-04-27 15:04:06,709 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7265.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:04:21,327 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7275.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:04:35,838 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7285.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:04:58,721 INFO [train.py:904] (4/8) Epoch 1, batch 7300, loss[loss=0.299, simple_loss=0.374, pruned_loss=0.112, over 16879.00 frames. ], tot_loss[loss=0.3356, simple_loss=0.3898, pruned_loss=0.1408, over 3081376.16 frames. ], batch size: 109, lr: 3.76e-02, grad_scale: 8.0 2023-04-27 15:05:13,205 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6820, 3.4139, 3.4817, 3.7521, 2.7906, 3.5679, 3.4909, 3.3890], device='cuda:4'), covar=tensor([0.0340, 0.0237, 0.0331, 0.0188, 0.1302, 0.0292, 0.0524, 0.0264], device='cuda:4'), in_proj_covar=tensor([0.0090, 0.0063, 0.0115, 0.0089, 0.0140, 0.0092, 0.0082, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:05:19,383 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.541e+02 5.301e+02 6.672e+02 8.069e+02 1.507e+03, threshold=1.334e+03, percent-clipped=6.0 2023-04-27 15:05:47,852 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0486, 4.5011, 3.8815, 1.4719, 4.6314, 4.5377, 3.1752, 3.7002], device='cuda:4'), covar=tensor([0.0497, 0.0107, 0.0316, 0.2782, 0.0070, 0.0075, 0.0345, 0.0276], device='cuda:4'), in_proj_covar=tensor([0.0085, 0.0066, 0.0065, 0.0150, 0.0065, 0.0057, 0.0072, 0.0082], device='cuda:4'), out_proj_covar=tensor([1.3407e-04, 1.0580e-04, 1.0953e-04, 2.2915e-04, 1.0936e-04, 9.6667e-05, 1.3020e-04, 1.3148e-04], device='cuda:4') 2023-04-27 15:05:49,057 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=7333.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:05:53,687 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7336.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:06:15,327 INFO [train.py:904] (4/8) Epoch 1, batch 7350, loss[loss=0.3021, simple_loss=0.3772, pruned_loss=0.1135, over 16850.00 frames. ], tot_loss[loss=0.3319, simple_loss=0.3869, pruned_loss=0.1385, over 3063207.16 frames. ], batch size: 83, lr: 3.75e-02, grad_scale: 8.0 2023-04-27 15:06:17,194 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6865, 3.2919, 3.0619, 4.0865, 2.8148, 3.7851, 2.7875, 2.5904], device='cuda:4'), covar=tensor([0.0338, 0.0368, 0.0310, 0.0269, 0.1417, 0.0249, 0.0745, 0.1505], device='cuda:4'), in_proj_covar=tensor([0.0110, 0.0109, 0.0091, 0.0121, 0.0193, 0.0104, 0.0131, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:07:21,277 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.39 vs. limit=5.0 2023-04-27 15:07:26,839 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.97 vs. limit=5.0 2023-04-27 15:07:30,952 INFO [train.py:904] (4/8) Epoch 1, batch 7400, loss[loss=0.3875, simple_loss=0.4231, pruned_loss=0.1759, over 15330.00 frames. ], tot_loss[loss=0.3346, simple_loss=0.3892, pruned_loss=0.14, over 3072608.45 frames. ], batch size: 192, lr: 3.74e-02, grad_scale: 8.0 2023-04-27 15:07:36,106 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7404.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 15:07:40,480 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7407.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:07:50,855 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.431e+02 4.999e+02 6.259e+02 7.546e+02 1.554e+03, threshold=1.252e+03, percent-clipped=1.0 2023-04-27 15:08:03,985 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-27 15:08:36,685 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3917, 5.0732, 4.8953, 5.0480, 5.0266, 5.4580, 5.2923, 5.0286], device='cuda:4'), covar=tensor([0.0659, 0.0867, 0.0850, 0.1112, 0.1587, 0.0614, 0.0669, 0.1502], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0200, 0.0177, 0.0181, 0.0221, 0.0169, 0.0161, 0.0240], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:08:52,556 INFO [train.py:904] (4/8) Epoch 1, batch 7450, loss[loss=0.3703, simple_loss=0.4118, pruned_loss=0.1644, over 16425.00 frames. ], tot_loss[loss=0.3362, simple_loss=0.3906, pruned_loss=0.141, over 3088699.12 frames. ], batch size: 75, lr: 3.73e-02, grad_scale: 8.0 2023-04-27 15:09:14,184 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3591, 4.7651, 4.6185, 4.7526, 4.6878, 5.1646, 5.0020, 4.7114], device='cuda:4'), covar=tensor([0.0766, 0.0965, 0.0958, 0.1139, 0.1798, 0.0601, 0.0693, 0.1421], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0203, 0.0179, 0.0182, 0.0223, 0.0172, 0.0163, 0.0240], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:09:22,511 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7468.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:10:15,193 INFO [train.py:904] (4/8) Epoch 1, batch 7500, loss[loss=0.3182, simple_loss=0.3823, pruned_loss=0.1271, over 16810.00 frames. ], tot_loss[loss=0.3382, simple_loss=0.3922, pruned_loss=0.1421, over 3081525.93 frames. ], batch size: 102, lr: 3.72e-02, grad_scale: 8.0 2023-04-27 15:10:33,066 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0492, 3.0103, 1.4209, 3.0207, 1.8995, 3.0429, 1.6073, 2.3155], device='cuda:4'), covar=tensor([0.0081, 0.0133, 0.1595, 0.0081, 0.0860, 0.0170, 0.1420, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0060, 0.0064, 0.0136, 0.0061, 0.0117, 0.0070, 0.0145, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:10:35,036 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.564e+02 5.230e+02 6.451e+02 8.258e+02 1.814e+03, threshold=1.290e+03, percent-clipped=2.0 2023-04-27 15:10:40,735 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6989, 4.5760, 4.1633, 3.8498, 4.5588, 2.9868, 4.1340, 4.4821], device='cuda:4'), covar=tensor([0.0073, 0.0072, 0.0083, 0.0325, 0.0050, 0.0639, 0.0091, 0.0100], device='cuda:4'), in_proj_covar=tensor([0.0047, 0.0037, 0.0054, 0.0077, 0.0040, 0.0078, 0.0051, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-27 15:10:43,266 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7519.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 15:11:06,777 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7534.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 15:11:31,890 INFO [train.py:904] (4/8) Epoch 1, batch 7550, loss[loss=0.3247, simple_loss=0.3788, pruned_loss=0.1353, over 16684.00 frames. ], tot_loss[loss=0.3375, simple_loss=0.3912, pruned_loss=0.142, over 3086849.03 frames. ], batch size: 89, lr: 3.72e-02, grad_scale: 4.0 2023-04-27 15:11:54,746 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7565.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:12:21,505 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=7582.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 15:12:50,411 INFO [train.py:904] (4/8) Epoch 1, batch 7600, loss[loss=0.3817, simple_loss=0.4133, pruned_loss=0.175, over 15325.00 frames. ], tot_loss[loss=0.3376, simple_loss=0.3908, pruned_loss=0.1421, over 3104334.77 frames. ], batch size: 190, lr: 3.71e-02, grad_scale: 8.0 2023-04-27 15:13:01,713 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2974, 1.5548, 1.6698, 1.2503, 2.0438, 2.0543, 2.2488, 2.0350], device='cuda:4'), covar=tensor([0.0051, 0.0434, 0.0190, 0.0338, 0.0129, 0.0225, 0.0093, 0.0149], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0068, 0.0048, 0.0051, 0.0043, 0.0053, 0.0033, 0.0037], device='cuda:4'), out_proj_covar=tensor([4.0777e-05, 1.0106e-04, 6.7564e-05, 7.4597e-05, 6.0919e-05, 7.4670e-05, 5.0007e-05, 5.5636e-05], device='cuda:4') 2023-04-27 15:13:10,042 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=7613.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:13:12,789 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.798e+02 5.452e+02 6.510e+02 8.243e+02 1.443e+03, threshold=1.302e+03, percent-clipped=3.0 2023-04-27 15:13:29,690 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7624.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:13:40,332 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7631.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:14:13,721 INFO [train.py:904] (4/8) Epoch 1, batch 7650, loss[loss=0.2952, simple_loss=0.3735, pruned_loss=0.1084, over 16469.00 frames. ], tot_loss[loss=0.338, simple_loss=0.3909, pruned_loss=0.1425, over 3102148.83 frames. ], batch size: 35, lr: 3.70e-02, grad_scale: 8.0 2023-04-27 15:14:24,543 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5554, 2.5043, 1.6926, 2.6503, 1.8891, 2.5502, 1.8202, 2.2778], device='cuda:4'), covar=tensor([0.0096, 0.0154, 0.1035, 0.0083, 0.0600, 0.0273, 0.1003, 0.0483], device='cuda:4'), in_proj_covar=tensor([0.0060, 0.0063, 0.0135, 0.0060, 0.0116, 0.0069, 0.0142, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:15:14,472 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7685.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:15:30,501 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3907, 3.1175, 3.2414, 3.4005, 2.9952, 3.2941, 3.1921, 3.0668], device='cuda:4'), covar=tensor([0.0253, 0.0171, 0.0169, 0.0125, 0.0591, 0.0174, 0.0498, 0.0199], device='cuda:4'), in_proj_covar=tensor([0.0091, 0.0065, 0.0118, 0.0094, 0.0145, 0.0095, 0.0085, 0.0098], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:15:39,928 INFO [train.py:904] (4/8) Epoch 1, batch 7700, loss[loss=0.3221, simple_loss=0.378, pruned_loss=0.1331, over 16414.00 frames. ], tot_loss[loss=0.3405, simple_loss=0.3922, pruned_loss=0.1444, over 3079636.55 frames. ], batch size: 146, lr: 3.69e-02, grad_scale: 8.0 2023-04-27 15:15:45,452 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7704.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 15:16:00,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.630e+02 5.474e+02 6.817e+02 8.632e+02 3.010e+03, threshold=1.363e+03, percent-clipped=1.0 2023-04-27 15:16:45,713 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7743.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:16:57,055 INFO [train.py:904] (4/8) Epoch 1, batch 7750, loss[loss=0.3769, simple_loss=0.4242, pruned_loss=0.1648, over 16229.00 frames. ], tot_loss[loss=0.3413, simple_loss=0.3928, pruned_loss=0.1449, over 3081052.72 frames. ], batch size: 165, lr: 3.68e-02, grad_scale: 8.0 2023-04-27 15:16:59,251 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=7752.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 15:17:16,057 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7763.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:17:18,223 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.01 vs. limit=2.0 2023-04-27 15:18:14,288 INFO [train.py:904] (4/8) Epoch 1, batch 7800, loss[loss=0.3195, simple_loss=0.3688, pruned_loss=0.1351, over 16707.00 frames. ], tot_loss[loss=0.3441, simple_loss=0.3943, pruned_loss=0.1469, over 3058952.65 frames. ], batch size: 134, lr: 3.67e-02, grad_scale: 8.0 2023-04-27 15:18:16,659 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7802.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:18:19,996 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7804.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:18:36,278 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.181e+02 4.983e+02 6.289e+02 7.590e+02 1.248e+03, threshold=1.258e+03, percent-clipped=0.0 2023-04-27 15:18:43,958 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7819.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 15:19:31,859 INFO [train.py:904] (4/8) Epoch 1, batch 7850, loss[loss=0.3338, simple_loss=0.3964, pruned_loss=0.1356, over 16832.00 frames. ], tot_loss[loss=0.3434, simple_loss=0.3949, pruned_loss=0.1459, over 3068736.18 frames. ], batch size: 102, lr: 3.66e-02, grad_scale: 8.0 2023-04-27 15:19:51,388 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7863.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:19:56,865 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=7867.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:20:49,083 INFO [train.py:904] (4/8) Epoch 1, batch 7900, loss[loss=0.3248, simple_loss=0.3925, pruned_loss=0.1286, over 16855.00 frames. ], tot_loss[loss=0.3395, simple_loss=0.3924, pruned_loss=0.1433, over 3085330.99 frames. ], batch size: 116, lr: 3.66e-02, grad_scale: 8.0 2023-04-27 15:21:11,994 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.293e+02 4.952e+02 6.401e+02 7.784e+02 1.850e+03, threshold=1.280e+03, percent-clipped=3.0 2023-04-27 15:21:17,307 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0825, 2.0162, 2.4169, 3.0244, 3.5101, 3.2019, 1.9159, 3.4657], device='cuda:4'), covar=tensor([0.0055, 0.0441, 0.0249, 0.0107, 0.0055, 0.0107, 0.0243, 0.0064], device='cuda:4'), in_proj_covar=tensor([0.0044, 0.0092, 0.0072, 0.0052, 0.0043, 0.0045, 0.0064, 0.0041], device='cuda:4'), out_proj_covar=tensor([8.1944e-05, 1.6527e-04, 1.3927e-04, 9.8238e-05, 7.7300e-05, 8.2784e-05, 1.1008e-04, 7.5769e-05], device='cuda:4') 2023-04-27 15:21:39,249 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7931.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:22:08,691 INFO [train.py:904] (4/8) Epoch 1, batch 7950, loss[loss=0.2692, simple_loss=0.3295, pruned_loss=0.1044, over 17065.00 frames. ], tot_loss[loss=0.3401, simple_loss=0.3924, pruned_loss=0.1439, over 3080575.07 frames. ], batch size: 53, lr: 3.65e-02, grad_scale: 8.0 2023-04-27 15:22:33,055 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.74 vs. limit=5.0 2023-04-27 15:22:54,878 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=7979.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:22:56,122 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7980.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:23:10,184 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1113, 2.6894, 2.5498, 1.6927, 2.7543, 2.7549, 2.2200, 2.5624], device='cuda:4'), covar=tensor([0.0464, 0.0120, 0.0219, 0.1514, 0.0119, 0.0080, 0.0371, 0.0209], device='cuda:4'), in_proj_covar=tensor([0.0088, 0.0064, 0.0066, 0.0147, 0.0064, 0.0054, 0.0073, 0.0084], device='cuda:4'), out_proj_covar=tensor([1.4408e-04, 1.0916e-04, 1.1345e-04, 2.3003e-04, 1.1290e-04, 9.6690e-05, 1.3586e-04, 1.3979e-04], device='cuda:4') 2023-04-27 15:23:30,885 INFO [train.py:904] (4/8) Epoch 1, batch 8000, loss[loss=0.3409, simple_loss=0.4014, pruned_loss=0.1402, over 16832.00 frames. ], tot_loss[loss=0.3411, simple_loss=0.3926, pruned_loss=0.1448, over 3070469.68 frames. ], batch size: 102, lr: 3.64e-02, grad_scale: 8.0 2023-04-27 15:23:51,330 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.520e+02 5.080e+02 6.216e+02 7.580e+02 1.182e+03, threshold=1.243e+03, percent-clipped=0.0 2023-04-27 15:24:45,921 INFO [train.py:904] (4/8) Epoch 1, batch 8050, loss[loss=0.3628, simple_loss=0.4113, pruned_loss=0.1571, over 16749.00 frames. ], tot_loss[loss=0.3408, simple_loss=0.3927, pruned_loss=0.1445, over 3075477.77 frames. ], batch size: 124, lr: 3.63e-02, grad_scale: 8.0 2023-04-27 15:24:55,378 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.65 vs. limit=2.0 2023-04-27 15:25:04,599 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8063.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:25:57,836 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=8099.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:26:00,725 INFO [train.py:904] (4/8) Epoch 1, batch 8100, loss[loss=0.3471, simple_loss=0.3876, pruned_loss=0.1533, over 11523.00 frames. ], tot_loss[loss=0.3387, simple_loss=0.3919, pruned_loss=0.1427, over 3088043.70 frames. ], batch size: 247, lr: 3.62e-02, grad_scale: 4.0 2023-04-27 15:26:15,159 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=8111.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:26:22,779 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.844e+02 5.411e+02 6.350e+02 7.571e+02 1.310e+03, threshold=1.270e+03, percent-clipped=1.0 2023-04-27 15:27:16,439 INFO [train.py:904] (4/8) Epoch 1, batch 8150, loss[loss=0.3839, simple_loss=0.4139, pruned_loss=0.177, over 11393.00 frames. ], tot_loss[loss=0.3346, simple_loss=0.3882, pruned_loss=0.1404, over 3091981.34 frames. ], batch size: 246, lr: 3.62e-02, grad_scale: 4.0 2023-04-27 15:27:27,358 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=8158.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:27:50,865 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1576, 2.8220, 2.4343, 3.4808, 2.4333, 3.2112, 2.5545, 2.2980], device='cuda:4'), covar=tensor([0.0342, 0.0329, 0.0307, 0.0258, 0.1102, 0.0240, 0.0583, 0.1078], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0116, 0.0095, 0.0131, 0.0201, 0.0115, 0.0137, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:28:23,106 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7329, 3.6154, 3.5192, 2.7328, 3.3950, 3.5073, 3.6832, 1.8988], device='cuda:4'), covar=tensor([0.1153, 0.0084, 0.0077, 0.0353, 0.0103, 0.0121, 0.0104, 0.0760], device='cuda:4'), in_proj_covar=tensor([0.0113, 0.0043, 0.0047, 0.0079, 0.0045, 0.0045, 0.0049, 0.0086], device='cuda:4'), out_proj_covar=tensor([2.3063e-04, 9.7484e-05, 1.0873e-04, 1.6986e-04, 1.0014e-04, 1.1041e-04, 1.0674e-04, 1.8107e-04], device='cuda:4') 2023-04-27 15:28:33,141 INFO [train.py:904] (4/8) Epoch 1, batch 8200, loss[loss=0.2994, simple_loss=0.3698, pruned_loss=0.1146, over 16855.00 frames. ], tot_loss[loss=0.3309, simple_loss=0.3847, pruned_loss=0.1385, over 3118073.07 frames. ], batch size: 109, lr: 3.61e-02, grad_scale: 4.0 2023-04-27 15:28:56,623 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.505e+02 5.388e+02 6.507e+02 7.835e+02 3.023e+03, threshold=1.301e+03, percent-clipped=3.0 2023-04-27 15:29:53,196 INFO [train.py:904] (4/8) Epoch 1, batch 8250, loss[loss=0.3022, simple_loss=0.3548, pruned_loss=0.1248, over 11890.00 frames. ], tot_loss[loss=0.3273, simple_loss=0.3828, pruned_loss=0.1359, over 3107847.62 frames. ], batch size: 247, lr: 3.60e-02, grad_scale: 4.0 2023-04-27 15:30:15,907 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=8265.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:30:22,721 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9586, 4.4082, 4.7485, 5.0213, 4.2640, 4.7989, 4.6648, 4.4903], device='cuda:4'), covar=tensor([0.0277, 0.0217, 0.0181, 0.0092, 0.0657, 0.0137, 0.0149, 0.0179], device='cuda:4'), in_proj_covar=tensor([0.0089, 0.0063, 0.0120, 0.0094, 0.0141, 0.0093, 0.0086, 0.0101], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:30:41,334 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8280.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:31:14,606 INFO [train.py:904] (4/8) Epoch 1, batch 8300, loss[loss=0.2903, simple_loss=0.3611, pruned_loss=0.1097, over 16708.00 frames. ], tot_loss[loss=0.3184, simple_loss=0.3771, pruned_loss=0.1298, over 3091345.22 frames. ], batch size: 57, lr: 3.59e-02, grad_scale: 4.0 2023-04-27 15:31:31,240 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7963, 4.2154, 4.6887, 4.8037, 4.1647, 4.6501, 4.5190, 4.3804], device='cuda:4'), covar=tensor([0.0294, 0.0215, 0.0164, 0.0093, 0.0699, 0.0151, 0.0159, 0.0197], device='cuda:4'), in_proj_covar=tensor([0.0088, 0.0063, 0.0120, 0.0093, 0.0141, 0.0092, 0.0085, 0.0100], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:31:40,147 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.807e+02 4.177e+02 5.020e+02 6.033e+02 1.438e+03, threshold=1.004e+03, percent-clipped=1.0 2023-04-27 15:31:56,754 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=8326.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:31:59,950 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=8328.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:32:36,034 INFO [train.py:904] (4/8) Epoch 1, batch 8350, loss[loss=0.3066, simple_loss=0.349, pruned_loss=0.132, over 11687.00 frames. ], tot_loss[loss=0.3105, simple_loss=0.3727, pruned_loss=0.1242, over 3079977.30 frames. ], batch size: 246, lr: 3.58e-02, grad_scale: 4.0 2023-04-27 15:33:55,086 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8399.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:33:57,931 INFO [train.py:904] (4/8) Epoch 1, batch 8400, loss[loss=0.2808, simple_loss=0.3554, pruned_loss=0.1031, over 16534.00 frames. ], tot_loss[loss=0.3041, simple_loss=0.3679, pruned_loss=0.1202, over 3061527.13 frames. ], batch size: 62, lr: 3.58e-02, grad_scale: 8.0 2023-04-27 15:34:21,541 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.723e+02 4.252e+02 5.243e+02 6.006e+02 1.213e+03, threshold=1.049e+03, percent-clipped=1.0 2023-04-27 15:34:23,955 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6492, 3.7025, 3.7427, 3.7387, 3.8446, 4.1471, 4.1396, 3.8230], device='cuda:4'), covar=tensor([0.1469, 0.1313, 0.0841, 0.1692, 0.2413, 0.0890, 0.0635, 0.2144], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0198, 0.0169, 0.0176, 0.0212, 0.0172, 0.0150, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-27 15:34:45,553 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-27 15:34:52,624 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.02 vs. limit=2.0 2023-04-27 15:35:12,364 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=8447.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:35:18,229 INFO [train.py:904] (4/8) Epoch 1, batch 8450, loss[loss=0.2689, simple_loss=0.3447, pruned_loss=0.09657, over 16288.00 frames. ], tot_loss[loss=0.3011, simple_loss=0.3662, pruned_loss=0.118, over 3065617.48 frames. ], batch size: 165, lr: 3.57e-02, grad_scale: 8.0 2023-04-27 15:35:30,870 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8458.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:36:39,471 INFO [train.py:904] (4/8) Epoch 1, batch 8500, loss[loss=0.261, simple_loss=0.3354, pruned_loss=0.0933, over 15289.00 frames. ], tot_loss[loss=0.2932, simple_loss=0.3602, pruned_loss=0.1131, over 3057247.22 frames. ], batch size: 190, lr: 3.56e-02, grad_scale: 8.0 2023-04-27 15:36:49,164 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=8506.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:37:04,602 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.878e+02 4.840e+02 6.056e+02 7.383e+02 1.593e+03, threshold=1.211e+03, percent-clipped=7.0 2023-04-27 15:37:47,730 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9214, 3.7182, 4.1825, 4.1964, 4.3630, 3.8046, 4.0079, 4.0848], device='cuda:4'), covar=tensor([0.0276, 0.0405, 0.0478, 0.0475, 0.0388, 0.0351, 0.0629, 0.0284], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0113, 0.0132, 0.0132, 0.0141, 0.0122, 0.0169, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:38:03,990 INFO [train.py:904] (4/8) Epoch 1, batch 8550, loss[loss=0.2848, simple_loss=0.3477, pruned_loss=0.111, over 11886.00 frames. ], tot_loss[loss=0.2898, simple_loss=0.3568, pruned_loss=0.1114, over 3032413.23 frames. ], batch size: 247, lr: 3.55e-02, grad_scale: 8.0 2023-04-27 15:38:59,243 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.71 vs. limit=2.0 2023-04-27 15:39:41,998 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-27 15:39:44,565 INFO [train.py:904] (4/8) Epoch 1, batch 8600, loss[loss=0.2866, simple_loss=0.3655, pruned_loss=0.1039, over 15460.00 frames. ], tot_loss[loss=0.2904, simple_loss=0.3587, pruned_loss=0.111, over 3032463.69 frames. ], batch size: 190, lr: 3.54e-02, grad_scale: 8.0 2023-04-27 15:40:17,873 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.023e+02 4.189e+02 5.220e+02 6.302e+02 1.296e+03, threshold=1.044e+03, percent-clipped=1.0 2023-04-27 15:40:28,414 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=8621.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:41:26,180 INFO [train.py:904] (4/8) Epoch 1, batch 8650, loss[loss=0.2615, simple_loss=0.3327, pruned_loss=0.09518, over 12491.00 frames. ], tot_loss[loss=0.2859, simple_loss=0.3558, pruned_loss=0.108, over 3043388.55 frames. ], batch size: 247, lr: 3.54e-02, grad_scale: 8.0 2023-04-27 15:41:27,547 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.48 vs. limit=5.0 2023-04-27 15:42:30,732 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-27 15:43:13,564 INFO [train.py:904] (4/8) Epoch 1, batch 8700, loss[loss=0.2811, simple_loss=0.3408, pruned_loss=0.1107, over 12278.00 frames. ], tot_loss[loss=0.2814, simple_loss=0.3515, pruned_loss=0.1056, over 3043974.88 frames. ], batch size: 250, lr: 3.53e-02, grad_scale: 8.0 2023-04-27 15:43:41,221 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.978e+02 4.653e+02 5.603e+02 6.924e+02 1.986e+03, threshold=1.121e+03, percent-clipped=4.0 2023-04-27 15:44:37,144 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7630, 3.6377, 1.3629, 3.5988, 2.0199, 3.6679, 1.5850, 2.6087], device='cuda:4'), covar=tensor([0.0052, 0.0091, 0.2054, 0.0069, 0.1037, 0.0198, 0.1676, 0.0669], device='cuda:4'), in_proj_covar=tensor([0.0063, 0.0068, 0.0151, 0.0066, 0.0127, 0.0076, 0.0154, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-04-27 15:44:49,919 INFO [train.py:904] (4/8) Epoch 1, batch 8750, loss[loss=0.3078, simple_loss=0.3792, pruned_loss=0.1182, over 15221.00 frames. ], tot_loss[loss=0.2792, simple_loss=0.3501, pruned_loss=0.1041, over 3044314.01 frames. ], batch size: 191, lr: 3.52e-02, grad_scale: 8.0 2023-04-27 15:44:54,450 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-27 15:45:14,453 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=8760.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:46:42,405 INFO [train.py:904] (4/8) Epoch 1, batch 8800, loss[loss=0.2663, simple_loss=0.3439, pruned_loss=0.09436, over 15246.00 frames. ], tot_loss[loss=0.2779, simple_loss=0.3491, pruned_loss=0.1034, over 3047046.67 frames. ], batch size: 190, lr: 3.51e-02, grad_scale: 8.0 2023-04-27 15:47:13,494 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.267e+02 4.433e+02 5.797e+02 7.279e+02 2.542e+03, threshold=1.159e+03, percent-clipped=4.0 2023-04-27 15:47:25,071 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=8821.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:47:33,316 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7009, 5.0408, 4.7876, 4.9705, 4.4749, 4.5222, 4.5514, 5.0324], device='cuda:4'), covar=tensor([0.0289, 0.0527, 0.0524, 0.0194, 0.0475, 0.0307, 0.0365, 0.0379], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0201, 0.0178, 0.0117, 0.0147, 0.0122, 0.0166, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 15:48:27,339 INFO [train.py:904] (4/8) Epoch 1, batch 8850, loss[loss=0.2264, simple_loss=0.309, pruned_loss=0.07193, over 12377.00 frames. ], tot_loss[loss=0.2773, simple_loss=0.3503, pruned_loss=0.1022, over 3023343.26 frames. ], batch size: 248, lr: 3.51e-02, grad_scale: 8.0 2023-04-27 15:48:31,111 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6293, 2.6110, 2.6276, 2.0282, 2.5747, 2.4819, 2.6185, 1.7590], device='cuda:4'), covar=tensor([0.1110, 0.0113, 0.0115, 0.0453, 0.0091, 0.0112, 0.0088, 0.0785], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0045, 0.0050, 0.0085, 0.0046, 0.0047, 0.0049, 0.0093], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 15:50:13,540 INFO [train.py:904] (4/8) Epoch 1, batch 8900, loss[loss=0.2862, simple_loss=0.3525, pruned_loss=0.1099, over 12970.00 frames. ], tot_loss[loss=0.2756, simple_loss=0.3499, pruned_loss=0.1006, over 3042762.14 frames. ], batch size: 248, lr: 3.50e-02, grad_scale: 8.0 2023-04-27 15:50:42,917 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.973e+02 4.392e+02 5.224e+02 6.190e+02 1.359e+03, threshold=1.045e+03, percent-clipped=1.0 2023-04-27 15:50:56,784 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8921.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:51:13,366 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=8927.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:51:13,583 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-27 15:52:19,949 INFO [train.py:904] (4/8) Epoch 1, batch 8950, loss[loss=0.2669, simple_loss=0.3465, pruned_loss=0.09364, over 16502.00 frames. ], tot_loss[loss=0.2752, simple_loss=0.3493, pruned_loss=0.1005, over 3064722.28 frames. ], batch size: 147, lr: 3.49e-02, grad_scale: 8.0 2023-04-27 15:53:00,567 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=8969.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:53:29,540 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2687, 2.7755, 2.6701, 1.7360, 2.8333, 2.7984, 2.4158, 2.6660], device='cuda:4'), covar=tensor([0.0506, 0.0105, 0.0206, 0.1520, 0.0123, 0.0093, 0.0390, 0.0235], device='cuda:4'), in_proj_covar=tensor([0.0092, 0.0064, 0.0061, 0.0140, 0.0059, 0.0059, 0.0075, 0.0084], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:53:38,176 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=8986.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:53:42,553 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=8988.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:54:08,281 INFO [train.py:904] (4/8) Epoch 1, batch 9000, loss[loss=0.2404, simple_loss=0.3223, pruned_loss=0.07928, over 16218.00 frames. ], tot_loss[loss=0.2703, simple_loss=0.3453, pruned_loss=0.09763, over 3088118.74 frames. ], batch size: 165, lr: 3.48e-02, grad_scale: 8.0 2023-04-27 15:54:08,282 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 15:54:19,192 INFO [train.py:938] (4/8) Epoch 1, validation: loss=0.2299, simple_loss=0.3267, pruned_loss=0.06658, over 944034.00 frames. 2023-04-27 15:54:19,193 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-27 15:54:46,196 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9013.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:54:51,598 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.333e+02 3.931e+02 4.892e+02 5.908e+02 1.148e+03, threshold=9.783e+02, percent-clipped=2.0 2023-04-27 15:55:56,670 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9047.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:56:02,939 INFO [train.py:904] (4/8) Epoch 1, batch 9050, loss[loss=0.3212, simple_loss=0.3735, pruned_loss=0.1344, over 12447.00 frames. ], tot_loss[loss=0.2737, simple_loss=0.3477, pruned_loss=0.0998, over 3084215.02 frames. ], batch size: 248, lr: 3.48e-02, grad_scale: 8.0 2023-04-27 15:56:38,564 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7312, 3.5040, 3.7083, 4.1134, 4.0142, 3.8350, 4.0495, 3.9125], device='cuda:4'), covar=tensor([0.0413, 0.0467, 0.0953, 0.0270, 0.0382, 0.0464, 0.0288, 0.0344], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0163, 0.0244, 0.0170, 0.0143, 0.0154, 0.0136, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:56:48,984 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9074.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:57:19,879 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1841, 1.6376, 2.3197, 2.8431, 3.1119, 3.0021, 2.0095, 3.2290], device='cuda:4'), covar=tensor([0.0038, 0.0591, 0.0301, 0.0111, 0.0049, 0.0165, 0.0313, 0.0062], device='cuda:4'), in_proj_covar=tensor([0.0048, 0.0092, 0.0072, 0.0054, 0.0041, 0.0044, 0.0070, 0.0040], device='cuda:4'), out_proj_covar=tensor([8.9407e-05, 1.6582e-04, 1.3853e-04, 1.0041e-04, 7.4036e-05, 8.2532e-05, 1.2110e-04, 7.4743e-05], device='cuda:4') 2023-04-27 15:57:45,071 INFO [train.py:904] (4/8) Epoch 1, batch 9100, loss[loss=0.317, simple_loss=0.3826, pruned_loss=0.1257, over 15306.00 frames. ], tot_loss[loss=0.274, simple_loss=0.3474, pruned_loss=0.1003, over 3091105.03 frames. ], batch size: 191, lr: 3.47e-02, grad_scale: 8.0 2023-04-27 15:57:50,415 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.09 vs. limit=5.0 2023-04-27 15:58:14,291 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.937e+02 4.687e+02 5.823e+02 7.250e+02 1.575e+03, threshold=1.165e+03, percent-clipped=5.0 2023-04-27 15:58:14,919 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9116.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:58:20,624 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4980, 4.7941, 4.8258, 4.8081, 4.8168, 5.2317, 5.0613, 4.7243], device='cuda:4'), covar=tensor([0.0557, 0.1042, 0.0822, 0.1166, 0.1746, 0.0769, 0.0615, 0.1526], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0208, 0.0178, 0.0182, 0.0221, 0.0182, 0.0151, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-27 15:59:40,678 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6764, 3.6029, 3.2982, 1.6751, 2.7102, 2.1357, 3.1526, 3.5746], device='cuda:4'), covar=tensor([0.0304, 0.0305, 0.0355, 0.2099, 0.0947, 0.1163, 0.0847, 0.0371], device='cuda:4'), in_proj_covar=tensor([0.0107, 0.0078, 0.0125, 0.0158, 0.0152, 0.0143, 0.0140, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 15:59:42,649 INFO [train.py:904] (4/8) Epoch 1, batch 9150, loss[loss=0.2566, simple_loss=0.3362, pruned_loss=0.08846, over 16941.00 frames. ], tot_loss[loss=0.2734, simple_loss=0.3479, pruned_loss=0.09946, over 3099597.32 frames. ], batch size: 109, lr: 3.46e-02, grad_scale: 8.0 2023-04-27 16:01:27,877 INFO [train.py:904] (4/8) Epoch 1, batch 9200, loss[loss=0.2626, simple_loss=0.32, pruned_loss=0.1026, over 12222.00 frames. ], tot_loss[loss=0.2683, simple_loss=0.3418, pruned_loss=0.09737, over 3099477.87 frames. ], batch size: 247, lr: 3.45e-02, grad_scale: 8.0 2023-04-27 16:01:54,927 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.438e+02 5.266e+02 6.436e+02 8.220e+02 2.129e+03, threshold=1.287e+03, percent-clipped=5.0 2023-04-27 16:02:57,950 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0086, 2.6714, 2.6121, 1.7262, 2.6819, 2.6955, 2.5299, 2.6057], device='cuda:4'), covar=tensor([0.0554, 0.0130, 0.0165, 0.1394, 0.0149, 0.0134, 0.0268, 0.0231], device='cuda:4'), in_proj_covar=tensor([0.0101, 0.0068, 0.0067, 0.0145, 0.0063, 0.0064, 0.0078, 0.0089], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:03:04,215 INFO [train.py:904] (4/8) Epoch 1, batch 9250, loss[loss=0.2611, simple_loss=0.3411, pruned_loss=0.09055, over 16319.00 frames. ], tot_loss[loss=0.2678, simple_loss=0.3409, pruned_loss=0.09733, over 3075252.58 frames. ], batch size: 146, lr: 3.45e-02, grad_scale: 8.0 2023-04-27 16:03:43,187 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0375, 4.0145, 3.6942, 1.5751, 2.8473, 2.3154, 3.2898, 4.1421], device='cuda:4'), covar=tensor([0.0256, 0.0301, 0.0355, 0.2248, 0.1036, 0.1326, 0.0976, 0.0321], device='cuda:4'), in_proj_covar=tensor([0.0108, 0.0081, 0.0126, 0.0159, 0.0151, 0.0144, 0.0139, 0.0077], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 16:04:16,907 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9283.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:04:58,121 INFO [train.py:904] (4/8) Epoch 1, batch 9300, loss[loss=0.2443, simple_loss=0.3204, pruned_loss=0.0841, over 16243.00 frames. ], tot_loss[loss=0.2641, simple_loss=0.3374, pruned_loss=0.09539, over 3066276.64 frames. ], batch size: 35, lr: 3.44e-02, grad_scale: 8.0 2023-04-27 16:05:07,011 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6665, 3.2980, 2.0405, 4.0256, 4.0461, 3.9937, 2.0687, 3.1418], device='cuda:4'), covar=tensor([0.2247, 0.0368, 0.1951, 0.0095, 0.0104, 0.0274, 0.1199, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0096, 0.0166, 0.0062, 0.0071, 0.0084, 0.0138, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:05:33,417 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.238e+02 3.939e+02 4.601e+02 5.465e+02 1.094e+03, threshold=9.201e+02, percent-clipped=0.0 2023-04-27 16:06:27,249 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9342.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:06:36,226 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1853, 3.1449, 2.9315, 1.8976, 2.6071, 2.1200, 2.8295, 3.2309], device='cuda:4'), covar=tensor([0.0263, 0.0389, 0.0372, 0.1691, 0.0830, 0.1096, 0.0686, 0.0273], device='cuda:4'), in_proj_covar=tensor([0.0108, 0.0082, 0.0126, 0.0158, 0.0149, 0.0141, 0.0136, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-04-27 16:06:44,327 INFO [train.py:904] (4/8) Epoch 1, batch 9350, loss[loss=0.2484, simple_loss=0.3277, pruned_loss=0.08448, over 16598.00 frames. ], tot_loss[loss=0.2636, simple_loss=0.3373, pruned_loss=0.09499, over 3078269.73 frames. ], batch size: 62, lr: 3.43e-02, grad_scale: 8.0 2023-04-27 16:07:24,444 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9369.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:08:26,592 INFO [train.py:904] (4/8) Epoch 1, batch 9400, loss[loss=0.2515, simple_loss=0.3186, pruned_loss=0.09215, over 11955.00 frames. ], tot_loss[loss=0.2625, simple_loss=0.3365, pruned_loss=0.09425, over 3058688.26 frames. ], batch size: 248, lr: 3.43e-02, grad_scale: 8.0 2023-04-27 16:08:54,851 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9414.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:08:56,911 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.254e+02 4.024e+02 4.732e+02 5.961e+02 1.474e+03, threshold=9.465e+02, percent-clipped=2.0 2023-04-27 16:08:57,986 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=9416.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:09:08,662 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1108, 4.8419, 4.9758, 5.1837, 4.5278, 5.0672, 4.9353, 4.6194], device='cuda:4'), covar=tensor([0.0306, 0.0151, 0.0131, 0.0076, 0.0450, 0.0119, 0.0106, 0.0167], device='cuda:4'), in_proj_covar=tensor([0.0085, 0.0062, 0.0112, 0.0087, 0.0136, 0.0086, 0.0078, 0.0095], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:10:08,169 INFO [train.py:904] (4/8) Epoch 1, batch 9450, loss[loss=0.2602, simple_loss=0.3345, pruned_loss=0.0929, over 16979.00 frames. ], tot_loss[loss=0.2641, simple_loss=0.3385, pruned_loss=0.09481, over 3057818.60 frames. ], batch size: 109, lr: 3.42e-02, grad_scale: 8.0 2023-04-27 16:10:33,964 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=9464.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:10:58,593 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9475.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:11:50,036 INFO [train.py:904] (4/8) Epoch 1, batch 9500, loss[loss=0.2691, simple_loss=0.3421, pruned_loss=0.09807, over 16940.00 frames. ], tot_loss[loss=0.2623, simple_loss=0.337, pruned_loss=0.09383, over 3055457.94 frames. ], batch size: 109, lr: 3.41e-02, grad_scale: 8.0 2023-04-27 16:12:06,468 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4702, 3.0197, 2.4687, 3.8193, 2.4546, 3.6238, 2.6373, 2.2815], device='cuda:4'), covar=tensor([0.0307, 0.0285, 0.0307, 0.0238, 0.1129, 0.0172, 0.0540, 0.1226], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0127, 0.0103, 0.0144, 0.0201, 0.0121, 0.0143, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:12:21,412 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.089e+02 4.155e+02 5.314e+02 6.779e+02 1.064e+03, threshold=1.063e+03, percent-clipped=4.0 2023-04-27 16:12:26,025 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2107, 3.1178, 3.1599, 3.4909, 3.3807, 3.2675, 3.3919, 3.3958], device='cuda:4'), covar=tensor([0.0359, 0.0335, 0.0872, 0.0336, 0.0417, 0.0633, 0.0369, 0.0293], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0166, 0.0249, 0.0173, 0.0144, 0.0152, 0.0139, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:13:37,008 INFO [train.py:904] (4/8) Epoch 1, batch 9550, loss[loss=0.2577, simple_loss=0.3307, pruned_loss=0.09234, over 12634.00 frames. ], tot_loss[loss=0.263, simple_loss=0.3375, pruned_loss=0.09427, over 3074524.17 frames. ], batch size: 247, lr: 3.41e-02, grad_scale: 8.0 2023-04-27 16:14:46,567 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=9583.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:14:48,721 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9584.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:15:18,626 INFO [train.py:904] (4/8) Epoch 1, batch 9600, loss[loss=0.3069, simple_loss=0.3553, pruned_loss=0.1292, over 12567.00 frames. ], tot_loss[loss=0.2658, simple_loss=0.3396, pruned_loss=0.09607, over 3052957.09 frames. ], batch size: 248, lr: 3.40e-02, grad_scale: 8.0 2023-04-27 16:15:48,620 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.034e+02 4.688e+02 5.707e+02 6.655e+02 1.542e+03, threshold=1.141e+03, percent-clipped=4.0 2023-04-27 16:16:20,297 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=9631.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:16:45,979 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=9642.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:16:53,240 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9645.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:17:08,555 INFO [train.py:904] (4/8) Epoch 1, batch 9650, loss[loss=0.2515, simple_loss=0.323, pruned_loss=0.09, over 12258.00 frames. ], tot_loss[loss=0.2667, simple_loss=0.3414, pruned_loss=0.096, over 3053930.22 frames. ], batch size: 248, lr: 3.39e-02, grad_scale: 8.0 2023-04-27 16:17:16,485 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7735, 1.5648, 1.8872, 2.4291, 2.7239, 2.5377, 1.5066, 2.5137], device='cuda:4'), covar=tensor([0.0056, 0.0452, 0.0238, 0.0138, 0.0055, 0.0141, 0.0304, 0.0063], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0090, 0.0073, 0.0056, 0.0044, 0.0045, 0.0071, 0.0042], device='cuda:4'), out_proj_covar=tensor([9.4174e-05, 1.6238e-04, 1.3836e-04, 1.0688e-04, 7.7578e-05, 8.4632e-05, 1.2404e-04, 7.5636e-05], device='cuda:4') 2023-04-27 16:17:21,196 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3998, 4.8818, 4.7666, 4.8136, 4.8194, 5.2102, 5.1075, 4.6774], device='cuda:4'), covar=tensor([0.0642, 0.1139, 0.0889, 0.1347, 0.1763, 0.0704, 0.0636, 0.1955], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0219, 0.0182, 0.0188, 0.0222, 0.0182, 0.0156, 0.0234], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:17:26,905 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9658.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:17:52,702 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=9669.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:18:35,171 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=9690.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:18:54,550 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5445, 3.4444, 3.1275, 3.0792, 2.3866, 2.1429, 3.5414, 3.9398], device='cuda:4'), covar=tensor([0.1675, 0.0638, 0.0850, 0.0407, 0.1340, 0.1269, 0.0219, 0.0086], device='cuda:4'), in_proj_covar=tensor([0.0218, 0.0192, 0.0214, 0.0132, 0.0172, 0.0165, 0.0146, 0.0079], device='cuda:4'), out_proj_covar=tensor([2.5564e-04, 2.2239e-04, 2.3357e-04, 1.5598e-04, 2.0554e-04, 2.0005e-04, 1.7547e-04, 9.9910e-05], device='cuda:4') 2023-04-27 16:18:57,043 INFO [train.py:904] (4/8) Epoch 1, batch 9700, loss[loss=0.2687, simple_loss=0.3461, pruned_loss=0.09559, over 16622.00 frames. ], tot_loss[loss=0.265, simple_loss=0.3399, pruned_loss=0.09509, over 3064752.17 frames. ], batch size: 75, lr: 3.38e-02, grad_scale: 8.0 2023-04-27 16:19:24,739 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.013e+02 4.098e+02 5.059e+02 6.165e+02 1.399e+03, threshold=1.012e+03, percent-clipped=2.0 2023-04-27 16:19:28,117 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=9717.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:19:33,567 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9719.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:19:34,844 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3149, 1.5256, 1.6629, 1.4700, 1.9421, 1.9297, 2.1629, 2.1607], device='cuda:4'), covar=tensor([0.0040, 0.0322, 0.0194, 0.0234, 0.0103, 0.0165, 0.0055, 0.0076], device='cuda:4'), in_proj_covar=tensor([0.0034, 0.0078, 0.0061, 0.0069, 0.0056, 0.0066, 0.0036, 0.0044], device='cuda:4'), out_proj_covar=tensor([4.6297e-05, 1.2131e-04, 9.3651e-05, 1.0583e-04, 8.6301e-05, 1.0214e-04, 5.6489e-05, 7.0836e-05], device='cuda:4') 2023-04-27 16:20:27,159 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9743.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:20:40,561 INFO [train.py:904] (4/8) Epoch 1, batch 9750, loss[loss=0.2331, simple_loss=0.3186, pruned_loss=0.07377, over 16891.00 frames. ], tot_loss[loss=0.2631, simple_loss=0.3381, pruned_loss=0.09402, over 3070795.17 frames. ], batch size: 102, lr: 3.38e-02, grad_scale: 8.0 2023-04-27 16:20:44,008 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-27 16:21:01,346 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.71 vs. limit=2.0 2023-04-27 16:21:17,328 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9770.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:22:19,290 INFO [train.py:904] (4/8) Epoch 1, batch 9800, loss[loss=0.2666, simple_loss=0.357, pruned_loss=0.08815, over 16751.00 frames. ], tot_loss[loss=0.2605, simple_loss=0.3371, pruned_loss=0.09191, over 3089188.09 frames. ], batch size: 124, lr: 3.37e-02, grad_scale: 8.0 2023-04-27 16:22:26,121 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9804.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:22:47,902 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.893e+02 4.491e+02 5.387e+02 6.683e+02 1.234e+03, threshold=1.077e+03, percent-clipped=5.0 2023-04-27 16:24:05,874 INFO [train.py:904] (4/8) Epoch 1, batch 9850, loss[loss=0.2809, simple_loss=0.354, pruned_loss=0.1039, over 16179.00 frames. ], tot_loss[loss=0.2613, simple_loss=0.3386, pruned_loss=0.09202, over 3090370.02 frames. ], batch size: 165, lr: 3.36e-02, grad_scale: 8.0 2023-04-27 16:25:58,729 INFO [train.py:904] (4/8) Epoch 1, batch 9900, loss[loss=0.2582, simple_loss=0.3424, pruned_loss=0.08694, over 15085.00 frames. ], tot_loss[loss=0.2613, simple_loss=0.3384, pruned_loss=0.09213, over 3057837.91 frames. ], batch size: 190, lr: 3.36e-02, grad_scale: 8.0 2023-04-27 16:26:31,897 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.967e+02 4.199e+02 5.185e+02 6.475e+02 1.120e+03, threshold=1.037e+03, percent-clipped=1.0 2023-04-27 16:27:32,793 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9940.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:27:57,263 INFO [train.py:904] (4/8) Epoch 1, batch 9950, loss[loss=0.2708, simple_loss=0.3504, pruned_loss=0.09558, over 15481.00 frames. ], tot_loss[loss=0.2628, simple_loss=0.3406, pruned_loss=0.09251, over 3064746.85 frames. ], batch size: 191, lr: 3.35e-02, grad_scale: 8.0 2023-04-27 16:29:07,094 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-27 16:30:02,393 INFO [train.py:904] (4/8) Epoch 1, batch 10000, loss[loss=0.2463, simple_loss=0.3273, pruned_loss=0.08261, over 15498.00 frames. ], tot_loss[loss=0.2601, simple_loss=0.338, pruned_loss=0.09111, over 3082511.70 frames. ], batch size: 191, lr: 3.34e-02, grad_scale: 8.0 2023-04-27 16:30:30,578 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10014.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:30:32,719 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.207e+02 4.345e+02 5.531e+02 6.703e+02 1.367e+03, threshold=1.106e+03, percent-clipped=3.0 2023-04-27 16:31:22,499 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10039.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:31:33,519 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.65 vs. limit=2.0 2023-04-27 16:31:43,673 INFO [train.py:904] (4/8) Epoch 1, batch 10050, loss[loss=0.2936, simple_loss=0.3607, pruned_loss=0.1133, over 12228.00 frames. ], tot_loss[loss=0.2603, simple_loss=0.3384, pruned_loss=0.0911, over 3077977.27 frames. ], batch size: 247, lr: 3.34e-02, grad_scale: 8.0 2023-04-27 16:32:06,769 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10062.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:32:23,041 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10070.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:33:16,682 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10099.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:33:18,935 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10100.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 16:33:19,581 INFO [train.py:904] (4/8) Epoch 1, batch 10100, loss[loss=0.2605, simple_loss=0.3375, pruned_loss=0.09172, over 16672.00 frames. ], tot_loss[loss=0.2615, simple_loss=0.3391, pruned_loss=0.09193, over 3077266.46 frames. ], batch size: 134, lr: 3.33e-02, grad_scale: 16.0 2023-04-27 16:33:49,432 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.152e+02 4.595e+02 5.786e+02 6.822e+02 2.551e+03, threshold=1.157e+03, percent-clipped=1.0 2023-04-27 16:33:53,662 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=10118.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:34:06,039 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10123.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:34:22,740 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0440, 3.6098, 2.7780, 4.3185, 2.7460, 4.2926, 3.0107, 2.4314], device='cuda:4'), covar=tensor([0.0268, 0.0256, 0.0270, 0.0216, 0.1179, 0.0153, 0.0531, 0.1392], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0133, 0.0106, 0.0153, 0.0205, 0.0127, 0.0149, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:35:05,143 INFO [train.py:904] (4/8) Epoch 2, batch 0, loss[loss=0.3261, simple_loss=0.3624, pruned_loss=0.1449, over 17229.00 frames. ], tot_loss[loss=0.3261, simple_loss=0.3624, pruned_loss=0.1449, over 17229.00 frames. ], batch size: 45, lr: 3.26e-02, grad_scale: 8.0 2023-04-27 16:35:05,143 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 16:35:12,742 INFO [train.py:938] (4/8) Epoch 2, validation: loss=0.2235, simple_loss=0.3221, pruned_loss=0.06242, over 944034.00 frames. 2023-04-27 16:35:12,743 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-27 16:35:36,834 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-27 16:35:46,409 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4916, 1.8373, 1.9455, 1.4611, 2.3038, 2.2641, 2.2287, 2.5901], device='cuda:4'), covar=tensor([0.0033, 0.0216, 0.0108, 0.0197, 0.0079, 0.0116, 0.0043, 0.0060], device='cuda:4'), in_proj_covar=tensor([0.0036, 0.0077, 0.0065, 0.0071, 0.0062, 0.0069, 0.0039, 0.0046], device='cuda:4'), out_proj_covar=tensor([5.0514e-05, 1.1865e-04, 9.9956e-05, 1.1038e-04, 9.5770e-05, 1.0476e-04, 5.9811e-05, 7.5628e-05], device='cuda:4') 2023-04-27 16:36:09,342 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-27 16:36:19,365 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-27 16:36:22,853 INFO [train.py:904] (4/8) Epoch 2, batch 50, loss[loss=0.2746, simple_loss=0.3435, pruned_loss=0.1028, over 17055.00 frames. ], tot_loss[loss=0.3236, simple_loss=0.3714, pruned_loss=0.1379, over 745823.03 frames. ], batch size: 50, lr: 3.25e-02, grad_scale: 4.0 2023-04-27 16:36:45,678 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.307e+02 5.068e+02 6.475e+02 7.712e+02 1.660e+03, threshold=1.295e+03, percent-clipped=4.0 2023-04-27 16:37:16,878 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10240.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:37:26,926 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.67 vs. limit=2.0 2023-04-27 16:37:30,888 INFO [train.py:904] (4/8) Epoch 2, batch 100, loss[loss=0.282, simple_loss=0.358, pruned_loss=0.103, over 17099.00 frames. ], tot_loss[loss=0.3106, simple_loss=0.3631, pruned_loss=0.129, over 1312844.88 frames. ], batch size: 49, lr: 3.25e-02, grad_scale: 4.0 2023-04-27 16:37:33,226 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.36 vs. limit=5.0 2023-04-27 16:38:22,579 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=10288.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:38:38,636 INFO [train.py:904] (4/8) Epoch 2, batch 150, loss[loss=0.2555, simple_loss=0.3278, pruned_loss=0.09155, over 17178.00 frames. ], tot_loss[loss=0.3074, simple_loss=0.3601, pruned_loss=0.1273, over 1747304.81 frames. ], batch size: 46, lr: 3.24e-02, grad_scale: 4.0 2023-04-27 16:38:56,007 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10314.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:39:01,494 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.010e+02 4.543e+02 5.695e+02 6.778e+02 1.541e+03, threshold=1.139e+03, percent-clipped=2.0 2023-04-27 16:39:14,171 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.72 vs. limit=2.0 2023-04-27 16:39:47,450 INFO [train.py:904] (4/8) Epoch 2, batch 200, loss[loss=0.2576, simple_loss=0.3252, pruned_loss=0.09503, over 17209.00 frames. ], tot_loss[loss=0.301, simple_loss=0.3563, pruned_loss=0.1229, over 2100924.60 frames. ], batch size: 44, lr: 3.23e-02, grad_scale: 4.0 2023-04-27 16:40:03,410 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=10362.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:40:15,730 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.01 vs. limit=5.0 2023-04-27 16:40:50,468 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10395.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:40:55,036 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10399.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:40:58,058 INFO [train.py:904] (4/8) Epoch 2, batch 250, loss[loss=0.2975, simple_loss=0.3635, pruned_loss=0.1157, over 17041.00 frames. ], tot_loss[loss=0.2983, simple_loss=0.3538, pruned_loss=0.1214, over 2365260.16 frames. ], batch size: 50, lr: 3.23e-02, grad_scale: 4.0 2023-04-27 16:41:11,022 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10409.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:41:22,650 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.115e+02 4.151e+02 5.077e+02 6.310e+02 1.196e+03, threshold=1.015e+03, percent-clipped=1.0 2023-04-27 16:41:22,934 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10418.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:42:03,850 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=10447.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:42:09,609 INFO [train.py:904] (4/8) Epoch 2, batch 300, loss[loss=0.2938, simple_loss=0.3422, pruned_loss=0.1227, over 16767.00 frames. ], tot_loss[loss=0.2938, simple_loss=0.3491, pruned_loss=0.1193, over 2574878.62 frames. ], batch size: 124, lr: 3.22e-02, grad_scale: 4.0 2023-04-27 16:42:14,797 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10454.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:42:36,304 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10470.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:42:55,658 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5496, 3.6176, 1.8625, 3.5268, 2.1592, 3.5404, 1.9522, 2.8020], device='cuda:4'), covar=tensor([0.0054, 0.0099, 0.1288, 0.0063, 0.0752, 0.0246, 0.1088, 0.0479], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0082, 0.0163, 0.0072, 0.0140, 0.0101, 0.0166, 0.0137], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0003, 0.0001, 0.0003, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 16:43:19,861 INFO [train.py:904] (4/8) Epoch 2, batch 350, loss[loss=0.2865, simple_loss=0.3366, pruned_loss=0.1182, over 16508.00 frames. ], tot_loss[loss=0.2891, simple_loss=0.3451, pruned_loss=0.1166, over 2738398.20 frames. ], batch size: 146, lr: 3.21e-02, grad_scale: 4.0 2023-04-27 16:43:39,856 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10515.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:43:42,893 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.580e+02 4.016e+02 4.975e+02 5.908e+02 1.216e+03, threshold=9.949e+02, percent-clipped=2.0 2023-04-27 16:44:27,070 INFO [train.py:904] (4/8) Epoch 2, batch 400, loss[loss=0.2499, simple_loss=0.3286, pruned_loss=0.0856, over 17111.00 frames. ], tot_loss[loss=0.2851, simple_loss=0.3426, pruned_loss=0.1138, over 2869715.26 frames. ], batch size: 49, lr: 3.21e-02, grad_scale: 8.0 2023-04-27 16:44:37,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4385, 4.6110, 4.5565, 4.6624, 4.5345, 5.1374, 4.9937, 4.6097], device='cuda:4'), covar=tensor([0.1028, 0.1154, 0.0984, 0.1334, 0.2215, 0.0742, 0.0679, 0.1748], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0260, 0.0225, 0.0222, 0.0278, 0.0221, 0.0188, 0.0287], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 16:45:36,484 INFO [train.py:904] (4/8) Epoch 2, batch 450, loss[loss=0.269, simple_loss=0.3164, pruned_loss=0.1108, over 16474.00 frames. ], tot_loss[loss=0.2814, simple_loss=0.3397, pruned_loss=0.1116, over 2976057.05 frames. ], batch size: 146, lr: 3.20e-02, grad_scale: 8.0 2023-04-27 16:45:52,889 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4410, 4.3916, 4.7733, 4.9158, 4.9566, 4.4077, 4.5094, 4.7074], device='cuda:4'), covar=tensor([0.0247, 0.0290, 0.0459, 0.0348, 0.0366, 0.0305, 0.0576, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0134, 0.0156, 0.0154, 0.0176, 0.0143, 0.0214, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-04-27 16:45:59,584 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.527e+02 4.134e+02 5.028e+02 5.858e+02 1.204e+03, threshold=1.006e+03, percent-clipped=2.0 2023-04-27 16:46:03,011 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9547, 4.5965, 4.8530, 5.3380, 5.3103, 4.6715, 5.4145, 5.1971], device='cuda:4'), covar=tensor([0.0352, 0.0437, 0.1039, 0.0292, 0.0365, 0.0487, 0.0216, 0.0265], device='cuda:4'), in_proj_covar=tensor([0.0206, 0.0225, 0.0339, 0.0232, 0.0195, 0.0191, 0.0172, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:46:15,711 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6950, 2.7838, 3.1731, 2.3689, 3.2458, 3.2271, 3.3623, 1.6312], device='cuda:4'), covar=tensor([0.1113, 0.0177, 0.0091, 0.0477, 0.0075, 0.0098, 0.0073, 0.0854], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0051, 0.0056, 0.0093, 0.0051, 0.0054, 0.0057, 0.0098], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 16:46:44,101 INFO [train.py:904] (4/8) Epoch 2, batch 500, loss[loss=0.2537, simple_loss=0.315, pruned_loss=0.09625, over 16816.00 frames. ], tot_loss[loss=0.2781, simple_loss=0.3372, pruned_loss=0.1095, over 3054353.77 frames. ], batch size: 83, lr: 3.20e-02, grad_scale: 8.0 2023-04-27 16:47:24,941 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-27 16:47:44,993 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10695.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:47:52,719 INFO [train.py:904] (4/8) Epoch 2, batch 550, loss[loss=0.2564, simple_loss=0.3194, pruned_loss=0.09671, over 16472.00 frames. ], tot_loss[loss=0.2764, simple_loss=0.3357, pruned_loss=0.1085, over 3116953.74 frames. ], batch size: 68, lr: 3.19e-02, grad_scale: 8.0 2023-04-27 16:48:17,046 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.652e+02 3.972e+02 4.860e+02 6.091e+02 1.021e+03, threshold=9.720e+02, percent-clipped=1.0 2023-04-27 16:48:17,358 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10718.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 16:48:51,487 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=10743.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:48:52,809 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10744.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:49:02,283 INFO [train.py:904] (4/8) Epoch 2, batch 600, loss[loss=0.2363, simple_loss=0.3137, pruned_loss=0.07948, over 17138.00 frames. ], tot_loss[loss=0.2754, simple_loss=0.3348, pruned_loss=0.108, over 3167468.88 frames. ], batch size: 47, lr: 3.18e-02, grad_scale: 8.0 2023-04-27 16:49:22,654 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10765.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:49:24,352 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=10766.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:49:28,398 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.74 vs. limit=2.0 2023-04-27 16:49:33,759 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10774.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:50:11,095 INFO [train.py:904] (4/8) Epoch 2, batch 650, loss[loss=0.2294, simple_loss=0.2979, pruned_loss=0.08045, over 16801.00 frames. ], tot_loss[loss=0.2751, simple_loss=0.3342, pruned_loss=0.108, over 3203620.75 frames. ], batch size: 39, lr: 3.18e-02, grad_scale: 8.0 2023-04-27 16:50:17,167 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10805.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:50:23,981 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10810.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 16:50:33,023 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.483e+02 4.152e+02 4.979e+02 6.118e+02 1.196e+03, threshold=9.959e+02, percent-clipped=5.0 2023-04-27 16:50:57,024 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10835.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:50:58,256 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4323, 5.0376, 4.9246, 4.5287, 5.1682, 2.8351, 4.8919, 5.2498], device='cuda:4'), covar=tensor([0.0053, 0.0071, 0.0067, 0.0293, 0.0051, 0.0879, 0.0058, 0.0070], device='cuda:4'), in_proj_covar=tensor([0.0057, 0.0049, 0.0068, 0.0091, 0.0051, 0.0102, 0.0066, 0.0068], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:51:19,283 INFO [train.py:904] (4/8) Epoch 2, batch 700, loss[loss=0.2958, simple_loss=0.3402, pruned_loss=0.1257, over 16856.00 frames. ], tot_loss[loss=0.2729, simple_loss=0.3326, pruned_loss=0.1066, over 3223545.06 frames. ], batch size: 116, lr: 3.17e-02, grad_scale: 8.0 2023-04-27 16:51:27,098 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.01 vs. limit=2.0 2023-04-27 16:52:06,310 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10885.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 16:52:20,078 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2749, 3.9168, 4.0234, 3.5209, 4.1579, 4.2515, 4.5677, 2.4704], device='cuda:4'), covar=tensor([0.0921, 0.0088, 0.0083, 0.0302, 0.0045, 0.0066, 0.0039, 0.0680], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0051, 0.0056, 0.0095, 0.0051, 0.0055, 0.0060, 0.0102], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 16:52:26,661 INFO [train.py:904] (4/8) Epoch 2, batch 750, loss[loss=0.2411, simple_loss=0.3222, pruned_loss=0.08003, over 17038.00 frames. ], tot_loss[loss=0.2719, simple_loss=0.3325, pruned_loss=0.1056, over 3251439.35 frames. ], batch size: 53, lr: 3.17e-02, grad_scale: 8.0 2023-04-27 16:52:50,206 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.205e+02 3.722e+02 4.836e+02 5.954e+02 8.013e+02, threshold=9.671e+02, percent-clipped=0.0 2023-04-27 16:53:27,975 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10946.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 16:53:33,379 INFO [train.py:904] (4/8) Epoch 2, batch 800, loss[loss=0.2485, simple_loss=0.3265, pruned_loss=0.08523, over 17120.00 frames. ], tot_loss[loss=0.2727, simple_loss=0.333, pruned_loss=0.1062, over 3271468.02 frames. ], batch size: 55, lr: 3.16e-02, grad_scale: 8.0 2023-04-27 16:53:43,012 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3712, 4.9802, 5.0212, 5.2037, 4.4644, 5.0906, 5.1226, 4.7687], device='cuda:4'), covar=tensor([0.0193, 0.0161, 0.0187, 0.0124, 0.0939, 0.0149, 0.0124, 0.0226], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0087, 0.0162, 0.0130, 0.0198, 0.0125, 0.0110, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 16:53:57,722 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9828, 5.6247, 5.5850, 5.6501, 5.7426, 6.0993, 6.0387, 5.7394], device='cuda:4'), covar=tensor([0.0536, 0.0977, 0.0904, 0.1248, 0.1819, 0.0692, 0.0622, 0.1512], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0274, 0.0231, 0.0226, 0.0289, 0.0227, 0.0194, 0.0297], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 16:54:24,706 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0043, 5.7239, 5.6035, 5.6236, 5.7580, 6.1666, 6.0423, 5.6107], device='cuda:4'), covar=tensor([0.0532, 0.1080, 0.0825, 0.1389, 0.2000, 0.0575, 0.0635, 0.1633], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0272, 0.0231, 0.0226, 0.0290, 0.0224, 0.0195, 0.0295], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 16:54:43,055 INFO [train.py:904] (4/8) Epoch 2, batch 850, loss[loss=0.2604, simple_loss=0.3179, pruned_loss=0.1015, over 16835.00 frames. ], tot_loss[loss=0.2707, simple_loss=0.3318, pruned_loss=0.1048, over 3272745.51 frames. ], batch size: 102, lr: 3.15e-02, grad_scale: 8.0 2023-04-27 16:55:06,012 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.379e+02 3.969e+02 4.868e+02 6.050e+02 1.097e+03, threshold=9.736e+02, percent-clipped=1.0 2023-04-27 16:55:49,289 INFO [train.py:904] (4/8) Epoch 2, batch 900, loss[loss=0.2841, simple_loss=0.334, pruned_loss=0.1171, over 15679.00 frames. ], tot_loss[loss=0.2691, simple_loss=0.3307, pruned_loss=0.1038, over 3288965.50 frames. ], batch size: 191, lr: 3.15e-02, grad_scale: 8.0 2023-04-27 16:56:04,579 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.79 vs. limit=2.0 2023-04-27 16:56:10,083 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11065.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:56:57,480 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11100.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:56:58,321 INFO [train.py:904] (4/8) Epoch 2, batch 950, loss[loss=0.2513, simple_loss=0.3156, pruned_loss=0.09356, over 16802.00 frames. ], tot_loss[loss=0.2679, simple_loss=0.3303, pruned_loss=0.1028, over 3305607.40 frames. ], batch size: 42, lr: 3.14e-02, grad_scale: 8.0 2023-04-27 16:57:10,361 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11110.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:57:13,954 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=11113.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:57:21,474 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.464e+02 3.927e+02 4.944e+02 5.832e+02 1.325e+03, threshold=9.889e+02, percent-clipped=4.0 2023-04-27 16:57:38,008 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11130.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:57:58,322 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0737, 2.0357, 1.7947, 1.7845, 2.5860, 2.4626, 2.6540, 2.6696], device='cuda:4'), covar=tensor([0.0061, 0.0136, 0.0098, 0.0152, 0.0050, 0.0094, 0.0042, 0.0049], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0081, 0.0070, 0.0077, 0.0069, 0.0077, 0.0044, 0.0053], device='cuda:4'), out_proj_covar=tensor([6.0770e-05, 1.2445e-04, 1.0765e-04, 1.2146e-04, 1.1050e-04, 1.2309e-04, 6.8681e-05, 8.8213e-05], device='cuda:4') 2023-04-27 16:58:06,962 INFO [train.py:904] (4/8) Epoch 2, batch 1000, loss[loss=0.24, simple_loss=0.315, pruned_loss=0.08252, over 17086.00 frames. ], tot_loss[loss=0.2674, simple_loss=0.3293, pruned_loss=0.1028, over 3308318.08 frames. ], batch size: 49, lr: 3.14e-02, grad_scale: 8.0 2023-04-27 16:58:16,902 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=11158.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:58:37,563 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11174.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:59:11,378 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-27 16:59:13,979 INFO [train.py:904] (4/8) Epoch 2, batch 1050, loss[loss=0.2692, simple_loss=0.3211, pruned_loss=0.1087, over 16720.00 frames. ], tot_loss[loss=0.2667, simple_loss=0.3287, pruned_loss=0.1023, over 3312560.44 frames. ], batch size: 134, lr: 3.13e-02, grad_scale: 8.0 2023-04-27 16:59:36,189 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.907e+02 3.907e+02 4.719e+02 5.702e+02 1.083e+03, threshold=9.439e+02, percent-clipped=2.0 2023-04-27 16:59:59,238 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-27 17:00:00,077 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11235.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:00:08,424 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11241.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 17:00:21,895 INFO [train.py:904] (4/8) Epoch 2, batch 1100, loss[loss=0.2711, simple_loss=0.3445, pruned_loss=0.09886, over 16729.00 frames. ], tot_loss[loss=0.2648, simple_loss=0.3273, pruned_loss=0.1012, over 3322064.69 frames. ], batch size: 57, lr: 3.12e-02, grad_scale: 8.0 2023-04-27 17:00:45,351 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.70 vs. limit=5.0 2023-04-27 17:00:53,173 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8582, 4.5980, 4.7518, 5.2322, 5.2423, 4.7366, 5.3383, 5.1037], device='cuda:4'), covar=tensor([0.0436, 0.0509, 0.1158, 0.0328, 0.0331, 0.0529, 0.0229, 0.0280], device='cuda:4'), in_proj_covar=tensor([0.0215, 0.0240, 0.0356, 0.0250, 0.0207, 0.0204, 0.0186, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:01:03,748 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3945, 4.9124, 4.8895, 4.4719, 5.1625, 2.3751, 4.8366, 5.2065], device='cuda:4'), covar=tensor([0.0050, 0.0083, 0.0077, 0.0326, 0.0043, 0.1004, 0.0080, 0.0085], device='cuda:4'), in_proj_covar=tensor([0.0060, 0.0051, 0.0071, 0.0096, 0.0054, 0.0103, 0.0070, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:01:26,584 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6187, 4.3703, 3.9273, 1.7753, 2.9552, 2.2719, 3.9179, 4.6404], device='cuda:4'), covar=tensor([0.0225, 0.0314, 0.0340, 0.1947, 0.0873, 0.1168, 0.0638, 0.0220], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0093, 0.0131, 0.0153, 0.0143, 0.0134, 0.0144, 0.0091], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 17:01:28,364 INFO [train.py:904] (4/8) Epoch 2, batch 1150, loss[loss=0.2454, simple_loss=0.323, pruned_loss=0.08389, over 16730.00 frames. ], tot_loss[loss=0.2634, simple_loss=0.3265, pruned_loss=0.1002, over 3317550.59 frames. ], batch size: 57, lr: 3.12e-02, grad_scale: 8.0 2023-04-27 17:01:52,665 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.647e+02 4.017e+02 4.974e+02 5.619e+02 1.017e+03, threshold=9.949e+02, percent-clipped=1.0 2023-04-27 17:02:36,113 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8604, 3.7643, 3.8906, 4.2708, 4.2100, 3.9718, 4.1029, 4.1193], device='cuda:4'), covar=tensor([0.0425, 0.0486, 0.1040, 0.0301, 0.0373, 0.0635, 0.0460, 0.0339], device='cuda:4'), in_proj_covar=tensor([0.0214, 0.0239, 0.0356, 0.0248, 0.0207, 0.0204, 0.0185, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:02:39,328 INFO [train.py:904] (4/8) Epoch 2, batch 1200, loss[loss=0.2736, simple_loss=0.3272, pruned_loss=0.1099, over 12509.00 frames. ], tot_loss[loss=0.2627, simple_loss=0.3256, pruned_loss=0.09991, over 3316173.80 frames. ], batch size: 246, lr: 3.11e-02, grad_scale: 8.0 2023-04-27 17:02:51,666 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0826, 1.9151, 2.0554, 1.9411, 2.7134, 2.6151, 3.6335, 3.2703], device='cuda:4'), covar=tensor([0.0023, 0.0166, 0.0124, 0.0166, 0.0081, 0.0118, 0.0021, 0.0051], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0081, 0.0071, 0.0079, 0.0070, 0.0079, 0.0044, 0.0053], device='cuda:4'), out_proj_covar=tensor([6.3736e-05, 1.2511e-04, 1.0941e-04, 1.2552e-04, 1.1121e-04, 1.2702e-04, 6.9019e-05, 8.9107e-05], device='cuda:4') 2023-04-27 17:03:46,826 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11400.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:03:47,487 INFO [train.py:904] (4/8) Epoch 2, batch 1250, loss[loss=0.3377, simple_loss=0.3709, pruned_loss=0.1523, over 16763.00 frames. ], tot_loss[loss=0.2635, simple_loss=0.3262, pruned_loss=0.1004, over 3326199.82 frames. ], batch size: 124, lr: 3.11e-02, grad_scale: 8.0 2023-04-27 17:04:05,542 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3155, 5.0450, 4.8756, 4.3955, 5.0767, 2.0898, 4.8971, 5.1982], device='cuda:4'), covar=tensor([0.0059, 0.0064, 0.0072, 0.0332, 0.0045, 0.1206, 0.0069, 0.0083], device='cuda:4'), in_proj_covar=tensor([0.0058, 0.0052, 0.0071, 0.0096, 0.0055, 0.0101, 0.0069, 0.0072], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:04:10,335 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.776e+02 4.393e+02 5.372e+02 6.605e+02 1.355e+03, threshold=1.074e+03, percent-clipped=5.0 2023-04-27 17:04:19,707 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-27 17:04:26,787 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11430.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:04:34,426 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2004, 5.5053, 5.0915, 5.4753, 4.7981, 4.7345, 5.0130, 5.5468], device='cuda:4'), covar=tensor([0.0421, 0.0582, 0.0771, 0.0256, 0.0530, 0.0416, 0.0464, 0.0445], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0282, 0.0244, 0.0165, 0.0189, 0.0161, 0.0225, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:04:39,303 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-27 17:04:49,846 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=11448.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:04:53,725 INFO [train.py:904] (4/8) Epoch 2, batch 1300, loss[loss=0.3097, simple_loss=0.3516, pruned_loss=0.1338, over 12336.00 frames. ], tot_loss[loss=0.2637, simple_loss=0.326, pruned_loss=0.1007, over 3322434.32 frames. ], batch size: 246, lr: 3.10e-02, grad_scale: 8.0 2023-04-27 17:05:30,714 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=11478.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:06:01,549 INFO [train.py:904] (4/8) Epoch 2, batch 1350, loss[loss=0.2751, simple_loss=0.3301, pruned_loss=0.11, over 16447.00 frames. ], tot_loss[loss=0.2637, simple_loss=0.3263, pruned_loss=0.1006, over 3316557.88 frames. ], batch size: 146, lr: 3.10e-02, grad_scale: 8.0 2023-04-27 17:06:24,047 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8392, 4.6975, 4.5288, 4.0737, 4.6501, 2.2681, 4.3159, 4.7526], device='cuda:4'), covar=tensor([0.0078, 0.0063, 0.0081, 0.0318, 0.0062, 0.1170, 0.0086, 0.0091], device='cuda:4'), in_proj_covar=tensor([0.0061, 0.0053, 0.0074, 0.0099, 0.0057, 0.0104, 0.0071, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:06:24,752 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.664e+02 4.249e+02 5.319e+02 6.551e+02 1.548e+03, threshold=1.064e+03, percent-clipped=4.0 2023-04-27 17:06:36,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1702, 4.1607, 1.9860, 4.1389, 2.5617, 4.1345, 2.4078, 2.9149], device='cuda:4'), covar=tensor([0.0060, 0.0140, 0.1518, 0.0056, 0.0841, 0.0227, 0.1180, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0078, 0.0093, 0.0166, 0.0078, 0.0153, 0.0114, 0.0172, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 17:06:41,096 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11530.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:06:45,058 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11532.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:06:57,529 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11541.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 17:07:09,714 INFO [train.py:904] (4/8) Epoch 2, batch 1400, loss[loss=0.2511, simple_loss=0.3026, pruned_loss=0.09979, over 16383.00 frames. ], tot_loss[loss=0.2634, simple_loss=0.3262, pruned_loss=0.1003, over 3322553.32 frames. ], batch size: 165, lr: 3.09e-02, grad_scale: 8.0 2023-04-27 17:07:14,172 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-27 17:08:03,053 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=11589.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 17:08:05,960 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7617, 3.8458, 4.2206, 3.6003, 4.0819, 4.0121, 4.4339, 2.1598], device='cuda:4'), covar=tensor([0.0753, 0.0076, 0.0059, 0.0295, 0.0063, 0.0091, 0.0056, 0.0856], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0051, 0.0057, 0.0099, 0.0051, 0.0056, 0.0061, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 17:08:08,120 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11593.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:08:19,413 INFO [train.py:904] (4/8) Epoch 2, batch 1450, loss[loss=0.2569, simple_loss=0.3081, pruned_loss=0.1028, over 16330.00 frames. ], tot_loss[loss=0.2639, simple_loss=0.3265, pruned_loss=0.1006, over 3318938.90 frames. ], batch size: 165, lr: 3.08e-02, grad_scale: 8.0 2023-04-27 17:08:43,788 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.475e+02 3.985e+02 4.876e+02 5.970e+02 8.943e+02, threshold=9.752e+02, percent-clipped=0.0 2023-04-27 17:09:06,435 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11635.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:09:25,418 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7892, 1.9291, 1.6204, 1.7888, 2.5328, 2.6510, 2.7435, 2.7337], device='cuda:4'), covar=tensor([0.0078, 0.0164, 0.0155, 0.0216, 0.0084, 0.0125, 0.0054, 0.0070], device='cuda:4'), in_proj_covar=tensor([0.0041, 0.0083, 0.0074, 0.0082, 0.0073, 0.0081, 0.0046, 0.0056], device='cuda:4'), out_proj_covar=tensor([6.4889e-05, 1.2736e-04, 1.1538e-04, 1.2978e-04, 1.1704e-04, 1.3103e-04, 7.2006e-05, 9.4650e-05], device='cuda:4') 2023-04-27 17:09:26,110 INFO [train.py:904] (4/8) Epoch 2, batch 1500, loss[loss=0.2636, simple_loss=0.3134, pruned_loss=0.1069, over 16868.00 frames. ], tot_loss[loss=0.2628, simple_loss=0.3254, pruned_loss=0.1001, over 3322972.15 frames. ], batch size: 96, lr: 3.08e-02, grad_scale: 8.0 2023-04-27 17:09:58,965 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0520, 4.0611, 2.3643, 4.6857, 4.7192, 4.3968, 2.5401, 3.6208], device='cuda:4'), covar=tensor([0.1741, 0.0315, 0.1521, 0.0138, 0.0171, 0.0337, 0.1040, 0.0471], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0108, 0.0162, 0.0068, 0.0089, 0.0099, 0.0148, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:10:30,029 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11696.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:10:33,618 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11699.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:10:36,151 INFO [train.py:904] (4/8) Epoch 2, batch 1550, loss[loss=0.3201, simple_loss=0.3587, pruned_loss=0.1407, over 16770.00 frames. ], tot_loss[loss=0.2665, simple_loss=0.328, pruned_loss=0.1024, over 3329840.30 frames. ], batch size: 134, lr: 3.07e-02, grad_scale: 8.0 2023-04-27 17:10:58,965 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.956e+02 4.303e+02 5.262e+02 6.472e+02 2.309e+03, threshold=1.052e+03, percent-clipped=9.0 2023-04-27 17:11:44,741 INFO [train.py:904] (4/8) Epoch 2, batch 1600, loss[loss=0.2655, simple_loss=0.3421, pruned_loss=0.09439, over 17045.00 frames. ], tot_loss[loss=0.2697, simple_loss=0.331, pruned_loss=0.1042, over 3328950.26 frames. ], batch size: 55, lr: 3.07e-02, grad_scale: 8.0 2023-04-27 17:11:56,574 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9727, 4.5222, 4.4232, 1.8424, 4.6680, 4.6423, 3.7016, 4.0574], device='cuda:4'), covar=tensor([0.0635, 0.0061, 0.0163, 0.1650, 0.0069, 0.0041, 0.0253, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0077, 0.0075, 0.0151, 0.0069, 0.0067, 0.0094, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0001, 0.0003, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:11:57,719 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11760.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:12:53,457 INFO [train.py:904] (4/8) Epoch 2, batch 1650, loss[loss=0.2249, simple_loss=0.3002, pruned_loss=0.07478, over 17218.00 frames. ], tot_loss[loss=0.2699, simple_loss=0.3319, pruned_loss=0.104, over 3328809.37 frames. ], batch size: 43, lr: 3.06e-02, grad_scale: 8.0 2023-04-27 17:13:13,213 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-27 17:13:15,793 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11816.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:13:17,808 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.782e+02 3.949e+02 4.848e+02 6.012e+02 1.253e+03, threshold=9.697e+02, percent-clipped=3.0 2023-04-27 17:13:19,509 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0066, 1.9873, 1.8719, 1.8587, 2.7168, 2.6027, 3.5367, 2.9232], device='cuda:4'), covar=tensor([0.0023, 0.0193, 0.0159, 0.0215, 0.0092, 0.0137, 0.0051, 0.0079], device='cuda:4'), in_proj_covar=tensor([0.0041, 0.0083, 0.0074, 0.0083, 0.0072, 0.0081, 0.0047, 0.0056], device='cuda:4'), out_proj_covar=tensor([6.5443e-05, 1.2718e-04, 1.1541e-04, 1.3138e-04, 1.1560e-04, 1.3178e-04, 7.4533e-05, 9.5508e-05], device='cuda:4') 2023-04-27 17:13:35,559 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11830.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:13:56,916 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11846.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:14:03,718 INFO [train.py:904] (4/8) Epoch 2, batch 1700, loss[loss=0.2677, simple_loss=0.3275, pruned_loss=0.104, over 16796.00 frames. ], tot_loss[loss=0.2709, simple_loss=0.3329, pruned_loss=0.1045, over 3330834.54 frames. ], batch size: 102, lr: 3.06e-02, grad_scale: 8.0 2023-04-27 17:14:40,628 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11877.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:14:41,657 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=11878.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:14:55,343 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11888.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:15:13,200 INFO [train.py:904] (4/8) Epoch 2, batch 1750, loss[loss=0.3825, simple_loss=0.4163, pruned_loss=0.1743, over 12183.00 frames. ], tot_loss[loss=0.2727, simple_loss=0.3345, pruned_loss=0.1055, over 3323695.34 frames. ], batch size: 246, lr: 3.05e-02, grad_scale: 8.0 2023-04-27 17:15:21,944 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11907.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:15:36,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.686e+02 4.328e+02 5.055e+02 5.950e+02 1.641e+03, threshold=1.011e+03, percent-clipped=5.0 2023-04-27 17:15:39,032 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.74 vs. limit=2.0 2023-04-27 17:15:39,640 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11920.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:16:17,846 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11948.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:16:21,481 INFO [train.py:904] (4/8) Epoch 2, batch 1800, loss[loss=0.3196, simple_loss=0.3572, pruned_loss=0.1411, over 16862.00 frames. ], tot_loss[loss=0.2748, simple_loss=0.3368, pruned_loss=0.1064, over 3320867.31 frames. ], batch size: 109, lr: 3.05e-02, grad_scale: 8.0 2023-04-27 17:17:03,312 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11981.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:17:07,149 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11984.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:17:15,725 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-27 17:17:16,283 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11991.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:17:33,761 INFO [train.py:904] (4/8) Epoch 2, batch 1850, loss[loss=0.2505, simple_loss=0.3224, pruned_loss=0.08931, over 16870.00 frames. ], tot_loss[loss=0.2751, simple_loss=0.3373, pruned_loss=0.1065, over 3319406.69 frames. ], batch size: 42, lr: 3.04e-02, grad_scale: 8.0 2023-04-27 17:17:35,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3479, 3.9515, 4.0533, 3.6102, 4.5333, 4.1826, 4.6725, 2.4130], device='cuda:4'), covar=tensor([0.0833, 0.0103, 0.0111, 0.0285, 0.0042, 0.0121, 0.0045, 0.0712], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0050, 0.0058, 0.0098, 0.0050, 0.0054, 0.0059, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002], device='cuda:4') 2023-04-27 17:17:45,627 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12009.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:17:57,887 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.566e+02 4.018e+02 4.936e+02 6.102e+02 1.131e+03, threshold=9.871e+02, percent-clipped=1.0 2023-04-27 17:18:35,109 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12045.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:18:43,196 INFO [train.py:904] (4/8) Epoch 2, batch 1900, loss[loss=0.3206, simple_loss=0.3823, pruned_loss=0.1295, over 11887.00 frames. ], tot_loss[loss=0.2724, simple_loss=0.3355, pruned_loss=0.1047, over 3315955.88 frames. ], batch size: 246, lr: 3.04e-02, grad_scale: 8.0 2023-04-27 17:18:49,559 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12055.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:19:09,520 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0444, 4.0557, 3.8941, 4.0266, 3.9812, 4.5874, 4.3578, 4.0162], device='cuda:4'), covar=tensor([0.1411, 0.1362, 0.1459, 0.1739, 0.2620, 0.0967, 0.1020, 0.2128], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0267, 0.0237, 0.0231, 0.0296, 0.0235, 0.0209, 0.0298], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:19:17,216 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=12076.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:19:32,929 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-27 17:19:51,452 INFO [train.py:904] (4/8) Epoch 2, batch 1950, loss[loss=0.2429, simple_loss=0.3163, pruned_loss=0.08475, over 17114.00 frames. ], tot_loss[loss=0.2713, simple_loss=0.3349, pruned_loss=0.1039, over 3311623.66 frames. ], batch size: 47, lr: 3.03e-02, grad_scale: 8.0 2023-04-27 17:20:14,619 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.438e+02 4.259e+02 5.225e+02 6.323e+02 1.109e+03, threshold=1.045e+03, percent-clipped=2.0 2023-04-27 17:20:40,772 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12137.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:20:59,924 INFO [train.py:904] (4/8) Epoch 2, batch 2000, loss[loss=0.2708, simple_loss=0.3218, pruned_loss=0.1099, over 16621.00 frames. ], tot_loss[loss=0.2703, simple_loss=0.3343, pruned_loss=0.1031, over 3317960.30 frames. ], batch size: 75, lr: 3.02e-02, grad_scale: 8.0 2023-04-27 17:21:18,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8789, 1.7245, 1.4776, 1.3307, 2.1914, 1.9788, 2.1404, 2.2565], device='cuda:4'), covar=tensor([0.0054, 0.0148, 0.0134, 0.0193, 0.0073, 0.0137, 0.0071, 0.0066], device='cuda:4'), in_proj_covar=tensor([0.0044, 0.0086, 0.0077, 0.0085, 0.0075, 0.0085, 0.0047, 0.0059], device='cuda:4'), out_proj_covar=tensor([7.1978e-05, 1.3366e-04, 1.1982e-04, 1.3393e-04, 1.2059e-04, 1.3838e-04, 7.5598e-05, 1.0162e-04], device='cuda:4') 2023-04-27 17:21:23,914 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0496, 4.2074, 2.6271, 5.1916, 5.0082, 4.5482, 2.0113, 4.0514], device='cuda:4'), covar=tensor([0.1620, 0.0303, 0.1379, 0.0054, 0.0149, 0.0316, 0.1232, 0.0405], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0109, 0.0164, 0.0067, 0.0095, 0.0102, 0.0148, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:21:28,808 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12172.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:21:51,917 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12188.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:22:09,025 INFO [train.py:904] (4/8) Epoch 2, batch 2050, loss[loss=0.2359, simple_loss=0.3185, pruned_loss=0.07666, over 17057.00 frames. ], tot_loss[loss=0.2711, simple_loss=0.3346, pruned_loss=0.1038, over 3320352.67 frames. ], batch size: 53, lr: 3.02e-02, grad_scale: 16.0 2023-04-27 17:22:11,061 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12202.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:22:32,914 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.569e+02 4.228e+02 5.159e+02 6.169e+02 1.128e+03, threshold=1.032e+03, percent-clipped=1.0 2023-04-27 17:22:58,999 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12236.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:23:20,403 INFO [train.py:904] (4/8) Epoch 2, batch 2100, loss[loss=0.248, simple_loss=0.3289, pruned_loss=0.08352, over 17269.00 frames. ], tot_loss[loss=0.2724, simple_loss=0.3356, pruned_loss=0.1046, over 3323925.02 frames. ], batch size: 52, lr: 3.01e-02, grad_scale: 16.0 2023-04-27 17:23:36,180 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-04-27 17:23:51,514 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9917, 2.8004, 2.1308, 3.3566, 3.1367, 3.3929, 1.8864, 2.6828], device='cuda:4'), covar=tensor([0.1194, 0.0289, 0.1193, 0.0070, 0.0166, 0.0207, 0.0958, 0.0492], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0106, 0.0161, 0.0065, 0.0093, 0.0099, 0.0144, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:23:53,735 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12276.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:23:58,091 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.21 vs. limit=5.0 2023-04-27 17:24:15,543 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12291.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:24:21,675 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-27 17:24:28,737 INFO [train.py:904] (4/8) Epoch 2, batch 2150, loss[loss=0.265, simple_loss=0.3463, pruned_loss=0.09182, over 16732.00 frames. ], tot_loss[loss=0.2725, simple_loss=0.3358, pruned_loss=0.1046, over 3324018.39 frames. ], batch size: 62, lr: 3.01e-02, grad_scale: 16.0 2023-04-27 17:24:33,511 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12304.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:24:51,834 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.440e+02 4.527e+02 5.287e+02 6.148e+02 1.100e+03, threshold=1.057e+03, percent-clipped=4.0 2023-04-27 17:24:52,316 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2057, 1.6884, 1.5062, 1.4323, 2.1819, 1.8575, 2.0687, 2.3159], device='cuda:4'), covar=tensor([0.0046, 0.0144, 0.0136, 0.0173, 0.0066, 0.0137, 0.0076, 0.0065], device='cuda:4'), in_proj_covar=tensor([0.0044, 0.0085, 0.0078, 0.0084, 0.0074, 0.0086, 0.0048, 0.0060], device='cuda:4'), out_proj_covar=tensor([7.2643e-05, 1.3093e-04, 1.2026e-04, 1.3315e-04, 1.1835e-04, 1.4045e-04, 7.6523e-05, 1.0206e-04], device='cuda:4') 2023-04-27 17:25:22,226 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12339.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:25:23,416 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12340.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:25:38,415 INFO [train.py:904] (4/8) Epoch 2, batch 2200, loss[loss=0.279, simple_loss=0.3334, pruned_loss=0.1123, over 16719.00 frames. ], tot_loss[loss=0.273, simple_loss=0.3361, pruned_loss=0.105, over 3318494.02 frames. ], batch size: 89, lr: 3.00e-02, grad_scale: 16.0 2023-04-27 17:25:44,357 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12355.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:26:27,529 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.96 vs. limit=2.0 2023-04-27 17:26:48,777 INFO [train.py:904] (4/8) Epoch 2, batch 2250, loss[loss=0.3194, simple_loss=0.367, pruned_loss=0.136, over 15573.00 frames. ], tot_loss[loss=0.2757, simple_loss=0.338, pruned_loss=0.1067, over 3311004.49 frames. ], batch size: 191, lr: 3.00e-02, grad_scale: 16.0 2023-04-27 17:26:51,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12403.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:27:12,082 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.514e+02 4.351e+02 5.087e+02 6.619e+02 9.905e+02, threshold=1.017e+03, percent-clipped=0.0 2023-04-27 17:27:32,276 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12432.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:27:51,322 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=12446.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:27:57,826 INFO [train.py:904] (4/8) Epoch 2, batch 2300, loss[loss=0.2334, simple_loss=0.3133, pruned_loss=0.07675, over 17119.00 frames. ], tot_loss[loss=0.2764, simple_loss=0.3391, pruned_loss=0.1069, over 3302149.04 frames. ], batch size: 48, lr: 2.99e-02, grad_scale: 8.0 2023-04-27 17:28:27,170 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5223, 4.3302, 4.3623, 4.7970, 4.7745, 4.3729, 4.6640, 4.6795], device='cuda:4'), covar=tensor([0.0422, 0.0451, 0.1187, 0.0351, 0.0334, 0.0707, 0.0491, 0.0322], device='cuda:4'), in_proj_covar=tensor([0.0230, 0.0263, 0.0384, 0.0275, 0.0220, 0.0212, 0.0198, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:28:27,185 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12472.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:29:06,504 INFO [train.py:904] (4/8) Epoch 2, batch 2350, loss[loss=0.2803, simple_loss=0.3396, pruned_loss=0.1105, over 16374.00 frames. ], tot_loss[loss=0.2787, simple_loss=0.3403, pruned_loss=0.1086, over 3294371.15 frames. ], batch size: 68, lr: 2.99e-02, grad_scale: 8.0 2023-04-27 17:29:07,831 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12502.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:29:15,039 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12507.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:29:31,565 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.397e+02 4.650e+02 5.724e+02 7.343e+02 1.924e+03, threshold=1.145e+03, percent-clipped=10.0 2023-04-27 17:29:33,872 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12520.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:30:14,668 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12550.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:30:15,611 INFO [train.py:904] (4/8) Epoch 2, batch 2400, loss[loss=0.293, simple_loss=0.3571, pruned_loss=0.1144, over 16471.00 frames. ], tot_loss[loss=0.2784, simple_loss=0.341, pruned_loss=0.1079, over 3307891.00 frames. ], batch size: 68, lr: 2.98e-02, grad_scale: 8.0 2023-04-27 17:30:52,415 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12576.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:31:26,490 INFO [train.py:904] (4/8) Epoch 2, batch 2450, loss[loss=0.2693, simple_loss=0.3302, pruned_loss=0.1042, over 16800.00 frames. ], tot_loss[loss=0.2773, simple_loss=0.3408, pruned_loss=0.1069, over 3309887.22 frames. ], batch size: 102, lr: 2.98e-02, grad_scale: 8.0 2023-04-27 17:31:31,833 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12604.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:31:51,020 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.702e+02 4.143e+02 4.943e+02 6.147e+02 1.117e+03, threshold=9.887e+02, percent-clipped=0.0 2023-04-27 17:31:59,227 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12624.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:32:13,830 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.16 vs. limit=2.0 2023-04-27 17:32:20,242 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12640.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:32:35,554 INFO [train.py:904] (4/8) Epoch 2, batch 2500, loss[loss=0.2914, simple_loss=0.3453, pruned_loss=0.1187, over 16711.00 frames. ], tot_loss[loss=0.2772, simple_loss=0.3412, pruned_loss=0.1066, over 3309989.15 frames. ], batch size: 124, lr: 2.97e-02, grad_scale: 8.0 2023-04-27 17:32:36,846 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12652.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:33:27,073 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12688.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:33:43,145 INFO [train.py:904] (4/8) Epoch 2, batch 2550, loss[loss=0.2939, simple_loss=0.3472, pruned_loss=0.1203, over 16627.00 frames. ], tot_loss[loss=0.2763, simple_loss=0.3406, pruned_loss=0.106, over 3316048.13 frames. ], batch size: 134, lr: 2.97e-02, grad_scale: 8.0 2023-04-27 17:33:59,921 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1134, 4.5815, 4.9287, 4.9363, 4.4075, 4.8186, 4.9964, 4.5098], device='cuda:4'), covar=tensor([0.0281, 0.0192, 0.0169, 0.0129, 0.0759, 0.0219, 0.0136, 0.0253], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0090, 0.0165, 0.0132, 0.0205, 0.0136, 0.0117, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:34:08,155 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.860e+02 4.439e+02 5.440e+02 6.842e+02 1.130e+03, threshold=1.088e+03, percent-clipped=2.0 2023-04-27 17:34:27,688 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12732.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:34:53,761 INFO [train.py:904] (4/8) Epoch 2, batch 2600, loss[loss=0.2157, simple_loss=0.2972, pruned_loss=0.06709, over 16813.00 frames. ], tot_loss[loss=0.2749, simple_loss=0.3397, pruned_loss=0.1051, over 3320498.83 frames. ], batch size: 42, lr: 2.96e-02, grad_scale: 8.0 2023-04-27 17:35:08,892 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=12763.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:35:23,825 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8166, 4.0533, 3.2998, 3.3771, 3.0311, 2.2489, 4.1731, 4.6332], device='cuda:4'), covar=tensor([0.1667, 0.0433, 0.0857, 0.0469, 0.1696, 0.1195, 0.0256, 0.0229], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0223, 0.0231, 0.0157, 0.0256, 0.0178, 0.0179, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:35:32,886 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=12780.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:35:40,511 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3849, 3.0611, 2.5975, 2.6399, 2.3010, 1.9199, 2.9932, 3.3290], device='cuda:4'), covar=tensor([0.1271, 0.0540, 0.0828, 0.0390, 0.1551, 0.1198, 0.0318, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0224, 0.0232, 0.0158, 0.0258, 0.0179, 0.0179, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:35:46,189 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8737, 1.7860, 2.2127, 2.5515, 2.7833, 2.7594, 1.6529, 2.7672], device='cuda:4'), covar=tensor([0.0055, 0.0274, 0.0163, 0.0132, 0.0036, 0.0077, 0.0204, 0.0051], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0107, 0.0090, 0.0080, 0.0054, 0.0054, 0.0086, 0.0054], device='cuda:4'), out_proj_covar=tensor([1.3124e-04, 1.9447e-04, 1.7232e-04, 1.5425e-04, 9.6813e-05, 1.0375e-04, 1.5053e-04, 1.0174e-04], device='cuda:4') 2023-04-27 17:36:01,522 INFO [train.py:904] (4/8) Epoch 2, batch 2650, loss[loss=0.327, simple_loss=0.3702, pruned_loss=0.1419, over 12500.00 frames. ], tot_loss[loss=0.2743, simple_loss=0.3395, pruned_loss=0.1046, over 3313544.50 frames. ], batch size: 247, lr: 2.96e-02, grad_scale: 8.0 2023-04-27 17:36:03,294 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12802.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:36:26,584 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 17:36:27,046 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.569e+02 4.221e+02 4.762e+02 5.639e+02 1.002e+03, threshold=9.524e+02, percent-clipped=0.0 2023-04-27 17:36:28,893 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-27 17:36:34,036 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12824.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 17:37:09,358 INFO [train.py:904] (4/8) Epoch 2, batch 2700, loss[loss=0.2519, simple_loss=0.3299, pruned_loss=0.08693, over 17139.00 frames. ], tot_loss[loss=0.2725, simple_loss=0.3389, pruned_loss=0.1031, over 3317906.16 frames. ], batch size: 47, lr: 2.95e-02, grad_scale: 8.0 2023-04-27 17:38:19,507 INFO [train.py:904] (4/8) Epoch 2, batch 2750, loss[loss=0.2302, simple_loss=0.3178, pruned_loss=0.07127, over 17130.00 frames. ], tot_loss[loss=0.2706, simple_loss=0.338, pruned_loss=0.1016, over 3319821.15 frames. ], batch size: 47, lr: 2.95e-02, grad_scale: 8.0 2023-04-27 17:38:41,652 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.401e+02 4.125e+02 4.810e+02 5.955e+02 1.093e+03, threshold=9.620e+02, percent-clipped=2.0 2023-04-27 17:39:26,263 INFO [train.py:904] (4/8) Epoch 2, batch 2800, loss[loss=0.2717, simple_loss=0.3284, pruned_loss=0.1075, over 16477.00 frames. ], tot_loss[loss=0.2711, simple_loss=0.338, pruned_loss=0.1021, over 3322414.86 frames. ], batch size: 146, lr: 2.94e-02, grad_scale: 8.0 2023-04-27 17:40:33,619 INFO [train.py:904] (4/8) Epoch 2, batch 2850, loss[loss=0.2953, simple_loss=0.3682, pruned_loss=0.1112, over 17135.00 frames. ], tot_loss[loss=0.2715, simple_loss=0.3376, pruned_loss=0.1027, over 3315469.21 frames. ], batch size: 48, lr: 2.94e-02, grad_scale: 8.0 2023-04-27 17:40:57,336 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.754e+02 4.789e+02 5.762e+02 6.825e+02 1.913e+03, threshold=1.152e+03, percent-clipped=8.0 2023-04-27 17:41:13,872 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5931, 4.4425, 4.4246, 1.8286, 4.6078, 4.6202, 3.3742, 3.7841], device='cuda:4'), covar=tensor([0.0740, 0.0062, 0.0140, 0.1394, 0.0069, 0.0046, 0.0257, 0.0190], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0079, 0.0078, 0.0148, 0.0071, 0.0068, 0.0095, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:41:18,156 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5560, 4.5315, 4.4783, 1.8658, 3.2365, 2.2349, 4.0911, 4.4892], device='cuda:4'), covar=tensor([0.0248, 0.0385, 0.0254, 0.1709, 0.0703, 0.1219, 0.0628, 0.0409], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0108, 0.0141, 0.0154, 0.0146, 0.0139, 0.0150, 0.0102], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 17:41:21,530 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5456, 5.8714, 5.4777, 5.7033, 5.0223, 4.8017, 5.2978, 5.9404], device='cuda:4'), covar=tensor([0.0442, 0.0556, 0.0800, 0.0350, 0.0547, 0.0413, 0.0413, 0.0433], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0307, 0.0271, 0.0189, 0.0211, 0.0177, 0.0243, 0.0210], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:41:41,130 INFO [train.py:904] (4/8) Epoch 2, batch 2900, loss[loss=0.3217, simple_loss=0.3565, pruned_loss=0.1434, over 16871.00 frames. ], tot_loss[loss=0.2713, simple_loss=0.3371, pruned_loss=0.1027, over 3325143.66 frames. ], batch size: 116, lr: 2.93e-02, grad_scale: 8.0 2023-04-27 17:42:20,729 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7559, 5.2509, 5.1408, 5.1297, 5.1038, 5.6745, 5.4972, 5.1519], device='cuda:4'), covar=tensor([0.0582, 0.0930, 0.0816, 0.1309, 0.1692, 0.0613, 0.0668, 0.1412], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0266, 0.0236, 0.0231, 0.0296, 0.0242, 0.0208, 0.0301], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:42:49,006 INFO [train.py:904] (4/8) Epoch 2, batch 2950, loss[loss=0.2824, simple_loss=0.3339, pruned_loss=0.1154, over 16641.00 frames. ], tot_loss[loss=0.2714, simple_loss=0.3365, pruned_loss=0.1032, over 3330019.94 frames. ], batch size: 134, lr: 2.93e-02, grad_scale: 8.0 2023-04-27 17:42:50,408 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=13102.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:43:13,803 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.070e+02 4.352e+02 4.979e+02 5.894e+02 1.092e+03, threshold=9.958e+02, percent-clipped=0.0 2023-04-27 17:43:14,145 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=13119.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 17:43:53,663 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=13150.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:43:54,563 INFO [train.py:904] (4/8) Epoch 2, batch 3000, loss[loss=0.443, simple_loss=0.4488, pruned_loss=0.2186, over 11904.00 frames. ], tot_loss[loss=0.2716, simple_loss=0.3366, pruned_loss=0.1033, over 3332403.92 frames. ], batch size: 246, lr: 2.92e-02, grad_scale: 8.0 2023-04-27 17:43:54,563 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 17:44:03,914 INFO [train.py:938] (4/8) Epoch 2, validation: loss=0.1858, simple_loss=0.2917, pruned_loss=0.04, over 944034.00 frames. 2023-04-27 17:44:03,915 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-27 17:44:52,578 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1876, 1.5008, 1.8984, 2.1173, 2.3196, 2.2784, 1.7075, 2.1767], device='cuda:4'), covar=tensor([0.0055, 0.0231, 0.0120, 0.0119, 0.0035, 0.0088, 0.0172, 0.0050], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0110, 0.0092, 0.0083, 0.0056, 0.0057, 0.0094, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 17:45:09,406 INFO [train.py:904] (4/8) Epoch 2, batch 3050, loss[loss=0.2954, simple_loss=0.3534, pruned_loss=0.1187, over 16545.00 frames. ], tot_loss[loss=0.2715, simple_loss=0.3363, pruned_loss=0.1034, over 3331241.80 frames. ], batch size: 68, lr: 2.92e-02, grad_scale: 8.0 2023-04-27 17:45:33,139 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.792e+02 4.585e+02 5.657e+02 6.871e+02 1.163e+03, threshold=1.131e+03, percent-clipped=2.0 2023-04-27 17:45:45,629 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9267, 1.9159, 1.6544, 1.7215, 2.4799, 2.5106, 2.7668, 2.6472], device='cuda:4'), covar=tensor([0.0052, 0.0123, 0.0124, 0.0147, 0.0057, 0.0095, 0.0042, 0.0053], device='cuda:4'), in_proj_covar=tensor([0.0045, 0.0088, 0.0081, 0.0085, 0.0076, 0.0086, 0.0054, 0.0063], device='cuda:4'), out_proj_covar=tensor([7.6458e-05, 1.3673e-04, 1.2446e-04, 1.3506e-04, 1.2544e-04, 1.4103e-04, 8.7187e-05, 1.0754e-04], device='cuda:4') 2023-04-27 17:46:15,658 INFO [train.py:904] (4/8) Epoch 2, batch 3100, loss[loss=0.2984, simple_loss=0.3492, pruned_loss=0.1238, over 16842.00 frames. ], tot_loss[loss=0.2708, simple_loss=0.335, pruned_loss=0.1032, over 3334085.91 frames. ], batch size: 102, lr: 2.91e-02, grad_scale: 8.0 2023-04-27 17:46:16,784 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-04-27 17:46:25,748 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1712, 1.7504, 1.7015, 1.6077, 2.0653, 2.1035, 2.3652, 2.3018], device='cuda:4'), covar=tensor([0.0036, 0.0112, 0.0101, 0.0128, 0.0053, 0.0104, 0.0043, 0.0051], device='cuda:4'), in_proj_covar=tensor([0.0045, 0.0088, 0.0082, 0.0085, 0.0077, 0.0087, 0.0053, 0.0063], device='cuda:4'), out_proj_covar=tensor([7.6253e-05, 1.3804e-04, 1.2705e-04, 1.3611e-04, 1.2697e-04, 1.4227e-04, 8.6698e-05, 1.0760e-04], device='cuda:4') 2023-04-27 17:47:22,153 INFO [train.py:904] (4/8) Epoch 2, batch 3150, loss[loss=0.2956, simple_loss=0.3581, pruned_loss=0.1165, over 16572.00 frames. ], tot_loss[loss=0.27, simple_loss=0.3347, pruned_loss=0.1027, over 3328216.16 frames. ], batch size: 68, lr: 2.91e-02, grad_scale: 8.0 2023-04-27 17:47:44,588 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.682e+02 3.986e+02 4.862e+02 6.258e+02 1.253e+03, threshold=9.723e+02, percent-clipped=2.0 2023-04-27 17:48:10,617 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1797, 4.0168, 4.3007, 3.2753, 4.2399, 4.2199, 4.5551, 2.5506], device='cuda:4'), covar=tensor([0.0814, 0.0046, 0.0052, 0.0339, 0.0035, 0.0066, 0.0033, 0.0586], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0048, 0.0056, 0.0102, 0.0053, 0.0057, 0.0060, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:48:28,291 INFO [train.py:904] (4/8) Epoch 2, batch 3200, loss[loss=0.2771, simple_loss=0.3268, pruned_loss=0.1137, over 16904.00 frames. ], tot_loss[loss=0.2683, simple_loss=0.3333, pruned_loss=0.1017, over 3327863.18 frames. ], batch size: 109, lr: 2.90e-02, grad_scale: 8.0 2023-04-27 17:49:34,225 INFO [train.py:904] (4/8) Epoch 2, batch 3250, loss[loss=0.2764, simple_loss=0.3347, pruned_loss=0.109, over 16849.00 frames. ], tot_loss[loss=0.2695, simple_loss=0.3342, pruned_loss=0.1024, over 3324702.95 frames. ], batch size: 90, lr: 2.90e-02, grad_scale: 8.0 2023-04-27 17:49:58,157 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.960e+02 4.105e+02 4.918e+02 6.860e+02 1.329e+03, threshold=9.835e+02, percent-clipped=3.0 2023-04-27 17:49:58,591 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=13419.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:50:17,310 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-27 17:50:42,672 INFO [train.py:904] (4/8) Epoch 2, batch 3300, loss[loss=0.3342, simple_loss=0.3707, pruned_loss=0.1488, over 16785.00 frames. ], tot_loss[loss=0.2702, simple_loss=0.3346, pruned_loss=0.1029, over 3314688.84 frames. ], batch size: 124, lr: 2.89e-02, grad_scale: 8.0 2023-04-27 17:51:03,261 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=13467.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:51:43,060 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.97 vs. limit=5.0 2023-04-27 17:51:48,084 INFO [train.py:904] (4/8) Epoch 2, batch 3350, loss[loss=0.2847, simple_loss=0.3392, pruned_loss=0.1151, over 16804.00 frames. ], tot_loss[loss=0.2705, simple_loss=0.3353, pruned_loss=0.1028, over 3310190.33 frames. ], batch size: 116, lr: 2.89e-02, grad_scale: 8.0 2023-04-27 17:51:53,728 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5646, 1.5622, 2.1715, 2.3464, 2.5208, 2.2422, 1.7786, 2.4977], device='cuda:4'), covar=tensor([0.0070, 0.0338, 0.0165, 0.0172, 0.0069, 0.0149, 0.0253, 0.0064], device='cuda:4'), in_proj_covar=tensor([0.0074, 0.0111, 0.0094, 0.0084, 0.0059, 0.0059, 0.0095, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 17:52:07,893 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5837, 5.9068, 5.6398, 5.8420, 5.1435, 4.7775, 5.3999, 6.0705], device='cuda:4'), covar=tensor([0.0450, 0.0529, 0.0599, 0.0278, 0.0456, 0.0384, 0.0412, 0.0414], device='cuda:4'), in_proj_covar=tensor([0.0221, 0.0310, 0.0272, 0.0192, 0.0207, 0.0182, 0.0243, 0.0214], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:52:13,283 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.954e+02 4.311e+02 5.188e+02 6.478e+02 1.182e+03, threshold=1.038e+03, percent-clipped=6.0 2023-04-27 17:52:43,040 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-27 17:52:56,206 INFO [train.py:904] (4/8) Epoch 2, batch 3400, loss[loss=0.2498, simple_loss=0.3338, pruned_loss=0.08294, over 17048.00 frames. ], tot_loss[loss=0.2697, simple_loss=0.3346, pruned_loss=0.1023, over 3308165.55 frames. ], batch size: 50, lr: 2.88e-02, grad_scale: 8.0 2023-04-27 17:54:05,273 INFO [train.py:904] (4/8) Epoch 2, batch 3450, loss[loss=0.298, simple_loss=0.3483, pruned_loss=0.1239, over 16488.00 frames. ], tot_loss[loss=0.2679, simple_loss=0.3333, pruned_loss=0.1012, over 3313772.33 frames. ], batch size: 68, lr: 2.88e-02, grad_scale: 8.0 2023-04-27 17:54:29,793 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.037e+02 4.547e+02 5.342e+02 6.587e+02 1.202e+03, threshold=1.068e+03, percent-clipped=2.0 2023-04-27 17:55:11,516 INFO [train.py:904] (4/8) Epoch 2, batch 3500, loss[loss=0.2473, simple_loss=0.3181, pruned_loss=0.08823, over 17204.00 frames. ], tot_loss[loss=0.2666, simple_loss=0.3314, pruned_loss=0.1008, over 3317385.35 frames. ], batch size: 44, lr: 2.88e-02, grad_scale: 8.0 2023-04-27 17:55:20,623 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7761, 3.9078, 2.9558, 3.4492, 2.9514, 2.2611, 4.1072, 4.5935], device='cuda:4'), covar=tensor([0.1587, 0.0519, 0.0955, 0.0386, 0.1717, 0.1174, 0.0248, 0.0186], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0222, 0.0230, 0.0158, 0.0252, 0.0175, 0.0179, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:56:21,450 INFO [train.py:904] (4/8) Epoch 2, batch 3550, loss[loss=0.2682, simple_loss=0.3239, pruned_loss=0.1062, over 16729.00 frames. ], tot_loss[loss=0.2647, simple_loss=0.3295, pruned_loss=0.09995, over 3326154.04 frames. ], batch size: 89, lr: 2.87e-02, grad_scale: 8.0 2023-04-27 17:56:32,051 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.68 vs. limit=2.0 2023-04-27 17:56:45,026 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.316e+02 4.227e+02 5.059e+02 5.919e+02 1.365e+03, threshold=1.012e+03, percent-clipped=3.0 2023-04-27 17:57:11,475 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=13737.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:57:17,546 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1777, 1.8904, 1.6380, 1.5503, 2.2880, 2.1774, 2.6132, 2.3685], device='cuda:4'), covar=tensor([0.0043, 0.0140, 0.0129, 0.0179, 0.0075, 0.0122, 0.0050, 0.0076], device='cuda:4'), in_proj_covar=tensor([0.0049, 0.0091, 0.0088, 0.0093, 0.0083, 0.0093, 0.0058, 0.0068], device='cuda:4'), out_proj_covar=tensor([8.3018e-05, 1.4222e-04, 1.3609e-04, 1.5052e-04, 1.3692e-04, 1.5256e-04, 9.5878e-05, 1.1691e-04], device='cuda:4') 2023-04-27 17:57:29,798 INFO [train.py:904] (4/8) Epoch 2, batch 3600, loss[loss=0.2979, simple_loss=0.3444, pruned_loss=0.1256, over 16900.00 frames. ], tot_loss[loss=0.2628, simple_loss=0.3277, pruned_loss=0.09897, over 3336495.74 frames. ], batch size: 109, lr: 2.87e-02, grad_scale: 8.0 2023-04-27 17:58:37,304 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=13798.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 17:58:40,991 INFO [train.py:904] (4/8) Epoch 2, batch 3650, loss[loss=0.2677, simple_loss=0.3158, pruned_loss=0.1098, over 16829.00 frames. ], tot_loss[loss=0.2635, simple_loss=0.3271, pruned_loss=0.09995, over 3318572.92 frames. ], batch size: 102, lr: 2.86e-02, grad_scale: 8.0 2023-04-27 17:59:00,184 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2071, 3.1568, 3.1709, 3.4356, 3.3974, 3.2701, 3.3481, 3.4596], device='cuda:4'), covar=tensor([0.0420, 0.0380, 0.0918, 0.0380, 0.0478, 0.0979, 0.0532, 0.0360], device='cuda:4'), in_proj_covar=tensor([0.0239, 0.0274, 0.0391, 0.0296, 0.0232, 0.0218, 0.0218, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:59:08,124 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.959e+02 4.047e+02 4.833e+02 5.657e+02 1.025e+03, threshold=9.667e+02, percent-clipped=2.0 2023-04-27 17:59:27,260 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.72 vs. limit=2.0 2023-04-27 17:59:52,524 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8578, 3.5945, 2.5130, 4.3804, 4.2105, 4.0871, 1.7598, 3.0181], device='cuda:4'), covar=tensor([0.1551, 0.0303, 0.1330, 0.0067, 0.0211, 0.0267, 0.1198, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0165, 0.0069, 0.0106, 0.0108, 0.0150, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 17:59:54,876 INFO [train.py:904] (4/8) Epoch 2, batch 3700, loss[loss=0.2701, simple_loss=0.3207, pruned_loss=0.1098, over 16443.00 frames. ], tot_loss[loss=0.2644, simple_loss=0.3257, pruned_loss=0.1016, over 3278408.24 frames. ], batch size: 146, lr: 2.86e-02, grad_scale: 8.0 2023-04-27 18:00:21,428 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4701, 3.9611, 3.7501, 1.8093, 3.9458, 3.9962, 3.2707, 3.2461], device='cuda:4'), covar=tensor([0.0698, 0.0075, 0.0196, 0.1422, 0.0084, 0.0058, 0.0297, 0.0277], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0081, 0.0077, 0.0147, 0.0075, 0.0069, 0.0100, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:00:23,977 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6605, 2.8496, 2.5365, 3.9243, 2.1508, 3.6368, 2.3966, 2.3029], device='cuda:4'), covar=tensor([0.0273, 0.0363, 0.0324, 0.0189, 0.1365, 0.0182, 0.0657, 0.1065], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0166, 0.0138, 0.0198, 0.0246, 0.0154, 0.0173, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:01:04,379 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6014, 1.7934, 2.8175, 3.2075, 3.6691, 3.3700, 2.0960, 3.5177], device='cuda:4'), covar=tensor([0.0035, 0.0339, 0.0131, 0.0093, 0.0026, 0.0075, 0.0214, 0.0040], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0112, 0.0093, 0.0084, 0.0057, 0.0060, 0.0093, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 18:01:07,709 INFO [train.py:904] (4/8) Epoch 2, batch 3750, loss[loss=0.2813, simple_loss=0.3472, pruned_loss=0.1077, over 17061.00 frames. ], tot_loss[loss=0.2668, simple_loss=0.3267, pruned_loss=0.1035, over 3272003.60 frames. ], batch size: 53, lr: 2.85e-02, grad_scale: 8.0 2023-04-27 18:01:33,975 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.812e+02 4.240e+02 4.945e+02 6.119e+02 1.020e+03, threshold=9.889e+02, percent-clipped=4.0 2023-04-27 18:01:58,477 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0328, 3.1984, 3.1158, 1.4865, 3.2402, 3.2653, 2.9047, 2.6223], device='cuda:4'), covar=tensor([0.0834, 0.0113, 0.0149, 0.1670, 0.0118, 0.0083, 0.0274, 0.0372], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0079, 0.0076, 0.0146, 0.0075, 0.0068, 0.0099, 0.0110], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:02:00,414 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8615, 3.1505, 3.5590, 2.7327, 3.4269, 3.5877, 3.6774, 1.7965], device='cuda:4'), covar=tensor([0.0608, 0.0143, 0.0055, 0.0279, 0.0061, 0.0044, 0.0044, 0.0544], device='cuda:4'), in_proj_covar=tensor([0.0113, 0.0049, 0.0056, 0.0103, 0.0052, 0.0053, 0.0059, 0.0102], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:02:24,609 INFO [train.py:904] (4/8) Epoch 2, batch 3800, loss[loss=0.2748, simple_loss=0.3398, pruned_loss=0.1049, over 16750.00 frames. ], tot_loss[loss=0.2699, simple_loss=0.3285, pruned_loss=0.1056, over 3260602.87 frames. ], batch size: 134, lr: 2.85e-02, grad_scale: 8.0 2023-04-27 18:03:39,937 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-27 18:03:40,205 INFO [train.py:904] (4/8) Epoch 2, batch 3850, loss[loss=0.2714, simple_loss=0.3199, pruned_loss=0.1115, over 16905.00 frames. ], tot_loss[loss=0.2682, simple_loss=0.3269, pruned_loss=0.1047, over 3263689.16 frames. ], batch size: 116, lr: 2.84e-02, grad_scale: 8.0 2023-04-27 18:04:07,044 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.317e+02 4.007e+02 4.656e+02 5.451e+02 1.276e+03, threshold=9.312e+02, percent-clipped=3.0 2023-04-27 18:04:53,516 INFO [train.py:904] (4/8) Epoch 2, batch 3900, loss[loss=0.2687, simple_loss=0.318, pruned_loss=0.1097, over 16864.00 frames. ], tot_loss[loss=0.2667, simple_loss=0.3255, pruned_loss=0.1039, over 3265076.65 frames. ], batch size: 109, lr: 2.84e-02, grad_scale: 8.0 2023-04-27 18:05:25,144 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14072.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:05:54,133 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0993, 5.1025, 4.6824, 4.2663, 4.9042, 2.3417, 4.6827, 4.8914], device='cuda:4'), covar=tensor([0.0040, 0.0046, 0.0071, 0.0258, 0.0046, 0.1082, 0.0061, 0.0087], device='cuda:4'), in_proj_covar=tensor([0.0064, 0.0057, 0.0081, 0.0102, 0.0064, 0.0107, 0.0077, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:05:55,292 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14093.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 18:06:05,397 INFO [train.py:904] (4/8) Epoch 2, batch 3950, loss[loss=0.3317, simple_loss=0.372, pruned_loss=0.1457, over 12461.00 frames. ], tot_loss[loss=0.2656, simple_loss=0.3239, pruned_loss=0.1036, over 3267321.87 frames. ], batch size: 246, lr: 2.83e-02, grad_scale: 8.0 2023-04-27 18:06:09,465 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14103.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:06:31,642 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.483e+02 3.936e+02 4.888e+02 5.915e+02 1.002e+03, threshold=9.776e+02, percent-clipped=1.0 2023-04-27 18:06:34,613 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14120.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 18:06:49,490 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7719, 1.7479, 2.0525, 2.7534, 2.7945, 2.6239, 1.7056, 2.5047], device='cuda:4'), covar=tensor([0.0043, 0.0279, 0.0170, 0.0112, 0.0037, 0.0103, 0.0218, 0.0055], device='cuda:4'), in_proj_covar=tensor([0.0072, 0.0111, 0.0095, 0.0084, 0.0058, 0.0061, 0.0092, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 18:06:52,447 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14133.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:07:17,632 INFO [train.py:904] (4/8) Epoch 2, batch 4000, loss[loss=0.2264, simple_loss=0.3005, pruned_loss=0.07618, over 16654.00 frames. ], tot_loss[loss=0.2655, simple_loss=0.3234, pruned_loss=0.1038, over 3274767.05 frames. ], batch size: 62, lr: 2.83e-02, grad_scale: 8.0 2023-04-27 18:07:31,553 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7083, 2.8054, 2.3831, 4.0884, 2.1294, 3.8139, 2.2581, 2.3394], device='cuda:4'), covar=tensor([0.0283, 0.0435, 0.0359, 0.0212, 0.1509, 0.0198, 0.0723, 0.1158], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0167, 0.0139, 0.0195, 0.0245, 0.0155, 0.0173, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:07:37,761 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14164.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:08:02,172 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14181.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 18:08:31,014 INFO [train.py:904] (4/8) Epoch 2, batch 4050, loss[loss=0.2575, simple_loss=0.3292, pruned_loss=0.09292, over 16341.00 frames. ], tot_loss[loss=0.26, simple_loss=0.3206, pruned_loss=0.09965, over 3283345.79 frames. ], batch size: 165, lr: 2.83e-02, grad_scale: 8.0 2023-04-27 18:08:50,651 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7814, 3.2071, 2.2702, 4.2647, 4.2426, 4.0368, 1.8129, 3.0924], device='cuda:4'), covar=tensor([0.1560, 0.0404, 0.1447, 0.0075, 0.0122, 0.0212, 0.1265, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0115, 0.0167, 0.0068, 0.0105, 0.0106, 0.0151, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:08:56,315 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.111e+02 3.521e+02 4.461e+02 5.736e+02 1.012e+03, threshold=8.922e+02, percent-clipped=2.0 2023-04-27 18:09:43,226 INFO [train.py:904] (4/8) Epoch 2, batch 4100, loss[loss=0.2604, simple_loss=0.3348, pruned_loss=0.09303, over 16414.00 frames. ], tot_loss[loss=0.2567, simple_loss=0.3196, pruned_loss=0.09687, over 3292767.44 frames. ], batch size: 146, lr: 2.82e-02, grad_scale: 8.0 2023-04-27 18:10:15,248 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3592, 4.2692, 1.4196, 4.4063, 2.5450, 4.1899, 1.6744, 2.8572], device='cuda:4'), covar=tensor([0.0038, 0.0091, 0.2168, 0.0036, 0.0819, 0.0201, 0.1882, 0.0622], device='cuda:4'), in_proj_covar=tensor([0.0077, 0.0096, 0.0166, 0.0079, 0.0152, 0.0125, 0.0172, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:10:59,418 INFO [train.py:904] (4/8) Epoch 2, batch 4150, loss[loss=0.2906, simple_loss=0.3597, pruned_loss=0.1107, over 17016.00 frames. ], tot_loss[loss=0.2661, simple_loss=0.3291, pruned_loss=0.1015, over 3264292.46 frames. ], batch size: 55, lr: 2.82e-02, grad_scale: 8.0 2023-04-27 18:11:25,251 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.427e+02 3.836e+02 4.498e+02 5.328e+02 1.203e+03, threshold=8.995e+02, percent-clipped=3.0 2023-04-27 18:12:05,597 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4569, 1.5968, 2.5885, 3.1461, 3.4398, 3.3600, 1.8070, 3.6680], device='cuda:4'), covar=tensor([0.0036, 0.0315, 0.0113, 0.0072, 0.0023, 0.0053, 0.0204, 0.0026], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0111, 0.0094, 0.0082, 0.0057, 0.0058, 0.0093, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 18:12:14,211 INFO [train.py:904] (4/8) Epoch 2, batch 4200, loss[loss=0.3087, simple_loss=0.3746, pruned_loss=0.1214, over 15259.00 frames. ], tot_loss[loss=0.2736, simple_loss=0.3377, pruned_loss=0.1047, over 3228902.23 frames. ], batch size: 190, lr: 2.81e-02, grad_scale: 8.0 2023-04-27 18:13:15,597 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=14393.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:13:26,254 INFO [train.py:904] (4/8) Epoch 2, batch 4250, loss[loss=0.2615, simple_loss=0.3307, pruned_loss=0.09616, over 15308.00 frames. ], tot_loss[loss=0.2766, simple_loss=0.3418, pruned_loss=0.1057, over 3208923.50 frames. ], batch size: 190, lr: 2.81e-02, grad_scale: 8.0 2023-04-27 18:13:44,184 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14413.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:13:45,288 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3739, 5.5784, 5.2606, 5.3775, 4.8238, 4.6074, 5.0857, 5.6473], device='cuda:4'), covar=tensor([0.0318, 0.0479, 0.0549, 0.0276, 0.0411, 0.0358, 0.0382, 0.0424], device='cuda:4'), in_proj_covar=tensor([0.0217, 0.0283, 0.0253, 0.0181, 0.0196, 0.0175, 0.0229, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:13:52,832 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.326e+02 4.073e+02 4.856e+02 5.593e+02 1.302e+03, threshold=9.713e+02, percent-clipped=3.0 2023-04-27 18:14:05,658 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14428.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:14:21,893 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2751, 3.8258, 3.8928, 1.7243, 4.1147, 4.0808, 3.3846, 3.3791], device='cuda:4'), covar=tensor([0.0761, 0.0091, 0.0112, 0.1552, 0.0043, 0.0040, 0.0206, 0.0290], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0078, 0.0074, 0.0149, 0.0072, 0.0072, 0.0100, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:14:24,083 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=14441.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:14:38,545 INFO [train.py:904] (4/8) Epoch 2, batch 4300, loss[loss=0.2876, simple_loss=0.3639, pruned_loss=0.1056, over 16753.00 frames. ], tot_loss[loss=0.2755, simple_loss=0.3429, pruned_loss=0.104, over 3210222.85 frames. ], batch size: 124, lr: 2.80e-02, grad_scale: 16.0 2023-04-27 18:14:49,276 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14459.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:15:10,455 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14474.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:15:12,712 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14476.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 18:15:28,836 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6821, 1.2351, 1.5112, 1.7203, 1.7031, 1.8431, 1.4341, 1.9092], device='cuda:4'), covar=tensor([0.0065, 0.0196, 0.0098, 0.0094, 0.0039, 0.0049, 0.0141, 0.0035], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0109, 0.0091, 0.0079, 0.0056, 0.0057, 0.0090, 0.0053], device='cuda:4'), out_proj_covar=tensor([1.2737e-04, 1.9548e-04, 1.7050e-04, 1.4917e-04, 9.9563e-05, 1.0617e-04, 1.5715e-04, 9.6508e-05], device='cuda:4') 2023-04-27 18:15:29,940 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14487.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:15:49,383 INFO [train.py:904] (4/8) Epoch 2, batch 4350, loss[loss=0.2633, simple_loss=0.3335, pruned_loss=0.09652, over 16418.00 frames. ], tot_loss[loss=0.2789, simple_loss=0.3466, pruned_loss=0.1056, over 3199455.45 frames. ], batch size: 35, lr: 2.80e-02, grad_scale: 16.0 2023-04-27 18:16:13,767 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.262e+02 4.011e+02 4.650e+02 6.201e+02 1.293e+03, threshold=9.301e+02, percent-clipped=5.0 2023-04-27 18:16:56,175 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14548.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:16:59,398 INFO [train.py:904] (4/8) Epoch 2, batch 4400, loss[loss=0.2957, simple_loss=0.3647, pruned_loss=0.1133, over 16361.00 frames. ], tot_loss[loss=0.2806, simple_loss=0.3486, pruned_loss=0.1063, over 3207399.75 frames. ], batch size: 146, lr: 2.79e-02, grad_scale: 16.0 2023-04-27 18:18:09,487 INFO [train.py:904] (4/8) Epoch 2, batch 4450, loss[loss=0.2752, simple_loss=0.3539, pruned_loss=0.09826, over 17002.00 frames. ], tot_loss[loss=0.2809, simple_loss=0.3506, pruned_loss=0.1056, over 3223530.32 frames. ], batch size: 55, lr: 2.79e-02, grad_scale: 16.0 2023-04-27 18:18:34,420 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.338e+02 3.415e+02 4.231e+02 5.024e+02 9.561e+02, threshold=8.461e+02, percent-clipped=1.0 2023-04-27 18:18:38,047 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7301, 4.3674, 4.1829, 1.6928, 3.0247, 2.5344, 3.9844, 4.6666], device='cuda:4'), covar=tensor([0.0223, 0.0348, 0.0317, 0.1833, 0.0781, 0.1009, 0.0581, 0.0278], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0103, 0.0145, 0.0155, 0.0145, 0.0136, 0.0149, 0.0093], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 18:18:51,026 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6644, 1.6876, 2.5295, 3.2182, 3.5076, 3.4927, 2.1717, 3.7340], device='cuda:4'), covar=tensor([0.0025, 0.0262, 0.0112, 0.0081, 0.0024, 0.0052, 0.0153, 0.0040], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0109, 0.0092, 0.0082, 0.0056, 0.0058, 0.0092, 0.0054], device='cuda:4'), out_proj_covar=tensor([1.2802e-04, 1.9620e-04, 1.7171e-04, 1.5429e-04, 9.8995e-05, 1.0780e-04, 1.6057e-04, 9.8132e-05], device='cuda:4') 2023-04-27 18:19:17,771 INFO [train.py:904] (4/8) Epoch 2, batch 4500, loss[loss=0.2568, simple_loss=0.3357, pruned_loss=0.08894, over 16990.00 frames. ], tot_loss[loss=0.2792, simple_loss=0.3494, pruned_loss=0.1045, over 3228734.74 frames. ], batch size: 41, lr: 2.79e-02, grad_scale: 16.0 2023-04-27 18:20:04,117 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8953, 4.2663, 3.3463, 3.1073, 3.3523, 2.4655, 4.6492, 5.2265], device='cuda:4'), covar=tensor([0.1840, 0.0463, 0.0871, 0.0571, 0.1855, 0.1103, 0.0220, 0.0050], device='cuda:4'), in_proj_covar=tensor([0.0252, 0.0219, 0.0233, 0.0164, 0.0279, 0.0180, 0.0187, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:20:07,586 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9418, 4.1703, 4.0246, 1.6991, 4.2733, 4.1394, 3.3985, 3.4633], device='cuda:4'), covar=tensor([0.0925, 0.0051, 0.0128, 0.1439, 0.0058, 0.0064, 0.0189, 0.0234], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0078, 0.0078, 0.0149, 0.0073, 0.0072, 0.0099, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:20:28,469 INFO [train.py:904] (4/8) Epoch 2, batch 4550, loss[loss=0.2736, simple_loss=0.342, pruned_loss=0.1026, over 16428.00 frames. ], tot_loss[loss=0.2788, simple_loss=0.3492, pruned_loss=0.1042, over 3241023.73 frames. ], batch size: 35, lr: 2.78e-02, grad_scale: 16.0 2023-04-27 18:20:54,773 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.252e+02 3.646e+02 4.280e+02 5.239e+02 1.001e+03, threshold=8.560e+02, percent-clipped=1.0 2023-04-27 18:21:08,369 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=14728.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:21:19,364 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.50 vs. limit=5.0 2023-04-27 18:21:25,005 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14740.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:21:26,953 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5042, 4.9776, 5.1380, 5.2952, 4.7195, 5.3129, 5.0578, 4.9948], device='cuda:4'), covar=tensor([0.0139, 0.0098, 0.0129, 0.0078, 0.0640, 0.0091, 0.0102, 0.0151], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0083, 0.0148, 0.0115, 0.0178, 0.0121, 0.0107, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:21:33,969 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0130, 2.5475, 2.4566, 3.2923, 3.2701, 3.1804, 1.9229, 2.6643], device='cuda:4'), covar=tensor([0.1227, 0.0307, 0.0974, 0.0066, 0.0176, 0.0244, 0.0971, 0.0532], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0114, 0.0163, 0.0063, 0.0096, 0.0105, 0.0149, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:21:42,830 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.70 vs. limit=5.0 2023-04-27 18:21:43,174 INFO [train.py:904] (4/8) Epoch 2, batch 4600, loss[loss=0.2821, simple_loss=0.3545, pruned_loss=0.1048, over 16458.00 frames. ], tot_loss[loss=0.2775, simple_loss=0.3489, pruned_loss=0.1031, over 3249522.52 frames. ], batch size: 68, lr: 2.78e-02, grad_scale: 16.0 2023-04-27 18:21:53,379 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=14759.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:22:10,142 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14769.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:22:17,945 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.65 vs. limit=2.0 2023-04-27 18:22:18,545 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=14776.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:22:18,603 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=14776.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 18:22:45,423 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9338, 3.1429, 3.1520, 1.5883, 3.3351, 3.2865, 2.8744, 2.9106], device='cuda:4'), covar=tensor([0.0865, 0.0104, 0.0151, 0.1411, 0.0067, 0.0058, 0.0281, 0.0282], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0081, 0.0080, 0.0151, 0.0073, 0.0072, 0.0103, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:22:54,477 INFO [train.py:904] (4/8) Epoch 2, batch 4650, loss[loss=0.2744, simple_loss=0.3477, pruned_loss=0.1005, over 16670.00 frames. ], tot_loss[loss=0.275, simple_loss=0.3465, pruned_loss=0.1017, over 3245688.64 frames. ], batch size: 134, lr: 2.77e-02, grad_scale: 16.0 2023-04-27 18:22:55,631 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14801.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:23:03,206 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=14807.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:23:08,630 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.01 vs. limit=2.0 2023-04-27 18:23:20,765 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.700e+02 3.221e+02 3.729e+02 4.286e+02 9.218e+02, threshold=7.459e+02, percent-clipped=2.0 2023-04-27 18:23:27,860 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=14824.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 18:23:54,900 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14843.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:24:06,746 INFO [train.py:904] (4/8) Epoch 2, batch 4700, loss[loss=0.2681, simple_loss=0.3298, pruned_loss=0.1032, over 17047.00 frames. ], tot_loss[loss=0.2726, simple_loss=0.3436, pruned_loss=0.1008, over 3235874.50 frames. ], batch size: 53, lr: 2.77e-02, grad_scale: 16.0 2023-04-27 18:24:29,055 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3661, 4.5798, 4.2783, 4.4494, 4.0178, 4.0580, 4.0582, 4.5737], device='cuda:4'), covar=tensor([0.0364, 0.0507, 0.0754, 0.0293, 0.0480, 0.0616, 0.0541, 0.0501], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0280, 0.0262, 0.0184, 0.0196, 0.0177, 0.0232, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:24:30,928 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1704, 4.0554, 4.0855, 1.5535, 4.2908, 4.1893, 3.3183, 3.4064], device='cuda:4'), covar=tensor([0.0928, 0.0083, 0.0127, 0.1708, 0.0077, 0.0050, 0.0315, 0.0314], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0084, 0.0084, 0.0155, 0.0075, 0.0075, 0.0106, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:24:50,707 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.86 vs. limit=5.0 2023-04-27 18:24:50,800 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.78 vs. limit=2.0 2023-04-27 18:25:20,206 INFO [train.py:904] (4/8) Epoch 2, batch 4750, loss[loss=0.2414, simple_loss=0.3154, pruned_loss=0.08368, over 16519.00 frames. ], tot_loss[loss=0.2705, simple_loss=0.3408, pruned_loss=0.1001, over 3210446.58 frames. ], batch size: 75, lr: 2.77e-02, grad_scale: 16.0 2023-04-27 18:25:28,242 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1836, 4.1619, 4.3274, 3.1891, 4.1980, 4.1104, 4.5157, 1.9800], device='cuda:4'), covar=tensor([0.0631, 0.0031, 0.0044, 0.0302, 0.0042, 0.0076, 0.0022, 0.0567], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0050, 0.0059, 0.0108, 0.0054, 0.0058, 0.0056, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:25:45,828 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.318e+02 3.402e+02 4.045e+02 5.011e+02 1.217e+03, threshold=8.089e+02, percent-clipped=2.0 2023-04-27 18:26:29,081 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.04 vs. limit=2.0 2023-04-27 18:26:30,650 INFO [train.py:904] (4/8) Epoch 2, batch 4800, loss[loss=0.2766, simple_loss=0.3461, pruned_loss=0.1036, over 16856.00 frames. ], tot_loss[loss=0.2669, simple_loss=0.3373, pruned_loss=0.09822, over 3206231.29 frames. ], batch size: 116, lr: 2.76e-02, grad_scale: 16.0 2023-04-27 18:27:43,242 INFO [train.py:904] (4/8) Epoch 2, batch 4850, loss[loss=0.2406, simple_loss=0.3305, pruned_loss=0.07537, over 16380.00 frames. ], tot_loss[loss=0.2688, simple_loss=0.3393, pruned_loss=0.0991, over 3172928.58 frames. ], batch size: 146, lr: 2.76e-02, grad_scale: 16.0 2023-04-27 18:28:11,112 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.137e+02 3.566e+02 4.230e+02 5.018e+02 9.509e+02, threshold=8.461e+02, percent-clipped=3.0 2023-04-27 18:28:37,090 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8146, 1.3406, 2.0608, 2.6188, 2.7911, 2.7613, 1.6298, 2.6603], device='cuda:4'), covar=tensor([0.0041, 0.0303, 0.0159, 0.0107, 0.0032, 0.0061, 0.0207, 0.0036], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0105, 0.0088, 0.0076, 0.0055, 0.0055, 0.0088, 0.0051], device='cuda:4'), out_proj_covar=tensor([1.2184e-04, 1.8694e-04, 1.6457e-04, 1.4368e-04, 9.4556e-05, 1.0184e-04, 1.5366e-04, 9.2202e-05], device='cuda:4') 2023-04-27 18:28:58,605 INFO [train.py:904] (4/8) Epoch 2, batch 4900, loss[loss=0.2831, simple_loss=0.3511, pruned_loss=0.1075, over 16774.00 frames. ], tot_loss[loss=0.2671, simple_loss=0.3388, pruned_loss=0.09773, over 3163123.36 frames. ], batch size: 83, lr: 2.75e-02, grad_scale: 16.0 2023-04-27 18:29:24,615 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=15069.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:30:04,907 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=15096.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:30:10,700 INFO [train.py:904] (4/8) Epoch 2, batch 4950, loss[loss=0.2662, simple_loss=0.3448, pruned_loss=0.09379, over 16393.00 frames. ], tot_loss[loss=0.2664, simple_loss=0.3386, pruned_loss=0.09714, over 3185172.58 frames. ], batch size: 75, lr: 2.75e-02, grad_scale: 16.0 2023-04-27 18:30:34,535 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=15117.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:30:37,411 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.417e+02 3.788e+02 4.518e+02 5.433e+02 9.876e+02, threshold=9.036e+02, percent-clipped=3.0 2023-04-27 18:31:06,506 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9616, 4.5749, 4.6704, 4.8626, 4.1776, 4.7236, 4.6461, 4.4469], device='cuda:4'), covar=tensor([0.0231, 0.0163, 0.0171, 0.0098, 0.0679, 0.0165, 0.0124, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0089, 0.0153, 0.0116, 0.0179, 0.0126, 0.0104, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:31:12,882 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=15143.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:31:22,555 INFO [train.py:904] (4/8) Epoch 2, batch 5000, loss[loss=0.3054, simple_loss=0.362, pruned_loss=0.1244, over 12101.00 frames. ], tot_loss[loss=0.268, simple_loss=0.3404, pruned_loss=0.09777, over 3187650.35 frames. ], batch size: 246, lr: 2.74e-02, grad_scale: 16.0 2023-04-27 18:32:21,457 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=15191.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:32:26,396 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=15195.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:32:34,518 INFO [train.py:904] (4/8) Epoch 2, batch 5050, loss[loss=0.2518, simple_loss=0.331, pruned_loss=0.08633, over 16449.00 frames. ], tot_loss[loss=0.2674, simple_loss=0.3405, pruned_loss=0.09714, over 3195504.70 frames. ], batch size: 68, lr: 2.74e-02, grad_scale: 16.0 2023-04-27 18:33:00,585 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.434e+02 3.699e+02 4.750e+02 5.849e+02 1.293e+03, threshold=9.501e+02, percent-clipped=2.0 2023-04-27 18:33:45,519 INFO [train.py:904] (4/8) Epoch 2, batch 5100, loss[loss=0.2469, simple_loss=0.3142, pruned_loss=0.08985, over 17215.00 frames. ], tot_loss[loss=0.2644, simple_loss=0.3378, pruned_loss=0.09546, over 3206215.21 frames. ], batch size: 45, lr: 2.74e-02, grad_scale: 16.0 2023-04-27 18:33:53,376 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=15256.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 18:34:01,332 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 18:34:14,802 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-27 18:34:55,623 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9847, 3.9401, 4.0597, 4.0700, 3.9564, 4.5465, 4.4869, 4.0500], device='cuda:4'), covar=tensor([0.1229, 0.1235, 0.0781, 0.1367, 0.2189, 0.0785, 0.0519, 0.1690], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0250, 0.0227, 0.0217, 0.0282, 0.0237, 0.0181, 0.0291], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:34:56,469 INFO [train.py:904] (4/8) Epoch 2, batch 5150, loss[loss=0.2743, simple_loss=0.3547, pruned_loss=0.09697, over 15543.00 frames. ], tot_loss[loss=0.2641, simple_loss=0.3387, pruned_loss=0.09475, over 3195344.77 frames. ], batch size: 191, lr: 2.73e-02, grad_scale: 16.0 2023-04-27 18:35:14,434 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7652, 2.9898, 2.5006, 4.1901, 2.1510, 3.8867, 2.4941, 2.5492], device='cuda:4'), covar=tensor([0.0296, 0.0381, 0.0338, 0.0184, 0.1397, 0.0188, 0.0616, 0.0988], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0172, 0.0145, 0.0201, 0.0250, 0.0161, 0.0180, 0.0231], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:35:22,256 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.454e+02 3.560e+02 4.122e+02 5.032e+02 1.152e+03, threshold=8.244e+02, percent-clipped=1.0 2023-04-27 18:35:41,998 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.73 vs. limit=5.0 2023-04-27 18:36:10,018 INFO [train.py:904] (4/8) Epoch 2, batch 5200, loss[loss=0.2499, simple_loss=0.3233, pruned_loss=0.08826, over 16460.00 frames. ], tot_loss[loss=0.2634, simple_loss=0.3374, pruned_loss=0.09472, over 3194880.86 frames. ], batch size: 68, lr: 2.73e-02, grad_scale: 16.0 2023-04-27 18:36:20,883 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2055, 3.0554, 2.7648, 1.8391, 2.6456, 2.1531, 2.8651, 3.1715], device='cuda:4'), covar=tensor([0.0255, 0.0330, 0.0427, 0.1483, 0.0594, 0.0890, 0.0517, 0.0241], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0104, 0.0151, 0.0154, 0.0145, 0.0140, 0.0149, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 18:36:30,274 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4607, 3.3151, 3.3099, 2.9183, 3.3066, 2.0243, 3.1662, 3.2113], device='cuda:4'), covar=tensor([0.0081, 0.0070, 0.0073, 0.0261, 0.0063, 0.1022, 0.0076, 0.0099], device='cuda:4'), in_proj_covar=tensor([0.0061, 0.0050, 0.0072, 0.0096, 0.0058, 0.0106, 0.0068, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:37:16,249 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=15396.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:37:23,464 INFO [train.py:904] (4/8) Epoch 2, batch 5250, loss[loss=0.2247, simple_loss=0.305, pruned_loss=0.07217, over 16666.00 frames. ], tot_loss[loss=0.2622, simple_loss=0.3348, pruned_loss=0.09477, over 3192875.86 frames. ], batch size: 62, lr: 2.72e-02, grad_scale: 16.0 2023-04-27 18:37:48,868 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.244e+02 3.713e+02 4.444e+02 5.638e+02 1.103e+03, threshold=8.887e+02, percent-clipped=2.0 2023-04-27 18:37:51,275 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 18:38:23,202 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=15444.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:38:24,973 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 18:38:32,784 INFO [train.py:904] (4/8) Epoch 2, batch 5300, loss[loss=0.2151, simple_loss=0.2923, pruned_loss=0.069, over 17220.00 frames. ], tot_loss[loss=0.2583, simple_loss=0.3307, pruned_loss=0.09294, over 3192380.19 frames. ], batch size: 45, lr: 2.72e-02, grad_scale: 16.0 2023-04-27 18:39:18,629 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9944, 1.5861, 2.1665, 2.8291, 2.9241, 2.8610, 1.5457, 2.8479], device='cuda:4'), covar=tensor([0.0045, 0.0289, 0.0153, 0.0104, 0.0029, 0.0067, 0.0220, 0.0049], device='cuda:4'), in_proj_covar=tensor([0.0073, 0.0109, 0.0091, 0.0082, 0.0058, 0.0059, 0.0092, 0.0053], device='cuda:4'), out_proj_covar=tensor([1.3118e-04, 1.9346e-04, 1.6903e-04, 1.5346e-04, 9.9348e-05, 1.0802e-04, 1.5921e-04, 9.4931e-05], device='cuda:4') 2023-04-27 18:39:30,812 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6236, 3.7131, 3.4621, 3.6487, 2.7086, 3.6088, 3.5217, 3.2462], device='cuda:4'), covar=tensor([0.0490, 0.0261, 0.0378, 0.0244, 0.1329, 0.0293, 0.0780, 0.0397], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0095, 0.0161, 0.0124, 0.0190, 0.0133, 0.0111, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:39:42,791 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9325, 3.9533, 3.4628, 3.1635, 3.2126, 2.4687, 4.1161, 4.6674], device='cuda:4'), covar=tensor([0.1527, 0.0538, 0.0806, 0.0503, 0.1535, 0.1006, 0.0289, 0.0124], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0219, 0.0231, 0.0168, 0.0262, 0.0176, 0.0191, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:39:43,336 INFO [train.py:904] (4/8) Epoch 2, batch 5350, loss[loss=0.2703, simple_loss=0.3411, pruned_loss=0.09975, over 16550.00 frames. ], tot_loss[loss=0.2552, simple_loss=0.3279, pruned_loss=0.09121, over 3201473.15 frames. ], batch size: 62, lr: 2.72e-02, grad_scale: 16.0 2023-04-27 18:40:09,982 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.455e+02 3.880e+02 4.934e+02 5.714e+02 1.287e+03, threshold=9.868e+02, percent-clipped=1.0 2023-04-27 18:40:56,836 INFO [train.py:904] (4/8) Epoch 2, batch 5400, loss[loss=0.2739, simple_loss=0.346, pruned_loss=0.1009, over 16747.00 frames. ], tot_loss[loss=0.259, simple_loss=0.3319, pruned_loss=0.09308, over 3214598.89 frames. ], batch size: 89, lr: 2.71e-02, grad_scale: 16.0 2023-04-27 18:40:57,153 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=15551.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 18:41:30,484 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-27 18:41:37,928 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3603, 4.3645, 3.9087, 1.8753, 2.9462, 2.5489, 3.7025, 4.3744], device='cuda:4'), covar=tensor([0.0250, 0.0332, 0.0420, 0.1703, 0.0810, 0.0953, 0.0718, 0.0351], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0104, 0.0149, 0.0151, 0.0145, 0.0137, 0.0147, 0.0095], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 18:42:14,657 INFO [train.py:904] (4/8) Epoch 2, batch 5450, loss[loss=0.2819, simple_loss=0.35, pruned_loss=0.1068, over 16747.00 frames. ], tot_loss[loss=0.2639, simple_loss=0.3362, pruned_loss=0.0958, over 3225743.65 frames. ], batch size: 83, lr: 2.71e-02, grad_scale: 16.0 2023-04-27 18:42:15,354 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3880, 4.3047, 1.6855, 4.4190, 2.7798, 4.4221, 2.1205, 2.8794], device='cuda:4'), covar=tensor([0.0034, 0.0127, 0.1831, 0.0045, 0.0784, 0.0175, 0.1356, 0.0658], device='cuda:4'), in_proj_covar=tensor([0.0075, 0.0100, 0.0169, 0.0077, 0.0158, 0.0127, 0.0174, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:42:23,175 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4669, 4.4401, 5.0697, 5.1519, 5.1921, 4.4553, 4.6462, 4.6398], device='cuda:4'), covar=tensor([0.0253, 0.0254, 0.0231, 0.0289, 0.0304, 0.0239, 0.0623, 0.0273], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0157, 0.0176, 0.0171, 0.0205, 0.0167, 0.0256, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 18:42:43,047 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.524e+02 4.170e+02 4.882e+02 6.060e+02 1.199e+03, threshold=9.765e+02, percent-clipped=3.0 2023-04-27 18:43:34,031 INFO [train.py:904] (4/8) Epoch 2, batch 5500, loss[loss=0.3056, simple_loss=0.37, pruned_loss=0.1206, over 16882.00 frames. ], tot_loss[loss=0.2787, simple_loss=0.3469, pruned_loss=0.1052, over 3172702.61 frames. ], batch size: 116, lr: 2.70e-02, grad_scale: 16.0 2023-04-27 18:44:50,715 INFO [train.py:904] (4/8) Epoch 2, batch 5550, loss[loss=0.4424, simple_loss=0.4564, pruned_loss=0.2142, over 11140.00 frames. ], tot_loss[loss=0.2934, simple_loss=0.3575, pruned_loss=0.1147, over 3129686.72 frames. ], batch size: 248, lr: 2.70e-02, grad_scale: 16.0 2023-04-27 18:45:14,299 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6919, 3.4907, 3.4205, 2.5196, 3.3451, 3.3825, 3.4126, 1.6433], device='cuda:4'), covar=tensor([0.0634, 0.0031, 0.0049, 0.0295, 0.0044, 0.0088, 0.0043, 0.0525], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0052, 0.0058, 0.0108, 0.0052, 0.0056, 0.0058, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:45:18,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0684, 4.6990, 4.8735, 5.0290, 4.2523, 4.9557, 4.8308, 4.6049], device='cuda:4'), covar=tensor([0.0268, 0.0176, 0.0173, 0.0112, 0.0847, 0.0161, 0.0122, 0.0193], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0090, 0.0149, 0.0117, 0.0182, 0.0122, 0.0106, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:45:19,404 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.692e+02 5.827e+02 6.984e+02 8.601e+02 1.757e+03, threshold=1.397e+03, percent-clipped=15.0 2023-04-27 18:45:42,282 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-27 18:46:11,044 INFO [train.py:904] (4/8) Epoch 2, batch 5600, loss[loss=0.3237, simple_loss=0.3824, pruned_loss=0.1325, over 16842.00 frames. ], tot_loss[loss=0.304, simple_loss=0.365, pruned_loss=0.1215, over 3100627.26 frames. ], batch size: 116, lr: 2.70e-02, grad_scale: 16.0 2023-04-27 18:46:39,100 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-27 18:46:40,409 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.96 vs. limit=5.0 2023-04-27 18:47:34,289 INFO [train.py:904] (4/8) Epoch 2, batch 5650, loss[loss=0.2904, simple_loss=0.3521, pruned_loss=0.1144, over 16711.00 frames. ], tot_loss[loss=0.312, simple_loss=0.371, pruned_loss=0.1265, over 3103353.98 frames. ], batch size: 57, lr: 2.69e-02, grad_scale: 16.0 2023-04-27 18:48:01,991 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.886e+02 5.674e+02 6.824e+02 8.519e+02 2.118e+03, threshold=1.365e+03, percent-clipped=2.0 2023-04-27 18:48:07,300 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.97 vs. limit=5.0 2023-04-27 18:48:53,333 INFO [train.py:904] (4/8) Epoch 2, batch 5700, loss[loss=0.3008, simple_loss=0.3697, pruned_loss=0.116, over 16761.00 frames. ], tot_loss[loss=0.3165, simple_loss=0.374, pruned_loss=0.1295, over 3076371.28 frames. ], batch size: 39, lr: 2.69e-02, grad_scale: 16.0 2023-04-27 18:48:53,700 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=15851.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:49:44,708 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=15883.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:50:09,228 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=15899.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:50:12,120 INFO [train.py:904] (4/8) Epoch 2, batch 5750, loss[loss=0.262, simple_loss=0.3405, pruned_loss=0.09175, over 16909.00 frames. ], tot_loss[loss=0.3202, simple_loss=0.3772, pruned_loss=0.1316, over 3044751.61 frames. ], batch size: 96, lr: 2.69e-02, grad_scale: 8.0 2023-04-27 18:50:15,967 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4762, 4.2571, 4.5499, 4.8016, 4.8181, 4.2699, 4.8403, 4.7575], device='cuda:4'), covar=tensor([0.0474, 0.0432, 0.0737, 0.0264, 0.0255, 0.0357, 0.0218, 0.0224], device='cuda:4'), in_proj_covar=tensor([0.0224, 0.0261, 0.0358, 0.0269, 0.0203, 0.0192, 0.0200, 0.0208], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:50:42,007 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.285e+02 4.919e+02 6.542e+02 8.153e+02 1.932e+03, threshold=1.308e+03, percent-clipped=2.0 2023-04-27 18:51:22,904 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=15944.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:51:33,504 INFO [train.py:904] (4/8) Epoch 2, batch 5800, loss[loss=0.2981, simple_loss=0.3677, pruned_loss=0.1143, over 16673.00 frames. ], tot_loss[loss=0.3197, simple_loss=0.3776, pruned_loss=0.1309, over 3028530.56 frames. ], batch size: 124, lr: 2.68e-02, grad_scale: 8.0 2023-04-27 18:52:56,020 INFO [train.py:904] (4/8) Epoch 2, batch 5850, loss[loss=0.281, simple_loss=0.3555, pruned_loss=0.1032, over 16526.00 frames. ], tot_loss[loss=0.316, simple_loss=0.3749, pruned_loss=0.1285, over 3041248.88 frames. ], batch size: 75, lr: 2.68e-02, grad_scale: 8.0 2023-04-27 18:52:57,324 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16001.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:53:25,313 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.722e+02 4.866e+02 5.967e+02 7.152e+02 1.262e+03, threshold=1.193e+03, percent-clipped=0.0 2023-04-27 18:53:44,868 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6575, 4.8542, 4.5500, 4.6877, 4.2908, 4.1762, 4.4270, 4.9146], device='cuda:4'), covar=tensor([0.0314, 0.0444, 0.0643, 0.0298, 0.0424, 0.0500, 0.0411, 0.0435], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0287, 0.0272, 0.0188, 0.0198, 0.0182, 0.0237, 0.0204], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:54:18,253 INFO [train.py:904] (4/8) Epoch 2, batch 5900, loss[loss=0.364, simple_loss=0.3994, pruned_loss=0.1643, over 11416.00 frames. ], tot_loss[loss=0.3142, simple_loss=0.3735, pruned_loss=0.1274, over 3049915.38 frames. ], batch size: 246, lr: 2.67e-02, grad_scale: 8.0 2023-04-27 18:54:40,574 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16062.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:55:17,991 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16085.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:55:42,172 INFO [train.py:904] (4/8) Epoch 2, batch 5950, loss[loss=0.2937, simple_loss=0.3694, pruned_loss=0.109, over 16870.00 frames. ], tot_loss[loss=0.3128, simple_loss=0.3746, pruned_loss=0.1255, over 3057666.59 frames. ], batch size: 96, lr: 2.67e-02, grad_scale: 8.0 2023-04-27 18:56:13,606 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.088e+02 4.747e+02 6.406e+02 8.221e+02 1.979e+03, threshold=1.281e+03, percent-clipped=5.0 2023-04-27 18:56:56,089 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16146.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:57:01,946 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16149.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:57:04,555 INFO [train.py:904] (4/8) Epoch 2, batch 6000, loss[loss=0.2821, simple_loss=0.3533, pruned_loss=0.1054, over 16860.00 frames. ], tot_loss[loss=0.311, simple_loss=0.3731, pruned_loss=0.1244, over 3081857.61 frames. ], batch size: 96, lr: 2.67e-02, grad_scale: 8.0 2023-04-27 18:57:04,555 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 18:57:15,924 INFO [train.py:938] (4/8) Epoch 2, validation: loss=0.2302, simple_loss=0.3372, pruned_loss=0.06166, over 944034.00 frames. 2023-04-27 18:57:15,925 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-27 18:57:22,668 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3745, 1.7234, 2.3282, 3.0581, 3.1804, 3.0898, 1.6863, 3.2690], device='cuda:4'), covar=tensor([0.0028, 0.0240, 0.0117, 0.0066, 0.0027, 0.0058, 0.0154, 0.0039], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0109, 0.0092, 0.0081, 0.0058, 0.0058, 0.0090, 0.0053], device='cuda:4'), out_proj_covar=tensor([1.2424e-04, 1.9273e-04, 1.6868e-04, 1.5108e-04, 9.9004e-05, 1.0447e-04, 1.5435e-04, 9.2965e-05], device='cuda:4') 2023-04-27 18:58:34,806 INFO [train.py:904] (4/8) Epoch 2, batch 6050, loss[loss=0.2893, simple_loss=0.3627, pruned_loss=0.1079, over 17252.00 frames. ], tot_loss[loss=0.3071, simple_loss=0.37, pruned_loss=0.1221, over 3089278.01 frames. ], batch size: 52, lr: 2.66e-02, grad_scale: 8.0 2023-04-27 18:58:48,685 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16210.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:59:04,137 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.028e+02 4.577e+02 5.718e+02 7.421e+02 1.323e+03, threshold=1.144e+03, percent-clipped=1.0 2023-04-27 18:59:19,187 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8486, 3.0117, 2.3023, 3.7810, 3.6993, 3.6088, 1.7792, 2.6945], device='cuda:4'), covar=tensor([0.1267, 0.0354, 0.1104, 0.0060, 0.0165, 0.0235, 0.1042, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0119, 0.0164, 0.0068, 0.0105, 0.0113, 0.0150, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:59:33,592 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16239.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:59:52,042 INFO [train.py:904] (4/8) Epoch 2, batch 6100, loss[loss=0.2892, simple_loss=0.3668, pruned_loss=0.1058, over 16690.00 frames. ], tot_loss[loss=0.3049, simple_loss=0.369, pruned_loss=0.1204, over 3096319.78 frames. ], batch size: 89, lr: 2.66e-02, grad_scale: 8.0 2023-04-27 19:00:02,431 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6164, 3.3665, 3.0041, 1.7717, 2.5327, 1.9797, 3.0606, 3.5127], device='cuda:4'), covar=tensor([0.0311, 0.0367, 0.0464, 0.1588, 0.0757, 0.1066, 0.0639, 0.0269], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0104, 0.0146, 0.0148, 0.0142, 0.0134, 0.0146, 0.0095], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 19:00:41,506 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2196, 1.3357, 1.8962, 2.1975, 2.1293, 2.2047, 1.3786, 2.0542], device='cuda:4'), covar=tensor([0.0048, 0.0204, 0.0098, 0.0074, 0.0040, 0.0049, 0.0156, 0.0038], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0106, 0.0091, 0.0079, 0.0058, 0.0057, 0.0091, 0.0053], device='cuda:4'), out_proj_covar=tensor([1.1936e-04, 1.8790e-04, 1.6704e-04, 1.4688e-04, 9.9711e-05, 1.0272e-04, 1.5506e-04, 9.1840e-05], device='cuda:4') 2023-04-27 19:01:09,730 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6684, 3.1413, 2.2346, 3.8959, 3.8021, 3.8121, 1.7093, 2.9744], device='cuda:4'), covar=tensor([0.1523, 0.0369, 0.1359, 0.0052, 0.0201, 0.0249, 0.1218, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0120, 0.0168, 0.0068, 0.0108, 0.0116, 0.0153, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:01:17,143 INFO [train.py:904] (4/8) Epoch 2, batch 6150, loss[loss=0.3148, simple_loss=0.373, pruned_loss=0.1283, over 16886.00 frames. ], tot_loss[loss=0.3026, simple_loss=0.3665, pruned_loss=0.1193, over 3104158.57 frames. ], batch size: 116, lr: 2.66e-02, grad_scale: 8.0 2023-04-27 19:01:45,849 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.137e+02 4.737e+02 5.827e+02 7.174e+02 1.739e+03, threshold=1.165e+03, percent-clipped=5.0 2023-04-27 19:01:53,493 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0766, 4.2083, 3.6474, 3.2442, 3.2294, 2.5053, 4.4172, 5.1040], device='cuda:4'), covar=tensor([0.1603, 0.0468, 0.0875, 0.0564, 0.1741, 0.1030, 0.0284, 0.0095], device='cuda:4'), in_proj_covar=tensor([0.0254, 0.0224, 0.0241, 0.0181, 0.0277, 0.0180, 0.0197, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:01:57,717 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-27 19:02:34,534 INFO [train.py:904] (4/8) Epoch 2, batch 6200, loss[loss=0.2546, simple_loss=0.3265, pruned_loss=0.09133, over 17026.00 frames. ], tot_loss[loss=0.3014, simple_loss=0.3646, pruned_loss=0.1191, over 3088331.00 frames. ], batch size: 55, lr: 2.65e-02, grad_scale: 8.0 2023-04-27 19:02:45,548 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16357.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:03:00,332 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.04 vs. limit=2.0 2023-04-27 19:03:06,775 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16371.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:03:51,369 INFO [train.py:904] (4/8) Epoch 2, batch 6250, loss[loss=0.2828, simple_loss=0.3593, pruned_loss=0.1031, over 17101.00 frames. ], tot_loss[loss=0.303, simple_loss=0.3659, pruned_loss=0.12, over 3097199.76 frames. ], batch size: 49, lr: 2.65e-02, grad_scale: 8.0 2023-04-27 19:04:18,705 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.383e+02 4.905e+02 5.910e+02 7.685e+02 1.899e+03, threshold=1.182e+03, percent-clipped=6.0 2023-04-27 19:04:34,822 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8612, 3.7425, 3.6912, 2.8164, 3.7271, 1.7851, 3.5539, 3.6197], device='cuda:4'), covar=tensor([0.0098, 0.0071, 0.0085, 0.0431, 0.0078, 0.1574, 0.0082, 0.0149], device='cuda:4'), in_proj_covar=tensor([0.0062, 0.0049, 0.0075, 0.0096, 0.0058, 0.0106, 0.0069, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:04:37,329 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16432.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:04:51,243 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16441.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:05:05,090 INFO [train.py:904] (4/8) Epoch 2, batch 6300, loss[loss=0.3496, simple_loss=0.3955, pruned_loss=0.1518, over 16700.00 frames. ], tot_loss[loss=0.3031, simple_loss=0.3665, pruned_loss=0.1198, over 3107859.35 frames. ], batch size: 62, lr: 2.64e-02, grad_scale: 8.0 2023-04-27 19:05:29,230 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-27 19:06:22,078 INFO [train.py:904] (4/8) Epoch 2, batch 6350, loss[loss=0.2662, simple_loss=0.339, pruned_loss=0.09674, over 16498.00 frames. ], tot_loss[loss=0.307, simple_loss=0.3684, pruned_loss=0.1228, over 3076750.96 frames. ], batch size: 68, lr: 2.64e-02, grad_scale: 8.0 2023-04-27 19:06:29,788 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16505.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:06:30,028 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3685, 4.1521, 3.8025, 1.7161, 2.8702, 2.5268, 3.6705, 4.1798], device='cuda:4'), covar=tensor([0.0231, 0.0367, 0.0385, 0.1842, 0.0771, 0.0952, 0.0619, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0109, 0.0155, 0.0156, 0.0150, 0.0138, 0.0153, 0.0101], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:06:52,066 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.677e+02 5.217e+02 6.547e+02 8.022e+02 1.954e+03, threshold=1.309e+03, percent-clipped=7.0 2023-04-27 19:07:21,292 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=16539.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:07:39,358 INFO [train.py:904] (4/8) Epoch 2, batch 6400, loss[loss=0.3814, simple_loss=0.414, pruned_loss=0.1744, over 11016.00 frames. ], tot_loss[loss=0.3069, simple_loss=0.3678, pruned_loss=0.123, over 3081327.84 frames. ], batch size: 248, lr: 2.64e-02, grad_scale: 8.0 2023-04-27 19:08:03,044 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.66 vs. limit=2.0 2023-04-27 19:08:34,561 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=16587.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:08:55,789 INFO [train.py:904] (4/8) Epoch 2, batch 6450, loss[loss=0.3052, simple_loss=0.3488, pruned_loss=0.1308, over 11589.00 frames. ], tot_loss[loss=0.302, simple_loss=0.3648, pruned_loss=0.1196, over 3102320.65 frames. ], batch size: 248, lr: 2.63e-02, grad_scale: 8.0 2023-04-27 19:09:25,707 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.052e+02 4.859e+02 5.740e+02 6.998e+02 1.216e+03, threshold=1.148e+03, percent-clipped=0.0 2023-04-27 19:10:13,782 INFO [train.py:904] (4/8) Epoch 2, batch 6500, loss[loss=0.2312, simple_loss=0.3104, pruned_loss=0.07599, over 16708.00 frames. ], tot_loss[loss=0.2979, simple_loss=0.361, pruned_loss=0.1174, over 3108233.02 frames. ], batch size: 62, lr: 2.63e-02, grad_scale: 8.0 2023-04-27 19:10:18,829 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9970, 3.8789, 4.0440, 4.3758, 4.4074, 3.8546, 4.3515, 4.2424], device='cuda:4'), covar=tensor([0.0586, 0.0497, 0.0988, 0.0342, 0.0351, 0.0639, 0.0321, 0.0398], device='cuda:4'), in_proj_covar=tensor([0.0238, 0.0270, 0.0376, 0.0282, 0.0216, 0.0200, 0.0215, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:10:24,426 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=16657.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:10:27,668 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-27 19:10:32,669 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9809, 3.7862, 3.3894, 1.6561, 2.7088, 2.1201, 3.2229, 3.9182], device='cuda:4'), covar=tensor([0.0311, 0.0382, 0.0420, 0.1921, 0.0822, 0.1163, 0.0790, 0.0330], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0107, 0.0150, 0.0153, 0.0148, 0.0138, 0.0149, 0.0098], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:11:32,129 INFO [train.py:904] (4/8) Epoch 2, batch 6550, loss[loss=0.3089, simple_loss=0.3912, pruned_loss=0.1133, over 16330.00 frames. ], tot_loss[loss=0.2991, simple_loss=0.3632, pruned_loss=0.1175, over 3126965.47 frames. ], batch size: 165, lr: 2.63e-02, grad_scale: 8.0 2023-04-27 19:11:37,690 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=16705.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:11:50,587 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-27 19:11:59,595 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.469e+02 4.692e+02 5.928e+02 7.455e+02 1.589e+03, threshold=1.186e+03, percent-clipped=4.0 2023-04-27 19:12:11,158 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16727.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:12:32,924 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=16741.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:12:47,617 INFO [train.py:904] (4/8) Epoch 2, batch 6600, loss[loss=0.2816, simple_loss=0.3541, pruned_loss=0.1045, over 16783.00 frames. ], tot_loss[loss=0.3022, simple_loss=0.3669, pruned_loss=0.1187, over 3132568.76 frames. ], batch size: 76, lr: 2.62e-02, grad_scale: 8.0 2023-04-27 19:13:09,850 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5853, 3.3689, 2.0351, 4.4218, 4.2847, 3.9700, 1.8042, 2.7579], device='cuda:4'), covar=tensor([0.1719, 0.0433, 0.1620, 0.0081, 0.0146, 0.0275, 0.1311, 0.0806], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0120, 0.0163, 0.0067, 0.0105, 0.0117, 0.0151, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:13:47,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=16789.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:14:06,684 INFO [train.py:904] (4/8) Epoch 2, batch 6650, loss[loss=0.2938, simple_loss=0.3581, pruned_loss=0.1147, over 16849.00 frames. ], tot_loss[loss=0.3033, simple_loss=0.3672, pruned_loss=0.1197, over 3128368.74 frames. ], batch size: 116, lr: 2.62e-02, grad_scale: 8.0 2023-04-27 19:14:12,023 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1216, 3.1093, 3.1795, 3.4036, 3.3975, 3.1666, 3.3730, 3.3525], device='cuda:4'), covar=tensor([0.0526, 0.0418, 0.0999, 0.0396, 0.0470, 0.0993, 0.0472, 0.0396], device='cuda:4'), in_proj_covar=tensor([0.0241, 0.0270, 0.0377, 0.0284, 0.0221, 0.0196, 0.0216, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:14:13,248 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=16805.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:14:35,746 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.480e+02 5.041e+02 6.003e+02 7.722e+02 1.259e+03, threshold=1.201e+03, percent-clipped=2.0 2023-04-27 19:14:59,412 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3848, 4.0392, 3.9478, 1.7206, 4.2138, 4.2153, 3.1857, 3.0726], device='cuda:4'), covar=tensor([0.0865, 0.0091, 0.0168, 0.1612, 0.0071, 0.0050, 0.0293, 0.0442], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0082, 0.0080, 0.0148, 0.0076, 0.0071, 0.0105, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:15:23,513 INFO [train.py:904] (4/8) Epoch 2, batch 6700, loss[loss=0.2697, simple_loss=0.3465, pruned_loss=0.09648, over 16433.00 frames. ], tot_loss[loss=0.3017, simple_loss=0.3654, pruned_loss=0.119, over 3132947.55 frames. ], batch size: 146, lr: 2.61e-02, grad_scale: 8.0 2023-04-27 19:15:26,815 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=16853.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:15:50,582 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16868.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:16:29,655 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0179, 3.8419, 3.6192, 1.6396, 2.8198, 2.2636, 3.4846, 3.9574], device='cuda:4'), covar=tensor([0.0352, 0.0422, 0.0394, 0.1772, 0.0757, 0.0957, 0.0614, 0.0346], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0107, 0.0152, 0.0157, 0.0152, 0.0139, 0.0151, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:16:38,390 INFO [train.py:904] (4/8) Epoch 2, batch 6750, loss[loss=0.2801, simple_loss=0.3466, pruned_loss=0.1068, over 16905.00 frames. ], tot_loss[loss=0.3025, simple_loss=0.3651, pruned_loss=0.1199, over 3129289.14 frames. ], batch size: 109, lr: 2.61e-02, grad_scale: 8.0 2023-04-27 19:16:47,206 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16906.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:17:07,556 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.622e+02 5.251e+02 6.489e+02 8.459e+02 1.507e+03, threshold=1.298e+03, percent-clipped=4.0 2023-04-27 19:17:22,259 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16929.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 19:17:53,310 INFO [train.py:904] (4/8) Epoch 2, batch 6800, loss[loss=0.3071, simple_loss=0.3699, pruned_loss=0.1222, over 16922.00 frames. ], tot_loss[loss=0.3031, simple_loss=0.3654, pruned_loss=0.1204, over 3111679.83 frames. ], batch size: 109, lr: 2.61e-02, grad_scale: 8.0 2023-04-27 19:18:18,746 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16967.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 19:19:04,182 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-27 19:19:10,217 INFO [train.py:904] (4/8) Epoch 2, batch 6850, loss[loss=0.2969, simple_loss=0.3878, pruned_loss=0.103, over 17290.00 frames. ], tot_loss[loss=0.3026, simple_loss=0.3662, pruned_loss=0.1195, over 3123703.69 frames. ], batch size: 52, lr: 2.60e-02, grad_scale: 8.0 2023-04-27 19:19:11,238 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.67 vs. limit=2.0 2023-04-27 19:19:28,367 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0010, 4.2386, 3.3724, 3.1573, 3.7208, 2.9780, 4.6536, 5.0557], device='cuda:4'), covar=tensor([0.1702, 0.0506, 0.0943, 0.0584, 0.1261, 0.0852, 0.0258, 0.0119], device='cuda:4'), in_proj_covar=tensor([0.0252, 0.0221, 0.0238, 0.0175, 0.0274, 0.0179, 0.0198, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:19:38,485 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.045e+02 4.695e+02 5.738e+02 7.185e+02 1.728e+03, threshold=1.148e+03, percent-clipped=4.0 2023-04-27 19:19:49,695 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=17027.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:20:17,323 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5502, 5.8431, 5.4861, 5.6603, 4.9232, 4.9345, 5.2675, 5.8969], device='cuda:4'), covar=tensor([0.0399, 0.0478, 0.0618, 0.0337, 0.0532, 0.0373, 0.0405, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0221, 0.0305, 0.0276, 0.0193, 0.0205, 0.0187, 0.0249, 0.0213], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:20:17,399 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1280, 4.7935, 4.9192, 4.9708, 4.3719, 4.9881, 4.9238, 4.6662], device='cuda:4'), covar=tensor([0.0344, 0.0257, 0.0182, 0.0103, 0.0728, 0.0200, 0.0157, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0094, 0.0151, 0.0119, 0.0178, 0.0127, 0.0108, 0.0140], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 19:20:24,230 INFO [train.py:904] (4/8) Epoch 2, batch 6900, loss[loss=0.3308, simple_loss=0.383, pruned_loss=0.1393, over 16201.00 frames. ], tot_loss[loss=0.3021, simple_loss=0.3679, pruned_loss=0.1181, over 3131738.78 frames. ], batch size: 165, lr: 2.60e-02, grad_scale: 8.0 2023-04-27 19:21:01,219 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=17075.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:21:24,844 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5450, 4.6135, 4.3625, 1.7146, 4.8498, 4.8922, 3.8880, 3.6115], device='cuda:4'), covar=tensor([0.0846, 0.0078, 0.0198, 0.1535, 0.0057, 0.0028, 0.0260, 0.0338], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0085, 0.0080, 0.0151, 0.0077, 0.0071, 0.0106, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 19:21:29,148 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5837, 3.3555, 2.9558, 1.7050, 2.4069, 2.1611, 3.0151, 3.4820], device='cuda:4'), covar=tensor([0.0296, 0.0398, 0.0488, 0.1511, 0.0845, 0.0912, 0.0584, 0.0318], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0104, 0.0149, 0.0151, 0.0145, 0.0134, 0.0145, 0.0101], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:21:30,967 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6729, 1.2462, 1.3230, 1.5788, 1.6419, 1.8202, 1.3977, 1.7618], device='cuda:4'), covar=tensor([0.0057, 0.0161, 0.0084, 0.0099, 0.0049, 0.0055, 0.0144, 0.0045], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0108, 0.0094, 0.0082, 0.0061, 0.0059, 0.0095, 0.0054], device='cuda:4'), out_proj_covar=tensor([1.2098e-04, 1.8968e-04, 1.6996e-04, 1.5055e-04, 1.0569e-04, 1.0459e-04, 1.6182e-04, 9.3593e-05], device='cuda:4') 2023-04-27 19:21:40,743 INFO [train.py:904] (4/8) Epoch 2, batch 6950, loss[loss=0.3579, simple_loss=0.3884, pruned_loss=0.1637, over 11687.00 frames. ], tot_loss[loss=0.3098, simple_loss=0.3723, pruned_loss=0.1236, over 3098214.35 frames. ], batch size: 248, lr: 2.60e-02, grad_scale: 8.0 2023-04-27 19:22:09,949 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.451e+02 5.662e+02 6.855e+02 8.747e+02 1.724e+03, threshold=1.371e+03, percent-clipped=6.0 2023-04-27 19:22:54,958 INFO [train.py:904] (4/8) Epoch 2, batch 7000, loss[loss=0.3474, simple_loss=0.3878, pruned_loss=0.1534, over 11762.00 frames. ], tot_loss[loss=0.3093, simple_loss=0.3724, pruned_loss=0.1231, over 3090568.00 frames. ], batch size: 248, lr: 2.59e-02, grad_scale: 8.0 2023-04-27 19:23:11,343 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5048, 3.7322, 2.9598, 2.7228, 2.9748, 2.2993, 3.9313, 4.5346], device='cuda:4'), covar=tensor([0.2018, 0.0657, 0.1082, 0.0740, 0.1673, 0.1082, 0.0323, 0.0179], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0224, 0.0237, 0.0177, 0.0268, 0.0182, 0.0195, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:23:53,115 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17189.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:24:04,851 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-27 19:24:10,871 INFO [train.py:904] (4/8) Epoch 2, batch 7050, loss[loss=0.3446, simple_loss=0.3988, pruned_loss=0.1453, over 15318.00 frames. ], tot_loss[loss=0.3096, simple_loss=0.3729, pruned_loss=0.1232, over 3068637.15 frames. ], batch size: 190, lr: 2.59e-02, grad_scale: 8.0 2023-04-27 19:24:37,722 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.062e+02 5.191e+02 6.345e+02 7.855e+02 1.482e+03, threshold=1.269e+03, percent-clipped=3.0 2023-04-27 19:24:43,981 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17224.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 19:24:47,448 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6486, 3.4030, 2.4176, 4.9949, 4.8293, 4.3568, 2.1145, 3.2752], device='cuda:4'), covar=tensor([0.1606, 0.0469, 0.1261, 0.0064, 0.0123, 0.0227, 0.1109, 0.0607], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0120, 0.0164, 0.0070, 0.0108, 0.0118, 0.0151, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:25:14,299 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.85 vs. limit=2.0 2023-04-27 19:25:22,962 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17250.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:25:23,667 INFO [train.py:904] (4/8) Epoch 2, batch 7100, loss[loss=0.3005, simple_loss=0.3704, pruned_loss=0.1152, over 15493.00 frames. ], tot_loss[loss=0.3089, simple_loss=0.3716, pruned_loss=0.1231, over 3049950.22 frames. ], batch size: 190, lr: 2.59e-02, grad_scale: 8.0 2023-04-27 19:25:40,700 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17262.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 19:26:38,658 INFO [train.py:904] (4/8) Epoch 2, batch 7150, loss[loss=0.3108, simple_loss=0.3683, pruned_loss=0.1267, over 15280.00 frames. ], tot_loss[loss=0.306, simple_loss=0.3686, pruned_loss=0.1217, over 3059866.90 frames. ], batch size: 190, lr: 2.58e-02, grad_scale: 8.0 2023-04-27 19:27:07,389 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.740e+02 5.248e+02 6.301e+02 7.700e+02 1.780e+03, threshold=1.260e+03, percent-clipped=1.0 2023-04-27 19:27:21,979 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-27 19:27:52,978 INFO [train.py:904] (4/8) Epoch 2, batch 7200, loss[loss=0.2961, simple_loss=0.3575, pruned_loss=0.1173, over 15271.00 frames. ], tot_loss[loss=0.3031, simple_loss=0.3663, pruned_loss=0.1199, over 3031569.45 frames. ], batch size: 191, lr: 2.58e-02, grad_scale: 8.0 2023-04-27 19:27:53,572 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1096, 2.2358, 1.8701, 1.9947, 2.7507, 2.5195, 3.3095, 3.0346], device='cuda:4'), covar=tensor([0.0016, 0.0170, 0.0183, 0.0197, 0.0093, 0.0143, 0.0025, 0.0063], device='cuda:4'), in_proj_covar=tensor([0.0045, 0.0098, 0.0099, 0.0101, 0.0088, 0.0101, 0.0053, 0.0071], device='cuda:4'), out_proj_covar=tensor([6.4273e-05, 1.5198e-04, 1.5002e-04, 1.5811e-04, 1.4244e-04, 1.6177e-04, 8.2570e-05, 1.1555e-04], device='cuda:4') 2023-04-27 19:28:10,376 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17362.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:28:11,678 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7512, 3.0834, 3.0474, 2.1391, 3.0227, 3.0394, 3.1266, 1.6352], device='cuda:4'), covar=tensor([0.0542, 0.0038, 0.0052, 0.0302, 0.0054, 0.0075, 0.0040, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0053, 0.0056, 0.0110, 0.0054, 0.0058, 0.0059, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 19:28:44,454 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17383.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 19:28:44,611 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2254, 2.6060, 2.2769, 3.5463, 1.9837, 3.3328, 2.3317, 2.1131], device='cuda:4'), covar=tensor([0.0297, 0.0423, 0.0352, 0.0207, 0.1422, 0.0197, 0.0681, 0.1156], device='cuda:4'), in_proj_covar=tensor([0.0216, 0.0186, 0.0158, 0.0217, 0.0266, 0.0172, 0.0188, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:29:13,591 INFO [train.py:904] (4/8) Epoch 2, batch 7250, loss[loss=0.2981, simple_loss=0.3534, pruned_loss=0.1214, over 16359.00 frames. ], tot_loss[loss=0.3001, simple_loss=0.3638, pruned_loss=0.1182, over 3040480.24 frames. ], batch size: 146, lr: 2.58e-02, grad_scale: 8.0 2023-04-27 19:29:23,350 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-27 19:29:42,538 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.618e+02 4.777e+02 5.726e+02 7.027e+02 1.873e+03, threshold=1.145e+03, percent-clipped=4.0 2023-04-27 19:29:47,317 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17423.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:30:19,158 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17444.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 19:30:29,420 INFO [train.py:904] (4/8) Epoch 2, batch 7300, loss[loss=0.3175, simple_loss=0.3787, pruned_loss=0.1281, over 16601.00 frames. ], tot_loss[loss=0.2973, simple_loss=0.3617, pruned_loss=0.1165, over 3062481.63 frames. ], batch size: 57, lr: 2.57e-02, grad_scale: 8.0 2023-04-27 19:31:46,485 INFO [train.py:904] (4/8) Epoch 2, batch 7350, loss[loss=0.2696, simple_loss=0.3427, pruned_loss=0.0982, over 16546.00 frames. ], tot_loss[loss=0.2972, simple_loss=0.361, pruned_loss=0.1167, over 3042810.43 frames. ], batch size: 75, lr: 2.57e-02, grad_scale: 8.0 2023-04-27 19:32:16,421 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.647e+02 4.932e+02 5.634e+02 7.018e+02 1.459e+03, threshold=1.127e+03, percent-clipped=2.0 2023-04-27 19:32:22,427 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=17524.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 19:32:57,422 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17545.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:33:06,149 INFO [train.py:904] (4/8) Epoch 2, batch 7400, loss[loss=0.2822, simple_loss=0.3595, pruned_loss=0.1025, over 17218.00 frames. ], tot_loss[loss=0.2974, simple_loss=0.3616, pruned_loss=0.1166, over 3071777.26 frames. ], batch size: 45, lr: 2.57e-02, grad_scale: 8.0 2023-04-27 19:33:25,351 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=17562.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:33:41,414 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=17572.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:33:47,729 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.40 vs. limit=5.0 2023-04-27 19:33:56,947 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4383, 1.4433, 1.6465, 2.3005, 2.4180, 2.3572, 1.4498, 2.3083], device='cuda:4'), covar=tensor([0.0050, 0.0245, 0.0144, 0.0089, 0.0036, 0.0055, 0.0183, 0.0039], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0111, 0.0095, 0.0080, 0.0061, 0.0059, 0.0098, 0.0055], device='cuda:4'), out_proj_covar=tensor([1.2084e-04, 1.9368e-04, 1.7202e-04, 1.4558e-04, 1.0387e-04, 1.0401e-04, 1.6654e-04, 9.4020e-05], device='cuda:4') 2023-04-27 19:34:27,196 INFO [train.py:904] (4/8) Epoch 2, batch 7450, loss[loss=0.3489, simple_loss=0.4088, pruned_loss=0.1445, over 15368.00 frames. ], tot_loss[loss=0.2995, simple_loss=0.363, pruned_loss=0.118, over 3077062.03 frames. ], batch size: 190, lr: 2.56e-02, grad_scale: 8.0 2023-04-27 19:34:43,752 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=17610.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:34:59,850 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.824e+02 5.406e+02 6.493e+02 7.703e+02 1.630e+03, threshold=1.299e+03, percent-clipped=5.0 2023-04-27 19:35:49,110 INFO [train.py:904] (4/8) Epoch 2, batch 7500, loss[loss=0.3566, simple_loss=0.3928, pruned_loss=0.1602, over 11402.00 frames. ], tot_loss[loss=0.2997, simple_loss=0.3642, pruned_loss=0.1176, over 3090965.98 frames. ], batch size: 248, lr: 2.56e-02, grad_scale: 8.0 2023-04-27 19:36:57,932 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4356, 3.4404, 1.5209, 3.4939, 2.3785, 3.5099, 1.8408, 2.6739], device='cuda:4'), covar=tensor([0.0044, 0.0175, 0.1560, 0.0037, 0.0693, 0.0261, 0.1197, 0.0509], device='cuda:4'), in_proj_covar=tensor([0.0073, 0.0111, 0.0177, 0.0074, 0.0161, 0.0140, 0.0179, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 19:36:59,679 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1555, 3.0845, 3.1937, 3.4226, 3.3737, 3.1992, 3.3095, 3.4177], device='cuda:4'), covar=tensor([0.0503, 0.0509, 0.0852, 0.0366, 0.0474, 0.1116, 0.0534, 0.0373], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0275, 0.0370, 0.0279, 0.0219, 0.0199, 0.0217, 0.0221], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:37:03,901 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4050, 3.3549, 3.3016, 2.8484, 3.3257, 2.1285, 3.1855, 3.1851], device='cuda:4'), covar=tensor([0.0071, 0.0058, 0.0073, 0.0236, 0.0064, 0.1019, 0.0075, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0062, 0.0050, 0.0076, 0.0095, 0.0057, 0.0108, 0.0068, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:37:05,020 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17700.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:37:05,673 INFO [train.py:904] (4/8) Epoch 2, batch 7550, loss[loss=0.288, simple_loss=0.345, pruned_loss=0.1155, over 16547.00 frames. ], tot_loss[loss=0.2988, simple_loss=0.3628, pruned_loss=0.1173, over 3082591.31 frames. ], batch size: 68, lr: 2.56e-02, grad_scale: 8.0 2023-04-27 19:37:11,531 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2282, 5.4494, 5.2107, 5.2713, 4.7252, 4.4309, 4.9667, 5.5267], device='cuda:4'), covar=tensor([0.0334, 0.0474, 0.0592, 0.0286, 0.0395, 0.0419, 0.0336, 0.0461], device='cuda:4'), in_proj_covar=tensor([0.0217, 0.0305, 0.0277, 0.0192, 0.0202, 0.0190, 0.0241, 0.0205], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:37:32,528 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17718.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:37:34,648 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.677e+02 5.137e+02 6.469e+02 8.298e+02 1.927e+03, threshold=1.294e+03, percent-clipped=3.0 2023-04-27 19:37:51,094 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17730.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:38:04,539 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17739.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 19:38:22,209 INFO [train.py:904] (4/8) Epoch 2, batch 7600, loss[loss=0.2777, simple_loss=0.3513, pruned_loss=0.1021, over 16748.00 frames. ], tot_loss[loss=0.2993, simple_loss=0.3627, pruned_loss=0.118, over 3066276.32 frames. ], batch size: 83, lr: 2.55e-02, grad_scale: 8.0 2023-04-27 19:38:37,892 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17761.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:39:25,113 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17791.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:39:27,564 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9727, 3.9447, 2.6108, 5.1596, 5.1003, 4.4511, 2.2470, 3.4107], device='cuda:4'), covar=tensor([0.1299, 0.0307, 0.1193, 0.0054, 0.0092, 0.0206, 0.0984, 0.0561], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0124, 0.0166, 0.0068, 0.0108, 0.0119, 0.0154, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:39:39,705 INFO [train.py:904] (4/8) Epoch 2, batch 7650, loss[loss=0.2846, simple_loss=0.3502, pruned_loss=0.1095, over 16525.00 frames. ], tot_loss[loss=0.2986, simple_loss=0.3625, pruned_loss=0.1173, over 3101140.82 frames. ], batch size: 75, lr: 2.55e-02, grad_scale: 8.0 2023-04-27 19:40:10,726 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.673e+02 5.378e+02 6.384e+02 7.645e+02 1.933e+03, threshold=1.277e+03, percent-clipped=1.0 2023-04-27 19:40:50,416 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=17845.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:40:58,872 INFO [train.py:904] (4/8) Epoch 2, batch 7700, loss[loss=0.4309, simple_loss=0.4492, pruned_loss=0.2063, over 11527.00 frames. ], tot_loss[loss=0.3002, simple_loss=0.3631, pruned_loss=0.1187, over 3101536.15 frames. ], batch size: 248, lr: 2.55e-02, grad_scale: 8.0 2023-04-27 19:41:53,362 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7154, 4.4869, 4.4563, 4.6313, 3.7168, 4.6465, 4.5643, 4.2148], device='cuda:4'), covar=tensor([0.0352, 0.0323, 0.0241, 0.0150, 0.0868, 0.0253, 0.0181, 0.0270], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0096, 0.0146, 0.0119, 0.0174, 0.0125, 0.0106, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 19:42:04,291 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=17893.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:42:09,405 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 19:42:16,830 INFO [train.py:904] (4/8) Epoch 2, batch 7750, loss[loss=0.2737, simple_loss=0.3481, pruned_loss=0.09961, over 16607.00 frames. ], tot_loss[loss=0.3004, simple_loss=0.3638, pruned_loss=0.1185, over 3112989.48 frames. ], batch size: 57, lr: 2.54e-02, grad_scale: 16.0 2023-04-27 19:42:46,601 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.308e+02 4.989e+02 6.319e+02 7.239e+02 1.280e+03, threshold=1.264e+03, percent-clipped=2.0 2023-04-27 19:42:49,601 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8600, 3.6853, 3.7313, 3.1844, 3.6828, 1.7461, 3.5607, 3.6230], device='cuda:4'), covar=tensor([0.0070, 0.0071, 0.0070, 0.0301, 0.0063, 0.1348, 0.0073, 0.0085], device='cuda:4'), in_proj_covar=tensor([0.0061, 0.0050, 0.0076, 0.0098, 0.0058, 0.0108, 0.0069, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:43:31,901 INFO [train.py:904] (4/8) Epoch 2, batch 7800, loss[loss=0.3252, simple_loss=0.3859, pruned_loss=0.1323, over 16233.00 frames. ], tot_loss[loss=0.3027, simple_loss=0.3654, pruned_loss=0.12, over 3120116.57 frames. ], batch size: 165, lr: 2.54e-02, grad_scale: 16.0 2023-04-27 19:43:47,055 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17960.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:44:52,420 INFO [train.py:904] (4/8) Epoch 2, batch 7850, loss[loss=0.264, simple_loss=0.3311, pruned_loss=0.09846, over 16283.00 frames. ], tot_loss[loss=0.3027, simple_loss=0.366, pruned_loss=0.1197, over 3116518.26 frames. ], batch size: 35, lr: 2.54e-02, grad_scale: 16.0 2023-04-27 19:45:18,572 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18018.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:45:21,253 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.805e+02 5.114e+02 6.121e+02 7.560e+02 1.271e+03, threshold=1.224e+03, percent-clipped=1.0 2023-04-27 19:45:22,948 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18021.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:45:48,864 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18039.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 19:46:05,625 INFO [train.py:904] (4/8) Epoch 2, batch 7900, loss[loss=0.2781, simple_loss=0.3519, pruned_loss=0.1022, over 17122.00 frames. ], tot_loss[loss=0.3022, simple_loss=0.3653, pruned_loss=0.1196, over 3108044.81 frames. ], batch size: 49, lr: 2.53e-02, grad_scale: 8.0 2023-04-27 19:46:13,023 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0519, 2.1372, 1.8152, 1.9436, 2.6375, 2.4300, 3.3852, 3.1063], device='cuda:4'), covar=tensor([0.0018, 0.0135, 0.0171, 0.0176, 0.0077, 0.0134, 0.0023, 0.0062], device='cuda:4'), in_proj_covar=tensor([0.0048, 0.0098, 0.0103, 0.0105, 0.0092, 0.0102, 0.0057, 0.0075], device='cuda:4'), out_proj_covar=tensor([6.8018e-05, 1.4971e-04, 1.5652e-04, 1.6249e-04, 1.4863e-04, 1.6268e-04, 8.8438e-05, 1.2229e-04], device='cuda:4') 2023-04-27 19:46:13,969 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18056.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:46:29,400 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=18066.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:46:31,911 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.48 vs. limit=5.0 2023-04-27 19:46:32,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3035, 3.2151, 2.5983, 2.6610, 2.4070, 1.8275, 3.2572, 3.7692], device='cuda:4'), covar=tensor([0.2020, 0.0632, 0.0990, 0.0702, 0.1708, 0.1484, 0.0387, 0.0248], device='cuda:4'), in_proj_covar=tensor([0.0260, 0.0227, 0.0241, 0.0185, 0.0280, 0.0186, 0.0203, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:46:54,733 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-27 19:47:03,304 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18086.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:47:05,058 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=18087.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 19:47:17,363 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8274, 4.0814, 3.8580, 3.9089, 3.5007, 3.8183, 3.7712, 4.0221], device='cuda:4'), covar=tensor([0.0434, 0.0560, 0.0629, 0.0326, 0.0478, 0.0594, 0.0482, 0.0575], device='cuda:4'), in_proj_covar=tensor([0.0223, 0.0310, 0.0280, 0.0196, 0.0209, 0.0198, 0.0256, 0.0212], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:47:25,420 INFO [train.py:904] (4/8) Epoch 2, batch 7950, loss[loss=0.258, simple_loss=0.3352, pruned_loss=0.09038, over 16762.00 frames. ], tot_loss[loss=0.3015, simple_loss=0.3644, pruned_loss=0.1193, over 3108258.32 frames. ], batch size: 83, lr: 2.53e-02, grad_scale: 8.0 2023-04-27 19:47:32,287 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0199, 1.6141, 1.5598, 1.4626, 1.8203, 1.6522, 1.8906, 1.8799], device='cuda:4'), covar=tensor([0.0016, 0.0079, 0.0105, 0.0108, 0.0059, 0.0090, 0.0037, 0.0059], device='cuda:4'), in_proj_covar=tensor([0.0049, 0.0100, 0.0104, 0.0106, 0.0094, 0.0103, 0.0058, 0.0076], device='cuda:4'), out_proj_covar=tensor([7.0238e-05, 1.5398e-04, 1.5778e-04, 1.6513e-04, 1.5151e-04, 1.6352e-04, 9.0687e-05, 1.2318e-04], device='cuda:4') 2023-04-27 19:47:56,298 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.877e+02 5.212e+02 6.244e+02 7.914e+02 2.379e+03, threshold=1.249e+03, percent-clipped=3.0 2023-04-27 19:48:42,375 INFO [train.py:904] (4/8) Epoch 2, batch 8000, loss[loss=0.406, simple_loss=0.4182, pruned_loss=0.1969, over 11084.00 frames. ], tot_loss[loss=0.3013, simple_loss=0.3639, pruned_loss=0.1193, over 3104932.98 frames. ], batch size: 248, lr: 2.53e-02, grad_scale: 8.0 2023-04-27 19:48:57,754 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.88 vs. limit=2.0 2023-04-27 19:49:10,039 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-27 19:49:16,873 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18174.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:49:18,065 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18175.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:49:56,233 INFO [train.py:904] (4/8) Epoch 2, batch 8050, loss[loss=0.3323, simple_loss=0.3725, pruned_loss=0.1461, over 11528.00 frames. ], tot_loss[loss=0.3025, simple_loss=0.3644, pruned_loss=0.1203, over 3067245.65 frames. ], batch size: 249, lr: 2.52e-02, grad_scale: 8.0 2023-04-27 19:50:24,924 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.258e+02 4.729e+02 5.978e+02 7.201e+02 1.227e+03, threshold=1.196e+03, percent-clipped=0.0 2023-04-27 19:50:29,187 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7756, 4.1227, 3.8685, 3.9799, 3.5666, 3.7926, 3.7953, 4.0073], device='cuda:4'), covar=tensor([0.0500, 0.0617, 0.0772, 0.0339, 0.0552, 0.0695, 0.0516, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0310, 0.0282, 0.0198, 0.0208, 0.0198, 0.0253, 0.0215], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:50:47,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18235.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:50:48,667 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18236.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:51:11,139 INFO [train.py:904] (4/8) Epoch 2, batch 8100, loss[loss=0.2986, simple_loss=0.3648, pruned_loss=0.1162, over 16660.00 frames. ], tot_loss[loss=0.3016, simple_loss=0.3643, pruned_loss=0.1194, over 3070742.42 frames. ], batch size: 57, lr: 2.52e-02, grad_scale: 8.0 2023-04-27 19:51:12,415 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6848, 3.7885, 3.6824, 2.4694, 3.6276, 3.5089, 3.8464, 2.0169], device='cuda:4'), covar=tensor([0.0572, 0.0025, 0.0036, 0.0277, 0.0045, 0.0086, 0.0023, 0.0425], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0054, 0.0058, 0.0108, 0.0053, 0.0059, 0.0057, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 19:52:29,024 INFO [train.py:904] (4/8) Epoch 2, batch 8150, loss[loss=0.3505, simple_loss=0.3823, pruned_loss=0.1594, over 11201.00 frames. ], tot_loss[loss=0.2981, simple_loss=0.3614, pruned_loss=0.1174, over 3075277.79 frames. ], batch size: 250, lr: 2.52e-02, grad_scale: 8.0 2023-04-27 19:52:35,776 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5500, 3.8826, 4.1081, 4.0979, 4.1225, 3.7512, 3.3528, 3.9793], device='cuda:4'), covar=tensor([0.0491, 0.0411, 0.0493, 0.0581, 0.0697, 0.0506, 0.1673, 0.0377], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0153, 0.0166, 0.0163, 0.0201, 0.0169, 0.0259, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 19:52:52,885 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18316.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:53:00,707 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.225e+02 5.223e+02 6.693e+02 9.034e+02 1.543e+03, threshold=1.339e+03, percent-clipped=7.0 2023-04-27 19:53:48,015 INFO [train.py:904] (4/8) Epoch 2, batch 8200, loss[loss=0.2978, simple_loss=0.3643, pruned_loss=0.1156, over 16384.00 frames. ], tot_loss[loss=0.295, simple_loss=0.3581, pruned_loss=0.1159, over 3090453.53 frames. ], batch size: 146, lr: 2.51e-02, grad_scale: 4.0 2023-04-27 19:53:57,186 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18356.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:54:46,849 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18386.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:55:10,294 INFO [train.py:904] (4/8) Epoch 2, batch 8250, loss[loss=0.2583, simple_loss=0.3228, pruned_loss=0.09692, over 12053.00 frames. ], tot_loss[loss=0.2923, simple_loss=0.3574, pruned_loss=0.1136, over 3078680.36 frames. ], batch size: 248, lr: 2.51e-02, grad_scale: 4.0 2023-04-27 19:55:11,275 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.87 vs. limit=5.0 2023-04-27 19:55:15,139 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=18404.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:55:32,666 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9025, 4.5373, 4.6874, 4.1235, 4.6964, 2.2889, 4.3398, 4.6829], device='cuda:4'), covar=tensor([0.0057, 0.0065, 0.0055, 0.0252, 0.0044, 0.1180, 0.0057, 0.0082], device='cuda:4'), in_proj_covar=tensor([0.0063, 0.0053, 0.0080, 0.0097, 0.0059, 0.0111, 0.0071, 0.0077], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:55:44,640 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.019e+02 4.548e+02 5.368e+02 6.842e+02 2.128e+03, threshold=1.074e+03, percent-clipped=3.0 2023-04-27 19:56:05,979 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=18434.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:56:33,571 INFO [train.py:904] (4/8) Epoch 2, batch 8300, loss[loss=0.2558, simple_loss=0.3467, pruned_loss=0.08243, over 16882.00 frames. ], tot_loss[loss=0.2843, simple_loss=0.3522, pruned_loss=0.1082, over 3077444.26 frames. ], batch size: 96, lr: 2.51e-02, grad_scale: 4.0 2023-04-27 19:56:48,991 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6758, 4.0106, 3.2754, 2.8225, 3.1746, 2.4949, 4.2165, 4.7249], device='cuda:4'), covar=tensor([0.1839, 0.0540, 0.0944, 0.0764, 0.1520, 0.1104, 0.0263, 0.0116], device='cuda:4'), in_proj_covar=tensor([0.0252, 0.0220, 0.0235, 0.0178, 0.0262, 0.0184, 0.0195, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:57:37,587 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18490.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:57:54,223 INFO [train.py:904] (4/8) Epoch 2, batch 8350, loss[loss=0.2732, simple_loss=0.3543, pruned_loss=0.09605, over 15331.00 frames. ], tot_loss[loss=0.2789, simple_loss=0.349, pruned_loss=0.1044, over 3063324.42 frames. ], batch size: 190, lr: 2.50e-02, grad_scale: 4.0 2023-04-27 19:58:06,643 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18508.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:58:28,818 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.204e+02 3.865e+02 4.910e+02 6.166e+02 1.131e+03, threshold=9.820e+02, percent-clipped=1.0 2023-04-27 19:58:42,168 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18530.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:58:44,076 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18531.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:59:15,560 INFO [train.py:904] (4/8) Epoch 2, batch 8400, loss[loss=0.2397, simple_loss=0.3218, pruned_loss=0.0788, over 16688.00 frames. ], tot_loss[loss=0.2733, simple_loss=0.3448, pruned_loss=0.1009, over 3057278.96 frames. ], batch size: 89, lr: 2.50e-02, grad_scale: 8.0 2023-04-27 19:59:16,099 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18551.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:59:45,548 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18569.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 20:00:07,836 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.32 vs. limit=5.0 2023-04-27 20:00:35,047 INFO [train.py:904] (4/8) Epoch 2, batch 8450, loss[loss=0.2434, simple_loss=0.3335, pruned_loss=0.07667, over 15360.00 frames. ], tot_loss[loss=0.2695, simple_loss=0.3424, pruned_loss=0.09829, over 3055442.52 frames. ], batch size: 190, lr: 2.50e-02, grad_scale: 8.0 2023-04-27 20:00:50,098 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8850, 4.6258, 4.6491, 4.1554, 4.6238, 1.9790, 4.3512, 4.6169], device='cuda:4'), covar=tensor([0.0045, 0.0043, 0.0051, 0.0180, 0.0041, 0.1290, 0.0056, 0.0072], device='cuda:4'), in_proj_covar=tensor([0.0061, 0.0051, 0.0079, 0.0091, 0.0059, 0.0111, 0.0071, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:01:00,267 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18616.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:01:09,358 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.749e+02 4.009e+02 5.012e+02 6.279e+02 1.629e+03, threshold=1.002e+03, percent-clipped=8.0 2023-04-27 20:01:11,310 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3801, 3.1647, 2.6380, 2.3768, 2.2363, 1.9038, 3.2864, 3.6637], device='cuda:4'), covar=tensor([0.1670, 0.0621, 0.0908, 0.0696, 0.1743, 0.1373, 0.0273, 0.0159], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0213, 0.0228, 0.0173, 0.0236, 0.0182, 0.0190, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:01:20,769 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.70 vs. limit=2.0 2023-04-27 20:01:55,557 INFO [train.py:904] (4/8) Epoch 2, batch 8500, loss[loss=0.2535, simple_loss=0.3266, pruned_loss=0.09025, over 16717.00 frames. ], tot_loss[loss=0.2629, simple_loss=0.3369, pruned_loss=0.09439, over 3047927.23 frames. ], batch size: 124, lr: 2.49e-02, grad_scale: 8.0 2023-04-27 20:02:17,759 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=18664.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:02:34,610 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18674.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:03:19,432 INFO [train.py:904] (4/8) Epoch 2, batch 8550, loss[loss=0.2446, simple_loss=0.3251, pruned_loss=0.08206, over 16706.00 frames. ], tot_loss[loss=0.2605, simple_loss=0.3345, pruned_loss=0.09323, over 3025176.26 frames. ], batch size: 57, lr: 2.49e-02, grad_scale: 4.0 2023-04-27 20:04:03,003 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.327e+02 4.014e+02 4.987e+02 6.688e+02 1.627e+03, threshold=9.973e+02, percent-clipped=3.0 2023-04-27 20:04:26,409 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18735.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:04:58,635 INFO [train.py:904] (4/8) Epoch 2, batch 8600, loss[loss=0.3016, simple_loss=0.37, pruned_loss=0.1166, over 15202.00 frames. ], tot_loss[loss=0.2611, simple_loss=0.3358, pruned_loss=0.09323, over 3008289.10 frames. ], batch size: 191, lr: 2.49e-02, grad_scale: 2.0 2023-04-27 20:06:30,387 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-27 20:06:37,154 INFO [train.py:904] (4/8) Epoch 2, batch 8650, loss[loss=0.2191, simple_loss=0.3116, pruned_loss=0.06336, over 15482.00 frames. ], tot_loss[loss=0.2567, simple_loss=0.3326, pruned_loss=0.09042, over 3011612.70 frames. ], batch size: 191, lr: 2.49e-02, grad_scale: 2.0 2023-04-27 20:07:28,747 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-27 20:07:29,095 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.199e+02 3.846e+02 4.801e+02 6.827e+02 6.621e+03, threshold=9.601e+02, percent-clipped=16.0 2023-04-27 20:07:44,441 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18830.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:07:46,462 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18831.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:07:48,350 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7218, 4.8488, 5.1923, 5.2784, 5.3558, 4.7481, 4.9474, 4.7173], device='cuda:4'), covar=tensor([0.0179, 0.0296, 0.0377, 0.0370, 0.0260, 0.0210, 0.0577, 0.0311], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0148, 0.0166, 0.0156, 0.0192, 0.0165, 0.0246, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-04-27 20:08:14,595 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18846.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:08:23,066 INFO [train.py:904] (4/8) Epoch 2, batch 8700, loss[loss=0.2291, simple_loss=0.3111, pruned_loss=0.07352, over 16483.00 frames. ], tot_loss[loss=0.2511, simple_loss=0.3276, pruned_loss=0.08726, over 3023129.45 frames. ], batch size: 68, lr: 2.48e-02, grad_scale: 2.0 2023-04-27 20:08:49,482 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18864.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 20:09:14,855 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=18878.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:09:17,053 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=18879.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:10:00,396 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-27 20:10:01,022 INFO [train.py:904] (4/8) Epoch 2, batch 8750, loss[loss=0.2922, simple_loss=0.3638, pruned_loss=0.1103, over 15371.00 frames. ], tot_loss[loss=0.2492, simple_loss=0.3268, pruned_loss=0.08575, over 3037144.66 frames. ], batch size: 191, lr: 2.48e-02, grad_scale: 2.0 2023-04-27 20:10:08,305 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.67 vs. limit=2.0 2023-04-27 20:10:45,261 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0762, 3.2201, 3.1617, 1.5330, 3.3940, 3.3559, 2.8653, 2.7046], device='cuda:4'), covar=tensor([0.0930, 0.0126, 0.0170, 0.1609, 0.0077, 0.0069, 0.0332, 0.0421], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0084, 0.0083, 0.0158, 0.0074, 0.0073, 0.0111, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 20:10:57,517 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.382e+02 3.853e+02 4.840e+02 6.218e+02 1.090e+03, threshold=9.680e+02, percent-clipped=4.0 2023-04-27 20:11:10,259 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18930.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:11:42,890 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4068, 2.9132, 2.4914, 2.3097, 2.2142, 1.9350, 2.8583, 3.1561], device='cuda:4'), covar=tensor([0.1388, 0.0532, 0.0871, 0.0682, 0.1584, 0.1360, 0.0328, 0.0203], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0220, 0.0234, 0.0179, 0.0220, 0.0183, 0.0193, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:11:53,566 INFO [train.py:904] (4/8) Epoch 2, batch 8800, loss[loss=0.262, simple_loss=0.3365, pruned_loss=0.09376, over 15387.00 frames. ], tot_loss[loss=0.2467, simple_loss=0.3249, pruned_loss=0.08429, over 3059608.10 frames. ], batch size: 191, lr: 2.48e-02, grad_scale: 4.0 2023-04-27 20:13:18,193 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18991.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:13:38,729 INFO [train.py:904] (4/8) Epoch 2, batch 8850, loss[loss=0.2396, simple_loss=0.328, pruned_loss=0.07565, over 16680.00 frames. ], tot_loss[loss=0.2465, simple_loss=0.3265, pruned_loss=0.08324, over 3044639.21 frames. ], batch size: 134, lr: 2.47e-02, grad_scale: 4.0 2023-04-27 20:14:28,277 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.418e+02 3.744e+02 4.791e+02 5.972e+02 1.134e+03, threshold=9.582e+02, percent-clipped=2.0 2023-04-27 20:14:43,445 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19030.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:15:19,200 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2651, 2.9911, 2.9035, 1.9307, 2.7419, 2.0172, 2.7739, 3.1153], device='cuda:4'), covar=tensor([0.0266, 0.0420, 0.0382, 0.1353, 0.0536, 0.0918, 0.0609, 0.0387], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0103, 0.0153, 0.0152, 0.0142, 0.0136, 0.0145, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:15:27,231 INFO [train.py:904] (4/8) Epoch 2, batch 8900, loss[loss=0.23, simple_loss=0.3117, pruned_loss=0.07417, over 16515.00 frames. ], tot_loss[loss=0.2452, simple_loss=0.3261, pruned_loss=0.0821, over 3043093.01 frames. ], batch size: 62, lr: 2.47e-02, grad_scale: 4.0 2023-04-27 20:17:32,215 INFO [train.py:904] (4/8) Epoch 2, batch 8950, loss[loss=0.2126, simple_loss=0.3043, pruned_loss=0.06044, over 16180.00 frames. ], tot_loss[loss=0.2454, simple_loss=0.3258, pruned_loss=0.0825, over 3059486.84 frames. ], batch size: 166, lr: 2.47e-02, grad_scale: 4.0 2023-04-27 20:17:50,896 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9205, 4.1356, 1.9524, 4.2170, 2.7573, 4.1775, 2.1884, 3.0126], device='cuda:4'), covar=tensor([0.0039, 0.0084, 0.1589, 0.0024, 0.0811, 0.0195, 0.1405, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0076, 0.0111, 0.0174, 0.0075, 0.0157, 0.0137, 0.0181, 0.0156], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 20:18:20,856 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.564e+02 4.048e+02 4.879e+02 5.848e+02 1.298e+03, threshold=9.758e+02, percent-clipped=2.0 2023-04-27 20:18:32,650 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=19129.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 20:19:11,846 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19146.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:19:21,317 INFO [train.py:904] (4/8) Epoch 2, batch 9000, loss[loss=0.2454, simple_loss=0.3112, pruned_loss=0.08983, over 12100.00 frames. ], tot_loss[loss=0.2412, simple_loss=0.3217, pruned_loss=0.08037, over 3050405.42 frames. ], batch size: 247, lr: 2.46e-02, grad_scale: 2.0 2023-04-27 20:19:21,317 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 20:19:31,139 INFO [train.py:938] (4/8) Epoch 2, validation: loss=0.2044, simple_loss=0.3047, pruned_loss=0.05205, over 944034.00 frames. 2023-04-27 20:19:31,140 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-27 20:19:46,225 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0613, 2.6827, 2.4216, 3.4148, 3.3178, 3.3227, 1.8793, 2.6941], device='cuda:4'), covar=tensor([0.1165, 0.0344, 0.0969, 0.0073, 0.0206, 0.0318, 0.1067, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0125, 0.0167, 0.0069, 0.0111, 0.0120, 0.0155, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:20:00,924 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19164.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 20:20:52,915 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=19190.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 20:21:02,681 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=19194.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:21:13,856 INFO [train.py:904] (4/8) Epoch 2, batch 9050, loss[loss=0.2018, simple_loss=0.2892, pruned_loss=0.05725, over 16902.00 frames. ], tot_loss[loss=0.2429, simple_loss=0.3233, pruned_loss=0.08122, over 3057973.13 frames. ], batch size: 102, lr: 2.46e-02, grad_scale: 2.0 2023-04-27 20:21:37,537 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=19212.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 20:22:01,199 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.755e+02 4.212e+02 4.597e+02 5.989e+02 9.578e+02, threshold=9.193e+02, percent-clipped=0.0 2023-04-27 20:22:10,980 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9288, 3.8222, 3.3685, 1.6252, 2.7238, 2.2549, 3.2072, 3.7359], device='cuda:4'), covar=tensor([0.0304, 0.0413, 0.0443, 0.1839, 0.0757, 0.1008, 0.0903, 0.0511], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0101, 0.0155, 0.0151, 0.0141, 0.0134, 0.0144, 0.0102], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:22:58,685 INFO [train.py:904] (4/8) Epoch 2, batch 9100, loss[loss=0.2599, simple_loss=0.3454, pruned_loss=0.08725, over 16221.00 frames. ], tot_loss[loss=0.2435, simple_loss=0.3235, pruned_loss=0.08175, over 3071628.33 frames. ], batch size: 165, lr: 2.46e-02, grad_scale: 2.0 2023-04-27 20:24:22,919 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19286.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:24:55,898 INFO [train.py:904] (4/8) Epoch 2, batch 9150, loss[loss=0.2353, simple_loss=0.3158, pruned_loss=0.07739, over 16925.00 frames. ], tot_loss[loss=0.243, simple_loss=0.324, pruned_loss=0.08106, over 3077908.77 frames. ], batch size: 96, lr: 2.46e-02, grad_scale: 2.0 2023-04-27 20:25:49,334 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.714e+02 4.631e+02 5.471e+02 6.935e+02 2.154e+03, threshold=1.094e+03, percent-clipped=5.0 2023-04-27 20:25:59,007 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19330.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:26:40,135 INFO [train.py:904] (4/8) Epoch 2, batch 9200, loss[loss=0.2358, simple_loss=0.3205, pruned_loss=0.07556, over 15251.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.3187, pruned_loss=0.07961, over 3057534.67 frames. ], batch size: 191, lr: 2.45e-02, grad_scale: 4.0 2023-04-27 20:27:30,641 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=19378.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:28:03,621 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=19394.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:28:16,032 INFO [train.py:904] (4/8) Epoch 2, batch 9250, loss[loss=0.2189, simple_loss=0.2948, pruned_loss=0.07149, over 12209.00 frames. ], tot_loss[loss=0.2382, simple_loss=0.3177, pruned_loss=0.07931, over 3054455.29 frames. ], batch size: 248, lr: 2.45e-02, grad_scale: 4.0 2023-04-27 20:29:05,883 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.723e+02 4.011e+02 4.767e+02 6.782e+02 2.707e+03, threshold=9.534e+02, percent-clipped=7.0 2023-04-27 20:29:58,368 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-27 20:30:06,015 INFO [train.py:904] (4/8) Epoch 2, batch 9300, loss[loss=0.2122, simple_loss=0.2958, pruned_loss=0.06428, over 16184.00 frames. ], tot_loss[loss=0.2345, simple_loss=0.3145, pruned_loss=0.07728, over 3051183.95 frames. ], batch size: 165, lr: 2.45e-02, grad_scale: 4.0 2023-04-27 20:30:14,338 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=19455.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:31:20,776 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19485.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 20:31:28,790 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 20:31:35,260 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4513, 1.6569, 2.2252, 2.9122, 3.3064, 3.1550, 1.7285, 2.7602], device='cuda:4'), covar=tensor([0.0026, 0.0209, 0.0122, 0.0079, 0.0023, 0.0052, 0.0172, 0.0066], device='cuda:4'), in_proj_covar=tensor([0.0072, 0.0110, 0.0094, 0.0081, 0.0063, 0.0059, 0.0098, 0.0055], device='cuda:4'), out_proj_covar=tensor([1.2197e-04, 1.8672e-04, 1.6528e-04, 1.4218e-04, 1.0468e-04, 9.7837e-05, 1.6325e-04, 9.0346e-05], device='cuda:4') 2023-04-27 20:31:49,019 INFO [train.py:904] (4/8) Epoch 2, batch 9350, loss[loss=0.2505, simple_loss=0.3227, pruned_loss=0.08917, over 16867.00 frames. ], tot_loss[loss=0.2343, simple_loss=0.3148, pruned_loss=0.0769, over 3075549.19 frames. ], batch size: 116, lr: 2.44e-02, grad_scale: 4.0 2023-04-27 20:32:24,103 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3488, 1.2979, 1.5214, 2.1584, 2.2808, 2.2702, 1.1756, 2.0429], device='cuda:4'), covar=tensor([0.0047, 0.0225, 0.0145, 0.0099, 0.0043, 0.0063, 0.0224, 0.0066], device='cuda:4'), in_proj_covar=tensor([0.0073, 0.0111, 0.0095, 0.0081, 0.0064, 0.0060, 0.0099, 0.0056], device='cuda:4'), out_proj_covar=tensor([1.2433e-04, 1.8834e-04, 1.6753e-04, 1.4301e-04, 1.0578e-04, 9.9089e-05, 1.6507e-04, 9.0923e-05], device='cuda:4') 2023-04-27 20:32:37,319 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.597e+02 3.883e+02 4.622e+02 6.098e+02 2.142e+03, threshold=9.245e+02, percent-clipped=3.0 2023-04-27 20:33:23,672 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.65 vs. limit=2.0 2023-04-27 20:33:28,460 INFO [train.py:904] (4/8) Epoch 2, batch 9400, loss[loss=0.2177, simple_loss=0.2942, pruned_loss=0.07056, over 12685.00 frames. ], tot_loss[loss=0.2344, simple_loss=0.3151, pruned_loss=0.07684, over 3075201.23 frames. ], batch size: 248, lr: 2.44e-02, grad_scale: 4.0 2023-04-27 20:34:39,320 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19586.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:35:08,265 INFO [train.py:904] (4/8) Epoch 2, batch 9450, loss[loss=0.2265, simple_loss=0.3037, pruned_loss=0.07465, over 16716.00 frames. ], tot_loss[loss=0.2356, simple_loss=0.3166, pruned_loss=0.07732, over 3065939.61 frames. ], batch size: 62, lr: 2.44e-02, grad_scale: 4.0 2023-04-27 20:35:42,255 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=19618.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:35:45,912 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-27 20:35:56,417 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.451e+02 4.122e+02 5.164e+02 6.390e+02 1.240e+03, threshold=1.033e+03, percent-clipped=6.0 2023-04-27 20:36:16,143 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=19634.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:36:48,219 INFO [train.py:904] (4/8) Epoch 2, batch 9500, loss[loss=0.2273, simple_loss=0.3124, pruned_loss=0.07108, over 16915.00 frames. ], tot_loss[loss=0.2332, simple_loss=0.3145, pruned_loss=0.07595, over 3078125.93 frames. ], batch size: 109, lr: 2.43e-02, grad_scale: 4.0 2023-04-27 20:37:22,084 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3933, 1.4613, 1.8433, 2.2074, 2.3998, 2.3671, 1.6111, 2.3918], device='cuda:4'), covar=tensor([0.0041, 0.0213, 0.0119, 0.0103, 0.0042, 0.0065, 0.0162, 0.0037], device='cuda:4'), in_proj_covar=tensor([0.0072, 0.0110, 0.0096, 0.0083, 0.0064, 0.0060, 0.0098, 0.0055], device='cuda:4'), out_proj_covar=tensor([1.2303e-04, 1.8726e-04, 1.6937e-04, 1.4606e-04, 1.0522e-04, 9.8436e-05, 1.6233e-04, 8.8071e-05], device='cuda:4') 2023-04-27 20:37:46,271 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=19679.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:37:59,645 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-27 20:38:34,554 INFO [train.py:904] (4/8) Epoch 2, batch 9550, loss[loss=0.2615, simple_loss=0.327, pruned_loss=0.098, over 12138.00 frames. ], tot_loss[loss=0.2334, simple_loss=0.3148, pruned_loss=0.07605, over 3092030.01 frames. ], batch size: 246, lr: 2.43e-02, grad_scale: 4.0 2023-04-27 20:39:23,641 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.975e+02 3.908e+02 4.753e+02 5.956e+02 1.328e+03, threshold=9.507e+02, percent-clipped=2.0 2023-04-27 20:40:12,588 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19750.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:40:13,477 INFO [train.py:904] (4/8) Epoch 2, batch 9600, loss[loss=0.2508, simple_loss=0.3168, pruned_loss=0.09239, over 11950.00 frames. ], tot_loss[loss=0.2357, simple_loss=0.3165, pruned_loss=0.07747, over 3067800.29 frames. ], batch size: 248, lr: 2.43e-02, grad_scale: 8.0 2023-04-27 20:40:19,129 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3168, 2.9114, 2.6312, 2.3733, 2.1788, 1.9716, 2.8841, 3.1966], device='cuda:4'), covar=tensor([0.1323, 0.0471, 0.0879, 0.0673, 0.1231, 0.1256, 0.0283, 0.0192], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0221, 0.0238, 0.0182, 0.0205, 0.0181, 0.0194, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:40:44,202 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3090, 3.2673, 2.7987, 2.2784, 2.3635, 1.8634, 3.4012, 3.8356], device='cuda:4'), covar=tensor([0.1866, 0.0648, 0.1045, 0.0872, 0.1353, 0.1284, 0.0356, 0.0198], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0220, 0.0236, 0.0182, 0.0204, 0.0180, 0.0194, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:41:22,041 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19785.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 20:41:32,506 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6693, 3.5100, 3.2496, 1.8040, 2.6641, 2.1374, 3.0604, 3.4204], device='cuda:4'), covar=tensor([0.0294, 0.0375, 0.0398, 0.1512, 0.0691, 0.0926, 0.0765, 0.0473], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0100, 0.0152, 0.0148, 0.0139, 0.0134, 0.0143, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:42:00,140 INFO [train.py:904] (4/8) Epoch 2, batch 9650, loss[loss=0.2263, simple_loss=0.305, pruned_loss=0.07382, over 12228.00 frames. ], tot_loss[loss=0.2388, simple_loss=0.3196, pruned_loss=0.07895, over 3047539.51 frames. ], batch size: 247, lr: 2.43e-02, grad_scale: 8.0 2023-04-27 20:42:54,373 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.395e+02 4.451e+02 5.149e+02 6.499e+02 1.556e+03, threshold=1.030e+03, percent-clipped=6.0 2023-04-27 20:43:09,972 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=19833.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 20:43:48,656 INFO [train.py:904] (4/8) Epoch 2, batch 9700, loss[loss=0.2284, simple_loss=0.3074, pruned_loss=0.07468, over 12586.00 frames. ], tot_loss[loss=0.2383, simple_loss=0.319, pruned_loss=0.07886, over 3059823.86 frames. ], batch size: 248, lr: 2.42e-02, grad_scale: 8.0 2023-04-27 20:45:33,548 INFO [train.py:904] (4/8) Epoch 2, batch 9750, loss[loss=0.212, simple_loss=0.3058, pruned_loss=0.05905, over 16276.00 frames. ], tot_loss[loss=0.2365, simple_loss=0.3167, pruned_loss=0.07822, over 3060950.57 frames. ], batch size: 165, lr: 2.42e-02, grad_scale: 8.0 2023-04-27 20:45:43,292 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0861, 3.2145, 3.1798, 1.5236, 3.3493, 3.2971, 2.9446, 2.7025], device='cuda:4'), covar=tensor([0.0781, 0.0099, 0.0144, 0.1328, 0.0079, 0.0069, 0.0275, 0.0337], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0080, 0.0076, 0.0146, 0.0071, 0.0069, 0.0106, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 20:46:21,255 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.379e+02 3.927e+02 4.879e+02 5.766e+02 1.284e+03, threshold=9.758e+02, percent-clipped=1.0 2023-04-27 20:46:26,745 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3458, 3.4468, 3.9760, 3.9841, 3.9672, 3.5306, 3.6667, 3.7511], device='cuda:4'), covar=tensor([0.0261, 0.0323, 0.0270, 0.0364, 0.0319, 0.0298, 0.0631, 0.0278], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0140, 0.0156, 0.0160, 0.0186, 0.0166, 0.0234, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-04-27 20:47:15,366 INFO [train.py:904] (4/8) Epoch 2, batch 9800, loss[loss=0.2262, simple_loss=0.3155, pruned_loss=0.06845, over 17249.00 frames. ], tot_loss[loss=0.2355, simple_loss=0.3167, pruned_loss=0.07711, over 3064323.89 frames. ], batch size: 52, lr: 2.42e-02, grad_scale: 8.0 2023-04-27 20:48:00,423 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19974.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:49:04,578 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6783, 5.0111, 4.7636, 4.7887, 4.3400, 4.2725, 4.5062, 5.0206], device='cuda:4'), covar=tensor([0.0361, 0.0452, 0.0582, 0.0338, 0.0404, 0.0595, 0.0380, 0.0501], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0294, 0.0255, 0.0183, 0.0196, 0.0188, 0.0232, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:49:05,346 INFO [train.py:904] (4/8) Epoch 2, batch 9850, loss[loss=0.241, simple_loss=0.3237, pruned_loss=0.07909, over 16707.00 frames. ], tot_loss[loss=0.2356, simple_loss=0.3178, pruned_loss=0.07674, over 3063171.28 frames. ], batch size: 134, lr: 2.41e-02, grad_scale: 8.0 2023-04-27 20:49:46,072 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20020.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:49:55,600 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.468e+02 4.069e+02 4.622e+02 5.694e+02 1.429e+03, threshold=9.244e+02, percent-clipped=4.0 2023-04-27 20:49:56,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9657, 2.2978, 2.0864, 3.2337, 2.0022, 3.0136, 2.3219, 1.8893], device='cuda:4'), covar=tensor([0.0290, 0.0574, 0.0351, 0.0227, 0.1397, 0.0188, 0.0639, 0.1229], device='cuda:4'), in_proj_covar=tensor([0.0218, 0.0200, 0.0162, 0.0223, 0.0269, 0.0173, 0.0196, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:50:58,120 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20050.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:50:58,949 INFO [train.py:904] (4/8) Epoch 2, batch 9900, loss[loss=0.2074, simple_loss=0.3091, pruned_loss=0.05292, over 16864.00 frames. ], tot_loss[loss=0.236, simple_loss=0.3182, pruned_loss=0.07692, over 3047178.08 frames. ], batch size: 96, lr: 2.41e-02, grad_scale: 8.0 2023-04-27 20:51:01,940 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5665, 4.4045, 4.3513, 4.0453, 4.2960, 1.4507, 3.9988, 4.3416], device='cuda:4'), covar=tensor([0.0034, 0.0032, 0.0046, 0.0117, 0.0040, 0.1397, 0.0048, 0.0062], device='cuda:4'), in_proj_covar=tensor([0.0058, 0.0050, 0.0074, 0.0081, 0.0056, 0.0111, 0.0067, 0.0072], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:52:00,674 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-27 20:52:09,557 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0851, 3.1877, 3.4531, 3.4534, 3.4722, 3.1449, 3.2541, 3.3045], device='cuda:4'), covar=tensor([0.0242, 0.0320, 0.0370, 0.0390, 0.0381, 0.0372, 0.0583, 0.0298], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0140, 0.0156, 0.0158, 0.0185, 0.0162, 0.0230, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-04-27 20:52:12,575 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20081.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:52:31,241 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20089.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:52:51,005 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=20098.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:52:56,913 INFO [train.py:904] (4/8) Epoch 2, batch 9950, loss[loss=0.1916, simple_loss=0.293, pruned_loss=0.04514, over 16712.00 frames. ], tot_loss[loss=0.237, simple_loss=0.3199, pruned_loss=0.07699, over 3051438.62 frames. ], batch size: 76, lr: 2.41e-02, grad_scale: 8.0 2023-04-27 20:53:12,182 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6644, 4.5524, 4.5232, 3.5683, 4.2473, 1.6555, 4.1390, 4.4185], device='cuda:4'), covar=tensor([0.0104, 0.0080, 0.0072, 0.0320, 0.0095, 0.1902, 0.0087, 0.0144], device='cuda:4'), in_proj_covar=tensor([0.0059, 0.0050, 0.0074, 0.0081, 0.0056, 0.0111, 0.0067, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:53:23,637 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20112.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:53:55,568 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.288e+02 4.201e+02 4.972e+02 6.369e+02 1.989e+03, threshold=9.943e+02, percent-clipped=8.0 2023-04-27 20:54:55,860 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20150.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:54:56,741 INFO [train.py:904] (4/8) Epoch 2, batch 10000, loss[loss=0.2383, simple_loss=0.3108, pruned_loss=0.08287, over 12999.00 frames. ], tot_loss[loss=0.2348, simple_loss=0.3176, pruned_loss=0.07597, over 3056631.61 frames. ], batch size: 250, lr: 2.41e-02, grad_scale: 8.0 2023-04-27 20:55:18,149 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-27 20:55:41,240 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20173.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:56:37,810 INFO [train.py:904] (4/8) Epoch 2, batch 10050, loss[loss=0.2555, simple_loss=0.3438, pruned_loss=0.08359, over 16440.00 frames. ], tot_loss[loss=0.2344, simple_loss=0.3177, pruned_loss=0.07557, over 3074919.61 frames. ], batch size: 146, lr: 2.40e-02, grad_scale: 8.0 2023-04-27 20:56:42,238 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8364, 3.7207, 3.6920, 3.8004, 3.3380, 3.7861, 3.6672, 3.5217], device='cuda:4'), covar=tensor([0.0325, 0.0208, 0.0176, 0.0125, 0.0546, 0.0211, 0.0473, 0.0236], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0094, 0.0140, 0.0116, 0.0167, 0.0125, 0.0098, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:56:44,597 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20204.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:57:04,471 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1175, 1.6463, 1.4191, 1.3909, 1.8688, 1.6788, 1.8853, 1.9200], device='cuda:4'), covar=tensor([0.0020, 0.0109, 0.0123, 0.0124, 0.0075, 0.0111, 0.0034, 0.0065], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0104, 0.0102, 0.0103, 0.0095, 0.0103, 0.0056, 0.0073], device='cuda:4'), out_proj_covar=tensor([6.7163e-05, 1.5716e-04, 1.4904e-04, 1.5458e-04, 1.4605e-04, 1.5801e-04, 8.1323e-05, 1.1383e-04], device='cuda:4') 2023-04-27 20:57:24,815 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.701e+02 4.201e+02 4.991e+02 6.585e+02 1.392e+03, threshold=9.982e+02, percent-clipped=3.0 2023-04-27 20:58:10,189 INFO [train.py:904] (4/8) Epoch 2, batch 10100, loss[loss=0.2407, simple_loss=0.3054, pruned_loss=0.08802, over 12354.00 frames. ], tot_loss[loss=0.2357, simple_loss=0.3187, pruned_loss=0.07639, over 3081405.99 frames. ], batch size: 246, lr: 2.40e-02, grad_scale: 8.0 2023-04-27 20:58:38,055 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20265.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:58:58,839 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20274.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:59:54,874 INFO [train.py:904] (4/8) Epoch 3, batch 0, loss[loss=0.4321, simple_loss=0.4396, pruned_loss=0.2123, over 16245.00 frames. ], tot_loss[loss=0.4321, simple_loss=0.4396, pruned_loss=0.2123, over 16245.00 frames. ], batch size: 165, lr: 2.28e-02, grad_scale: 8.0 2023-04-27 20:59:54,874 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 21:00:02,299 INFO [train.py:938] (4/8) Epoch 3, validation: loss=0.2012, simple_loss=0.3019, pruned_loss=0.05024, over 944034.00 frames. 2023-04-27 21:00:02,300 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-27 21:00:17,839 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-27 21:00:31,871 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=20322.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:00:38,799 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.937e+02 4.805e+02 5.776e+02 7.144e+02 1.042e+03, threshold=1.155e+03, percent-clipped=3.0 2023-04-27 21:01:12,891 INFO [train.py:904] (4/8) Epoch 3, batch 50, loss[loss=0.2489, simple_loss=0.3257, pruned_loss=0.08602, over 16813.00 frames. ], tot_loss[loss=0.2998, simple_loss=0.3544, pruned_loss=0.1226, over 748327.57 frames. ], batch size: 42, lr: 2.28e-02, grad_scale: 2.0 2023-04-27 21:01:45,733 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20376.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:02:02,845 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20388.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:02:19,630 INFO [train.py:904] (4/8) Epoch 3, batch 100, loss[loss=0.2949, simple_loss=0.3404, pruned_loss=0.1247, over 16858.00 frames. ], tot_loss[loss=0.2803, simple_loss=0.3418, pruned_loss=0.1094, over 1321057.64 frames. ], batch size: 116, lr: 2.27e-02, grad_scale: 2.0 2023-04-27 21:02:43,324 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20418.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:02:54,676 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.729e+02 4.335e+02 5.443e+02 6.351e+02 1.481e+03, threshold=1.089e+03, percent-clipped=3.0 2023-04-27 21:03:18,755 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20445.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:03:24,368 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20449.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:03:27,024 INFO [train.py:904] (4/8) Epoch 3, batch 150, loss[loss=0.2296, simple_loss=0.3049, pruned_loss=0.07717, over 17106.00 frames. ], tot_loss[loss=0.2716, simple_loss=0.3357, pruned_loss=0.1038, over 1776768.77 frames. ], batch size: 47, lr: 2.27e-02, grad_scale: 2.0 2023-04-27 21:03:32,734 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1960, 4.3122, 4.7477, 4.6723, 4.7443, 4.2717, 4.0459, 4.2871], device='cuda:4'), covar=tensor([0.0420, 0.0361, 0.0399, 0.0518, 0.0466, 0.0439, 0.0935, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0158, 0.0178, 0.0177, 0.0207, 0.0183, 0.0268, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-27 21:03:49,781 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20468.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:03:55,867 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-27 21:04:05,874 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20479.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:04:35,430 INFO [train.py:904] (4/8) Epoch 3, batch 200, loss[loss=0.2565, simple_loss=0.3101, pruned_loss=0.1015, over 16773.00 frames. ], tot_loss[loss=0.2678, simple_loss=0.3328, pruned_loss=0.1014, over 2119538.85 frames. ], batch size: 102, lr: 2.27e-02, grad_scale: 2.0 2023-04-27 21:05:09,760 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.165e+02 3.891e+02 4.616e+02 5.990e+02 1.220e+03, threshold=9.233e+02, percent-clipped=2.0 2023-04-27 21:05:43,634 INFO [train.py:904] (4/8) Epoch 3, batch 250, loss[loss=0.2697, simple_loss=0.3284, pruned_loss=0.1055, over 15371.00 frames. ], tot_loss[loss=0.2658, simple_loss=0.3296, pruned_loss=0.101, over 2382444.07 frames. ], batch size: 190, lr: 2.27e-02, grad_scale: 2.0 2023-04-27 21:05:58,416 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20560.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:06:00,890 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1562, 3.9292, 1.9337, 4.0553, 2.6464, 4.0936, 1.9934, 3.0345], device='cuda:4'), covar=tensor([0.0040, 0.0189, 0.1287, 0.0038, 0.0646, 0.0264, 0.1148, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0078, 0.0123, 0.0172, 0.0077, 0.0159, 0.0148, 0.0179, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-27 21:06:21,006 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3628, 3.5609, 3.5601, 1.6552, 3.6758, 3.6515, 3.0777, 2.7538], device='cuda:4'), covar=tensor([0.0694, 0.0091, 0.0133, 0.1349, 0.0067, 0.0067, 0.0336, 0.0426], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0081, 0.0078, 0.0144, 0.0073, 0.0069, 0.0109, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:06:35,590 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20588.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:06:54,934 INFO [train.py:904] (4/8) Epoch 3, batch 300, loss[loss=0.2376, simple_loss=0.3106, pruned_loss=0.08232, over 16748.00 frames. ], tot_loss[loss=0.2622, simple_loss=0.3269, pruned_loss=0.09879, over 2586465.56 frames. ], batch size: 57, lr: 2.26e-02, grad_scale: 2.0 2023-04-27 21:06:59,070 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-27 21:07:29,011 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.643e+02 3.919e+02 4.746e+02 5.726e+02 1.058e+03, threshold=9.491e+02, percent-clipped=1.0 2023-04-27 21:07:59,147 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20649.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:08:01,102 INFO [train.py:904] (4/8) Epoch 3, batch 350, loss[loss=0.2424, simple_loss=0.3189, pruned_loss=0.08292, over 17262.00 frames. ], tot_loss[loss=0.2573, simple_loss=0.3225, pruned_loss=0.09604, over 2740212.59 frames. ], batch size: 52, lr: 2.26e-02, grad_scale: 2.0 2023-04-27 21:08:36,947 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20676.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:08:47,650 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0588, 3.6675, 3.4874, 1.4438, 3.5858, 3.5616, 3.1089, 2.6375], device='cuda:4'), covar=tensor([0.0970, 0.0077, 0.0169, 0.1369, 0.0117, 0.0097, 0.0304, 0.0419], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0083, 0.0080, 0.0147, 0.0075, 0.0072, 0.0110, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:09:09,381 INFO [train.py:904] (4/8) Epoch 3, batch 400, loss[loss=0.2417, simple_loss=0.3184, pruned_loss=0.08252, over 17070.00 frames. ], tot_loss[loss=0.2557, simple_loss=0.3211, pruned_loss=0.09518, over 2871171.88 frames. ], batch size: 55, lr: 2.26e-02, grad_scale: 4.0 2023-04-27 21:09:41,231 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=20724.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:09:45,245 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.989e+02 4.408e+02 5.242e+02 6.457e+02 1.269e+03, threshold=1.048e+03, percent-clipped=5.0 2023-04-27 21:10:08,714 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20744.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:10:10,669 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20745.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:10:17,859 INFO [train.py:904] (4/8) Epoch 3, batch 450, loss[loss=0.3039, simple_loss=0.3578, pruned_loss=0.125, over 16490.00 frames. ], tot_loss[loss=0.2545, simple_loss=0.3198, pruned_loss=0.09458, over 2981194.28 frames. ], batch size: 146, lr: 2.25e-02, grad_scale: 4.0 2023-04-27 21:10:41,875 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20768.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:10:49,452 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20774.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:11:14,832 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=20793.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:11:24,055 INFO [train.py:904] (4/8) Epoch 3, batch 500, loss[loss=0.2251, simple_loss=0.2941, pruned_loss=0.07807, over 16519.00 frames. ], tot_loss[loss=0.2515, simple_loss=0.3176, pruned_loss=0.09275, over 3057521.88 frames. ], batch size: 75, lr: 2.25e-02, grad_scale: 4.0 2023-04-27 21:11:46,146 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=20816.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:12:00,585 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.032e+02 4.223e+02 4.982e+02 5.898e+02 1.253e+03, threshold=9.964e+02, percent-clipped=1.0 2023-04-27 21:12:34,370 INFO [train.py:904] (4/8) Epoch 3, batch 550, loss[loss=0.2837, simple_loss=0.3278, pruned_loss=0.1198, over 16900.00 frames. ], tot_loss[loss=0.2503, simple_loss=0.317, pruned_loss=0.09174, over 3121219.59 frames. ], batch size: 109, lr: 2.25e-02, grad_scale: 4.0 2023-04-27 21:12:45,417 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20860.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:13:05,204 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6093, 4.6715, 4.5024, 2.0151, 4.6700, 4.8014, 3.4670, 4.0239], device='cuda:4'), covar=tensor([0.0761, 0.0077, 0.0197, 0.1176, 0.0036, 0.0028, 0.0301, 0.0219], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0081, 0.0080, 0.0147, 0.0074, 0.0072, 0.0113, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:13:17,518 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-27 21:13:40,832 INFO [train.py:904] (4/8) Epoch 3, batch 600, loss[loss=0.307, simple_loss=0.3334, pruned_loss=0.1403, over 16888.00 frames. ], tot_loss[loss=0.2509, simple_loss=0.3165, pruned_loss=0.09271, over 3168538.20 frames. ], batch size: 116, lr: 2.25e-02, grad_scale: 4.0 2023-04-27 21:13:50,522 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=20908.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:13:58,170 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7505, 4.8494, 4.7312, 2.0322, 4.8225, 4.9587, 3.6031, 4.2168], device='cuda:4'), covar=tensor([0.0693, 0.0073, 0.0187, 0.1225, 0.0055, 0.0029, 0.0311, 0.0220], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0081, 0.0079, 0.0147, 0.0074, 0.0072, 0.0113, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:14:15,688 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.677e+02 4.207e+02 5.314e+02 6.523e+02 1.813e+03, threshold=1.063e+03, percent-clipped=4.0 2023-04-27 21:14:38,950 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20944.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:14:48,030 INFO [train.py:904] (4/8) Epoch 3, batch 650, loss[loss=0.2232, simple_loss=0.3081, pruned_loss=0.06912, over 17283.00 frames. ], tot_loss[loss=0.248, simple_loss=0.3145, pruned_loss=0.09073, over 3206142.80 frames. ], batch size: 52, lr: 2.24e-02, grad_scale: 4.0 2023-04-27 21:15:53,449 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.28 vs. limit=5.0 2023-04-27 21:15:57,310 INFO [train.py:904] (4/8) Epoch 3, batch 700, loss[loss=0.2602, simple_loss=0.3339, pruned_loss=0.09323, over 17032.00 frames. ], tot_loss[loss=0.2484, simple_loss=0.3147, pruned_loss=0.09105, over 3232521.02 frames. ], batch size: 55, lr: 2.24e-02, grad_scale: 4.0 2023-04-27 21:15:58,092 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.88 vs. limit=5.0 2023-04-27 21:16:02,596 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-27 21:16:12,641 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9669, 3.7303, 3.8775, 4.2054, 4.2549, 3.8273, 4.1721, 4.1937], device='cuda:4'), covar=tensor([0.0535, 0.0636, 0.1134, 0.0414, 0.0395, 0.0924, 0.0440, 0.0350], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0338, 0.0456, 0.0344, 0.0262, 0.0239, 0.0256, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:16:20,800 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4450, 4.1306, 4.2732, 4.6521, 4.7301, 4.1770, 4.7054, 4.6535], device='cuda:4'), covar=tensor([0.0458, 0.0614, 0.1058, 0.0348, 0.0320, 0.0599, 0.0296, 0.0310], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0338, 0.0457, 0.0344, 0.0263, 0.0239, 0.0256, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:16:30,865 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.136e+02 3.749e+02 4.303e+02 5.331e+02 1.033e+03, threshold=8.606e+02, percent-clipped=0.0 2023-04-27 21:16:38,519 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1939, 5.5261, 5.1735, 5.2898, 4.7601, 4.5616, 5.0132, 5.5772], device='cuda:4'), covar=tensor([0.0410, 0.0554, 0.0776, 0.0368, 0.0518, 0.0524, 0.0410, 0.0489], device='cuda:4'), in_proj_covar=tensor([0.0259, 0.0373, 0.0325, 0.0227, 0.0245, 0.0225, 0.0291, 0.0247], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:16:40,847 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1138, 3.3255, 3.2777, 1.5136, 3.3198, 3.3549, 2.8390, 2.8291], device='cuda:4'), covar=tensor([0.0692, 0.0081, 0.0136, 0.1304, 0.0077, 0.0069, 0.0309, 0.0319], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0080, 0.0078, 0.0147, 0.0074, 0.0072, 0.0112, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:16:55,503 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21044.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:17:03,823 INFO [train.py:904] (4/8) Epoch 3, batch 750, loss[loss=0.1875, simple_loss=0.2761, pruned_loss=0.04947, over 17245.00 frames. ], tot_loss[loss=0.2491, simple_loss=0.3154, pruned_loss=0.09139, over 3255508.98 frames. ], batch size: 44, lr: 2.24e-02, grad_scale: 4.0 2023-04-27 21:17:06,028 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21052.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:17:16,738 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-27 21:17:35,030 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21074.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:17:55,688 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5553, 4.4026, 4.3170, 4.4306, 4.0120, 4.3708, 4.3038, 4.1445], device='cuda:4'), covar=tensor([0.0258, 0.0173, 0.0163, 0.0106, 0.0617, 0.0164, 0.0310, 0.0250], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0119, 0.0182, 0.0148, 0.0220, 0.0161, 0.0131, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:17:57,838 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=21092.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:18:10,817 INFO [train.py:904] (4/8) Epoch 3, batch 800, loss[loss=0.2397, simple_loss=0.307, pruned_loss=0.08621, over 15857.00 frames. ], tot_loss[loss=0.249, simple_loss=0.3157, pruned_loss=0.09114, over 3274366.77 frames. ], batch size: 35, lr: 2.24e-02, grad_scale: 8.0 2023-04-27 21:18:27,398 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21113.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:18:34,912 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21118.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:18:40,095 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=21122.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:18:47,522 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.745e+02 4.034e+02 4.669e+02 5.861e+02 1.175e+03, threshold=9.338e+02, percent-clipped=5.0 2023-04-27 21:19:08,179 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5473, 5.9119, 5.5195, 5.7234, 5.1661, 5.0260, 5.3852, 5.9880], device='cuda:4'), covar=tensor([0.0432, 0.0575, 0.0952, 0.0358, 0.0606, 0.0439, 0.0523, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0251, 0.0373, 0.0318, 0.0222, 0.0241, 0.0220, 0.0283, 0.0244], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:19:20,165 INFO [train.py:904] (4/8) Epoch 3, batch 850, loss[loss=0.2256, simple_loss=0.2977, pruned_loss=0.07674, over 17173.00 frames. ], tot_loss[loss=0.2468, simple_loss=0.314, pruned_loss=0.0898, over 3286113.00 frames. ], batch size: 46, lr: 2.23e-02, grad_scale: 8.0 2023-04-27 21:19:58,146 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21179.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:20:27,600 INFO [train.py:904] (4/8) Epoch 3, batch 900, loss[loss=0.1728, simple_loss=0.2501, pruned_loss=0.0478, over 16827.00 frames. ], tot_loss[loss=0.2431, simple_loss=0.3119, pruned_loss=0.0872, over 3302087.10 frames. ], batch size: 39, lr: 2.23e-02, grad_scale: 8.0 2023-04-27 21:21:03,013 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.660e+02 3.821e+02 4.761e+02 5.619e+02 1.173e+03, threshold=9.521e+02, percent-clipped=3.0 2023-04-27 21:21:03,615 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9043, 4.3041, 3.2262, 2.9447, 3.0618, 2.2621, 4.5205, 4.8371], device='cuda:4'), covar=tensor([0.2023, 0.0621, 0.1200, 0.0846, 0.2464, 0.1365, 0.0311, 0.0241], device='cuda:4'), in_proj_covar=tensor([0.0261, 0.0237, 0.0251, 0.0196, 0.0265, 0.0188, 0.0211, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:21:26,684 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21244.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:21:30,210 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-27 21:21:35,849 INFO [train.py:904] (4/8) Epoch 3, batch 950, loss[loss=0.2257, simple_loss=0.3052, pruned_loss=0.0731, over 17236.00 frames. ], tot_loss[loss=0.2428, simple_loss=0.3118, pruned_loss=0.08688, over 3312867.29 frames. ], batch size: 44, lr: 2.23e-02, grad_scale: 8.0 2023-04-27 21:22:34,056 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=21292.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:22:45,620 INFO [train.py:904] (4/8) Epoch 3, batch 1000, loss[loss=0.2181, simple_loss=0.294, pruned_loss=0.07113, over 17214.00 frames. ], tot_loss[loss=0.2423, simple_loss=0.3104, pruned_loss=0.08712, over 3320143.48 frames. ], batch size: 45, lr: 2.23e-02, grad_scale: 8.0 2023-04-27 21:23:01,290 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21312.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:23:01,635 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-27 21:23:21,765 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.770e+02 3.937e+02 4.742e+02 6.159e+02 9.838e+02, threshold=9.485e+02, percent-clipped=2.0 2023-04-27 21:23:24,013 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21328.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:23:55,373 INFO [train.py:904] (4/8) Epoch 3, batch 1050, loss[loss=0.2283, simple_loss=0.2973, pruned_loss=0.07964, over 16851.00 frames. ], tot_loss[loss=0.2414, simple_loss=0.3095, pruned_loss=0.08664, over 3328016.89 frames. ], batch size: 42, lr: 2.22e-02, grad_scale: 8.0 2023-04-27 21:24:25,334 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21373.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:24:46,811 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21389.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:25:02,814 INFO [train.py:904] (4/8) Epoch 3, batch 1100, loss[loss=0.2598, simple_loss=0.314, pruned_loss=0.1028, over 12181.00 frames. ], tot_loss[loss=0.2412, simple_loss=0.3087, pruned_loss=0.08688, over 3313762.61 frames. ], batch size: 246, lr: 2.22e-02, grad_scale: 4.0 2023-04-27 21:25:12,308 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21408.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:25:38,414 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.639e+02 4.233e+02 5.080e+02 6.558e+02 1.311e+03, threshold=1.016e+03, percent-clipped=3.0 2023-04-27 21:26:10,401 INFO [train.py:904] (4/8) Epoch 3, batch 1150, loss[loss=0.2347, simple_loss=0.3118, pruned_loss=0.07873, over 16733.00 frames. ], tot_loss[loss=0.2386, simple_loss=0.3073, pruned_loss=0.08495, over 3323115.52 frames. ], batch size: 62, lr: 2.22e-02, grad_scale: 4.0 2023-04-27 21:26:35,942 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7267, 2.2720, 2.0227, 2.9911, 1.9347, 2.8903, 2.1797, 1.9163], device='cuda:4'), covar=tensor([0.0339, 0.0781, 0.0476, 0.0325, 0.1579, 0.0285, 0.0965, 0.1621], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0216, 0.0181, 0.0245, 0.0285, 0.0187, 0.0207, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:26:42,689 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21474.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:26:47,007 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3858, 4.2186, 3.6182, 1.9955, 2.9983, 2.3857, 3.6765, 4.1363], device='cuda:4'), covar=tensor([0.0241, 0.0412, 0.0476, 0.1487, 0.0678, 0.0987, 0.0619, 0.0542], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0116, 0.0153, 0.0144, 0.0139, 0.0133, 0.0147, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 21:27:19,790 INFO [train.py:904] (4/8) Epoch 3, batch 1200, loss[loss=0.2675, simple_loss=0.3246, pruned_loss=0.1052, over 16747.00 frames. ], tot_loss[loss=0.2372, simple_loss=0.3065, pruned_loss=0.08394, over 3326139.81 frames. ], batch size: 124, lr: 2.22e-02, grad_scale: 8.0 2023-04-27 21:27:29,101 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1530, 1.4170, 1.8807, 2.2287, 2.2507, 2.4213, 1.3356, 2.2985], device='cuda:4'), covar=tensor([0.0061, 0.0208, 0.0093, 0.0081, 0.0052, 0.0047, 0.0172, 0.0035], device='cuda:4'), in_proj_covar=tensor([0.0084, 0.0120, 0.0105, 0.0100, 0.0078, 0.0067, 0.0107, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:27:56,781 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.205e+02 3.899e+02 4.622e+02 5.754e+02 9.760e+02, threshold=9.243e+02, percent-clipped=0.0 2023-04-27 21:28:19,418 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6412, 2.7652, 1.7124, 2.7717, 1.9969, 2.7852, 1.9484, 2.4160], device='cuda:4'), covar=tensor([0.0099, 0.0217, 0.1121, 0.0077, 0.0694, 0.0473, 0.0930, 0.0505], device='cuda:4'), in_proj_covar=tensor([0.0087, 0.0131, 0.0170, 0.0083, 0.0158, 0.0161, 0.0180, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-27 21:28:27,749 INFO [train.py:904] (4/8) Epoch 3, batch 1250, loss[loss=0.2389, simple_loss=0.306, pruned_loss=0.08592, over 16610.00 frames. ], tot_loss[loss=0.2379, simple_loss=0.3071, pruned_loss=0.08439, over 3326955.03 frames. ], batch size: 62, lr: 2.21e-02, grad_scale: 8.0 2023-04-27 21:29:30,655 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21595.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:29:39,771 INFO [train.py:904] (4/8) Epoch 3, batch 1300, loss[loss=0.27, simple_loss=0.3362, pruned_loss=0.1019, over 16783.00 frames. ], tot_loss[loss=0.2387, simple_loss=0.3075, pruned_loss=0.08498, over 3326402.23 frames. ], batch size: 62, lr: 2.21e-02, grad_scale: 8.0 2023-04-27 21:29:56,019 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21613.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:30:17,377 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.341e+02 3.774e+02 4.586e+02 5.366e+02 8.286e+02, threshold=9.173e+02, percent-clipped=0.0 2023-04-27 21:30:49,953 INFO [train.py:904] (4/8) Epoch 3, batch 1350, loss[loss=0.2117, simple_loss=0.2877, pruned_loss=0.06782, over 16980.00 frames. ], tot_loss[loss=0.2377, simple_loss=0.3068, pruned_loss=0.08436, over 3323341.56 frames. ], batch size: 41, lr: 2.21e-02, grad_scale: 8.0 2023-04-27 21:30:56,228 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21656.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:31:13,016 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21668.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:31:20,953 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21674.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:31:28,699 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.65 vs. limit=2.0 2023-04-27 21:31:35,253 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21684.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:31:58,964 INFO [train.py:904] (4/8) Epoch 3, batch 1400, loss[loss=0.2068, simple_loss=0.2956, pruned_loss=0.059, over 17024.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.307, pruned_loss=0.08537, over 3322647.35 frames. ], batch size: 50, lr: 2.21e-02, grad_scale: 8.0 2023-04-27 21:32:09,165 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21708.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:32:27,537 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.65 vs. limit=2.0 2023-04-27 21:32:35,074 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.932e+02 4.012e+02 4.714e+02 6.047e+02 1.220e+03, threshold=9.428e+02, percent-clipped=2.0 2023-04-27 21:32:49,202 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2341, 5.0824, 5.0037, 5.0224, 4.4212, 4.9653, 5.0664, 4.6126], device='cuda:4'), covar=tensor([0.0312, 0.0131, 0.0169, 0.0126, 0.0940, 0.0221, 0.0173, 0.0299], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0128, 0.0195, 0.0159, 0.0230, 0.0171, 0.0138, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:33:07,161 INFO [train.py:904] (4/8) Epoch 3, batch 1450, loss[loss=0.1995, simple_loss=0.2791, pruned_loss=0.05996, over 17219.00 frames. ], tot_loss[loss=0.238, simple_loss=0.3068, pruned_loss=0.08465, over 3327094.37 frames. ], batch size: 44, lr: 2.21e-02, grad_scale: 8.0 2023-04-27 21:33:13,877 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=21756.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:33:20,792 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2617, 1.7259, 2.3612, 3.1977, 3.0663, 3.5364, 2.0604, 3.0833], device='cuda:4'), covar=tensor([0.0049, 0.0231, 0.0138, 0.0077, 0.0049, 0.0050, 0.0159, 0.0067], device='cuda:4'), in_proj_covar=tensor([0.0084, 0.0119, 0.0105, 0.0099, 0.0077, 0.0066, 0.0106, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:33:24,006 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21763.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:33:39,361 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21774.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:34:14,645 INFO [train.py:904] (4/8) Epoch 3, batch 1500, loss[loss=0.2344, simple_loss=0.3059, pruned_loss=0.08145, over 16532.00 frames. ], tot_loss[loss=0.2365, simple_loss=0.3055, pruned_loss=0.08377, over 3324103.76 frames. ], batch size: 68, lr: 2.20e-02, grad_scale: 8.0 2023-04-27 21:34:43,974 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=21822.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:34:47,235 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21824.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:34:47,542 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.66 vs. limit=2.0 2023-04-27 21:34:52,714 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.241e+02 3.626e+02 4.234e+02 5.507e+02 9.675e+02, threshold=8.468e+02, percent-clipped=1.0 2023-04-27 21:34:54,267 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 21:35:23,729 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21850.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:35:24,397 INFO [train.py:904] (4/8) Epoch 3, batch 1550, loss[loss=0.2553, simple_loss=0.337, pruned_loss=0.08678, over 17083.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.3072, pruned_loss=0.08528, over 3316337.08 frames. ], batch size: 50, lr: 2.20e-02, grad_scale: 8.0 2023-04-27 21:35:41,982 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0065, 1.7327, 2.1277, 3.0195, 3.2383, 3.4552, 1.5225, 3.0701], device='cuda:4'), covar=tensor([0.0057, 0.0201, 0.0129, 0.0079, 0.0035, 0.0049, 0.0179, 0.0058], device='cuda:4'), in_proj_covar=tensor([0.0085, 0.0121, 0.0104, 0.0100, 0.0078, 0.0068, 0.0106, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:36:31,484 INFO [train.py:904] (4/8) Epoch 3, batch 1600, loss[loss=0.2797, simple_loss=0.3536, pruned_loss=0.1029, over 16765.00 frames. ], tot_loss[loss=0.2411, simple_loss=0.3094, pruned_loss=0.08641, over 3311999.26 frames. ], batch size: 57, lr: 2.20e-02, grad_scale: 8.0 2023-04-27 21:36:46,718 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21911.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:37:02,840 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7358, 3.7559, 1.9021, 3.8372, 2.4699, 3.7615, 1.9990, 2.8177], device='cuda:4'), covar=tensor([0.0052, 0.0183, 0.1294, 0.0052, 0.0709, 0.0311, 0.1075, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0087, 0.0136, 0.0171, 0.0084, 0.0157, 0.0165, 0.0179, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-27 21:37:07,919 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.625e+02 3.946e+02 4.752e+02 6.203e+02 1.123e+03, threshold=9.504e+02, percent-clipped=2.0 2023-04-27 21:37:38,900 INFO [train.py:904] (4/8) Epoch 3, batch 1650, loss[loss=0.2314, simple_loss=0.3242, pruned_loss=0.06924, over 17058.00 frames. ], tot_loss[loss=0.2436, simple_loss=0.311, pruned_loss=0.08813, over 3315744.10 frames. ], batch size: 50, lr: 2.20e-02, grad_scale: 8.0 2023-04-27 21:37:39,884 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21951.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:38:01,237 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21968.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:38:03,296 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21969.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:38:04,696 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21970.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:38:24,229 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21984.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:38:50,890 INFO [train.py:904] (4/8) Epoch 3, batch 1700, loss[loss=0.2617, simple_loss=0.3397, pruned_loss=0.09184, over 16692.00 frames. ], tot_loss[loss=0.2436, simple_loss=0.3123, pruned_loss=0.08748, over 3324388.27 frames. ], batch size: 62, lr: 2.19e-02, grad_scale: 4.0 2023-04-27 21:39:05,877 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 21:39:08,812 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8311, 3.8813, 2.7811, 5.0754, 4.9189, 4.3146, 1.6776, 3.2697], device='cuda:4'), covar=tensor([0.1310, 0.0363, 0.1088, 0.0062, 0.0201, 0.0297, 0.1271, 0.0634], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0130, 0.0162, 0.0073, 0.0142, 0.0135, 0.0154, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:39:11,660 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22016.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:39:28,009 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.555e+02 4.242e+02 5.121e+02 6.013e+02 1.262e+03, threshold=1.024e+03, percent-clipped=2.0 2023-04-27 21:39:30,829 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22031.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:39:33,270 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22032.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:39:57,285 INFO [train.py:904] (4/8) Epoch 3, batch 1750, loss[loss=0.2193, simple_loss=0.3027, pruned_loss=0.06792, over 17144.00 frames. ], tot_loss[loss=0.2445, simple_loss=0.3137, pruned_loss=0.08763, over 3324121.99 frames. ], batch size: 49, lr: 2.19e-02, grad_scale: 4.0 2023-04-27 21:41:06,920 INFO [train.py:904] (4/8) Epoch 3, batch 1800, loss[loss=0.2961, simple_loss=0.3534, pruned_loss=0.1194, over 16885.00 frames. ], tot_loss[loss=0.2459, simple_loss=0.3159, pruned_loss=0.08802, over 3323876.74 frames. ], batch size: 116, lr: 2.19e-02, grad_scale: 4.0 2023-04-27 21:41:30,482 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22119.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:41:44,204 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.299e+02 3.837e+02 4.558e+02 5.944e+02 1.243e+03, threshold=9.115e+02, percent-clipped=4.0 2023-04-27 21:41:45,847 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22130.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:42:00,813 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9696, 5.5462, 5.5690, 5.3629, 5.4156, 6.0055, 5.6842, 5.3688], device='cuda:4'), covar=tensor([0.0551, 0.1344, 0.1004, 0.1258, 0.1963, 0.0735, 0.0801, 0.1734], device='cuda:4'), in_proj_covar=tensor([0.0229, 0.0337, 0.0303, 0.0285, 0.0371, 0.0323, 0.0256, 0.0384], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:42:14,095 INFO [train.py:904] (4/8) Epoch 3, batch 1850, loss[loss=0.2409, simple_loss=0.322, pruned_loss=0.07992, over 16642.00 frames. ], tot_loss[loss=0.2455, simple_loss=0.3156, pruned_loss=0.08767, over 3326983.88 frames. ], batch size: 62, lr: 2.19e-02, grad_scale: 4.0 2023-04-27 21:42:47,495 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22176.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:43:09,370 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22191.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:43:22,338 INFO [train.py:904] (4/8) Epoch 3, batch 1900, loss[loss=0.2218, simple_loss=0.2909, pruned_loss=0.0764, over 16890.00 frames. ], tot_loss[loss=0.2445, simple_loss=0.3145, pruned_loss=0.08725, over 3331514.81 frames. ], batch size: 90, lr: 2.18e-02, grad_scale: 4.0 2023-04-27 21:43:31,012 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22206.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:43:42,639 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22215.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:44:02,290 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.402e+02 3.901e+02 4.998e+02 5.919e+02 1.840e+03, threshold=9.997e+02, percent-clipped=6.0 2023-04-27 21:44:09,795 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9948, 4.6844, 4.8994, 5.2902, 5.3440, 4.5573, 5.3866, 5.3160], device='cuda:4'), covar=tensor([0.0547, 0.0613, 0.1024, 0.0334, 0.0358, 0.0535, 0.0249, 0.0249], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0363, 0.0493, 0.0373, 0.0280, 0.0262, 0.0281, 0.0296], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:44:12,190 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22237.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:44:12,204 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22237.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:44:32,423 INFO [train.py:904] (4/8) Epoch 3, batch 1950, loss[loss=0.2475, simple_loss=0.334, pruned_loss=0.08055, over 17116.00 frames. ], tot_loss[loss=0.2433, simple_loss=0.3141, pruned_loss=0.08624, over 3327514.88 frames. ], batch size: 47, lr: 2.18e-02, grad_scale: 4.0 2023-04-27 21:44:32,823 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22251.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:44:55,419 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-27 21:44:58,150 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22269.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:45:06,184 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22276.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:45:35,314 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22298.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:45:36,245 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22299.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:45:38,311 INFO [train.py:904] (4/8) Epoch 3, batch 2000, loss[loss=0.2231, simple_loss=0.3015, pruned_loss=0.07233, over 17177.00 frames. ], tot_loss[loss=0.2431, simple_loss=0.3141, pruned_loss=0.08603, over 3323036.74 frames. ], batch size: 46, lr: 2.18e-02, grad_scale: 8.0 2023-04-27 21:46:01,054 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22317.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:46:13,065 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22326.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:46:17,541 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.887e+02 3.938e+02 4.847e+02 6.098e+02 1.059e+03, threshold=9.693e+02, percent-clipped=1.0 2023-04-27 21:46:34,379 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8985, 2.7837, 2.5534, 4.3225, 2.0010, 4.2594, 2.3451, 2.6132], device='cuda:4'), covar=tensor([0.0277, 0.0595, 0.0394, 0.0186, 0.1635, 0.0174, 0.0787, 0.1235], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0216, 0.0184, 0.0243, 0.0285, 0.0188, 0.0206, 0.0277], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:46:46,960 INFO [train.py:904] (4/8) Epoch 3, batch 2050, loss[loss=0.2361, simple_loss=0.3206, pruned_loss=0.07578, over 17047.00 frames. ], tot_loss[loss=0.2417, simple_loss=0.3129, pruned_loss=0.08527, over 3328734.41 frames. ], batch size: 50, lr: 2.18e-02, grad_scale: 8.0 2023-04-27 21:47:31,199 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6238, 1.5990, 2.0769, 2.4555, 2.5114, 2.5422, 1.5410, 2.6232], device='cuda:4'), covar=tensor([0.0049, 0.0193, 0.0125, 0.0096, 0.0055, 0.0073, 0.0181, 0.0030], device='cuda:4'), in_proj_covar=tensor([0.0086, 0.0124, 0.0109, 0.0102, 0.0084, 0.0070, 0.0110, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:47:33,541 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22385.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:47:55,914 INFO [train.py:904] (4/8) Epoch 3, batch 2100, loss[loss=0.2961, simple_loss=0.3504, pruned_loss=0.1209, over 16847.00 frames. ], tot_loss[loss=0.2455, simple_loss=0.3151, pruned_loss=0.08797, over 3326731.46 frames. ], batch size: 116, lr: 2.17e-02, grad_scale: 8.0 2023-04-27 21:48:20,602 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22419.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:48:20,801 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5135, 4.2739, 3.8173, 1.7116, 2.6701, 2.0290, 3.6859, 4.3232], device='cuda:4'), covar=tensor([0.0320, 0.0647, 0.0523, 0.1871, 0.0947, 0.1228, 0.0865, 0.0831], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0120, 0.0152, 0.0143, 0.0135, 0.0127, 0.0146, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 21:48:31,785 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.621e+02 3.997e+02 4.789e+02 6.119e+02 1.821e+03, threshold=9.579e+02, percent-clipped=3.0 2023-04-27 21:48:55,080 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22446.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:49:00,598 INFO [train.py:904] (4/8) Epoch 3, batch 2150, loss[loss=0.2286, simple_loss=0.3123, pruned_loss=0.07242, over 17121.00 frames. ], tot_loss[loss=0.2463, simple_loss=0.3158, pruned_loss=0.08844, over 3315226.91 frames. ], batch size: 47, lr: 2.17e-02, grad_scale: 8.0 2023-04-27 21:49:10,550 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9844, 4.0030, 1.6275, 3.9896, 2.5610, 3.9367, 1.8270, 2.9540], device='cuda:4'), covar=tensor([0.0057, 0.0175, 0.1655, 0.0049, 0.0699, 0.0320, 0.1482, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0089, 0.0139, 0.0175, 0.0084, 0.0162, 0.0172, 0.0184, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-27 21:49:22,488 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22467.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:49:47,664 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22486.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:50:08,035 INFO [train.py:904] (4/8) Epoch 3, batch 2200, loss[loss=0.2658, simple_loss=0.3257, pruned_loss=0.103, over 16823.00 frames. ], tot_loss[loss=0.247, simple_loss=0.3167, pruned_loss=0.08863, over 3317735.27 frames. ], batch size: 124, lr: 2.17e-02, grad_scale: 4.0 2023-04-27 21:50:14,155 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22506.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:50:46,033 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.515e+02 4.185e+02 4.921e+02 5.938e+02 1.054e+03, threshold=9.842e+02, percent-clipped=2.0 2023-04-27 21:50:48,908 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22532.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:51:15,029 INFO [train.py:904] (4/8) Epoch 3, batch 2250, loss[loss=0.2778, simple_loss=0.3318, pruned_loss=0.112, over 16406.00 frames. ], tot_loss[loss=0.2471, simple_loss=0.3161, pruned_loss=0.08901, over 3321669.39 frames. ], batch size: 146, lr: 2.17e-02, grad_scale: 4.0 2023-04-27 21:51:18,565 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22554.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:51:42,459 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22571.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:51:54,795 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.00 vs. limit=5.0 2023-04-27 21:52:09,941 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22593.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:52:20,011 INFO [train.py:904] (4/8) Epoch 3, batch 2300, loss[loss=0.2167, simple_loss=0.2946, pruned_loss=0.06944, over 17219.00 frames. ], tot_loss[loss=0.2476, simple_loss=0.3173, pruned_loss=0.08893, over 3313353.61 frames. ], batch size: 45, lr: 2.17e-02, grad_scale: 4.0 2023-04-27 21:52:57,346 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22626.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:53:01,630 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.100e+02 3.931e+02 4.776e+02 5.724e+02 1.077e+03, threshold=9.553e+02, percent-clipped=1.0 2023-04-27 21:53:12,075 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22639.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:53:22,042 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 21:53:30,359 INFO [train.py:904] (4/8) Epoch 3, batch 2350, loss[loss=0.2569, simple_loss=0.327, pruned_loss=0.09338, over 16789.00 frames. ], tot_loss[loss=0.2481, simple_loss=0.3174, pruned_loss=0.08934, over 3317284.37 frames. ], batch size: 76, lr: 2.16e-02, grad_scale: 4.0 2023-04-27 21:53:40,505 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0994, 4.0786, 4.4871, 4.5458, 4.4996, 4.0839, 4.1824, 4.1219], device='cuda:4'), covar=tensor([0.0244, 0.0347, 0.0323, 0.0310, 0.0373, 0.0281, 0.0667, 0.0330], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0177, 0.0195, 0.0189, 0.0232, 0.0192, 0.0300, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 21:54:01,193 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22674.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:54:25,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5404, 4.3127, 4.2238, 1.9326, 2.9540, 2.4031, 3.9673, 4.2281], device='cuda:4'), covar=tensor([0.0242, 0.0491, 0.0322, 0.1420, 0.0669, 0.0901, 0.0626, 0.0559], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0125, 0.0157, 0.0147, 0.0138, 0.0132, 0.0150, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 21:54:35,057 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22700.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 21:54:35,724 INFO [train.py:904] (4/8) Epoch 3, batch 2400, loss[loss=0.2609, simple_loss=0.337, pruned_loss=0.09243, over 17001.00 frames. ], tot_loss[loss=0.2499, simple_loss=0.3195, pruned_loss=0.09015, over 3312484.23 frames. ], batch size: 50, lr: 2.16e-02, grad_scale: 8.0 2023-04-27 21:54:52,099 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5337, 2.4727, 2.4427, 4.0329, 1.9335, 3.6894, 2.2733, 2.3452], device='cuda:4'), covar=tensor([0.0304, 0.0630, 0.0379, 0.0168, 0.1531, 0.0195, 0.0810, 0.1027], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0219, 0.0183, 0.0247, 0.0289, 0.0190, 0.0209, 0.0280], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:55:17,391 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.811e+02 3.861e+02 4.414e+02 5.506e+02 1.152e+03, threshold=8.829e+02, percent-clipped=5.0 2023-04-27 21:55:32,305 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22741.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:55:46,917 INFO [train.py:904] (4/8) Epoch 3, batch 2450, loss[loss=0.211, simple_loss=0.2872, pruned_loss=0.06743, over 15974.00 frames. ], tot_loss[loss=0.2507, simple_loss=0.3206, pruned_loss=0.09043, over 3309618.41 frames. ], batch size: 35, lr: 2.16e-02, grad_scale: 8.0 2023-04-27 21:56:03,216 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2534, 4.9693, 5.1617, 5.4999, 5.6123, 4.7868, 5.5140, 5.5028], device='cuda:4'), covar=tensor([0.0528, 0.0483, 0.1078, 0.0318, 0.0287, 0.0415, 0.0285, 0.0329], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0357, 0.0486, 0.0367, 0.0274, 0.0259, 0.0284, 0.0292], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:56:33,239 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22786.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:56:33,415 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8029, 2.8040, 2.2938, 3.8767, 3.6036, 3.6264, 1.6912, 2.7505], device='cuda:4'), covar=tensor([0.1148, 0.0442, 0.1061, 0.0071, 0.0197, 0.0295, 0.1042, 0.0625], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0127, 0.0159, 0.0076, 0.0141, 0.0135, 0.0151, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:56:37,071 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1569, 1.6533, 2.3136, 2.8141, 2.8748, 3.1989, 1.8103, 2.9200], device='cuda:4'), covar=tensor([0.0041, 0.0183, 0.0108, 0.0081, 0.0043, 0.0051, 0.0145, 0.0037], device='cuda:4'), in_proj_covar=tensor([0.0087, 0.0120, 0.0105, 0.0102, 0.0082, 0.0071, 0.0108, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:56:46,144 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22794.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:56:55,061 INFO [train.py:904] (4/8) Epoch 3, batch 2500, loss[loss=0.2367, simple_loss=0.3211, pruned_loss=0.07613, over 16673.00 frames. ], tot_loss[loss=0.2503, simple_loss=0.3203, pruned_loss=0.09018, over 3312895.30 frames. ], batch size: 62, lr: 2.16e-02, grad_scale: 8.0 2023-04-27 21:57:27,680 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22825.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:57:33,675 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.850e+02 4.203e+02 4.842e+02 6.402e+02 1.699e+03, threshold=9.683e+02, percent-clipped=7.0 2023-04-27 21:57:36,885 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22832.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:57:41,074 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22834.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:58:03,441 INFO [train.py:904] (4/8) Epoch 3, batch 2550, loss[loss=0.2038, simple_loss=0.2731, pruned_loss=0.06722, over 15867.00 frames. ], tot_loss[loss=0.2508, simple_loss=0.3203, pruned_loss=0.09063, over 3316037.12 frames. ], batch size: 35, lr: 2.15e-02, grad_scale: 8.0 2023-04-27 21:58:10,925 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22855.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:58:31,421 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22871.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:58:44,551 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22880.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:58:52,568 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22886.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:59:01,458 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22893.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:59:13,834 INFO [train.py:904] (4/8) Epoch 3, batch 2600, loss[loss=0.2301, simple_loss=0.3077, pruned_loss=0.07623, over 16036.00 frames. ], tot_loss[loss=0.2492, simple_loss=0.3196, pruned_loss=0.0894, over 3317168.88 frames. ], batch size: 35, lr: 2.15e-02, grad_scale: 8.0 2023-04-27 21:59:39,353 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22919.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:59:53,532 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.483e+02 3.680e+02 4.418e+02 5.171e+02 1.031e+03, threshold=8.837e+02, percent-clipped=1.0 2023-04-27 22:00:09,150 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=22941.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:00:22,556 INFO [train.py:904] (4/8) Epoch 3, batch 2650, loss[loss=0.2171, simple_loss=0.2959, pruned_loss=0.06916, over 16097.00 frames. ], tot_loss[loss=0.2484, simple_loss=0.3194, pruned_loss=0.08876, over 3322893.16 frames. ], batch size: 35, lr: 2.15e-02, grad_scale: 8.0 2023-04-27 22:01:22,674 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22995.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 22:01:30,281 INFO [train.py:904] (4/8) Epoch 3, batch 2700, loss[loss=0.2699, simple_loss=0.3435, pruned_loss=0.09817, over 16443.00 frames. ], tot_loss[loss=0.2502, simple_loss=0.3212, pruned_loss=0.08963, over 3310074.92 frames. ], batch size: 75, lr: 2.15e-02, grad_scale: 8.0 2023-04-27 22:02:09,771 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.756e+02 4.226e+02 4.976e+02 6.032e+02 3.495e+03, threshold=9.952e+02, percent-clipped=7.0 2023-04-27 22:02:24,001 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23041.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:02:37,961 INFO [train.py:904] (4/8) Epoch 3, batch 2750, loss[loss=0.2879, simple_loss=0.3386, pruned_loss=0.1186, over 16737.00 frames. ], tot_loss[loss=0.2497, simple_loss=0.3212, pruned_loss=0.0891, over 3316895.84 frames. ], batch size: 134, lr: 2.15e-02, grad_scale: 8.0 2023-04-27 22:03:28,907 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=23089.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:03:45,505 INFO [train.py:904] (4/8) Epoch 3, batch 2800, loss[loss=0.182, simple_loss=0.2638, pruned_loss=0.0501, over 16992.00 frames. ], tot_loss[loss=0.2491, simple_loss=0.3203, pruned_loss=0.08894, over 3309350.27 frames. ], batch size: 41, lr: 2.14e-02, grad_scale: 8.0 2023-04-27 22:03:55,461 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=23108.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:04:25,542 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.361e+02 3.403e+02 4.548e+02 5.612e+02 1.011e+03, threshold=9.095e+02, percent-clipped=1.0 2023-04-27 22:04:54,772 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23150.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:04:55,605 INFO [train.py:904] (4/8) Epoch 3, batch 2850, loss[loss=0.2193, simple_loss=0.2861, pruned_loss=0.07625, over 16797.00 frames. ], tot_loss[loss=0.248, simple_loss=0.3193, pruned_loss=0.08837, over 3309899.54 frames. ], batch size: 102, lr: 2.14e-02, grad_scale: 8.0 2023-04-27 22:05:20,388 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=23169.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:35,809 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0315, 3.8058, 2.5510, 4.7952, 4.6147, 4.2808, 1.8609, 3.3489], device='cuda:4'), covar=tensor([0.1157, 0.0316, 0.1107, 0.0054, 0.0220, 0.0304, 0.1119, 0.0494], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0130, 0.0163, 0.0076, 0.0145, 0.0141, 0.0155, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:05:36,684 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23181.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:49,423 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3185, 3.1354, 1.6013, 3.2618, 2.2250, 3.2824, 1.8032, 2.6365], device='cuda:4'), covar=tensor([0.0071, 0.0405, 0.1340, 0.0076, 0.0723, 0.0521, 0.1168, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0091, 0.0140, 0.0173, 0.0087, 0.0162, 0.0172, 0.0182, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-27 22:06:03,310 INFO [train.py:904] (4/8) Epoch 3, batch 2900, loss[loss=0.2288, simple_loss=0.284, pruned_loss=0.08683, over 16770.00 frames. ], tot_loss[loss=0.2469, simple_loss=0.3174, pruned_loss=0.0882, over 3319225.63 frames. ], batch size: 83, lr: 2.14e-02, grad_scale: 8.0 2023-04-27 22:06:25,816 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=23217.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:06:42,326 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=23229.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:06:43,091 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.672e+02 4.009e+02 4.959e+02 5.927e+02 1.397e+03, threshold=9.918e+02, percent-clipped=5.0 2023-04-27 22:07:12,200 INFO [train.py:904] (4/8) Epoch 3, batch 2950, loss[loss=0.2952, simple_loss=0.3507, pruned_loss=0.1199, over 15318.00 frames. ], tot_loss[loss=0.2481, simple_loss=0.3176, pruned_loss=0.08928, over 3304860.28 frames. ], batch size: 190, lr: 2.14e-02, grad_scale: 8.0 2023-04-27 22:07:28,932 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-27 22:07:49,177 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=23278.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:08:07,123 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=23290.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:08:12,625 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23295.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:08:19,867 INFO [train.py:904] (4/8) Epoch 3, batch 3000, loss[loss=0.2893, simple_loss=0.3455, pruned_loss=0.1165, over 16274.00 frames. ], tot_loss[loss=0.248, simple_loss=0.3173, pruned_loss=0.08933, over 3315366.83 frames. ], batch size: 165, lr: 2.13e-02, grad_scale: 8.0 2023-04-27 22:08:19,867 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 22:08:30,493 INFO [train.py:938] (4/8) Epoch 3, validation: loss=0.1721, simple_loss=0.279, pruned_loss=0.03262, over 944034.00 frames. 2023-04-27 22:08:30,494 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-27 22:08:47,006 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.78 vs. limit=5.0 2023-04-27 22:09:10,134 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.558e+02 3.966e+02 4.777e+02 5.754e+02 1.756e+03, threshold=9.554e+02, percent-clipped=1.0 2023-04-27 22:09:16,278 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3844, 2.2254, 1.8353, 2.0641, 2.7901, 2.7347, 3.7717, 3.1399], device='cuda:4'), covar=tensor([0.0021, 0.0135, 0.0167, 0.0145, 0.0083, 0.0114, 0.0038, 0.0067], device='cuda:4'), in_proj_covar=tensor([0.0061, 0.0117, 0.0115, 0.0116, 0.0108, 0.0117, 0.0078, 0.0094], device='cuda:4'), out_proj_covar=tensor([8.8797e-05, 1.7088e-04, 1.6286e-04, 1.6835e-04, 1.6358e-04, 1.7432e-04, 1.1619e-04, 1.4325e-04], device='cuda:4') 2023-04-27 22:09:25,505 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=23343.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:09:37,532 INFO [train.py:904] (4/8) Epoch 3, batch 3050, loss[loss=0.2605, simple_loss=0.319, pruned_loss=0.101, over 16424.00 frames. ], tot_loss[loss=0.2482, simple_loss=0.3175, pruned_loss=0.08941, over 3316651.37 frames. ], batch size: 146, lr: 2.13e-02, grad_scale: 4.0 2023-04-27 22:10:44,552 INFO [train.py:904] (4/8) Epoch 3, batch 3100, loss[loss=0.2583, simple_loss=0.3078, pruned_loss=0.1044, over 16886.00 frames. ], tot_loss[loss=0.2479, simple_loss=0.3175, pruned_loss=0.08921, over 3312496.09 frames. ], batch size: 109, lr: 2.13e-02, grad_scale: 4.0 2023-04-27 22:10:46,454 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.68 vs. limit=2.0 2023-04-27 22:10:50,714 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0450, 5.5926, 5.5091, 5.5011, 5.5077, 6.0858, 5.8721, 5.6676], device='cuda:4'), covar=tensor([0.0549, 0.1039, 0.1038, 0.1277, 0.2062, 0.0753, 0.0639, 0.1591], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0328, 0.0305, 0.0288, 0.0375, 0.0326, 0.0260, 0.0384], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 22:11:12,199 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 22:11:28,255 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.500e+02 3.880e+02 4.394e+02 5.487e+02 1.419e+03, threshold=8.788e+02, percent-clipped=5.0 2023-04-27 22:11:53,265 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23450.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:11:54,090 INFO [train.py:904] (4/8) Epoch 3, batch 3150, loss[loss=0.214, simple_loss=0.2907, pruned_loss=0.06861, over 16851.00 frames. ], tot_loss[loss=0.2462, simple_loss=0.3158, pruned_loss=0.08832, over 3320303.00 frames. ], batch size: 42, lr: 2.13e-02, grad_scale: 4.0 2023-04-27 22:12:10,296 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9082, 4.7327, 5.4380, 5.3927, 5.4094, 5.0365, 4.9130, 4.8654], device='cuda:4'), covar=tensor([0.0209, 0.0267, 0.0277, 0.0358, 0.0321, 0.0191, 0.0695, 0.0272], device='cuda:4'), in_proj_covar=tensor([0.0204, 0.0193, 0.0208, 0.0211, 0.0254, 0.0209, 0.0320, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 22:12:12,234 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23464.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:12:35,075 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23481.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:12:44,061 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=23488.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:12:58,216 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=23498.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:01,358 INFO [train.py:904] (4/8) Epoch 3, batch 3200, loss[loss=0.2521, simple_loss=0.3169, pruned_loss=0.0937, over 16537.00 frames. ], tot_loss[loss=0.2442, simple_loss=0.314, pruned_loss=0.08722, over 3325992.15 frames. ], batch size: 75, lr: 2.13e-02, grad_scale: 8.0 2023-04-27 22:13:39,246 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=23529.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:42,226 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.397e+02 3.601e+02 4.370e+02 5.311e+02 9.274e+02, threshold=8.739e+02, percent-clipped=1.0 2023-04-27 22:13:51,660 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.77 vs. limit=2.0 2023-04-27 22:14:06,495 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=23549.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 22:14:08,943 INFO [train.py:904] (4/8) Epoch 3, batch 3250, loss[loss=0.2257, simple_loss=0.3084, pruned_loss=0.07148, over 16686.00 frames. ], tot_loss[loss=0.2429, simple_loss=0.3135, pruned_loss=0.08617, over 3335621.34 frames. ], batch size: 57, lr: 2.12e-02, grad_scale: 8.0 2023-04-27 22:14:38,902 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23573.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:56,593 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23585.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:15:18,373 INFO [train.py:904] (4/8) Epoch 3, batch 3300, loss[loss=0.24, simple_loss=0.3052, pruned_loss=0.08734, over 16851.00 frames. ], tot_loss[loss=0.2429, simple_loss=0.3144, pruned_loss=0.08575, over 3333558.39 frames. ], batch size: 96, lr: 2.12e-02, grad_scale: 8.0 2023-04-27 22:15:57,197 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.549e+02 3.983e+02 4.910e+02 5.801e+02 1.123e+03, threshold=9.819e+02, percent-clipped=5.0 2023-04-27 22:16:24,693 INFO [train.py:904] (4/8) Epoch 3, batch 3350, loss[loss=0.2397, simple_loss=0.3128, pruned_loss=0.08327, over 16786.00 frames. ], tot_loss[loss=0.2448, simple_loss=0.3159, pruned_loss=0.0869, over 3326693.22 frames. ], batch size: 83, lr: 2.12e-02, grad_scale: 8.0 2023-04-27 22:16:33,701 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.79 vs. limit=5.0 2023-04-27 22:17:33,461 INFO [train.py:904] (4/8) Epoch 3, batch 3400, loss[loss=0.2715, simple_loss=0.3296, pruned_loss=0.1067, over 16292.00 frames. ], tot_loss[loss=0.2442, simple_loss=0.3153, pruned_loss=0.08657, over 3326650.81 frames. ], batch size: 165, lr: 2.12e-02, grad_scale: 8.0 2023-04-27 22:17:40,994 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8666, 4.8246, 5.5005, 5.5477, 5.4808, 5.0027, 5.0956, 4.9983], device='cuda:4'), covar=tensor([0.0218, 0.0248, 0.0219, 0.0224, 0.0356, 0.0227, 0.0598, 0.0241], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0191, 0.0208, 0.0203, 0.0251, 0.0208, 0.0308, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 22:17:56,610 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8717, 4.1526, 3.2961, 2.5309, 3.1221, 2.0997, 4.3245, 4.5682], device='cuda:4'), covar=tensor([0.1723, 0.0542, 0.1035, 0.1075, 0.1827, 0.1420, 0.0286, 0.0273], device='cuda:4'), in_proj_covar=tensor([0.0260, 0.0241, 0.0253, 0.0206, 0.0283, 0.0188, 0.0213, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:18:13,370 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.692e+02 3.881e+02 4.662e+02 5.457e+02 1.013e+03, threshold=9.324e+02, percent-clipped=1.0 2023-04-27 22:18:33,439 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8310, 2.6827, 2.6810, 4.4162, 2.0133, 3.9456, 2.4077, 2.5800], device='cuda:4'), covar=tensor([0.0324, 0.0685, 0.0388, 0.0152, 0.1781, 0.0241, 0.0856, 0.1251], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0220, 0.0186, 0.0249, 0.0291, 0.0198, 0.0213, 0.0285], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:18:40,209 INFO [train.py:904] (4/8) Epoch 3, batch 3450, loss[loss=0.2417, simple_loss=0.2952, pruned_loss=0.09406, over 16871.00 frames. ], tot_loss[loss=0.2423, simple_loss=0.3131, pruned_loss=0.08578, over 3322268.10 frames. ], batch size: 90, lr: 2.11e-02, grad_scale: 8.0 2023-04-27 22:18:56,632 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-27 22:18:59,329 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23764.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:19:03,602 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-27 22:19:47,198 INFO [train.py:904] (4/8) Epoch 3, batch 3500, loss[loss=0.2258, simple_loss=0.3091, pruned_loss=0.0713, over 17060.00 frames. ], tot_loss[loss=0.241, simple_loss=0.312, pruned_loss=0.08501, over 3331206.20 frames. ], batch size: 53, lr: 2.11e-02, grad_scale: 8.0 2023-04-27 22:20:04,595 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=23812.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:20:31,310 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.165e+02 3.876e+02 4.693e+02 5.719e+02 1.171e+03, threshold=9.385e+02, percent-clipped=5.0 2023-04-27 22:20:48,797 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23844.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 22:20:50,699 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9753, 4.2876, 4.3729, 1.9402, 4.7396, 4.7495, 3.5071, 3.5112], device='cuda:4'), covar=tensor([0.0664, 0.0129, 0.0220, 0.1308, 0.0050, 0.0036, 0.0296, 0.0368], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0080, 0.0080, 0.0140, 0.0076, 0.0074, 0.0111, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:20:55,024 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-27 22:20:59,086 INFO [train.py:904] (4/8) Epoch 3, batch 3550, loss[loss=0.2734, simple_loss=0.3289, pruned_loss=0.1089, over 16364.00 frames. ], tot_loss[loss=0.2394, simple_loss=0.3107, pruned_loss=0.08401, over 3321863.73 frames. ], batch size: 146, lr: 2.11e-02, grad_scale: 8.0 2023-04-27 22:21:03,310 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.14 vs. limit=5.0 2023-04-27 22:21:29,304 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23873.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:21:43,167 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-27 22:21:45,227 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23885.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:22:05,751 INFO [train.py:904] (4/8) Epoch 3, batch 3600, loss[loss=0.2483, simple_loss=0.3278, pruned_loss=0.08438, over 17117.00 frames. ], tot_loss[loss=0.2388, simple_loss=0.3096, pruned_loss=0.084, over 3322370.09 frames. ], batch size: 48, lr: 2.11e-02, grad_scale: 8.0 2023-04-27 22:22:08,712 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8959, 2.7956, 2.5253, 4.3905, 2.0863, 4.1523, 2.4284, 2.6307], device='cuda:4'), covar=tensor([0.0281, 0.0653, 0.0416, 0.0173, 0.1631, 0.0206, 0.0860, 0.1149], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0224, 0.0188, 0.0254, 0.0293, 0.0198, 0.0215, 0.0286], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:22:33,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=23921.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:22:47,018 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.441e+02 3.888e+02 4.838e+02 6.309e+02 1.139e+03, threshold=9.677e+02, percent-clipped=4.0 2023-04-27 22:22:50,341 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=23933.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:23:14,450 INFO [train.py:904] (4/8) Epoch 3, batch 3650, loss[loss=0.2434, simple_loss=0.2992, pruned_loss=0.09382, over 15362.00 frames. ], tot_loss[loss=0.2381, simple_loss=0.3082, pruned_loss=0.08405, over 3310676.62 frames. ], batch size: 190, lr: 2.11e-02, grad_scale: 8.0 2023-04-27 22:24:29,888 INFO [train.py:904] (4/8) Epoch 3, batch 3700, loss[loss=0.2532, simple_loss=0.3103, pruned_loss=0.09802, over 16271.00 frames. ], tot_loss[loss=0.2397, simple_loss=0.3075, pruned_loss=0.08593, over 3286911.52 frames. ], batch size: 165, lr: 2.10e-02, grad_scale: 8.0 2023-04-27 22:24:53,074 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7560, 3.6576, 3.8050, 3.7671, 3.7015, 4.1218, 3.9569, 3.6050], device='cuda:4'), covar=tensor([0.1466, 0.1212, 0.0879, 0.1491, 0.1878, 0.0940, 0.0777, 0.1986], device='cuda:4'), in_proj_covar=tensor([0.0230, 0.0321, 0.0300, 0.0284, 0.0361, 0.0316, 0.0257, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 22:25:13,794 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.162e+02 3.909e+02 4.503e+02 5.530e+02 1.035e+03, threshold=9.006e+02, percent-clipped=1.0 2023-04-27 22:25:14,886 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7978, 5.2033, 4.7938, 4.9943, 4.4873, 4.3819, 4.6766, 5.1879], device='cuda:4'), covar=tensor([0.0423, 0.0437, 0.0732, 0.0309, 0.0461, 0.0582, 0.0412, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0264, 0.0373, 0.0322, 0.0232, 0.0247, 0.0234, 0.0295, 0.0260], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:25:22,594 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24037.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:25:42,970 INFO [train.py:904] (4/8) Epoch 3, batch 3750, loss[loss=0.2421, simple_loss=0.3278, pruned_loss=0.07816, over 17143.00 frames. ], tot_loss[loss=0.2415, simple_loss=0.3083, pruned_loss=0.08733, over 3283177.12 frames. ], batch size: 47, lr: 2.10e-02, grad_scale: 8.0 2023-04-27 22:26:46,931 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24098.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:26:50,659 INFO [train.py:904] (4/8) Epoch 3, batch 3800, loss[loss=0.2432, simple_loss=0.3144, pruned_loss=0.08595, over 15583.00 frames. ], tot_loss[loss=0.243, simple_loss=0.3089, pruned_loss=0.08857, over 3289435.52 frames. ], batch size: 191, lr: 2.10e-02, grad_scale: 8.0 2023-04-27 22:27:34,019 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.430e+02 3.668e+02 4.510e+02 5.659e+02 1.360e+03, threshold=9.019e+02, percent-clipped=5.0 2023-04-27 22:27:52,787 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=24144.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 22:28:01,891 INFO [train.py:904] (4/8) Epoch 3, batch 3850, loss[loss=0.2558, simple_loss=0.3167, pruned_loss=0.09738, over 16460.00 frames. ], tot_loss[loss=0.2436, simple_loss=0.3087, pruned_loss=0.0892, over 3290266.26 frames. ], batch size: 146, lr: 2.10e-02, grad_scale: 8.0 2023-04-27 22:28:25,368 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5734, 5.4867, 5.3159, 5.4144, 4.8411, 5.3357, 5.1896, 5.0584], device='cuda:4'), covar=tensor([0.0216, 0.0118, 0.0131, 0.0090, 0.0660, 0.0174, 0.0139, 0.0218], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0129, 0.0179, 0.0150, 0.0212, 0.0162, 0.0131, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 22:29:01,359 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=24192.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:29:13,623 INFO [train.py:904] (4/8) Epoch 3, batch 3900, loss[loss=0.2177, simple_loss=0.2847, pruned_loss=0.07537, over 16774.00 frames. ], tot_loss[loss=0.2423, simple_loss=0.3072, pruned_loss=0.08869, over 3290289.48 frames. ], batch size: 83, lr: 2.10e-02, grad_scale: 8.0 2023-04-27 22:29:35,478 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-04-27 22:29:50,305 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9555, 3.7817, 1.8474, 3.8597, 2.7469, 3.9357, 1.8981, 2.9444], device='cuda:4'), covar=tensor([0.0048, 0.0342, 0.1453, 0.0046, 0.0614, 0.0298, 0.1441, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0091, 0.0138, 0.0170, 0.0082, 0.0161, 0.0168, 0.0181, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-27 22:29:54,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5188, 4.2174, 4.3847, 4.6583, 4.7729, 4.1243, 4.6470, 4.6941], device='cuda:4'), covar=tensor([0.0499, 0.0575, 0.1035, 0.0389, 0.0326, 0.0623, 0.0509, 0.0306], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0349, 0.0454, 0.0361, 0.0266, 0.0250, 0.0283, 0.0283], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:29:56,950 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.402e+02 3.854e+02 4.579e+02 5.596e+02 1.788e+03, threshold=9.157e+02, percent-clipped=5.0 2023-04-27 22:30:25,197 INFO [train.py:904] (4/8) Epoch 3, batch 3950, loss[loss=0.2589, simple_loss=0.3116, pruned_loss=0.1031, over 16698.00 frames. ], tot_loss[loss=0.241, simple_loss=0.3053, pruned_loss=0.0884, over 3289819.73 frames. ], batch size: 134, lr: 2.09e-02, grad_scale: 8.0 2023-04-27 22:30:47,958 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3674, 4.3312, 1.6616, 4.2515, 2.7585, 4.4423, 2.1564, 3.0884], device='cuda:4'), covar=tensor([0.0037, 0.0120, 0.2033, 0.0043, 0.0740, 0.0155, 0.1512, 0.0656], device='cuda:4'), in_proj_covar=tensor([0.0090, 0.0137, 0.0171, 0.0083, 0.0161, 0.0168, 0.0179, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-27 22:31:34,849 INFO [train.py:904] (4/8) Epoch 3, batch 4000, loss[loss=0.238, simple_loss=0.3095, pruned_loss=0.08323, over 16899.00 frames. ], tot_loss[loss=0.2417, simple_loss=0.3056, pruned_loss=0.08895, over 3288324.52 frames. ], batch size: 116, lr: 2.09e-02, grad_scale: 8.0 2023-04-27 22:32:17,081 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.388e+02 3.319e+02 4.445e+02 5.556e+02 1.324e+03, threshold=8.889e+02, percent-clipped=2.0 2023-04-27 22:32:45,200 INFO [train.py:904] (4/8) Epoch 3, batch 4050, loss[loss=0.2159, simple_loss=0.2897, pruned_loss=0.07105, over 16786.00 frames. ], tot_loss[loss=0.239, simple_loss=0.3049, pruned_loss=0.08659, over 3283404.20 frames. ], batch size: 39, lr: 2.09e-02, grad_scale: 8.0 2023-04-27 22:33:46,328 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=24393.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:33:58,259 INFO [train.py:904] (4/8) Epoch 3, batch 4100, loss[loss=0.2928, simple_loss=0.3582, pruned_loss=0.1137, over 15374.00 frames. ], tot_loss[loss=0.2378, simple_loss=0.3055, pruned_loss=0.08502, over 3286444.75 frames. ], batch size: 190, lr: 2.09e-02, grad_scale: 8.0 2023-04-27 22:34:42,995 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.218e+02 2.972e+02 3.823e+02 4.924e+02 9.607e+02, threshold=7.646e+02, percent-clipped=2.0 2023-04-27 22:35:13,089 INFO [train.py:904] (4/8) Epoch 3, batch 4150, loss[loss=0.3333, simple_loss=0.3771, pruned_loss=0.1447, over 11055.00 frames. ], tot_loss[loss=0.246, simple_loss=0.3138, pruned_loss=0.08912, over 3261620.91 frames. ], batch size: 248, lr: 2.09e-02, grad_scale: 8.0 2023-04-27 22:36:10,456 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8453, 3.3214, 2.7677, 4.7961, 4.5543, 4.0927, 1.9330, 3.3490], device='cuda:4'), covar=tensor([0.1339, 0.0446, 0.1008, 0.0045, 0.0126, 0.0216, 0.1202, 0.0569], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0128, 0.0160, 0.0071, 0.0132, 0.0132, 0.0151, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:36:27,611 INFO [train.py:904] (4/8) Epoch 3, batch 4200, loss[loss=0.3068, simple_loss=0.3603, pruned_loss=0.1266, over 11772.00 frames. ], tot_loss[loss=0.2529, simple_loss=0.3218, pruned_loss=0.09199, over 3228035.81 frames. ], batch size: 247, lr: 2.08e-02, grad_scale: 8.0 2023-04-27 22:37:10,951 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.707e+02 3.802e+02 4.415e+02 5.397e+02 1.092e+03, threshold=8.829e+02, percent-clipped=9.0 2023-04-27 22:37:35,678 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24547.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:37:40,125 INFO [train.py:904] (4/8) Epoch 3, batch 4250, loss[loss=0.2176, simple_loss=0.3016, pruned_loss=0.06685, over 15366.00 frames. ], tot_loss[loss=0.2554, simple_loss=0.3253, pruned_loss=0.09272, over 3190868.21 frames. ], batch size: 190, lr: 2.08e-02, grad_scale: 8.0 2023-04-27 22:38:04,945 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1411, 1.3466, 1.6596, 2.1708, 2.1980, 2.3737, 1.4720, 2.3820], device='cuda:4'), covar=tensor([0.0052, 0.0187, 0.0114, 0.0084, 0.0061, 0.0052, 0.0149, 0.0050], device='cuda:4'), in_proj_covar=tensor([0.0084, 0.0121, 0.0106, 0.0099, 0.0086, 0.0065, 0.0110, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 22:38:06,821 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24569.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:38:53,671 INFO [train.py:904] (4/8) Epoch 3, batch 4300, loss[loss=0.2592, simple_loss=0.3358, pruned_loss=0.09137, over 17016.00 frames. ], tot_loss[loss=0.2543, simple_loss=0.326, pruned_loss=0.09133, over 3183624.25 frames. ], batch size: 50, lr: 2.08e-02, grad_scale: 8.0 2023-04-27 22:39:04,977 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24608.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:39:37,595 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24630.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:39:38,216 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.269e+02 3.576e+02 4.392e+02 5.357e+02 9.860e+02, threshold=8.785e+02, percent-clipped=3.0 2023-04-27 22:40:06,091 INFO [train.py:904] (4/8) Epoch 3, batch 4350, loss[loss=0.2535, simple_loss=0.3352, pruned_loss=0.08587, over 16442.00 frames. ], tot_loss[loss=0.2586, simple_loss=0.3305, pruned_loss=0.09334, over 3172588.17 frames. ], batch size: 146, lr: 2.08e-02, grad_scale: 8.0 2023-04-27 22:41:10,555 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=24693.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:41:21,565 INFO [train.py:904] (4/8) Epoch 3, batch 4400, loss[loss=0.2368, simple_loss=0.3185, pruned_loss=0.07754, over 17020.00 frames. ], tot_loss[loss=0.2606, simple_loss=0.3328, pruned_loss=0.09419, over 3180806.32 frames. ], batch size: 41, lr: 2.08e-02, grad_scale: 8.0 2023-04-27 22:41:30,251 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2464, 4.0987, 3.4273, 1.6578, 2.9728, 2.2794, 3.5127, 3.8167], device='cuda:4'), covar=tensor([0.0200, 0.0311, 0.0427, 0.1615, 0.0603, 0.0861, 0.0580, 0.0461], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0117, 0.0156, 0.0146, 0.0138, 0.0131, 0.0145, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 22:42:05,388 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.377e+02 3.698e+02 4.447e+02 5.386e+02 9.920e+02, threshold=8.895e+02, percent-clipped=4.0 2023-04-27 22:42:21,153 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=24741.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:42:34,827 INFO [train.py:904] (4/8) Epoch 3, batch 4450, loss[loss=0.2558, simple_loss=0.3443, pruned_loss=0.08367, over 16872.00 frames. ], tot_loss[loss=0.2616, simple_loss=0.3353, pruned_loss=0.09388, over 3189856.57 frames. ], batch size: 102, lr: 2.07e-02, grad_scale: 8.0 2023-04-27 22:42:55,214 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24765.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:43:47,147 INFO [train.py:904] (4/8) Epoch 3, batch 4500, loss[loss=0.2606, simple_loss=0.3377, pruned_loss=0.09182, over 16801.00 frames. ], tot_loss[loss=0.2604, simple_loss=0.3345, pruned_loss=0.09313, over 3182815.31 frames. ], batch size: 83, lr: 2.07e-02, grad_scale: 8.0 2023-04-27 22:44:05,003 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-27 22:44:22,414 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24826.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:44:27,878 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2196, 2.9493, 2.8191, 1.8992, 2.6211, 2.0548, 2.7702, 2.8786], device='cuda:4'), covar=tensor([0.0275, 0.0411, 0.0358, 0.1315, 0.0592, 0.0838, 0.0539, 0.0474], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0117, 0.0157, 0.0146, 0.0139, 0.0131, 0.0144, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 22:44:29,086 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.999e+02 3.148e+02 3.848e+02 4.477e+02 7.288e+02, threshold=7.696e+02, percent-clipped=0.0 2023-04-27 22:44:50,476 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2091, 1.1608, 1.5258, 1.9973, 2.0862, 2.2513, 1.3605, 2.1059], device='cuda:4'), covar=tensor([0.0039, 0.0211, 0.0107, 0.0110, 0.0065, 0.0052, 0.0181, 0.0047], device='cuda:4'), in_proj_covar=tensor([0.0083, 0.0124, 0.0108, 0.0100, 0.0087, 0.0067, 0.0115, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-27 22:44:52,822 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24848.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:44:56,540 INFO [train.py:904] (4/8) Epoch 3, batch 4550, loss[loss=0.2911, simple_loss=0.3611, pruned_loss=0.1106, over 16711.00 frames. ], tot_loss[loss=0.2604, simple_loss=0.3345, pruned_loss=0.09312, over 3184553.38 frames. ], batch size: 62, lr: 2.07e-02, grad_scale: 8.0 2023-04-27 22:45:03,293 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4946, 3.7135, 3.9454, 3.8888, 3.9737, 3.6498, 3.3254, 3.6654], device='cuda:4'), covar=tensor([0.0408, 0.0338, 0.0455, 0.0648, 0.0517, 0.0395, 0.1229, 0.0401], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0168, 0.0184, 0.0186, 0.0223, 0.0184, 0.0285, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-27 22:45:52,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1624, 5.0153, 4.8822, 4.9963, 4.2951, 4.9490, 4.7618, 4.5519], device='cuda:4'), covar=tensor([0.0214, 0.0108, 0.0150, 0.0095, 0.0804, 0.0151, 0.0144, 0.0251], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0115, 0.0165, 0.0136, 0.0194, 0.0145, 0.0122, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:46:07,316 INFO [train.py:904] (4/8) Epoch 3, batch 4600, loss[loss=0.2759, simple_loss=0.3547, pruned_loss=0.0986, over 16908.00 frames. ], tot_loss[loss=0.2603, simple_loss=0.3353, pruned_loss=0.09262, over 3195971.60 frames. ], batch size: 96, lr: 2.07e-02, grad_scale: 8.0 2023-04-27 22:46:10,656 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=24903.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:46:20,500 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24909.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:46:44,269 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=24925.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:46:52,469 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.050e+02 3.022e+02 3.683e+02 4.480e+02 7.885e+02, threshold=7.365e+02, percent-clipped=1.0 2023-04-27 22:47:20,338 INFO [train.py:904] (4/8) Epoch 3, batch 4650, loss[loss=0.2258, simple_loss=0.3055, pruned_loss=0.073, over 16682.00 frames. ], tot_loss[loss=0.259, simple_loss=0.3341, pruned_loss=0.092, over 3201632.66 frames. ], batch size: 62, lr: 2.07e-02, grad_scale: 8.0 2023-04-27 22:47:21,930 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1231, 2.4118, 2.4098, 3.2352, 3.0528, 3.1038, 1.8715, 2.7412], device='cuda:4'), covar=tensor([0.1075, 0.0394, 0.0867, 0.0076, 0.0166, 0.0297, 0.0997, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0128, 0.0163, 0.0069, 0.0128, 0.0133, 0.0151, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:48:24,106 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24995.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:48:32,747 INFO [train.py:904] (4/8) Epoch 3, batch 4700, loss[loss=0.2826, simple_loss=0.3499, pruned_loss=0.1077, over 11803.00 frames. ], tot_loss[loss=0.2563, simple_loss=0.3316, pruned_loss=0.0905, over 3200902.09 frames. ], batch size: 248, lr: 2.06e-02, grad_scale: 4.0 2023-04-27 22:49:17,307 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.399e+02 3.338e+02 4.097e+02 4.784e+02 1.007e+03, threshold=8.194e+02, percent-clipped=3.0 2023-04-27 22:49:45,101 INFO [train.py:904] (4/8) Epoch 3, batch 4750, loss[loss=0.2842, simple_loss=0.3431, pruned_loss=0.1126, over 11615.00 frames. ], tot_loss[loss=0.2517, simple_loss=0.3269, pruned_loss=0.08822, over 3218305.71 frames. ], batch size: 248, lr: 2.06e-02, grad_scale: 4.0 2023-04-27 22:49:52,653 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=25056.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 22:50:47,186 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6073, 3.6240, 2.8709, 2.5212, 2.6840, 2.1112, 3.7628, 4.1054], device='cuda:4'), covar=tensor([0.1885, 0.0582, 0.1132, 0.1007, 0.1818, 0.1197, 0.0349, 0.0241], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0236, 0.0251, 0.0207, 0.0292, 0.0188, 0.0214, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:50:58,608 INFO [train.py:904] (4/8) Epoch 3, batch 4800, loss[loss=0.2725, simple_loss=0.3551, pruned_loss=0.0949, over 16798.00 frames. ], tot_loss[loss=0.2481, simple_loss=0.3231, pruned_loss=0.08657, over 3209277.61 frames. ], batch size: 124, lr: 2.06e-02, grad_scale: 8.0 2023-04-27 22:51:28,790 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=25121.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:51:47,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.867e+02 2.955e+02 3.547e+02 4.621e+02 1.014e+03, threshold=7.094e+02, percent-clipped=1.0 2023-04-27 22:52:13,130 INFO [train.py:904] (4/8) Epoch 3, batch 4850, loss[loss=0.3385, simple_loss=0.3805, pruned_loss=0.1483, over 12325.00 frames. ], tot_loss[loss=0.2483, simple_loss=0.3241, pruned_loss=0.08629, over 3200715.92 frames. ], batch size: 248, lr: 2.06e-02, grad_scale: 2.0 2023-04-27 22:52:59,254 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5769, 5.9186, 5.5809, 5.8258, 5.2410, 4.6702, 5.5507, 6.0760], device='cuda:4'), covar=tensor([0.0398, 0.0476, 0.0703, 0.0277, 0.0440, 0.0436, 0.0323, 0.0503], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0347, 0.0307, 0.0220, 0.0225, 0.0221, 0.0277, 0.0243], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:53:25,028 INFO [train.py:904] (4/8) Epoch 3, batch 4900, loss[loss=0.2821, simple_loss=0.3589, pruned_loss=0.1026, over 15355.00 frames. ], tot_loss[loss=0.2471, simple_loss=0.3235, pruned_loss=0.08537, over 3192287.32 frames. ], batch size: 191, lr: 2.06e-02, grad_scale: 2.0 2023-04-27 22:53:28,504 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25203.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:53:29,419 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=25204.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:53:59,399 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25225.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:54:10,581 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.158e+02 3.416e+02 4.003e+02 5.207e+02 1.200e+03, threshold=8.007e+02, percent-clipped=10.0 2023-04-27 22:54:35,314 INFO [train.py:904] (4/8) Epoch 3, batch 4950, loss[loss=0.31, simple_loss=0.3684, pruned_loss=0.1259, over 11946.00 frames. ], tot_loss[loss=0.2463, simple_loss=0.3231, pruned_loss=0.08477, over 3201581.63 frames. ], batch size: 247, lr: 2.05e-02, grad_scale: 2.0 2023-04-27 22:54:36,192 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=25251.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:55:08,063 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=25273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:55:48,336 INFO [train.py:904] (4/8) Epoch 3, batch 5000, loss[loss=0.2777, simple_loss=0.3446, pruned_loss=0.1054, over 12038.00 frames. ], tot_loss[loss=0.247, simple_loss=0.3243, pruned_loss=0.08489, over 3205576.19 frames. ], batch size: 247, lr: 2.05e-02, grad_scale: 2.0 2023-04-27 22:56:35,246 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.358e+02 3.469e+02 4.106e+02 5.107e+02 1.084e+03, threshold=8.211e+02, percent-clipped=3.0 2023-04-27 22:56:59,650 INFO [train.py:904] (4/8) Epoch 3, batch 5050, loss[loss=0.2659, simple_loss=0.3409, pruned_loss=0.09543, over 16453.00 frames. ], tot_loss[loss=0.2464, simple_loss=0.3244, pruned_loss=0.08419, over 3215326.16 frames. ], batch size: 146, lr: 2.05e-02, grad_scale: 2.0 2023-04-27 22:57:00,009 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=25351.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 22:58:08,628 INFO [train.py:904] (4/8) Epoch 3, batch 5100, loss[loss=0.2173, simple_loss=0.3051, pruned_loss=0.06476, over 16287.00 frames. ], tot_loss[loss=0.2438, simple_loss=0.3216, pruned_loss=0.08294, over 3214554.84 frames. ], batch size: 165, lr: 2.05e-02, grad_scale: 2.0 2023-04-27 22:58:38,823 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25421.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:58:48,403 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=25428.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 22:58:57,532 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.562e+02 3.320e+02 3.999e+02 5.208e+02 7.552e+02, threshold=7.999e+02, percent-clipped=0.0 2023-04-27 22:59:23,204 INFO [train.py:904] (4/8) Epoch 3, batch 5150, loss[loss=0.2772, simple_loss=0.3534, pruned_loss=0.1006, over 16200.00 frames. ], tot_loss[loss=0.2429, simple_loss=0.3221, pruned_loss=0.08188, over 3219952.03 frames. ], batch size: 165, lr: 2.05e-02, grad_scale: 2.0 2023-04-27 22:59:50,321 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=25469.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:00:19,386 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=25489.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 23:00:35,982 INFO [train.py:904] (4/8) Epoch 3, batch 5200, loss[loss=0.2735, simple_loss=0.3403, pruned_loss=0.1034, over 12238.00 frames. ], tot_loss[loss=0.2428, simple_loss=0.3213, pruned_loss=0.0822, over 3217285.50 frames. ], batch size: 246, lr: 2.04e-02, grad_scale: 4.0 2023-04-27 23:00:36,667 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.60 vs. limit=2.0 2023-04-27 23:00:40,428 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25504.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:01:22,143 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.363e+02 3.525e+02 4.137e+02 5.123e+02 8.624e+02, threshold=8.275e+02, percent-clipped=4.0 2023-04-27 23:01:43,242 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4211, 4.3613, 4.3725, 3.0577, 3.9992, 4.2439, 4.2693, 2.4843], device='cuda:4'), covar=tensor([0.0317, 0.0019, 0.0018, 0.0202, 0.0037, 0.0043, 0.0026, 0.0271], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0053, 0.0060, 0.0111, 0.0053, 0.0065, 0.0058, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:01:45,790 INFO [train.py:904] (4/8) Epoch 3, batch 5250, loss[loss=0.2224, simple_loss=0.3049, pruned_loss=0.06995, over 16676.00 frames. ], tot_loss[loss=0.2414, simple_loss=0.3185, pruned_loss=0.08217, over 3213330.89 frames. ], batch size: 76, lr: 2.04e-02, grad_scale: 4.0 2023-04-27 23:01:47,881 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=25552.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:02:03,328 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-27 23:02:56,098 INFO [train.py:904] (4/8) Epoch 3, batch 5300, loss[loss=0.1947, simple_loss=0.2752, pruned_loss=0.0571, over 16541.00 frames. ], tot_loss[loss=0.2386, simple_loss=0.3151, pruned_loss=0.08103, over 3204665.90 frames. ], batch size: 68, lr: 2.04e-02, grad_scale: 4.0 2023-04-27 23:03:43,235 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.802e+02 3.089e+02 3.787e+02 4.480e+02 7.662e+02, threshold=7.574e+02, percent-clipped=0.0 2023-04-27 23:04:08,022 INFO [train.py:904] (4/8) Epoch 3, batch 5350, loss[loss=0.2896, simple_loss=0.3453, pruned_loss=0.1169, over 12280.00 frames. ], tot_loss[loss=0.2384, simple_loss=0.3147, pruned_loss=0.08106, over 3197989.44 frames. ], batch size: 246, lr: 2.04e-02, grad_scale: 4.0 2023-04-27 23:04:08,350 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25651.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 23:05:17,182 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=25699.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:05:19,279 INFO [train.py:904] (4/8) Epoch 3, batch 5400, loss[loss=0.2411, simple_loss=0.3154, pruned_loss=0.08341, over 16573.00 frames. ], tot_loss[loss=0.2406, simple_loss=0.3175, pruned_loss=0.08187, over 3200714.34 frames. ], batch size: 62, lr: 2.04e-02, grad_scale: 4.0 2023-04-27 23:06:07,787 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.583e+02 3.733e+02 4.548e+02 5.512e+02 9.876e+02, threshold=9.097e+02, percent-clipped=3.0 2023-04-27 23:06:34,330 INFO [train.py:904] (4/8) Epoch 3, batch 5450, loss[loss=0.3205, simple_loss=0.3847, pruned_loss=0.1281, over 16263.00 frames. ], tot_loss[loss=0.2452, simple_loss=0.3216, pruned_loss=0.08439, over 3203501.68 frames. ], batch size: 165, lr: 2.03e-02, grad_scale: 4.0 2023-04-27 23:07:24,721 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=25784.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:07:49,275 INFO [train.py:904] (4/8) Epoch 3, batch 5500, loss[loss=0.331, simple_loss=0.3877, pruned_loss=0.1371, over 15492.00 frames. ], tot_loss[loss=0.257, simple_loss=0.3313, pruned_loss=0.09137, over 3188727.28 frames. ], batch size: 191, lr: 2.03e-02, grad_scale: 4.0 2023-04-27 23:08:36,491 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2901, 5.5995, 5.3035, 5.3588, 4.8072, 4.4690, 5.0000, 5.7463], device='cuda:4'), covar=tensor([0.0411, 0.0506, 0.0730, 0.0342, 0.0530, 0.0587, 0.0486, 0.0538], device='cuda:4'), in_proj_covar=tensor([0.0259, 0.0364, 0.0321, 0.0231, 0.0236, 0.0232, 0.0296, 0.0252], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:08:39,221 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.596e+02 5.216e+02 6.145e+02 8.695e+02 2.860e+03, threshold=1.229e+03, percent-clipped=22.0 2023-04-27 23:09:06,205 INFO [train.py:904] (4/8) Epoch 3, batch 5550, loss[loss=0.3416, simple_loss=0.3967, pruned_loss=0.1432, over 16497.00 frames. ], tot_loss[loss=0.2693, simple_loss=0.3407, pruned_loss=0.09894, over 3177123.62 frames. ], batch size: 68, lr: 2.03e-02, grad_scale: 4.0 2023-04-27 23:09:08,596 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9147, 3.0528, 3.3957, 3.3955, 3.3747, 3.0982, 3.1736, 3.2524], device='cuda:4'), covar=tensor([0.0321, 0.0418, 0.0360, 0.0420, 0.0447, 0.0361, 0.0828, 0.0325], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0167, 0.0184, 0.0182, 0.0225, 0.0188, 0.0286, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-27 23:09:16,301 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2707, 4.5603, 4.3028, 4.3282, 3.9080, 3.9301, 4.1237, 4.5752], device='cuda:4'), covar=tensor([0.0538, 0.0609, 0.0747, 0.0385, 0.0527, 0.0877, 0.0544, 0.0667], device='cuda:4'), in_proj_covar=tensor([0.0260, 0.0363, 0.0321, 0.0230, 0.0236, 0.0231, 0.0295, 0.0251], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:09:18,255 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=25858.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:10:25,184 INFO [train.py:904] (4/8) Epoch 3, batch 5600, loss[loss=0.3651, simple_loss=0.4092, pruned_loss=0.1605, over 15272.00 frames. ], tot_loss[loss=0.2806, simple_loss=0.3481, pruned_loss=0.1065, over 3108209.44 frames. ], batch size: 190, lr: 2.03e-02, grad_scale: 8.0 2023-04-27 23:10:56,277 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=25919.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 23:11:21,466 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.427e+02 5.585e+02 6.873e+02 8.559e+02 2.132e+03, threshold=1.375e+03, percent-clipped=5.0 2023-04-27 23:11:48,522 INFO [train.py:904] (4/8) Epoch 3, batch 5650, loss[loss=0.3814, simple_loss=0.4164, pruned_loss=0.1732, over 11268.00 frames. ], tot_loss[loss=0.2903, simple_loss=0.3551, pruned_loss=0.1127, over 3076736.06 frames. ], batch size: 248, lr: 2.03e-02, grad_scale: 8.0 2023-04-27 23:13:10,043 INFO [train.py:904] (4/8) Epoch 3, batch 5700, loss[loss=0.2743, simple_loss=0.3468, pruned_loss=0.1009, over 17160.00 frames. ], tot_loss[loss=0.292, simple_loss=0.3565, pruned_loss=0.1138, over 3086417.99 frames. ], batch size: 46, lr: 2.02e-02, grad_scale: 8.0 2023-04-27 23:13:23,363 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-27 23:14:00,538 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.120e+02 5.064e+02 6.111e+02 7.661e+02 1.195e+03, threshold=1.222e+03, percent-clipped=0.0 2023-04-27 23:14:27,438 INFO [train.py:904] (4/8) Epoch 3, batch 5750, loss[loss=0.3586, simple_loss=0.3871, pruned_loss=0.1651, over 11147.00 frames. ], tot_loss[loss=0.2958, simple_loss=0.3601, pruned_loss=0.1158, over 3064380.89 frames. ], batch size: 248, lr: 2.02e-02, grad_scale: 8.0 2023-04-27 23:15:22,065 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26084.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:15:47,176 INFO [train.py:904] (4/8) Epoch 3, batch 5800, loss[loss=0.311, simple_loss=0.3555, pruned_loss=0.1333, over 12071.00 frames. ], tot_loss[loss=0.2955, simple_loss=0.3604, pruned_loss=0.1153, over 3045172.84 frames. ], batch size: 247, lr: 2.02e-02, grad_scale: 8.0 2023-04-27 23:15:48,971 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26102.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:16:36,164 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=26132.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:16:37,960 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7043, 4.5621, 4.4787, 4.5445, 3.7876, 4.6954, 4.6791, 4.2084], device='cuda:4'), covar=tensor([0.0425, 0.0251, 0.0230, 0.0172, 0.1062, 0.0270, 0.0218, 0.0366], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0129, 0.0170, 0.0143, 0.0202, 0.0159, 0.0127, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:16:38,588 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.751e+02 5.049e+02 6.109e+02 8.114e+02 1.629e+03, threshold=1.222e+03, percent-clipped=2.0 2023-04-27 23:17:05,088 INFO [train.py:904] (4/8) Epoch 3, batch 5850, loss[loss=0.3312, simple_loss=0.37, pruned_loss=0.1462, over 11474.00 frames. ], tot_loss[loss=0.2915, simple_loss=0.3575, pruned_loss=0.1128, over 3062481.39 frames. ], batch size: 250, lr: 2.02e-02, grad_scale: 8.0 2023-04-27 23:17:09,953 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4918, 4.7765, 4.7233, 3.4287, 4.2757, 4.5283, 4.7908, 2.7726], device='cuda:4'), covar=tensor([0.0306, 0.0012, 0.0015, 0.0169, 0.0021, 0.0043, 0.0009, 0.0259], device='cuda:4'), in_proj_covar=tensor([0.0109, 0.0051, 0.0058, 0.0107, 0.0050, 0.0062, 0.0055, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:17:24,104 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26163.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:17:31,340 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26168.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:18:26,955 INFO [train.py:904] (4/8) Epoch 3, batch 5900, loss[loss=0.2154, simple_loss=0.3033, pruned_loss=0.06372, over 16692.00 frames. ], tot_loss[loss=0.2898, simple_loss=0.3561, pruned_loss=0.1117, over 3063829.16 frames. ], batch size: 89, lr: 2.02e-02, grad_scale: 8.0 2023-04-27 23:18:32,204 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4827, 5.8019, 5.5151, 5.5903, 5.1476, 4.7481, 5.3705, 5.9684], device='cuda:4'), covar=tensor([0.0462, 0.0603, 0.0944, 0.0372, 0.0512, 0.0526, 0.0505, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0378, 0.0333, 0.0241, 0.0239, 0.0241, 0.0302, 0.0263], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:18:51,856 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26214.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 23:19:15,300 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26229.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:19:22,627 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.765e+02 4.492e+02 5.249e+02 6.969e+02 1.603e+03, threshold=1.050e+03, percent-clipped=2.0 2023-04-27 23:19:45,728 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-27 23:19:47,823 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26250.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:19:48,574 INFO [train.py:904] (4/8) Epoch 3, batch 5950, loss[loss=0.2764, simple_loss=0.3454, pruned_loss=0.1037, over 17228.00 frames. ], tot_loss[loss=0.2875, simple_loss=0.356, pruned_loss=0.1095, over 3072401.74 frames. ], batch size: 45, lr: 2.02e-02, grad_scale: 8.0 2023-04-27 23:21:07,922 INFO [train.py:904] (4/8) Epoch 3, batch 6000, loss[loss=0.297, simple_loss=0.3568, pruned_loss=0.1186, over 15367.00 frames. ], tot_loss[loss=0.2865, simple_loss=0.3548, pruned_loss=0.109, over 3066680.47 frames. ], batch size: 190, lr: 2.01e-02, grad_scale: 8.0 2023-04-27 23:21:07,922 INFO [train.py:929] (4/8) Computing validation loss 2023-04-27 23:21:18,890 INFO [train.py:938] (4/8) Epoch 3, validation: loss=0.2097, simple_loss=0.3184, pruned_loss=0.05055, over 944034.00 frames. 2023-04-27 23:21:18,891 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-27 23:21:34,612 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26311.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:22:03,878 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-27 23:22:07,342 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.939e+02 4.548e+02 5.519e+02 7.491e+02 1.813e+03, threshold=1.104e+03, percent-clipped=4.0 2023-04-27 23:22:36,223 INFO [train.py:904] (4/8) Epoch 3, batch 6050, loss[loss=0.2485, simple_loss=0.3314, pruned_loss=0.08281, over 16533.00 frames. ], tot_loss[loss=0.2836, simple_loss=0.3523, pruned_loss=0.1074, over 3088494.95 frames. ], batch size: 62, lr: 2.01e-02, grad_scale: 8.0 2023-04-27 23:23:51,473 INFO [train.py:904] (4/8) Epoch 3, batch 6100, loss[loss=0.2571, simple_loss=0.3331, pruned_loss=0.09051, over 17232.00 frames. ], tot_loss[loss=0.2813, simple_loss=0.351, pruned_loss=0.1058, over 3093517.44 frames. ], batch size: 45, lr: 2.01e-02, grad_scale: 8.0 2023-04-27 23:24:12,699 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26414.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:24:42,456 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.827e+02 4.226e+02 5.750e+02 6.713e+02 1.587e+03, threshold=1.150e+03, percent-clipped=3.0 2023-04-27 23:25:06,814 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-27 23:25:11,827 INFO [train.py:904] (4/8) Epoch 3, batch 6150, loss[loss=0.2363, simple_loss=0.3129, pruned_loss=0.07981, over 17252.00 frames. ], tot_loss[loss=0.2773, simple_loss=0.3475, pruned_loss=0.1035, over 3115988.71 frames. ], batch size: 52, lr: 2.01e-02, grad_scale: 4.0 2023-04-27 23:25:18,076 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2507, 4.2658, 4.1047, 3.4661, 4.0611, 1.6408, 3.8942, 4.0367], device='cuda:4'), covar=tensor([0.0060, 0.0043, 0.0069, 0.0288, 0.0058, 0.1432, 0.0070, 0.0086], device='cuda:4'), in_proj_covar=tensor([0.0073, 0.0061, 0.0095, 0.0112, 0.0071, 0.0120, 0.0083, 0.0093], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:25:23,316 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26458.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:25:48,834 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2069, 3.9296, 3.8761, 1.5770, 4.1447, 4.1417, 2.9520, 3.1972], device='cuda:4'), covar=tensor([0.0775, 0.0077, 0.0130, 0.1363, 0.0039, 0.0034, 0.0301, 0.0354], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0079, 0.0077, 0.0140, 0.0071, 0.0068, 0.0111, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-27 23:25:50,284 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26475.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:26:28,208 INFO [train.py:904] (4/8) Epoch 3, batch 6200, loss[loss=0.302, simple_loss=0.3484, pruned_loss=0.1278, over 12053.00 frames. ], tot_loss[loss=0.2759, simple_loss=0.3457, pruned_loss=0.1031, over 3122676.48 frames. ], batch size: 246, lr: 2.01e-02, grad_scale: 4.0 2023-04-27 23:26:48,658 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26514.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:27:03,422 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26524.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:27:18,519 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.570e+02 4.817e+02 5.994e+02 8.159e+02 2.733e+03, threshold=1.199e+03, percent-clipped=9.0 2023-04-27 23:27:41,889 INFO [train.py:904] (4/8) Epoch 3, batch 6250, loss[loss=0.3092, simple_loss=0.3595, pruned_loss=0.1295, over 11515.00 frames. ], tot_loss[loss=0.278, simple_loss=0.3466, pruned_loss=0.1047, over 3086484.64 frames. ], batch size: 246, lr: 2.00e-02, grad_scale: 4.0 2023-04-27 23:27:58,187 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=26562.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:28:54,857 INFO [train.py:904] (4/8) Epoch 3, batch 6300, loss[loss=0.2695, simple_loss=0.3368, pruned_loss=0.1011, over 16623.00 frames. ], tot_loss[loss=0.2781, simple_loss=0.347, pruned_loss=0.1046, over 3095201.01 frames. ], batch size: 62, lr: 2.00e-02, grad_scale: 4.0 2023-04-27 23:29:02,860 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26606.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:29:48,249 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.782e+02 4.322e+02 5.254e+02 6.822e+02 1.499e+03, threshold=1.051e+03, percent-clipped=4.0 2023-04-27 23:29:51,360 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4676, 4.5326, 1.9257, 4.5600, 2.7114, 4.6454, 2.2013, 3.1011], device='cuda:4'), covar=tensor([0.0051, 0.0125, 0.1631, 0.0022, 0.0755, 0.0205, 0.1340, 0.0585], device='cuda:4'), in_proj_covar=tensor([0.0089, 0.0131, 0.0174, 0.0081, 0.0162, 0.0164, 0.0179, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-27 23:29:53,201 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9203, 4.6828, 4.7179, 4.7169, 4.1818, 4.6879, 4.6893, 4.3853], device='cuda:4'), covar=tensor([0.0329, 0.0255, 0.0160, 0.0111, 0.0709, 0.0222, 0.0175, 0.0307], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0128, 0.0166, 0.0138, 0.0196, 0.0159, 0.0124, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:30:11,907 INFO [train.py:904] (4/8) Epoch 3, batch 6350, loss[loss=0.2936, simple_loss=0.3583, pruned_loss=0.1144, over 16348.00 frames. ], tot_loss[loss=0.2803, simple_loss=0.3482, pruned_loss=0.1062, over 3107897.99 frames. ], batch size: 146, lr: 2.00e-02, grad_scale: 4.0 2023-04-27 23:30:30,773 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.11 vs. limit=5.0 2023-04-27 23:31:25,029 INFO [train.py:904] (4/8) Epoch 3, batch 6400, loss[loss=0.2599, simple_loss=0.333, pruned_loss=0.09341, over 16223.00 frames. ], tot_loss[loss=0.2829, simple_loss=0.3497, pruned_loss=0.1081, over 3085130.48 frames. ], batch size: 165, lr: 2.00e-02, grad_scale: 8.0 2023-04-27 23:31:46,123 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7079, 2.6576, 2.2991, 4.1098, 1.7532, 3.6935, 2.2330, 2.4111], device='cuda:4'), covar=tensor([0.0354, 0.0785, 0.0562, 0.0208, 0.2016, 0.0291, 0.1083, 0.1420], device='cuda:4'), in_proj_covar=tensor([0.0258, 0.0237, 0.0200, 0.0262, 0.0312, 0.0214, 0.0225, 0.0298], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:32:06,291 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9162, 3.6500, 3.8269, 4.0934, 4.1346, 3.7080, 4.1403, 4.1242], device='cuda:4'), covar=tensor([0.0621, 0.0711, 0.1195, 0.0428, 0.0393, 0.0976, 0.0451, 0.0369], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0352, 0.0457, 0.0349, 0.0261, 0.0247, 0.0287, 0.0287], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:32:12,498 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.974e+02 5.048e+02 6.146e+02 7.949e+02 2.287e+03, threshold=1.229e+03, percent-clipped=11.0 2023-04-27 23:32:35,287 INFO [train.py:904] (4/8) Epoch 3, batch 6450, loss[loss=0.2451, simple_loss=0.3285, pruned_loss=0.08087, over 16783.00 frames. ], tot_loss[loss=0.2792, simple_loss=0.3477, pruned_loss=0.1054, over 3106708.16 frames. ], batch size: 89, lr: 2.00e-02, grad_scale: 8.0 2023-04-27 23:32:47,402 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26758.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:33:05,688 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26770.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:33:53,753 INFO [train.py:904] (4/8) Epoch 3, batch 6500, loss[loss=0.2511, simple_loss=0.323, pruned_loss=0.08956, over 16747.00 frames. ], tot_loss[loss=0.2758, simple_loss=0.3443, pruned_loss=0.1036, over 3120113.55 frames. ], batch size: 76, lr: 2.00e-02, grad_scale: 8.0 2023-04-27 23:34:01,035 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=26806.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:34:29,522 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26824.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:34:44,550 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.692e+02 4.621e+02 5.604e+02 6.997e+02 1.424e+03, threshold=1.121e+03, percent-clipped=2.0 2023-04-27 23:34:44,993 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1367, 5.3542, 4.9902, 5.2074, 4.7045, 4.5442, 4.8468, 5.4504], device='cuda:4'), covar=tensor([0.0457, 0.0551, 0.0964, 0.0346, 0.0539, 0.0539, 0.0498, 0.0607], device='cuda:4'), in_proj_covar=tensor([0.0276, 0.0382, 0.0341, 0.0245, 0.0246, 0.0245, 0.0303, 0.0265], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:35:12,110 INFO [train.py:904] (4/8) Epoch 3, batch 6550, loss[loss=0.3802, simple_loss=0.4129, pruned_loss=0.1737, over 11775.00 frames. ], tot_loss[loss=0.2778, simple_loss=0.3472, pruned_loss=0.1042, over 3117561.38 frames. ], batch size: 248, lr: 1.99e-02, grad_scale: 8.0 2023-04-27 23:35:45,473 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=26872.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:36:24,767 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2051, 2.3233, 2.1550, 3.5422, 1.7112, 3.1643, 1.9802, 2.1265], device='cuda:4'), covar=tensor([0.0480, 0.0965, 0.0624, 0.0311, 0.2329, 0.0416, 0.1288, 0.1730], device='cuda:4'), in_proj_covar=tensor([0.0257, 0.0234, 0.0198, 0.0258, 0.0309, 0.0212, 0.0225, 0.0296], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:36:25,008 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.68 vs. limit=2.0 2023-04-27 23:36:28,904 INFO [train.py:904] (4/8) Epoch 3, batch 6600, loss[loss=0.2725, simple_loss=0.3487, pruned_loss=0.09815, over 16889.00 frames. ], tot_loss[loss=0.2804, simple_loss=0.3501, pruned_loss=0.1053, over 3104924.37 frames. ], batch size: 42, lr: 1.99e-02, grad_scale: 8.0 2023-04-27 23:36:33,623 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26904.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:36:37,649 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26906.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:37:19,055 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7751, 3.8247, 4.2158, 4.1629, 4.2047, 3.8354, 3.9110, 3.8964], device='cuda:4'), covar=tensor([0.0248, 0.0297, 0.0341, 0.0409, 0.0394, 0.0290, 0.0717, 0.0363], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0169, 0.0189, 0.0185, 0.0225, 0.0191, 0.0290, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-27 23:37:20,959 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.349e+02 4.809e+02 5.887e+02 7.395e+02 2.123e+03, threshold=1.177e+03, percent-clipped=7.0 2023-04-27 23:37:45,744 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26950.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:37:46,399 INFO [train.py:904] (4/8) Epoch 3, batch 6650, loss[loss=0.2782, simple_loss=0.3429, pruned_loss=0.1067, over 15363.00 frames. ], tot_loss[loss=0.2823, simple_loss=0.3512, pruned_loss=0.1067, over 3103393.77 frames. ], batch size: 190, lr: 1.99e-02, grad_scale: 8.0 2023-04-27 23:37:51,226 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=26954.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:38:09,005 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26965.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:38:10,508 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-27 23:38:38,386 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.52 vs. limit=5.0 2023-04-27 23:39:03,588 INFO [train.py:904] (4/8) Epoch 3, batch 6700, loss[loss=0.2664, simple_loss=0.335, pruned_loss=0.09891, over 16642.00 frames. ], tot_loss[loss=0.2806, simple_loss=0.3491, pruned_loss=0.106, over 3113762.15 frames. ], batch size: 57, lr: 1.99e-02, grad_scale: 8.0 2023-04-27 23:39:04,132 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7962, 3.6489, 3.2200, 1.7179, 2.5895, 2.1202, 3.1485, 3.5284], device='cuda:4'), covar=tensor([0.0279, 0.0440, 0.0496, 0.1572, 0.0742, 0.0918, 0.0805, 0.0546], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0115, 0.0155, 0.0142, 0.0135, 0.0128, 0.0144, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-27 23:39:19,076 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=27011.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:39:57,110 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.121e+02 4.872e+02 5.906e+02 7.141e+02 1.703e+03, threshold=1.181e+03, percent-clipped=1.0 2023-04-27 23:40:21,184 INFO [train.py:904] (4/8) Epoch 3, batch 6750, loss[loss=0.3352, simple_loss=0.377, pruned_loss=0.1467, over 12031.00 frames. ], tot_loss[loss=0.2821, simple_loss=0.3494, pruned_loss=0.1074, over 3096412.47 frames. ], batch size: 247, lr: 1.99e-02, grad_scale: 8.0 2023-04-27 23:40:51,196 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=27070.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:41:37,189 INFO [train.py:904] (4/8) Epoch 3, batch 6800, loss[loss=0.2508, simple_loss=0.3379, pruned_loss=0.08184, over 16939.00 frames. ], tot_loss[loss=0.2811, simple_loss=0.3488, pruned_loss=0.1067, over 3095456.50 frames. ], batch size: 109, lr: 1.98e-02, grad_scale: 8.0 2023-04-27 23:42:03,929 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=27118.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:42:18,367 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8955, 3.1393, 3.4409, 3.4200, 3.3970, 3.0856, 3.1873, 3.2914], device='cuda:4'), covar=tensor([0.0310, 0.0409, 0.0331, 0.0399, 0.0451, 0.0383, 0.0731, 0.0334], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0172, 0.0187, 0.0182, 0.0228, 0.0189, 0.0289, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-27 23:42:22,337 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4545, 3.4100, 3.2353, 3.3244, 2.9742, 3.3666, 3.1795, 3.2266], device='cuda:4'), covar=tensor([0.0323, 0.0202, 0.0183, 0.0136, 0.0544, 0.0181, 0.0843, 0.0257], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0128, 0.0167, 0.0140, 0.0201, 0.0159, 0.0126, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:42:31,272 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.734e+02 4.279e+02 5.417e+02 7.296e+02 1.152e+03, threshold=1.083e+03, percent-clipped=0.0 2023-04-27 23:42:39,140 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7004, 5.0662, 4.7521, 4.7385, 4.4026, 4.2695, 4.5446, 5.1015], device='cuda:4'), covar=tensor([0.0505, 0.0502, 0.0749, 0.0411, 0.0458, 0.0628, 0.0481, 0.0545], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0372, 0.0331, 0.0239, 0.0240, 0.0240, 0.0298, 0.0263], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:42:51,797 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=27148.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:42:55,393 INFO [train.py:904] (4/8) Epoch 3, batch 6850, loss[loss=0.2757, simple_loss=0.3708, pruned_loss=0.09025, over 16468.00 frames. ], tot_loss[loss=0.2813, simple_loss=0.35, pruned_loss=0.1063, over 3112921.57 frames. ], batch size: 146, lr: 1.98e-02, grad_scale: 8.0 2023-04-27 23:42:55,805 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4552, 4.5836, 4.5993, 4.5745, 4.4025, 5.1364, 4.7675, 4.4304], device='cuda:4'), covar=tensor([0.0859, 0.1270, 0.1102, 0.1544, 0.2321, 0.0883, 0.0983, 0.1848], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0320, 0.0301, 0.0275, 0.0362, 0.0325, 0.0264, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:42:59,125 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6119, 3.7809, 3.5585, 5.3661, 5.2679, 4.9887, 2.7499, 3.9686], device='cuda:4'), covar=tensor([0.0958, 0.0358, 0.0587, 0.0032, 0.0102, 0.0122, 0.0766, 0.0418], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0130, 0.0160, 0.0069, 0.0133, 0.0140, 0.0149, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 23:44:10,166 INFO [train.py:904] (4/8) Epoch 3, batch 6900, loss[loss=0.343, simple_loss=0.3939, pruned_loss=0.1461, over 15303.00 frames. ], tot_loss[loss=0.2806, simple_loss=0.3517, pruned_loss=0.1047, over 3128700.20 frames. ], batch size: 191, lr: 1.98e-02, grad_scale: 8.0 2023-04-27 23:44:22,736 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=27209.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:45:02,578 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.150e+02 4.514e+02 5.203e+02 6.904e+02 1.170e+03, threshold=1.041e+03, percent-clipped=2.0 2023-04-27 23:45:28,454 INFO [train.py:904] (4/8) Epoch 3, batch 6950, loss[loss=0.2591, simple_loss=0.3385, pruned_loss=0.08981, over 16884.00 frames. ], tot_loss[loss=0.2835, simple_loss=0.3536, pruned_loss=0.1067, over 3132139.48 frames. ], batch size: 96, lr: 1.98e-02, grad_scale: 8.0 2023-04-27 23:45:41,656 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=27260.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 23:46:43,370 INFO [train.py:904] (4/8) Epoch 3, batch 7000, loss[loss=0.2612, simple_loss=0.3523, pruned_loss=0.08509, over 16837.00 frames. ], tot_loss[loss=0.2827, simple_loss=0.3535, pruned_loss=0.1059, over 3120163.53 frames. ], batch size: 83, lr: 1.98e-02, grad_scale: 8.0 2023-04-27 23:46:51,663 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=27306.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:47:35,868 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.986e+02 5.041e+02 6.145e+02 8.860e+02 1.565e+03, threshold=1.229e+03, percent-clipped=7.0 2023-04-27 23:48:01,107 INFO [train.py:904] (4/8) Epoch 3, batch 7050, loss[loss=0.2665, simple_loss=0.349, pruned_loss=0.09197, over 16309.00 frames. ], tot_loss[loss=0.2837, simple_loss=0.3549, pruned_loss=0.1062, over 3121371.49 frames. ], batch size: 146, lr: 1.98e-02, grad_scale: 8.0 2023-04-27 23:49:19,631 INFO [train.py:904] (4/8) Epoch 3, batch 7100, loss[loss=0.2964, simple_loss=0.3677, pruned_loss=0.1126, over 17124.00 frames. ], tot_loss[loss=0.2821, simple_loss=0.3526, pruned_loss=0.1058, over 3126634.45 frames. ], batch size: 48, lr: 1.97e-02, grad_scale: 8.0 2023-04-27 23:50:12,101 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.055e+02 4.788e+02 6.087e+02 7.690e+02 2.114e+03, threshold=1.217e+03, percent-clipped=3.0 2023-04-27 23:50:36,340 INFO [train.py:904] (4/8) Epoch 3, batch 7150, loss[loss=0.2741, simple_loss=0.3424, pruned_loss=0.1029, over 16872.00 frames. ], tot_loss[loss=0.2806, simple_loss=0.3501, pruned_loss=0.1055, over 3111936.49 frames. ], batch size: 109, lr: 1.97e-02, grad_scale: 8.0 2023-04-27 23:51:51,150 INFO [train.py:904] (4/8) Epoch 3, batch 7200, loss[loss=0.2428, simple_loss=0.3236, pruned_loss=0.08097, over 15274.00 frames. ], tot_loss[loss=0.278, simple_loss=0.348, pruned_loss=0.104, over 3096835.89 frames. ], batch size: 190, lr: 1.97e-02, grad_scale: 8.0 2023-04-27 23:51:55,799 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=27504.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:52:45,515 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.403e+02 3.669e+02 4.524e+02 6.083e+02 1.066e+03, threshold=9.047e+02, percent-clipped=1.0 2023-04-27 23:53:12,423 INFO [train.py:904] (4/8) Epoch 3, batch 7250, loss[loss=0.2518, simple_loss=0.3216, pruned_loss=0.09103, over 16136.00 frames. ], tot_loss[loss=0.2739, simple_loss=0.3445, pruned_loss=0.1016, over 3114383.30 frames. ], batch size: 165, lr: 1.97e-02, grad_scale: 8.0 2023-04-27 23:53:26,457 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=27560.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:54:02,256 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4083, 4.5673, 1.5018, 4.6816, 2.8761, 4.6598, 2.2278, 3.0665], device='cuda:4'), covar=tensor([0.0055, 0.0115, 0.1670, 0.0023, 0.0641, 0.0224, 0.1221, 0.0494], device='cuda:4'), in_proj_covar=tensor([0.0087, 0.0130, 0.0170, 0.0079, 0.0160, 0.0161, 0.0178, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-27 23:54:26,523 INFO [train.py:904] (4/8) Epoch 3, batch 7300, loss[loss=0.2765, simple_loss=0.3474, pruned_loss=0.1029, over 16661.00 frames. ], tot_loss[loss=0.2713, simple_loss=0.3424, pruned_loss=0.1001, over 3121560.29 frames. ], batch size: 134, lr: 1.97e-02, grad_scale: 8.0 2023-04-27 23:54:35,477 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=27606.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:54:38,057 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=27608.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:55:17,368 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.622e+02 4.400e+02 5.943e+02 7.946e+02 1.369e+03, threshold=1.189e+03, percent-clipped=13.0 2023-04-27 23:55:40,973 INFO [train.py:904] (4/8) Epoch 3, batch 7350, loss[loss=0.3318, simple_loss=0.3687, pruned_loss=0.1474, over 10857.00 frames. ], tot_loss[loss=0.2716, simple_loss=0.3423, pruned_loss=0.1004, over 3107266.60 frames. ], batch size: 246, lr: 1.97e-02, grad_scale: 8.0 2023-04-27 23:55:46,793 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=27654.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:56:38,598 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3391, 5.2332, 5.0679, 3.9888, 5.1063, 1.9421, 4.8450, 5.0804], device='cuda:4'), covar=tensor([0.0059, 0.0042, 0.0059, 0.0343, 0.0049, 0.1383, 0.0065, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0073, 0.0062, 0.0094, 0.0110, 0.0069, 0.0120, 0.0082, 0.0093], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:56:59,424 INFO [train.py:904] (4/8) Epoch 3, batch 7400, loss[loss=0.2652, simple_loss=0.3382, pruned_loss=0.0961, over 16850.00 frames. ], tot_loss[loss=0.2735, simple_loss=0.344, pruned_loss=0.1015, over 3111545.62 frames. ], batch size: 102, lr: 1.96e-02, grad_scale: 4.0 2023-04-27 23:57:20,344 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=27714.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:57:42,205 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3632, 4.1569, 3.6082, 1.7708, 2.8102, 2.2978, 3.6796, 4.0251], device='cuda:4'), covar=tensor([0.0224, 0.0473, 0.0560, 0.1794, 0.0837, 0.0999, 0.0566, 0.0580], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0122, 0.0161, 0.0150, 0.0143, 0.0134, 0.0152, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-27 23:57:55,129 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.894e+02 4.696e+02 5.728e+02 6.981e+02 1.392e+03, threshold=1.146e+03, percent-clipped=2.0 2023-04-27 23:58:18,284 INFO [train.py:904] (4/8) Epoch 3, batch 7450, loss[loss=0.3147, simple_loss=0.3628, pruned_loss=0.1333, over 11492.00 frames. ], tot_loss[loss=0.2761, simple_loss=0.3458, pruned_loss=0.1032, over 3102254.93 frames. ], batch size: 246, lr: 1.96e-02, grad_scale: 4.0 2023-04-27 23:58:58,352 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=27775.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:59:39,385 INFO [train.py:904] (4/8) Epoch 3, batch 7500, loss[loss=0.2645, simple_loss=0.3378, pruned_loss=0.09563, over 16927.00 frames. ], tot_loss[loss=0.2768, simple_loss=0.3471, pruned_loss=0.1033, over 3104464.50 frames. ], batch size: 109, lr: 1.96e-02, grad_scale: 4.0 2023-04-27 23:59:44,226 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=27804.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:00:09,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0125, 4.7589, 4.9740, 5.2851, 5.3170, 4.6663, 5.3727, 5.1812], device='cuda:4'), covar=tensor([0.0573, 0.0570, 0.0916, 0.0329, 0.0319, 0.0422, 0.0263, 0.0389], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0361, 0.0462, 0.0357, 0.0271, 0.0255, 0.0292, 0.0293], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:00:33,135 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.883e+02 4.957e+02 5.924e+02 7.629e+02 1.550e+03, threshold=1.185e+03, percent-clipped=6.0 2023-04-28 00:00:49,076 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5141, 4.3243, 4.2988, 4.3727, 3.9065, 4.4121, 4.3263, 4.1204], device='cuda:4'), covar=tensor([0.0334, 0.0210, 0.0165, 0.0123, 0.0587, 0.0187, 0.0210, 0.0261], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0129, 0.0165, 0.0139, 0.0193, 0.0156, 0.0123, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:00:55,939 INFO [train.py:904] (4/8) Epoch 3, batch 7550, loss[loss=0.2825, simple_loss=0.3452, pruned_loss=0.1099, over 15269.00 frames. ], tot_loss[loss=0.2763, simple_loss=0.3459, pruned_loss=0.1034, over 3098168.58 frames. ], batch size: 191, lr: 1.96e-02, grad_scale: 4.0 2023-04-28 00:00:58,762 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=27852.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:01:32,053 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9793, 1.6136, 1.4051, 1.4389, 1.7585, 1.6846, 1.7815, 1.8950], device='cuda:4'), covar=tensor([0.0022, 0.0090, 0.0123, 0.0119, 0.0062, 0.0096, 0.0046, 0.0058], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0118, 0.0123, 0.0121, 0.0113, 0.0122, 0.0078, 0.0098], device='cuda:4'), out_proj_covar=tensor([7.2548e-05, 1.6729e-04, 1.6853e-04, 1.6907e-04, 1.6317e-04, 1.7405e-04, 1.0956e-04, 1.4193e-04], device='cuda:4') 2023-04-28 00:02:09,943 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-28 00:02:13,359 INFO [train.py:904] (4/8) Epoch 3, batch 7600, loss[loss=0.3607, simple_loss=0.3917, pruned_loss=0.1648, over 11173.00 frames. ], tot_loss[loss=0.2764, simple_loss=0.3454, pruned_loss=0.1037, over 3104047.38 frames. ], batch size: 247, lr: 1.96e-02, grad_scale: 8.0 2023-04-28 00:03:08,896 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.035e+02 4.816e+02 6.068e+02 7.688e+02 1.978e+03, threshold=1.214e+03, percent-clipped=6.0 2023-04-28 00:03:31,741 INFO [train.py:904] (4/8) Epoch 3, batch 7650, loss[loss=0.2719, simple_loss=0.3476, pruned_loss=0.09807, over 16572.00 frames. ], tot_loss[loss=0.2782, simple_loss=0.3468, pruned_loss=0.1048, over 3109868.39 frames. ], batch size: 57, lr: 1.96e-02, grad_scale: 8.0 2023-04-28 00:03:58,714 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-28 00:04:45,520 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0640, 3.8762, 4.0350, 4.3086, 4.3428, 3.9131, 4.3408, 4.3420], device='cuda:4'), covar=tensor([0.0610, 0.0593, 0.1025, 0.0403, 0.0416, 0.0711, 0.0400, 0.0370], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0357, 0.0454, 0.0350, 0.0268, 0.0251, 0.0286, 0.0289], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:04:52,687 INFO [train.py:904] (4/8) Epoch 3, batch 7700, loss[loss=0.274, simple_loss=0.3482, pruned_loss=0.09987, over 16815.00 frames. ], tot_loss[loss=0.2794, simple_loss=0.3472, pruned_loss=0.1058, over 3093342.60 frames. ], batch size: 102, lr: 1.95e-02, grad_scale: 8.0 2023-04-28 00:04:59,411 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2987, 3.3440, 1.6298, 3.3501, 2.2478, 3.4920, 1.6466, 2.5554], device='cuda:4'), covar=tensor([0.0080, 0.0219, 0.1461, 0.0064, 0.0791, 0.0333, 0.1439, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0089, 0.0134, 0.0175, 0.0081, 0.0164, 0.0163, 0.0181, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 00:05:46,975 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.821e+02 4.680e+02 5.750e+02 6.748e+02 1.214e+03, threshold=1.150e+03, percent-clipped=1.0 2023-04-28 00:06:11,242 INFO [train.py:904] (4/8) Epoch 3, batch 7750, loss[loss=0.3393, simple_loss=0.381, pruned_loss=0.1488, over 11234.00 frames. ], tot_loss[loss=0.2801, simple_loss=0.3479, pruned_loss=0.1062, over 3085384.58 frames. ], batch size: 248, lr: 1.95e-02, grad_scale: 8.0 2023-04-28 00:06:40,880 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28070.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 00:07:28,338 INFO [train.py:904] (4/8) Epoch 3, batch 7800, loss[loss=0.3828, simple_loss=0.4101, pruned_loss=0.1778, over 11489.00 frames. ], tot_loss[loss=0.2829, simple_loss=0.3497, pruned_loss=0.108, over 3074291.85 frames. ], batch size: 248, lr: 1.95e-02, grad_scale: 8.0 2023-04-28 00:08:22,843 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.157e+02 4.949e+02 5.902e+02 7.570e+02 1.555e+03, threshold=1.180e+03, percent-clipped=4.0 2023-04-28 00:08:45,060 INFO [train.py:904] (4/8) Epoch 3, batch 7850, loss[loss=0.3473, simple_loss=0.397, pruned_loss=0.1488, over 11491.00 frames. ], tot_loss[loss=0.2838, simple_loss=0.351, pruned_loss=0.1083, over 3052482.13 frames. ], batch size: 248, lr: 1.95e-02, grad_scale: 8.0 2023-04-28 00:08:55,226 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-28 00:10:00,843 INFO [train.py:904] (4/8) Epoch 3, batch 7900, loss[loss=0.2963, simple_loss=0.3675, pruned_loss=0.1126, over 16199.00 frames. ], tot_loss[loss=0.2816, simple_loss=0.3493, pruned_loss=0.1069, over 3062873.49 frames. ], batch size: 165, lr: 1.95e-02, grad_scale: 8.0 2023-04-28 00:10:45,244 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28229.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:10:55,796 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.692e+02 4.334e+02 4.942e+02 5.999e+02 2.073e+03, threshold=9.884e+02, percent-clipped=3.0 2023-04-28 00:11:18,498 INFO [train.py:904] (4/8) Epoch 3, batch 7950, loss[loss=0.3522, simple_loss=0.3899, pruned_loss=0.1572, over 11748.00 frames. ], tot_loss[loss=0.2819, simple_loss=0.3493, pruned_loss=0.1073, over 3048627.93 frames. ], batch size: 246, lr: 1.94e-02, grad_scale: 8.0 2023-04-28 00:11:19,561 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1432, 4.9912, 4.8667, 4.9925, 4.3863, 5.0003, 4.9212, 4.6798], device='cuda:4'), covar=tensor([0.0367, 0.0224, 0.0188, 0.0138, 0.0802, 0.0225, 0.0163, 0.0333], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0133, 0.0170, 0.0142, 0.0203, 0.0162, 0.0126, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 00:11:26,015 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28255.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 00:12:13,418 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28287.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:12:18,753 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28290.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:12:33,514 INFO [train.py:904] (4/8) Epoch 3, batch 8000, loss[loss=0.2593, simple_loss=0.3327, pruned_loss=0.09295, over 16563.00 frames. ], tot_loss[loss=0.2823, simple_loss=0.3494, pruned_loss=0.1076, over 3055627.13 frames. ], batch size: 62, lr: 1.94e-02, grad_scale: 8.0 2023-04-28 00:12:56,308 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28316.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 00:13:27,063 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.455e+02 4.065e+02 5.202e+02 6.886e+02 1.574e+03, threshold=1.040e+03, percent-clipped=4.0 2023-04-28 00:13:46,476 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28348.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:13:49,305 INFO [train.py:904] (4/8) Epoch 3, batch 8050, loss[loss=0.2738, simple_loss=0.3503, pruned_loss=0.09866, over 16337.00 frames. ], tot_loss[loss=0.2815, simple_loss=0.3486, pruned_loss=0.1072, over 3045046.38 frames. ], batch size: 146, lr: 1.94e-02, grad_scale: 8.0 2023-04-28 00:14:18,866 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=28370.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 00:14:29,443 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0563, 3.9802, 4.5039, 4.4861, 4.4784, 4.1084, 4.1391, 4.0185], device='cuda:4'), covar=tensor([0.0226, 0.0261, 0.0249, 0.0335, 0.0394, 0.0231, 0.0673, 0.0312], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0178, 0.0192, 0.0193, 0.0231, 0.0198, 0.0294, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 00:15:05,403 INFO [train.py:904] (4/8) Epoch 3, batch 8100, loss[loss=0.2995, simple_loss=0.3474, pruned_loss=0.1258, over 11832.00 frames. ], tot_loss[loss=0.2801, simple_loss=0.348, pruned_loss=0.1061, over 3059599.79 frames. ], batch size: 246, lr: 1.94e-02, grad_scale: 8.0 2023-04-28 00:15:30,570 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=28418.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 00:15:57,041 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.864e+02 4.688e+02 5.875e+02 7.578e+02 1.406e+03, threshold=1.175e+03, percent-clipped=5.0 2023-04-28 00:16:20,001 INFO [train.py:904] (4/8) Epoch 3, batch 8150, loss[loss=0.2704, simple_loss=0.3408, pruned_loss=0.09999, over 15305.00 frames. ], tot_loss[loss=0.277, simple_loss=0.3453, pruned_loss=0.1043, over 3071726.86 frames. ], batch size: 191, lr: 1.94e-02, grad_scale: 8.0 2023-04-28 00:17:15,190 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28486.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:17:23,223 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28492.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:17:36,817 INFO [train.py:904] (4/8) Epoch 3, batch 8200, loss[loss=0.2538, simple_loss=0.338, pruned_loss=0.08478, over 16774.00 frames. ], tot_loss[loss=0.274, simple_loss=0.3423, pruned_loss=0.1028, over 3089311.86 frames. ], batch size: 83, lr: 1.94e-02, grad_scale: 8.0 2023-04-28 00:18:12,541 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9620, 2.6511, 2.3875, 4.6565, 1.9828, 4.1802, 2.6804, 2.5165], device='cuda:4'), covar=tensor([0.0338, 0.0943, 0.0619, 0.0178, 0.2165, 0.0294, 0.0996, 0.1638], device='cuda:4'), in_proj_covar=tensor([0.0269, 0.0245, 0.0205, 0.0272, 0.0322, 0.0218, 0.0235, 0.0309], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:18:31,948 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.657e+02 4.728e+02 5.816e+02 7.305e+02 1.524e+03, threshold=1.163e+03, percent-clipped=3.0 2023-04-28 00:18:50,500 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28547.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:18:55,850 INFO [train.py:904] (4/8) Epoch 3, batch 8250, loss[loss=0.2621, simple_loss=0.3416, pruned_loss=0.09125, over 15246.00 frames. ], tot_loss[loss=0.2715, simple_loss=0.3418, pruned_loss=0.1006, over 3075966.13 frames. ], batch size: 190, lr: 1.94e-02, grad_scale: 8.0 2023-04-28 00:19:00,224 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28553.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:19:51,757 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28585.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:20:16,983 INFO [train.py:904] (4/8) Epoch 3, batch 8300, loss[loss=0.2472, simple_loss=0.3097, pruned_loss=0.09232, over 12097.00 frames. ], tot_loss[loss=0.2645, simple_loss=0.3371, pruned_loss=0.09596, over 3063076.51 frames. ], batch size: 246, lr: 1.93e-02, grad_scale: 8.0 2023-04-28 00:20:33,879 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28611.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 00:21:14,648 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.491e+02 3.724e+02 4.538e+02 5.868e+02 1.330e+03, threshold=9.076e+02, percent-clipped=2.0 2023-04-28 00:21:27,050 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28643.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:21:39,705 INFO [train.py:904] (4/8) Epoch 3, batch 8350, loss[loss=0.2326, simple_loss=0.3243, pruned_loss=0.0704, over 16184.00 frames. ], tot_loss[loss=0.2601, simple_loss=0.335, pruned_loss=0.09265, over 3066865.43 frames. ], batch size: 165, lr: 1.93e-02, grad_scale: 8.0 2023-04-28 00:22:50,777 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7179, 4.3930, 4.6852, 4.9513, 5.0251, 4.4266, 5.0750, 5.0136], device='cuda:4'), covar=tensor([0.0590, 0.0660, 0.1054, 0.0360, 0.0355, 0.0467, 0.0329, 0.0288], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0352, 0.0447, 0.0346, 0.0264, 0.0253, 0.0280, 0.0284], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:23:00,530 INFO [train.py:904] (4/8) Epoch 3, batch 8400, loss[loss=0.2116, simple_loss=0.2938, pruned_loss=0.06467, over 15468.00 frames. ], tot_loss[loss=0.2546, simple_loss=0.3304, pruned_loss=0.08936, over 3040504.23 frames. ], batch size: 191, lr: 1.93e-02, grad_scale: 8.0 2023-04-28 00:23:58,229 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.191e+02 3.936e+02 4.649e+02 5.360e+02 9.184e+02, threshold=9.299e+02, percent-clipped=1.0 2023-04-28 00:24:20,120 INFO [train.py:904] (4/8) Epoch 3, batch 8450, loss[loss=0.2619, simple_loss=0.3396, pruned_loss=0.0921, over 16694.00 frames. ], tot_loss[loss=0.2516, simple_loss=0.3282, pruned_loss=0.08751, over 3040269.26 frames. ], batch size: 124, lr: 1.93e-02, grad_scale: 4.0 2023-04-28 00:25:42,010 INFO [train.py:904] (4/8) Epoch 3, batch 8500, loss[loss=0.199, simple_loss=0.2875, pruned_loss=0.05525, over 16741.00 frames. ], tot_loss[loss=0.2448, simple_loss=0.3226, pruned_loss=0.08353, over 3028708.74 frames. ], batch size: 83, lr: 1.93e-02, grad_scale: 4.0 2023-04-28 00:26:40,691 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.012e+02 3.668e+02 4.634e+02 5.862e+02 2.485e+03, threshold=9.268e+02, percent-clipped=6.0 2023-04-28 00:26:50,001 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28842.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:27:00,694 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28848.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:27:05,698 INFO [train.py:904] (4/8) Epoch 3, batch 8550, loss[loss=0.2393, simple_loss=0.3312, pruned_loss=0.0737, over 16760.00 frames. ], tot_loss[loss=0.2422, simple_loss=0.3201, pruned_loss=0.08221, over 3034103.20 frames. ], batch size: 124, lr: 1.93e-02, grad_scale: 4.0 2023-04-28 00:27:26,036 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28861.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:28:09,048 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0064, 3.7962, 2.6293, 5.1239, 4.9302, 4.5703, 2.4430, 3.1193], device='cuda:4'), covar=tensor([0.1351, 0.0449, 0.1195, 0.0068, 0.0163, 0.0325, 0.1069, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0133, 0.0161, 0.0071, 0.0134, 0.0143, 0.0153, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-04-28 00:28:12,810 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=28885.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:28:46,470 INFO [train.py:904] (4/8) Epoch 3, batch 8600, loss[loss=0.2057, simple_loss=0.2963, pruned_loss=0.05759, over 16503.00 frames. ], tot_loss[loss=0.2408, simple_loss=0.3202, pruned_loss=0.08072, over 3047643.74 frames. ], batch size: 68, lr: 1.92e-02, grad_scale: 4.0 2023-04-28 00:29:07,396 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=28911.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 00:29:29,394 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5614, 3.3512, 3.2986, 3.7483, 3.7473, 3.4622, 3.8564, 3.7811], device='cuda:4'), covar=tensor([0.0817, 0.0886, 0.1889, 0.0761, 0.0834, 0.1163, 0.0615, 0.0739], device='cuda:4'), in_proj_covar=tensor([0.0277, 0.0345, 0.0438, 0.0338, 0.0257, 0.0244, 0.0280, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:29:30,852 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28922.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:29:50,787 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=28933.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:29:57,553 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.629e+02 3.755e+02 4.530e+02 5.746e+02 8.888e+02, threshold=9.061e+02, percent-clipped=0.0 2023-04-28 00:30:11,245 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=28943.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:30:25,441 INFO [train.py:904] (4/8) Epoch 3, batch 8650, loss[loss=0.2294, simple_loss=0.3172, pruned_loss=0.07082, over 16366.00 frames. ], tot_loss[loss=0.2364, simple_loss=0.3171, pruned_loss=0.07789, over 3051117.27 frames. ], batch size: 146, lr: 1.92e-02, grad_scale: 4.0 2023-04-28 00:30:41,308 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.23 vs. limit=5.0 2023-04-28 00:30:45,828 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=28959.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 00:31:52,872 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=28991.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:32:12,024 INFO [train.py:904] (4/8) Epoch 3, batch 8700, loss[loss=0.2272, simple_loss=0.3058, pruned_loss=0.07431, over 16829.00 frames. ], tot_loss[loss=0.2326, simple_loss=0.3135, pruned_loss=0.07586, over 3044622.32 frames. ], batch size: 124, lr: 1.92e-02, grad_scale: 4.0 2023-04-28 00:32:15,164 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29002.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:32:23,772 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-28 00:33:18,499 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.924e+02 3.379e+02 3.929e+02 4.901e+02 7.493e+02, threshold=7.858e+02, percent-clipped=0.0 2023-04-28 00:33:46,755 INFO [train.py:904] (4/8) Epoch 3, batch 8750, loss[loss=0.2042, simple_loss=0.2857, pruned_loss=0.0613, over 12362.00 frames. ], tot_loss[loss=0.2317, simple_loss=0.3131, pruned_loss=0.07518, over 3031326.34 frames. ], batch size: 248, lr: 1.92e-02, grad_scale: 4.0 2023-04-28 00:34:19,064 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29063.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:34:51,596 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-28 00:35:07,380 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2047, 3.9542, 3.9383, 1.7997, 3.1036, 2.3688, 3.5630, 3.8901], device='cuda:4'), covar=tensor([0.0248, 0.0443, 0.0318, 0.1574, 0.0607, 0.0852, 0.0610, 0.0574], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0112, 0.0151, 0.0146, 0.0134, 0.0130, 0.0142, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 00:35:38,457 INFO [train.py:904] (4/8) Epoch 3, batch 8800, loss[loss=0.2721, simple_loss=0.3441, pruned_loss=0.1, over 13056.00 frames. ], tot_loss[loss=0.2291, simple_loss=0.3109, pruned_loss=0.07367, over 3032325.61 frames. ], batch size: 250, lr: 1.92e-02, grad_scale: 8.0 2023-04-28 00:36:03,776 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7276, 2.3406, 2.3818, 4.2299, 1.7035, 3.7491, 2.3095, 2.1468], device='cuda:4'), covar=tensor([0.0325, 0.1077, 0.0524, 0.0157, 0.2295, 0.0292, 0.1053, 0.1674], device='cuda:4'), in_proj_covar=tensor([0.0258, 0.0242, 0.0202, 0.0259, 0.0315, 0.0213, 0.0229, 0.0291], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:36:43,041 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 00:36:52,045 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.420e+02 3.673e+02 4.669e+02 5.868e+02 1.100e+03, threshold=9.339e+02, percent-clipped=8.0 2023-04-28 00:37:04,020 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29142.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:37:17,025 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29148.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:37:21,268 INFO [train.py:904] (4/8) Epoch 3, batch 8850, loss[loss=0.2323, simple_loss=0.3304, pruned_loss=0.06709, over 16385.00 frames. ], tot_loss[loss=0.2296, simple_loss=0.3134, pruned_loss=0.0729, over 3029425.99 frames. ], batch size: 146, lr: 1.92e-02, grad_scale: 8.0 2023-04-28 00:38:34,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7029, 1.8648, 1.6085, 1.6060, 2.3856, 2.1692, 2.5778, 2.5578], device='cuda:4'), covar=tensor([0.0018, 0.0179, 0.0179, 0.0205, 0.0087, 0.0139, 0.0050, 0.0069], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0125, 0.0123, 0.0125, 0.0113, 0.0123, 0.0076, 0.0099], device='cuda:4'), out_proj_covar=tensor([6.9048e-05, 1.7426e-04, 1.6625e-04, 1.7298e-04, 1.5912e-04, 1.7327e-04, 1.0472e-04, 1.3998e-04], device='cuda:4') 2023-04-28 00:38:45,756 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=29190.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:38:57,573 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=29196.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:39:06,873 INFO [train.py:904] (4/8) Epoch 3, batch 8900, loss[loss=0.2189, simple_loss=0.3067, pruned_loss=0.06554, over 16293.00 frames. ], tot_loss[loss=0.2279, simple_loss=0.3124, pruned_loss=0.07176, over 3026414.05 frames. ], batch size: 165, lr: 1.91e-02, grad_scale: 8.0 2023-04-28 00:39:40,209 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=29217.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:40:35,243 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29236.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:40:36,245 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.258e+02 3.766e+02 4.506e+02 5.745e+02 1.186e+03, threshold=9.012e+02, percent-clipped=1.0 2023-04-28 00:41:11,852 INFO [train.py:904] (4/8) Epoch 3, batch 8950, loss[loss=0.2071, simple_loss=0.2982, pruned_loss=0.058, over 16488.00 frames. ], tot_loss[loss=0.2288, simple_loss=0.3127, pruned_loss=0.0724, over 3049976.60 frames. ], batch size: 68, lr: 1.91e-02, grad_scale: 8.0 2023-04-28 00:42:54,125 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29297.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:43:00,671 INFO [train.py:904] (4/8) Epoch 3, batch 9000, loss[loss=0.2126, simple_loss=0.298, pruned_loss=0.06364, over 16795.00 frames. ], tot_loss[loss=0.2253, simple_loss=0.3093, pruned_loss=0.07062, over 3054442.74 frames. ], batch size: 124, lr: 1.91e-02, grad_scale: 8.0 2023-04-28 00:43:00,672 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 00:43:11,837 INFO [train.py:938] (4/8) Epoch 3, validation: loss=0.1886, simple_loss=0.2904, pruned_loss=0.04341, over 944034.00 frames. 2023-04-28 00:43:11,839 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 00:43:20,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7505, 5.0362, 5.1653, 5.0692, 5.1022, 5.5479, 5.2462, 4.9796], device='cuda:4'), covar=tensor([0.0643, 0.1233, 0.1017, 0.1481, 0.1897, 0.0813, 0.0969, 0.1965], device='cuda:4'), in_proj_covar=tensor([0.0219, 0.0304, 0.0289, 0.0269, 0.0347, 0.0317, 0.0252, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:44:05,746 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7972, 2.7473, 1.5115, 2.8015, 2.2393, 2.7986, 1.7860, 2.4497], device='cuda:4'), covar=tensor([0.0087, 0.0259, 0.1270, 0.0059, 0.0537, 0.0479, 0.1241, 0.0508], device='cuda:4'), in_proj_covar=tensor([0.0087, 0.0126, 0.0173, 0.0077, 0.0153, 0.0152, 0.0181, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 00:44:26,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.245e+02 3.595e+02 4.170e+02 5.160e+02 8.461e+02, threshold=8.339e+02, percent-clipped=0.0 2023-04-28 00:44:57,071 INFO [train.py:904] (4/8) Epoch 3, batch 9050, loss[loss=0.2323, simple_loss=0.3086, pruned_loss=0.07798, over 16165.00 frames. ], tot_loss[loss=0.2265, simple_loss=0.3103, pruned_loss=0.07135, over 3069429.75 frames. ], batch size: 165, lr: 1.91e-02, grad_scale: 8.0 2023-04-28 00:45:12,126 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=29358.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:45:25,793 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.71 vs. limit=5.0 2023-04-28 00:46:12,764 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7457, 3.5099, 3.3115, 3.9332, 3.9777, 3.5374, 3.9699, 3.9191], device='cuda:4'), covar=tensor([0.0651, 0.0753, 0.1895, 0.0650, 0.0599, 0.1306, 0.0643, 0.0664], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0354, 0.0444, 0.0348, 0.0261, 0.0249, 0.0277, 0.0287], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:46:41,635 INFO [train.py:904] (4/8) Epoch 3, batch 9100, loss[loss=0.2196, simple_loss=0.3161, pruned_loss=0.06157, over 16880.00 frames. ], tot_loss[loss=0.2268, simple_loss=0.3098, pruned_loss=0.07192, over 3066475.34 frames. ], batch size: 102, lr: 1.91e-02, grad_scale: 8.0 2023-04-28 00:46:50,537 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8243, 3.6767, 4.2758, 4.2215, 4.2743, 3.8269, 4.0129, 3.8836], device='cuda:4'), covar=tensor([0.0238, 0.0346, 0.0355, 0.0431, 0.0326, 0.0282, 0.0618, 0.0302], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0169, 0.0176, 0.0182, 0.0213, 0.0188, 0.0273, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-04-28 00:48:08,505 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.097e+02 4.371e+02 5.293e+02 6.986e+02 1.215e+03, threshold=1.059e+03, percent-clipped=12.0 2023-04-28 00:48:40,640 INFO [train.py:904] (4/8) Epoch 3, batch 9150, loss[loss=0.2115, simple_loss=0.2988, pruned_loss=0.06205, over 15448.00 frames. ], tot_loss[loss=0.2259, simple_loss=0.3101, pruned_loss=0.07087, over 3066624.09 frames. ], batch size: 191, lr: 1.91e-02, grad_scale: 8.0 2023-04-28 00:48:43,970 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6463, 1.8792, 1.3391, 1.4319, 2.3821, 2.1187, 2.5773, 2.6278], device='cuda:4'), covar=tensor([0.0017, 0.0120, 0.0178, 0.0176, 0.0067, 0.0106, 0.0037, 0.0049], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0122, 0.0123, 0.0124, 0.0114, 0.0123, 0.0075, 0.0098], device='cuda:4'), out_proj_covar=tensor([6.8294e-05, 1.6949e-04, 1.6661e-04, 1.6981e-04, 1.5927e-04, 1.7229e-04, 1.0249e-04, 1.3836e-04], device='cuda:4') 2023-04-28 00:50:24,877 INFO [train.py:904] (4/8) Epoch 3, batch 9200, loss[loss=0.2097, simple_loss=0.2844, pruned_loss=0.0675, over 11876.00 frames. ], tot_loss[loss=0.2223, simple_loss=0.305, pruned_loss=0.06977, over 3054735.44 frames. ], batch size: 247, lr: 1.90e-02, grad_scale: 8.0 2023-04-28 00:50:54,920 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29517.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:51:32,420 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.365e+02 4.043e+02 4.899e+02 6.737e+02 1.438e+03, threshold=9.797e+02, percent-clipped=1.0 2023-04-28 00:52:01,382 INFO [train.py:904] (4/8) Epoch 3, batch 9250, loss[loss=0.2019, simple_loss=0.2908, pruned_loss=0.05646, over 16659.00 frames. ], tot_loss[loss=0.2231, simple_loss=0.3057, pruned_loss=0.07026, over 3058227.96 frames. ], batch size: 134, lr: 1.90e-02, grad_scale: 8.0 2023-04-28 00:52:02,501 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29551.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:52:29,837 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=29565.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:53:31,901 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=29592.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:53:50,315 INFO [train.py:904] (4/8) Epoch 3, batch 9300, loss[loss=0.2027, simple_loss=0.283, pruned_loss=0.0612, over 17026.00 frames. ], tot_loss[loss=0.2205, simple_loss=0.3032, pruned_loss=0.06891, over 3036512.75 frames. ], batch size: 41, lr: 1.90e-02, grad_scale: 4.0 2023-04-28 00:54:15,988 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29612.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:54:58,346 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6905, 2.6894, 2.6444, 1.9020, 2.6091, 2.5096, 2.6601, 1.8819], device='cuda:4'), covar=tensor([0.0274, 0.0028, 0.0043, 0.0201, 0.0043, 0.0047, 0.0034, 0.0262], device='cuda:4'), in_proj_covar=tensor([0.0111, 0.0052, 0.0056, 0.0108, 0.0053, 0.0059, 0.0056, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 00:55:11,308 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.186e+02 3.534e+02 4.207e+02 4.979e+02 1.086e+03, threshold=8.414e+02, percent-clipped=1.0 2023-04-28 00:55:35,075 INFO [train.py:904] (4/8) Epoch 3, batch 9350, loss[loss=0.238, simple_loss=0.3253, pruned_loss=0.07539, over 15205.00 frames. ], tot_loss[loss=0.2203, simple_loss=0.3029, pruned_loss=0.06881, over 3043526.92 frames. ], batch size: 190, lr: 1.90e-02, grad_scale: 4.0 2023-04-28 00:55:51,538 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29658.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:56:00,152 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29662.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:56:45,906 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 00:57:16,961 INFO [train.py:904] (4/8) Epoch 3, batch 9400, loss[loss=0.2442, simple_loss=0.3279, pruned_loss=0.08027, over 15455.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.3027, pruned_loss=0.06879, over 3036355.76 frames. ], batch size: 191, lr: 1.90e-02, grad_scale: 4.0 2023-04-28 00:57:27,902 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=29706.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:57:50,641 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4703, 4.3948, 4.2995, 3.8614, 4.2993, 1.7377, 4.0824, 4.2034], device='cuda:4'), covar=tensor([0.0038, 0.0037, 0.0057, 0.0150, 0.0041, 0.1420, 0.0054, 0.0089], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0060, 0.0094, 0.0094, 0.0068, 0.0120, 0.0082, 0.0090], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:58:01,489 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29723.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 00:58:32,710 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.403e+02 3.993e+02 5.001e+02 6.201e+02 1.324e+03, threshold=1.000e+03, percent-clipped=7.0 2023-04-28 00:58:58,180 INFO [train.py:904] (4/8) Epoch 3, batch 9450, loss[loss=0.2466, simple_loss=0.3197, pruned_loss=0.0867, over 12543.00 frames. ], tot_loss[loss=0.2227, simple_loss=0.3056, pruned_loss=0.06993, over 3035206.86 frames. ], batch size: 249, lr: 1.90e-02, grad_scale: 4.0 2023-04-28 01:00:19,635 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-28 01:00:37,284 INFO [train.py:904] (4/8) Epoch 3, batch 9500, loss[loss=0.224, simple_loss=0.3092, pruned_loss=0.06943, over 16439.00 frames. ], tot_loss[loss=0.2213, simple_loss=0.3046, pruned_loss=0.06904, over 3043553.82 frames. ], batch size: 146, lr: 1.90e-02, grad_scale: 4.0 2023-04-28 01:00:41,516 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29802.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:01:23,695 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.60 vs. limit=2.0 2023-04-28 01:01:45,335 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2384, 2.2264, 2.1416, 3.5574, 1.7847, 3.2312, 2.2065, 1.9482], device='cuda:4'), covar=tensor([0.0373, 0.1084, 0.0623, 0.0236, 0.2145, 0.0300, 0.1151, 0.1738], device='cuda:4'), in_proj_covar=tensor([0.0261, 0.0248, 0.0206, 0.0265, 0.0312, 0.0217, 0.0235, 0.0289], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:01:50,571 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-28 01:01:50,890 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.435e+02 3.682e+02 4.369e+02 5.544e+02 1.182e+03, threshold=8.738e+02, percent-clipped=2.0 2023-04-28 01:01:54,371 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0618, 4.1715, 4.1922, 4.2482, 4.2400, 4.6910, 4.4769, 4.2133], device='cuda:4'), covar=tensor([0.1387, 0.1390, 0.1194, 0.1432, 0.2331, 0.0874, 0.0964, 0.1847], device='cuda:4'), in_proj_covar=tensor([0.0216, 0.0300, 0.0290, 0.0262, 0.0350, 0.0313, 0.0245, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:02:22,322 INFO [train.py:904] (4/8) Epoch 3, batch 9550, loss[loss=0.2342, simple_loss=0.3089, pruned_loss=0.07979, over 12495.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.3048, pruned_loss=0.06951, over 3034867.91 frames. ], batch size: 248, lr: 1.89e-02, grad_scale: 4.0 2023-04-28 01:02:49,783 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29863.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:03:49,011 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29892.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:04:04,157 INFO [train.py:904] (4/8) Epoch 3, batch 9600, loss[loss=0.2265, simple_loss=0.3146, pruned_loss=0.06915, over 16484.00 frames. ], tot_loss[loss=0.2245, simple_loss=0.3076, pruned_loss=0.07075, over 3051175.65 frames. ], batch size: 68, lr: 1.89e-02, grad_scale: 8.0 2023-04-28 01:04:16,860 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=29907.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:04:37,525 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7743, 3.8751, 4.3197, 4.2761, 4.2245, 3.9049, 3.9432, 3.8612], device='cuda:4'), covar=tensor([0.0238, 0.0247, 0.0267, 0.0332, 0.0338, 0.0232, 0.0587, 0.0313], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0160, 0.0166, 0.0170, 0.0201, 0.0180, 0.0257, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-04-28 01:05:18,577 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.324e+02 3.589e+02 4.321e+02 5.133e+02 1.002e+03, threshold=8.641e+02, percent-clipped=3.0 2023-04-28 01:05:23,702 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=29940.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:05:51,213 INFO [train.py:904] (4/8) Epoch 3, batch 9650, loss[loss=0.2469, simple_loss=0.329, pruned_loss=0.08234, over 15225.00 frames. ], tot_loss[loss=0.2264, simple_loss=0.31, pruned_loss=0.07143, over 3040547.78 frames. ], batch size: 190, lr: 1.89e-02, grad_scale: 8.0 2023-04-28 01:06:22,038 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7501, 2.5870, 2.2278, 3.9953, 3.7878, 3.8874, 1.5246, 2.9088], device='cuda:4'), covar=tensor([0.1287, 0.0556, 0.1127, 0.0061, 0.0139, 0.0262, 0.1381, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0134, 0.0161, 0.0067, 0.0129, 0.0143, 0.0154, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-04-28 01:07:41,652 INFO [train.py:904] (4/8) Epoch 3, batch 9700, loss[loss=0.2113, simple_loss=0.2888, pruned_loss=0.0669, over 12203.00 frames. ], tot_loss[loss=0.225, simple_loss=0.3086, pruned_loss=0.07073, over 3051348.49 frames. ], batch size: 248, lr: 1.89e-02, grad_scale: 8.0 2023-04-28 01:08:16,000 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30018.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 01:08:53,393 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9779, 3.3926, 3.4391, 2.4272, 3.3943, 3.6135, 3.5196, 1.8056], device='cuda:4'), covar=tensor([0.0368, 0.0025, 0.0052, 0.0239, 0.0033, 0.0031, 0.0029, 0.0353], device='cuda:4'), in_proj_covar=tensor([0.0109, 0.0052, 0.0055, 0.0106, 0.0053, 0.0059, 0.0054, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0002, 0.0001, 0.0003], device='cuda:4') 2023-04-28 01:08:59,978 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.947e+02 3.446e+02 4.381e+02 5.437e+02 9.929e+02, threshold=8.763e+02, percent-clipped=3.0 2023-04-28 01:09:24,291 INFO [train.py:904] (4/8) Epoch 3, batch 9750, loss[loss=0.222, simple_loss=0.2985, pruned_loss=0.07277, over 12403.00 frames. ], tot_loss[loss=0.2243, simple_loss=0.3072, pruned_loss=0.07066, over 3054710.84 frames. ], batch size: 248, lr: 1.89e-02, grad_scale: 8.0 2023-04-28 01:11:02,931 INFO [train.py:904] (4/8) Epoch 3, batch 9800, loss[loss=0.2007, simple_loss=0.3027, pruned_loss=0.04936, over 16831.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.3061, pruned_loss=0.06888, over 3068878.48 frames. ], batch size: 96, lr: 1.89e-02, grad_scale: 8.0 2023-04-28 01:11:48,275 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30125.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:12:14,772 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.212e+02 3.552e+02 4.452e+02 5.972e+02 1.192e+03, threshold=8.904e+02, percent-clipped=4.0 2023-04-28 01:12:47,470 INFO [train.py:904] (4/8) Epoch 3, batch 9850, loss[loss=0.2182, simple_loss=0.3047, pruned_loss=0.06585, over 16376.00 frames. ], tot_loss[loss=0.222, simple_loss=0.3073, pruned_loss=0.06837, over 3080982.90 frames. ], batch size: 146, lr: 1.88e-02, grad_scale: 8.0 2023-04-28 01:13:02,424 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30158.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:13:48,762 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-28 01:14:06,711 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30186.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:14:23,323 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.36 vs. limit=5.0 2023-04-28 01:14:39,079 INFO [train.py:904] (4/8) Epoch 3, batch 9900, loss[loss=0.2301, simple_loss=0.3152, pruned_loss=0.07246, over 15270.00 frames. ], tot_loss[loss=0.2224, simple_loss=0.3079, pruned_loss=0.06847, over 3085494.26 frames. ], batch size: 191, lr: 1.88e-02, grad_scale: 8.0 2023-04-28 01:14:54,602 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30207.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:14:59,716 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30209.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:15:39,763 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6144, 1.4181, 1.8813, 2.3577, 2.3645, 2.5259, 1.5164, 2.5204], device='cuda:4'), covar=tensor([0.0046, 0.0199, 0.0127, 0.0090, 0.0067, 0.0057, 0.0165, 0.0051], device='cuda:4'), in_proj_covar=tensor([0.0091, 0.0124, 0.0109, 0.0101, 0.0098, 0.0072, 0.0115, 0.0062], device='cuda:4'), out_proj_covar=tensor([1.3548e-04, 1.8807e-04, 1.6950e-04, 1.5555e-04, 1.4740e-04, 1.0704e-04, 1.7412e-04, 9.1509e-05], device='cuda:4') 2023-04-28 01:16:05,826 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.521e+02 3.778e+02 4.995e+02 6.247e+02 1.716e+03, threshold=9.990e+02, percent-clipped=10.0 2023-04-28 01:16:35,566 INFO [train.py:904] (4/8) Epoch 3, batch 9950, loss[loss=0.2205, simple_loss=0.3091, pruned_loss=0.06601, over 16710.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.3106, pruned_loss=0.0689, over 3098887.82 frames. ], batch size: 76, lr: 1.88e-02, grad_scale: 8.0 2023-04-28 01:16:47,150 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=30255.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:17:00,592 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-28 01:17:23,142 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30270.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:18:37,068 INFO [train.py:904] (4/8) Epoch 3, batch 10000, loss[loss=0.2114, simple_loss=0.2909, pruned_loss=0.06598, over 12573.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.3086, pruned_loss=0.06824, over 3088545.82 frames. ], batch size: 247, lr: 1.88e-02, grad_scale: 8.0 2023-04-28 01:19:12,396 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30318.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:19:40,428 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 01:19:55,680 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.113e+02 3.684e+02 4.714e+02 5.759e+02 1.067e+03, threshold=9.428e+02, percent-clipped=2.0 2023-04-28 01:20:19,871 INFO [train.py:904] (4/8) Epoch 3, batch 10050, loss[loss=0.2635, simple_loss=0.3544, pruned_loss=0.08628, over 16519.00 frames. ], tot_loss[loss=0.222, simple_loss=0.3081, pruned_loss=0.06794, over 3065502.26 frames. ], batch size: 147, lr: 1.88e-02, grad_scale: 8.0 2023-04-28 01:20:38,346 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30360.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:20:50,749 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=30366.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:20:51,186 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-28 01:21:54,399 INFO [train.py:904] (4/8) Epoch 3, batch 10100, loss[loss=0.1946, simple_loss=0.2815, pruned_loss=0.05388, over 16398.00 frames. ], tot_loss[loss=0.2235, simple_loss=0.3091, pruned_loss=0.06897, over 3069519.32 frames. ], batch size: 147, lr: 1.88e-02, grad_scale: 8.0 2023-04-28 01:22:37,213 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30421.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:23:00,423 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.870e+02 3.919e+02 4.872e+02 6.072e+02 1.469e+03, threshold=9.744e+02, percent-clipped=6.0 2023-04-28 01:23:38,627 INFO [train.py:904] (4/8) Epoch 4, batch 0, loss[loss=0.399, simple_loss=0.3948, pruned_loss=0.2016, over 16780.00 frames. ], tot_loss[loss=0.399, simple_loss=0.3948, pruned_loss=0.2016, over 16780.00 frames. ], batch size: 83, lr: 1.75e-02, grad_scale: 8.0 2023-04-28 01:23:38,627 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 01:23:46,515 INFO [train.py:938] (4/8) Epoch 4, validation: loss=0.188, simple_loss=0.2904, pruned_loss=0.04284, over 944034.00 frames. 2023-04-28 01:23:46,516 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 01:23:57,932 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30458.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:24:29,163 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30481.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:24:44,592 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9149, 4.6526, 4.9104, 5.3299, 5.3359, 4.6856, 5.4152, 5.2239], device='cuda:4'), covar=tensor([0.0652, 0.0655, 0.1187, 0.0373, 0.0377, 0.0427, 0.0359, 0.0406], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0363, 0.0470, 0.0352, 0.0277, 0.0256, 0.0292, 0.0295], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:24:56,021 INFO [train.py:904] (4/8) Epoch 4, batch 50, loss[loss=0.255, simple_loss=0.323, pruned_loss=0.09354, over 16487.00 frames. ], tot_loss[loss=0.2703, simple_loss=0.3347, pruned_loss=0.103, over 749816.45 frames. ], batch size: 75, lr: 1.75e-02, grad_scale: 1.0 2023-04-28 01:25:02,246 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=30506.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:25:16,777 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7310, 2.6599, 2.5413, 4.2427, 2.0845, 3.6425, 2.2751, 2.4684], device='cuda:4'), covar=tensor([0.0358, 0.0924, 0.0517, 0.0200, 0.1873, 0.0338, 0.1126, 0.1346], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0257, 0.0211, 0.0273, 0.0322, 0.0224, 0.0239, 0.0308], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:25:49,875 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.617e+02 4.176e+02 5.221e+02 6.159e+02 1.340e+03, threshold=1.044e+03, percent-clipped=3.0 2023-04-28 01:25:58,156 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4999, 3.3705, 2.6473, 2.3056, 2.4067, 2.0339, 3.4238, 3.5612], device='cuda:4'), covar=tensor([0.1599, 0.0538, 0.0980, 0.1025, 0.1674, 0.1200, 0.0334, 0.0447], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0244, 0.0256, 0.0215, 0.0245, 0.0198, 0.0218, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:26:02,161 INFO [train.py:904] (4/8) Epoch 4, batch 100, loss[loss=0.2284, simple_loss=0.3099, pruned_loss=0.07346, over 17130.00 frames. ], tot_loss[loss=0.2561, simple_loss=0.3235, pruned_loss=0.09433, over 1325186.53 frames. ], batch size: 49, lr: 1.75e-02, grad_scale: 1.0 2023-04-28 01:26:10,441 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 01:26:23,715 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30565.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:26:54,687 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8164, 2.4906, 2.3120, 2.1610, 3.0688, 3.1026, 3.8039, 3.4020], device='cuda:4'), covar=tensor([0.0011, 0.0144, 0.0165, 0.0194, 0.0089, 0.0110, 0.0039, 0.0069], device='cuda:4'), in_proj_covar=tensor([0.0058, 0.0128, 0.0128, 0.0127, 0.0117, 0.0127, 0.0083, 0.0102], device='cuda:4'), out_proj_covar=tensor([7.3018e-05, 1.7627e-04, 1.7005e-04, 1.7237e-04, 1.6248e-04, 1.7481e-04, 1.1135e-04, 1.3982e-04], device='cuda:4') 2023-04-28 01:26:57,451 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-28 01:27:11,703 INFO [train.py:904] (4/8) Epoch 4, batch 150, loss[loss=0.2078, simple_loss=0.2825, pruned_loss=0.06652, over 17242.00 frames. ], tot_loss[loss=0.2536, simple_loss=0.3211, pruned_loss=0.09304, over 1765697.68 frames. ], batch size: 44, lr: 1.75e-02, grad_scale: 1.0 2023-04-28 01:28:04,889 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.524e+02 3.830e+02 4.808e+02 5.813e+02 1.621e+03, threshold=9.616e+02, percent-clipped=3.0 2023-04-28 01:28:12,154 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6224, 1.4636, 1.8894, 2.3919, 2.4882, 2.5124, 1.3691, 2.5821], device='cuda:4'), covar=tensor([0.0058, 0.0179, 0.0132, 0.0090, 0.0062, 0.0079, 0.0169, 0.0042], device='cuda:4'), in_proj_covar=tensor([0.0095, 0.0128, 0.0114, 0.0106, 0.0102, 0.0077, 0.0118, 0.0065], device='cuda:4'), out_proj_covar=tensor([1.4168e-04, 1.9452e-04, 1.7710e-04, 1.6227e-04, 1.5374e-04, 1.1365e-04, 1.7755e-04, 9.7605e-05], device='cuda:4') 2023-04-28 01:28:19,174 INFO [train.py:904] (4/8) Epoch 4, batch 200, loss[loss=0.2442, simple_loss=0.3241, pruned_loss=0.08219, over 16697.00 frames. ], tot_loss[loss=0.2515, simple_loss=0.3199, pruned_loss=0.09152, over 2116819.66 frames. ], batch size: 62, lr: 1.75e-02, grad_scale: 1.0 2023-04-28 01:28:23,987 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30655.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:28:25,173 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1096, 1.7265, 2.1622, 2.8970, 2.8180, 3.1667, 1.5862, 3.0156], device='cuda:4'), covar=tensor([0.0045, 0.0177, 0.0139, 0.0080, 0.0071, 0.0059, 0.0165, 0.0045], device='cuda:4'), in_proj_covar=tensor([0.0095, 0.0128, 0.0113, 0.0105, 0.0102, 0.0077, 0.0118, 0.0065], device='cuda:4'), out_proj_covar=tensor([1.4172e-04, 1.9393e-04, 1.7559e-04, 1.6150e-04, 1.5326e-04, 1.1341e-04, 1.7663e-04, 9.7482e-05], device='cuda:4') 2023-04-28 01:28:44,137 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0801, 2.0039, 1.7226, 1.9001, 2.7079, 2.4861, 3.2256, 2.9672], device='cuda:4'), covar=tensor([0.0025, 0.0197, 0.0207, 0.0200, 0.0103, 0.0157, 0.0065, 0.0096], device='cuda:4'), in_proj_covar=tensor([0.0059, 0.0130, 0.0128, 0.0127, 0.0118, 0.0129, 0.0085, 0.0104], device='cuda:4'), out_proj_covar=tensor([7.5269e-05, 1.7825e-04, 1.7076e-04, 1.7267e-04, 1.6320e-04, 1.7848e-04, 1.1467e-04, 1.4345e-04], device='cuda:4') 2023-04-28 01:29:08,201 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9071, 4.7581, 4.6982, 4.6743, 4.2498, 4.7538, 4.7701, 4.4584], device='cuda:4'), covar=tensor([0.0343, 0.0235, 0.0159, 0.0143, 0.0737, 0.0214, 0.0226, 0.0302], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0147, 0.0188, 0.0156, 0.0222, 0.0178, 0.0135, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 01:29:26,979 INFO [train.py:904] (4/8) Epoch 4, batch 250, loss[loss=0.2254, simple_loss=0.3132, pruned_loss=0.06882, over 17040.00 frames. ], tot_loss[loss=0.25, simple_loss=0.3179, pruned_loss=0.09104, over 2385203.44 frames. ], batch size: 50, lr: 1.75e-02, grad_scale: 1.0 2023-04-28 01:29:48,571 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30716.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:29:48,716 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30716.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:29:49,820 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0753, 5.4114, 5.0496, 5.2392, 4.7500, 4.5068, 4.8755, 5.4341], device='cuda:4'), covar=tensor([0.0521, 0.0556, 0.0803, 0.0373, 0.0543, 0.0648, 0.0562, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0425, 0.0367, 0.0273, 0.0276, 0.0275, 0.0336, 0.0295], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:30:21,642 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.145e+02 3.686e+02 4.713e+02 5.844e+02 8.594e+02, threshold=9.427e+02, percent-clipped=0.0 2023-04-28 01:30:36,157 INFO [train.py:904] (4/8) Epoch 4, batch 300, loss[loss=0.2015, simple_loss=0.2843, pruned_loss=0.05938, over 17211.00 frames. ], tot_loss[loss=0.2464, simple_loss=0.3148, pruned_loss=0.08898, over 2583679.83 frames. ], batch size: 43, lr: 1.75e-02, grad_scale: 1.0 2023-04-28 01:31:16,650 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30781.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:31:43,552 INFO [train.py:904] (4/8) Epoch 4, batch 350, loss[loss=0.2488, simple_loss=0.3145, pruned_loss=0.09157, over 16622.00 frames. ], tot_loss[loss=0.2427, simple_loss=0.3113, pruned_loss=0.08707, over 2746491.78 frames. ], batch size: 62, lr: 1.74e-02, grad_scale: 1.0 2023-04-28 01:32:03,573 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2840, 3.3349, 1.5602, 3.5123, 2.4695, 3.4738, 1.7574, 2.6997], device='cuda:4'), covar=tensor([0.0091, 0.0233, 0.1439, 0.0077, 0.0666, 0.0276, 0.1162, 0.0503], device='cuda:4'), in_proj_covar=tensor([0.0097, 0.0140, 0.0169, 0.0080, 0.0157, 0.0164, 0.0176, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 01:32:20,667 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=30829.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:32:38,673 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.414e+02 3.784e+02 4.816e+02 5.903e+02 1.455e+03, threshold=9.632e+02, percent-clipped=4.0 2023-04-28 01:32:40,428 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4140, 3.1588, 3.9504, 2.7328, 3.8667, 3.8427, 3.9086, 2.1922], device='cuda:4'), covar=tensor([0.0311, 0.0102, 0.0029, 0.0220, 0.0026, 0.0063, 0.0023, 0.0272], device='cuda:4'), in_proj_covar=tensor([0.0110, 0.0057, 0.0058, 0.0111, 0.0056, 0.0063, 0.0056, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 01:32:51,066 INFO [train.py:904] (4/8) Epoch 4, batch 400, loss[loss=0.237, simple_loss=0.3013, pruned_loss=0.0864, over 16512.00 frames. ], tot_loss[loss=0.24, simple_loss=0.3079, pruned_loss=0.08598, over 2872188.59 frames. ], batch size: 68, lr: 1.74e-02, grad_scale: 2.0 2023-04-28 01:32:55,341 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1084, 1.5141, 2.2155, 2.8058, 2.7240, 3.2464, 1.6488, 3.0651], device='cuda:4'), covar=tensor([0.0052, 0.0205, 0.0134, 0.0098, 0.0075, 0.0078, 0.0198, 0.0062], device='cuda:4'), in_proj_covar=tensor([0.0098, 0.0129, 0.0114, 0.0107, 0.0103, 0.0080, 0.0121, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 01:33:11,852 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30865.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:33:22,062 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-28 01:33:31,649 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.80 vs. limit=5.0 2023-04-28 01:33:39,083 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-28 01:33:43,483 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0976, 3.7895, 3.1310, 5.2648, 5.1126, 4.5905, 2.1218, 3.4887], device='cuda:4'), covar=tensor([0.1066, 0.0383, 0.0836, 0.0059, 0.0155, 0.0250, 0.1002, 0.0528], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0132, 0.0160, 0.0069, 0.0148, 0.0146, 0.0150, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 01:34:01,546 INFO [train.py:904] (4/8) Epoch 4, batch 450, loss[loss=0.2023, simple_loss=0.2843, pruned_loss=0.06015, over 17257.00 frames. ], tot_loss[loss=0.2367, simple_loss=0.3055, pruned_loss=0.08392, over 2977260.01 frames. ], batch size: 45, lr: 1.74e-02, grad_scale: 2.0 2023-04-28 01:34:05,061 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30903.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:34:17,422 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=30913.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:34:31,566 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-28 01:34:35,324 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30925.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:34:45,912 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30934.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:34:56,469 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.382e+02 3.455e+02 4.330e+02 5.569e+02 1.727e+03, threshold=8.660e+02, percent-clipped=4.0 2023-04-28 01:35:07,406 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4449, 4.4375, 3.8089, 1.8748, 2.9672, 2.1633, 3.7433, 4.1641], device='cuda:4'), covar=tensor([0.0283, 0.0394, 0.0483, 0.1538, 0.0717, 0.1060, 0.0609, 0.0739], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0123, 0.0156, 0.0145, 0.0136, 0.0130, 0.0140, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 01:35:09,303 INFO [train.py:904] (4/8) Epoch 4, batch 500, loss[loss=0.182, simple_loss=0.2724, pruned_loss=0.04583, over 16805.00 frames. ], tot_loss[loss=0.2333, simple_loss=0.3031, pruned_loss=0.08174, over 3059158.60 frames. ], batch size: 42, lr: 1.74e-02, grad_scale: 2.0 2023-04-28 01:35:09,770 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7752, 4.5816, 4.5404, 2.0312, 4.6809, 4.8345, 3.3225, 3.6597], device='cuda:4'), covar=tensor([0.0740, 0.0078, 0.0135, 0.1221, 0.0050, 0.0038, 0.0304, 0.0353], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0081, 0.0080, 0.0144, 0.0074, 0.0074, 0.0112, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:35:28,300 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30964.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:35:58,987 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30986.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:36:09,056 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30995.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:36:17,824 INFO [train.py:904] (4/8) Epoch 4, batch 550, loss[loss=0.2574, simple_loss=0.3104, pruned_loss=0.1022, over 16846.00 frames. ], tot_loss[loss=0.2311, simple_loss=0.3021, pruned_loss=0.08007, over 3123160.44 frames. ], batch size: 116, lr: 1.74e-02, grad_scale: 2.0 2023-04-28 01:36:33,157 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31011.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:36:33,340 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31011.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:36:39,297 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31016.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:37:13,481 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.455e+02 3.696e+02 4.486e+02 5.421e+02 1.042e+03, threshold=8.971e+02, percent-clipped=3.0 2023-04-28 01:37:28,711 INFO [train.py:904] (4/8) Epoch 4, batch 600, loss[loss=0.1898, simple_loss=0.2658, pruned_loss=0.05693, over 17058.00 frames. ], tot_loss[loss=0.2303, simple_loss=0.3006, pruned_loss=0.07998, over 3170710.76 frames. ], batch size: 41, lr: 1.74e-02, grad_scale: 2.0 2023-04-28 01:37:46,831 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=31064.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:37:50,550 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-28 01:37:58,212 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31072.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 01:38:36,813 INFO [train.py:904] (4/8) Epoch 4, batch 650, loss[loss=0.2224, simple_loss=0.2869, pruned_loss=0.079, over 12321.00 frames. ], tot_loss[loss=0.2301, simple_loss=0.3, pruned_loss=0.08006, over 3195250.04 frames. ], batch size: 246, lr: 1.74e-02, grad_scale: 2.0 2023-04-28 01:39:30,534 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.918e+02 3.618e+02 4.311e+02 5.742e+02 1.399e+03, threshold=8.622e+02, percent-clipped=6.0 2023-04-28 01:39:36,684 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.99 vs. limit=5.0 2023-04-28 01:39:38,422 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-28 01:39:45,330 INFO [train.py:904] (4/8) Epoch 4, batch 700, loss[loss=0.2271, simple_loss=0.294, pruned_loss=0.08012, over 16780.00 frames. ], tot_loss[loss=0.2283, simple_loss=0.2992, pruned_loss=0.07868, over 3230211.14 frames. ], batch size: 102, lr: 1.73e-02, grad_scale: 2.0 2023-04-28 01:40:20,522 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6358, 4.6058, 4.6218, 3.9565, 4.5350, 2.0051, 4.3754, 4.6233], device='cuda:4'), covar=tensor([0.0070, 0.0063, 0.0081, 0.0326, 0.0061, 0.1438, 0.0080, 0.0131], device='cuda:4'), in_proj_covar=tensor([0.0081, 0.0073, 0.0112, 0.0120, 0.0080, 0.0130, 0.0097, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 01:40:53,737 INFO [train.py:904] (4/8) Epoch 4, batch 750, loss[loss=0.2239, simple_loss=0.2937, pruned_loss=0.07701, over 16525.00 frames. ], tot_loss[loss=0.2289, simple_loss=0.2997, pruned_loss=0.07901, over 3241413.81 frames. ], batch size: 68, lr: 1.73e-02, grad_scale: 2.0 2023-04-28 01:41:48,308 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.232e+02 3.368e+02 4.224e+02 5.033e+02 8.537e+02, threshold=8.448e+02, percent-clipped=0.0 2023-04-28 01:42:01,623 INFO [train.py:904] (4/8) Epoch 4, batch 800, loss[loss=0.2517, simple_loss=0.3065, pruned_loss=0.0984, over 16894.00 frames. ], tot_loss[loss=0.228, simple_loss=0.2991, pruned_loss=0.07843, over 3261273.65 frames. ], batch size: 96, lr: 1.73e-02, grad_scale: 4.0 2023-04-28 01:42:14,287 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31259.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:42:44,510 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31281.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:42:55,439 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31290.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:43:09,876 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-04-28 01:43:11,602 INFO [train.py:904] (4/8) Epoch 4, batch 850, loss[loss=0.2186, simple_loss=0.2983, pruned_loss=0.06946, over 16716.00 frames. ], tot_loss[loss=0.2273, simple_loss=0.2986, pruned_loss=0.07802, over 3281137.99 frames. ], batch size: 62, lr: 1.73e-02, grad_scale: 4.0 2023-04-28 01:43:24,706 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31311.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:44:04,054 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.57 vs. limit=5.0 2023-04-28 01:44:07,361 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.752e+02 3.699e+02 4.562e+02 5.706e+02 1.340e+03, threshold=9.123e+02, percent-clipped=5.0 2023-04-28 01:44:19,704 INFO [train.py:904] (4/8) Epoch 4, batch 900, loss[loss=0.1869, simple_loss=0.262, pruned_loss=0.05595, over 16762.00 frames. ], tot_loss[loss=0.2268, simple_loss=0.2986, pruned_loss=0.07749, over 3289184.17 frames. ], batch size: 39, lr: 1.73e-02, grad_scale: 4.0 2023-04-28 01:44:32,830 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=31359.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:44:42,741 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31367.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 01:45:31,311 INFO [train.py:904] (4/8) Epoch 4, batch 950, loss[loss=0.2045, simple_loss=0.2918, pruned_loss=0.0586, over 17040.00 frames. ], tot_loss[loss=0.2262, simple_loss=0.2981, pruned_loss=0.07713, over 3295632.48 frames. ], batch size: 50, lr: 1.73e-02, grad_scale: 4.0 2023-04-28 01:45:41,894 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31409.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:46:22,519 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.83 vs. limit=5.0 2023-04-28 01:46:26,065 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.766e+02 3.401e+02 4.040e+02 4.698e+02 9.697e+02, threshold=8.080e+02, percent-clipped=2.0 2023-04-28 01:46:38,678 INFO [train.py:904] (4/8) Epoch 4, batch 1000, loss[loss=0.2121, simple_loss=0.2962, pruned_loss=0.06394, over 17136.00 frames. ], tot_loss[loss=0.225, simple_loss=0.2964, pruned_loss=0.07682, over 3311100.46 frames. ], batch size: 47, lr: 1.73e-02, grad_scale: 4.0 2023-04-28 01:46:57,996 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.67 vs. limit=2.0 2023-04-28 01:47:06,246 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31470.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:47:24,953 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.96 vs. limit=5.0 2023-04-28 01:47:34,959 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3582, 4.5117, 4.8694, 4.8356, 4.7981, 4.5012, 4.4271, 4.2651], device='cuda:4'), covar=tensor([0.0222, 0.0364, 0.0279, 0.0360, 0.0360, 0.0256, 0.0695, 0.0382], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0211, 0.0218, 0.0216, 0.0258, 0.0230, 0.0321, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 01:47:36,222 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31491.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 01:47:37,492 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2113, 1.6939, 2.2998, 2.8294, 2.7402, 3.6366, 1.8226, 3.1526], device='cuda:4'), covar=tensor([0.0058, 0.0195, 0.0130, 0.0102, 0.0088, 0.0045, 0.0166, 0.0047], device='cuda:4'), in_proj_covar=tensor([0.0101, 0.0128, 0.0113, 0.0109, 0.0104, 0.0079, 0.0121, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 01:47:48,997 INFO [train.py:904] (4/8) Epoch 4, batch 1050, loss[loss=0.2402, simple_loss=0.3177, pruned_loss=0.08135, over 16805.00 frames. ], tot_loss[loss=0.2255, simple_loss=0.2966, pruned_loss=0.07716, over 3304835.09 frames. ], batch size: 57, lr: 1.72e-02, grad_scale: 4.0 2023-04-28 01:48:32,803 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9243, 3.7411, 3.8589, 4.1702, 4.1740, 3.7080, 4.0056, 4.1740], device='cuda:4'), covar=tensor([0.0666, 0.0643, 0.1179, 0.0404, 0.0377, 0.1306, 0.1170, 0.0383], device='cuda:4'), in_proj_covar=tensor([0.0351, 0.0421, 0.0566, 0.0437, 0.0328, 0.0314, 0.0342, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:48:37,479 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5268, 4.5467, 5.1675, 5.1676, 5.0915, 4.6937, 4.6728, 4.4726], device='cuda:4'), covar=tensor([0.0243, 0.0323, 0.0279, 0.0319, 0.0338, 0.0263, 0.0612, 0.0300], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0211, 0.0219, 0.0216, 0.0255, 0.0231, 0.0322, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 01:48:45,339 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.288e+02 3.437e+02 4.188e+02 5.032e+02 1.579e+03, threshold=8.377e+02, percent-clipped=2.0 2023-04-28 01:49:00,844 INFO [train.py:904] (4/8) Epoch 4, batch 1100, loss[loss=0.2732, simple_loss=0.3194, pruned_loss=0.1135, over 16932.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.2953, pruned_loss=0.07652, over 3313642.87 frames. ], batch size: 109, lr: 1.72e-02, grad_scale: 4.0 2023-04-28 01:49:02,446 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31552.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:49:11,033 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31559.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:49:41,971 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31581.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:49:55,696 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31590.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:50:09,609 INFO [train.py:904] (4/8) Epoch 4, batch 1150, loss[loss=0.2146, simple_loss=0.2817, pruned_loss=0.07372, over 16859.00 frames. ], tot_loss[loss=0.223, simple_loss=0.2945, pruned_loss=0.07575, over 3315119.84 frames. ], batch size: 116, lr: 1.72e-02, grad_scale: 4.0 2023-04-28 01:50:18,984 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=31607.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:50:47,367 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9853, 4.6934, 5.0006, 5.3240, 5.3917, 4.6940, 5.3709, 5.3820], device='cuda:4'), covar=tensor([0.0754, 0.0727, 0.1234, 0.0403, 0.0365, 0.0428, 0.0419, 0.0311], device='cuda:4'), in_proj_covar=tensor([0.0352, 0.0424, 0.0572, 0.0446, 0.0330, 0.0314, 0.0348, 0.0355], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:50:49,047 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=31629.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:50:56,258 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1589, 3.1815, 1.6985, 3.2644, 2.4062, 3.2994, 1.9603, 2.5739], device='cuda:4'), covar=tensor([0.0131, 0.0242, 0.1413, 0.0081, 0.0676, 0.0329, 0.1168, 0.0563], device='cuda:4'), in_proj_covar=tensor([0.0105, 0.0146, 0.0175, 0.0083, 0.0162, 0.0176, 0.0184, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 01:51:00,549 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=31638.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:51:04,257 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.256e+02 3.893e+02 4.649e+02 5.891e+02 1.405e+03, threshold=9.297e+02, percent-clipped=5.0 2023-04-28 01:51:19,067 INFO [train.py:904] (4/8) Epoch 4, batch 1200, loss[loss=0.2517, simple_loss=0.3333, pruned_loss=0.08508, over 17053.00 frames. ], tot_loss[loss=0.223, simple_loss=0.2942, pruned_loss=0.07593, over 3313094.66 frames. ], batch size: 50, lr: 1.72e-02, grad_scale: 8.0 2023-04-28 01:51:37,405 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3521, 3.0976, 3.7033, 2.7314, 3.4880, 3.7216, 3.6519, 1.7799], device='cuda:4'), covar=tensor([0.0254, 0.0098, 0.0031, 0.0165, 0.0036, 0.0036, 0.0036, 0.0271], device='cuda:4'), in_proj_covar=tensor([0.0110, 0.0057, 0.0058, 0.0108, 0.0058, 0.0062, 0.0057, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 01:51:40,992 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31667.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 01:52:27,377 INFO [train.py:904] (4/8) Epoch 4, batch 1250, loss[loss=0.1877, simple_loss=0.2641, pruned_loss=0.05566, over 16838.00 frames. ], tot_loss[loss=0.2243, simple_loss=0.2947, pruned_loss=0.07693, over 3322691.25 frames. ], batch size: 39, lr: 1.72e-02, grad_scale: 8.0 2023-04-28 01:52:31,938 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31704.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:52:47,962 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=31715.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:53:22,524 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.448e+02 3.415e+02 4.132e+02 5.466e+02 7.512e+02, threshold=8.265e+02, percent-clipped=0.0 2023-04-28 01:53:35,322 INFO [train.py:904] (4/8) Epoch 4, batch 1300, loss[loss=0.2483, simple_loss=0.3055, pruned_loss=0.09554, over 16726.00 frames. ], tot_loss[loss=0.2232, simple_loss=0.2945, pruned_loss=0.07595, over 3312250.95 frames. ], batch size: 134, lr: 1.72e-02, grad_scale: 8.0 2023-04-28 01:53:54,659 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31765.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:53:54,886 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31765.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:54:26,629 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31788.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:54:44,581 INFO [train.py:904] (4/8) Epoch 4, batch 1350, loss[loss=0.1842, simple_loss=0.2672, pruned_loss=0.0506, over 17192.00 frames. ], tot_loss[loss=0.2218, simple_loss=0.2943, pruned_loss=0.07463, over 3323849.67 frames. ], batch size: 46, lr: 1.72e-02, grad_scale: 8.0 2023-04-28 01:54:50,137 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8167, 4.9012, 5.4180, 5.3935, 5.3831, 4.8791, 4.8944, 4.7598], device='cuda:4'), covar=tensor([0.0189, 0.0246, 0.0245, 0.0348, 0.0300, 0.0223, 0.0649, 0.0232], device='cuda:4'), in_proj_covar=tensor([0.0213, 0.0211, 0.0221, 0.0214, 0.0259, 0.0232, 0.0326, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 01:55:00,206 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1974, 2.1466, 1.5706, 1.8852, 2.5922, 2.6602, 2.8097, 2.7501], device='cuda:4'), covar=tensor([0.0053, 0.0144, 0.0186, 0.0188, 0.0079, 0.0096, 0.0062, 0.0078], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0134, 0.0130, 0.0129, 0.0123, 0.0131, 0.0097, 0.0110], device='cuda:4'), out_proj_covar=tensor([9.0016e-05, 1.8282e-04, 1.7068e-04, 1.7291e-04, 1.6949e-04, 1.7794e-04, 1.3076e-04, 1.5093e-04], device='cuda:4') 2023-04-28 01:55:14,552 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31823.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:55:32,558 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 01:55:39,665 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.154e+02 3.282e+02 4.037e+02 5.139e+02 9.544e+02, threshold=8.075e+02, percent-clipped=2.0 2023-04-28 01:55:48,234 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31847.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 01:55:50,750 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31849.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:55:52,765 INFO [train.py:904] (4/8) Epoch 4, batch 1400, loss[loss=0.2097, simple_loss=0.2783, pruned_loss=0.07055, over 15594.00 frames. ], tot_loss[loss=0.2224, simple_loss=0.2946, pruned_loss=0.07507, over 3315020.08 frames. ], batch size: 190, lr: 1.72e-02, grad_scale: 8.0 2023-04-28 01:56:32,723 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31878.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:56:40,472 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31884.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:57:02,740 INFO [train.py:904] (4/8) Epoch 4, batch 1450, loss[loss=0.2418, simple_loss=0.2909, pruned_loss=0.09632, over 15545.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.2928, pruned_loss=0.07367, over 3324754.79 frames. ], batch size: 191, lr: 1.71e-02, grad_scale: 4.0 2023-04-28 01:57:22,443 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31915.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:57:56,594 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31939.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:58:00,353 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.224e+02 3.023e+02 3.906e+02 5.046e+02 9.760e+02, threshold=7.812e+02, percent-clipped=3.0 2023-04-28 01:58:12,696 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31950.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:58:13,457 INFO [train.py:904] (4/8) Epoch 4, batch 1500, loss[loss=0.1879, simple_loss=0.2752, pruned_loss=0.05032, over 17245.00 frames. ], tot_loss[loss=0.2205, simple_loss=0.293, pruned_loss=0.07399, over 3329122.92 frames. ], batch size: 45, lr: 1.71e-02, grad_scale: 4.0 2023-04-28 01:58:23,586 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.89 vs. limit=5.0 2023-04-28 01:58:47,444 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31976.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:59:26,618 INFO [train.py:904] (4/8) Epoch 4, batch 1550, loss[loss=0.2142, simple_loss=0.3008, pruned_loss=0.06383, over 17024.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.2945, pruned_loss=0.0754, over 3327615.88 frames. ], batch size: 53, lr: 1.71e-02, grad_scale: 4.0 2023-04-28 01:59:33,140 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-28 01:59:40,598 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=32011.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:00:09,098 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4985, 5.7704, 5.4254, 5.5977, 5.1119, 4.8213, 5.2621, 5.8710], device='cuda:4'), covar=tensor([0.0598, 0.0647, 0.0808, 0.0351, 0.0555, 0.0502, 0.0519, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0453, 0.0384, 0.0282, 0.0288, 0.0285, 0.0352, 0.0309], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:00:21,778 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.322e+02 3.491e+02 4.515e+02 5.293e+02 8.780e+02, threshold=9.030e+02, percent-clipped=2.0 2023-04-28 02:00:34,386 INFO [train.py:904] (4/8) Epoch 4, batch 1600, loss[loss=0.2167, simple_loss=0.3032, pruned_loss=0.0651, over 17282.00 frames. ], tot_loss[loss=0.2245, simple_loss=0.2964, pruned_loss=0.07632, over 3332156.10 frames. ], batch size: 52, lr: 1.71e-02, grad_scale: 8.0 2023-04-28 02:00:46,071 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32060.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:00:52,711 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32065.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:01:40,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8259, 5.1017, 4.7503, 4.9394, 4.5670, 4.3350, 4.6381, 5.1074], device='cuda:4'), covar=tensor([0.0456, 0.0601, 0.0858, 0.0357, 0.0538, 0.0689, 0.0531, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0453, 0.0386, 0.0282, 0.0288, 0.0285, 0.0353, 0.0312], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:01:41,676 INFO [train.py:904] (4/8) Epoch 4, batch 1650, loss[loss=0.2231, simple_loss=0.2892, pruned_loss=0.07849, over 16822.00 frames. ], tot_loss[loss=0.2267, simple_loss=0.2986, pruned_loss=0.07745, over 3332574.55 frames. ], batch size: 116, lr: 1.71e-02, grad_scale: 8.0 2023-04-28 02:01:58,772 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=32113.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:02:37,463 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.433e+02 3.425e+02 4.282e+02 5.685e+02 1.161e+03, threshold=8.563e+02, percent-clipped=3.0 2023-04-28 02:02:40,714 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32144.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:02:44,926 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32147.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 02:02:50,055 INFO [train.py:904] (4/8) Epoch 4, batch 1700, loss[loss=0.2252, simple_loss=0.3194, pruned_loss=0.06548, over 17268.00 frames. ], tot_loss[loss=0.2286, simple_loss=0.3004, pruned_loss=0.07835, over 3337650.10 frames. ], batch size: 52, lr: 1.71e-02, grad_scale: 8.0 2023-04-28 02:03:09,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1617, 5.8396, 5.8803, 5.6352, 5.7849, 6.0901, 5.9524, 5.6538], device='cuda:4'), covar=tensor([0.0625, 0.1225, 0.1045, 0.1527, 0.2060, 0.0813, 0.0908, 0.1764], device='cuda:4'), in_proj_covar=tensor([0.0262, 0.0369, 0.0342, 0.0312, 0.0418, 0.0373, 0.0293, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:03:10,429 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-28 02:03:29,666 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32179.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:03:53,202 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=32195.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:04:01,130 INFO [train.py:904] (4/8) Epoch 4, batch 1750, loss[loss=0.1748, simple_loss=0.2602, pruned_loss=0.04468, over 17282.00 frames. ], tot_loss[loss=0.228, simple_loss=0.3005, pruned_loss=0.07776, over 3340672.20 frames. ], batch size: 45, lr: 1.71e-02, grad_scale: 8.0 2023-04-28 02:04:48,326 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32234.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:04:58,762 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.044e+02 3.415e+02 4.164e+02 5.127e+02 8.958e+02, threshold=8.329e+02, percent-clipped=1.0 2023-04-28 02:05:11,464 INFO [train.py:904] (4/8) Epoch 4, batch 1800, loss[loss=0.2594, simple_loss=0.3165, pruned_loss=0.1012, over 16795.00 frames. ], tot_loss[loss=0.2304, simple_loss=0.3027, pruned_loss=0.079, over 3316023.18 frames. ], batch size: 124, lr: 1.71e-02, grad_scale: 8.0 2023-04-28 02:05:39,578 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32271.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:05:59,120 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1684, 3.1730, 1.5128, 3.2436, 2.2622, 3.2837, 1.6447, 2.5155], device='cuda:4'), covar=tensor([0.0104, 0.0243, 0.1372, 0.0084, 0.0675, 0.0301, 0.1169, 0.0495], device='cuda:4'), in_proj_covar=tensor([0.0105, 0.0148, 0.0173, 0.0084, 0.0160, 0.0176, 0.0182, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 02:06:04,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3750, 5.3058, 5.1188, 5.1347, 4.6721, 5.1951, 5.1202, 4.8193], device='cuda:4'), covar=tensor([0.0307, 0.0170, 0.0155, 0.0123, 0.0789, 0.0192, 0.0150, 0.0331], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0162, 0.0208, 0.0175, 0.0248, 0.0195, 0.0151, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:06:18,944 INFO [train.py:904] (4/8) Epoch 4, batch 1850, loss[loss=0.218, simple_loss=0.2965, pruned_loss=0.06978, over 16784.00 frames. ], tot_loss[loss=0.23, simple_loss=0.3029, pruned_loss=0.0786, over 3324027.50 frames. ], batch size: 102, lr: 1.70e-02, grad_scale: 4.0 2023-04-28 02:06:26,551 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32306.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:07:08,016 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3703, 2.6290, 2.4708, 4.8338, 1.9813, 4.1863, 2.4735, 2.4288], device='cuda:4'), covar=tensor([0.0322, 0.1262, 0.0663, 0.0190, 0.2473, 0.0381, 0.1294, 0.2029], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0270, 0.0223, 0.0284, 0.0333, 0.0249, 0.0247, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:07:18,154 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.462e+02 3.616e+02 4.272e+02 5.393e+02 1.345e+03, threshold=8.544e+02, percent-clipped=8.0 2023-04-28 02:07:29,364 INFO [train.py:904] (4/8) Epoch 4, batch 1900, loss[loss=0.2123, simple_loss=0.2849, pruned_loss=0.06989, over 16757.00 frames. ], tot_loss[loss=0.2301, simple_loss=0.3031, pruned_loss=0.07848, over 3310573.89 frames. ], batch size: 83, lr: 1.70e-02, grad_scale: 4.0 2023-04-28 02:07:43,161 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32360.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:08:23,693 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-28 02:08:41,339 INFO [train.py:904] (4/8) Epoch 4, batch 1950, loss[loss=0.2295, simple_loss=0.3071, pruned_loss=0.07599, over 17096.00 frames. ], tot_loss[loss=0.2291, simple_loss=0.3026, pruned_loss=0.07779, over 3309851.83 frames. ], batch size: 53, lr: 1.70e-02, grad_scale: 4.0 2023-04-28 02:08:50,795 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=32408.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:09:16,833 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7937, 4.8109, 4.6202, 3.7435, 4.6572, 2.0056, 4.4425, 4.6829], device='cuda:4'), covar=tensor([0.0059, 0.0045, 0.0092, 0.0347, 0.0062, 0.1416, 0.0082, 0.0106], device='cuda:4'), in_proj_covar=tensor([0.0086, 0.0078, 0.0115, 0.0127, 0.0085, 0.0129, 0.0104, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:09:37,980 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.345e+02 3.206e+02 3.800e+02 4.937e+02 1.195e+03, threshold=7.601e+02, percent-clipped=2.0 2023-04-28 02:09:39,417 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32444.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:09:47,991 INFO [train.py:904] (4/8) Epoch 4, batch 2000, loss[loss=0.206, simple_loss=0.2914, pruned_loss=0.0603, over 17098.00 frames. ], tot_loss[loss=0.2277, simple_loss=0.3014, pruned_loss=0.07694, over 3312376.59 frames. ], batch size: 49, lr: 1.70e-02, grad_scale: 8.0 2023-04-28 02:10:13,073 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0116, 4.7261, 4.8971, 5.2700, 5.3713, 4.5255, 5.3397, 5.3081], device='cuda:4'), covar=tensor([0.0731, 0.0694, 0.1385, 0.0434, 0.0418, 0.0592, 0.0393, 0.0342], device='cuda:4'), in_proj_covar=tensor([0.0357, 0.0430, 0.0577, 0.0445, 0.0331, 0.0325, 0.0352, 0.0362], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:10:28,152 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32479.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:10:45,039 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=32492.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:10:58,292 INFO [train.py:904] (4/8) Epoch 4, batch 2050, loss[loss=0.2476, simple_loss=0.316, pruned_loss=0.08963, over 16330.00 frames. ], tot_loss[loss=0.2286, simple_loss=0.3022, pruned_loss=0.07751, over 3317142.75 frames. ], batch size: 165, lr: 1.70e-02, grad_scale: 8.0 2023-04-28 02:11:34,832 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=32527.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:11:44,258 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32534.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:11:57,877 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.075e+02 3.479e+02 4.226e+02 5.349e+02 1.056e+03, threshold=8.453e+02, percent-clipped=7.0 2023-04-28 02:12:07,699 INFO [train.py:904] (4/8) Epoch 4, batch 2100, loss[loss=0.1899, simple_loss=0.2703, pruned_loss=0.05479, over 17252.00 frames. ], tot_loss[loss=0.2308, simple_loss=0.3038, pruned_loss=0.07893, over 3315359.42 frames. ], batch size: 45, lr: 1.70e-02, grad_scale: 8.0 2023-04-28 02:12:36,405 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32571.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:12:50,998 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=32582.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:13:17,653 INFO [train.py:904] (4/8) Epoch 4, batch 2150, loss[loss=0.2065, simple_loss=0.2827, pruned_loss=0.06513, over 16101.00 frames. ], tot_loss[loss=0.2315, simple_loss=0.3047, pruned_loss=0.0791, over 3325455.50 frames. ], batch size: 35, lr: 1.70e-02, grad_scale: 8.0 2023-04-28 02:13:25,209 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32606.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:13:42,052 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=32619.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:14:15,065 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.513e+02 3.522e+02 4.278e+02 5.461e+02 9.146e+02, threshold=8.556e+02, percent-clipped=3.0 2023-04-28 02:14:26,829 INFO [train.py:904] (4/8) Epoch 4, batch 2200, loss[loss=0.2432, simple_loss=0.3196, pruned_loss=0.0834, over 16553.00 frames. ], tot_loss[loss=0.2336, simple_loss=0.3063, pruned_loss=0.0804, over 3320816.63 frames. ], batch size: 62, lr: 1.70e-02, grad_scale: 8.0 2023-04-28 02:14:30,432 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=32654.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:15:34,037 INFO [train.py:904] (4/8) Epoch 4, batch 2250, loss[loss=0.2481, simple_loss=0.3238, pruned_loss=0.08617, over 16743.00 frames. ], tot_loss[loss=0.2338, simple_loss=0.3063, pruned_loss=0.08062, over 3320246.32 frames. ], batch size: 57, lr: 1.69e-02, grad_scale: 8.0 2023-04-28 02:16:32,021 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.152e+02 3.558e+02 4.554e+02 5.446e+02 8.366e+02, threshold=9.108e+02, percent-clipped=0.0 2023-04-28 02:16:42,203 INFO [train.py:904] (4/8) Epoch 4, batch 2300, loss[loss=0.238, simple_loss=0.2978, pruned_loss=0.08904, over 16842.00 frames. ], tot_loss[loss=0.2326, simple_loss=0.3055, pruned_loss=0.07983, over 3326476.35 frames. ], batch size: 116, lr: 1.69e-02, grad_scale: 4.0 2023-04-28 02:17:20,522 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2882, 1.7568, 1.4860, 1.5099, 2.1433, 1.9189, 2.0710, 2.3179], device='cuda:4'), covar=tensor([0.0051, 0.0138, 0.0174, 0.0181, 0.0080, 0.0132, 0.0070, 0.0079], device='cuda:4'), in_proj_covar=tensor([0.0077, 0.0137, 0.0136, 0.0136, 0.0131, 0.0140, 0.0104, 0.0118], device='cuda:4'), out_proj_covar=tensor([9.9643e-05, 1.8462e-04, 1.7721e-04, 1.8025e-04, 1.7780e-04, 1.8998e-04, 1.3940e-04, 1.6203e-04], device='cuda:4') 2023-04-28 02:17:51,377 INFO [train.py:904] (4/8) Epoch 4, batch 2350, loss[loss=0.2396, simple_loss=0.3189, pruned_loss=0.08013, over 17165.00 frames. ], tot_loss[loss=0.2323, simple_loss=0.3052, pruned_loss=0.07972, over 3323527.58 frames. ], batch size: 46, lr: 1.69e-02, grad_scale: 4.0 2023-04-28 02:18:48,982 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.340e+02 3.303e+02 3.984e+02 5.110e+02 1.375e+03, threshold=7.969e+02, percent-clipped=3.0 2023-04-28 02:18:57,191 INFO [train.py:904] (4/8) Epoch 4, batch 2400, loss[loss=0.2563, simple_loss=0.3191, pruned_loss=0.09674, over 16725.00 frames. ], tot_loss[loss=0.2317, simple_loss=0.3049, pruned_loss=0.07921, over 3326419.22 frames. ], batch size: 134, lr: 1.69e-02, grad_scale: 8.0 2023-04-28 02:19:49,154 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.61 vs. limit=5.0 2023-04-28 02:20:06,098 INFO [train.py:904] (4/8) Epoch 4, batch 2450, loss[loss=0.2565, simple_loss=0.327, pruned_loss=0.093, over 16477.00 frames. ], tot_loss[loss=0.2326, simple_loss=0.306, pruned_loss=0.07955, over 3317580.36 frames. ], batch size: 146, lr: 1.69e-02, grad_scale: 8.0 2023-04-28 02:20:18,834 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.10 vs. limit=5.0 2023-04-28 02:21:03,694 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.564e+02 3.644e+02 4.364e+02 5.421e+02 9.049e+02, threshold=8.728e+02, percent-clipped=5.0 2023-04-28 02:21:13,283 INFO [train.py:904] (4/8) Epoch 4, batch 2500, loss[loss=0.2541, simple_loss=0.3479, pruned_loss=0.08017, over 17026.00 frames. ], tot_loss[loss=0.2303, simple_loss=0.3044, pruned_loss=0.0781, over 3325100.98 frames. ], batch size: 55, lr: 1.69e-02, grad_scale: 8.0 2023-04-28 02:21:34,442 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=32966.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:22:21,139 INFO [train.py:904] (4/8) Epoch 4, batch 2550, loss[loss=0.2095, simple_loss=0.2845, pruned_loss=0.06728, over 16993.00 frames. ], tot_loss[loss=0.2317, simple_loss=0.3055, pruned_loss=0.07894, over 3330013.02 frames. ], batch size: 41, lr: 1.69e-02, grad_scale: 8.0 2023-04-28 02:22:29,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1354, 3.8594, 3.3563, 2.0607, 2.7058, 2.2365, 3.6204, 3.6655], device='cuda:4'), covar=tensor([0.0252, 0.0428, 0.0478, 0.1345, 0.0694, 0.0860, 0.0479, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0130, 0.0153, 0.0142, 0.0133, 0.0126, 0.0140, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 02:22:57,658 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33027.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:23:20,059 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.536e+02 3.626e+02 4.489e+02 5.750e+02 1.218e+03, threshold=8.978e+02, percent-clipped=4.0 2023-04-28 02:23:30,980 INFO [train.py:904] (4/8) Epoch 4, batch 2600, loss[loss=0.1842, simple_loss=0.2729, pruned_loss=0.04774, over 16836.00 frames. ], tot_loss[loss=0.2322, simple_loss=0.3064, pruned_loss=0.07903, over 3321905.18 frames. ], batch size: 42, lr: 1.69e-02, grad_scale: 8.0 2023-04-28 02:23:31,244 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2419, 4.1222, 4.1297, 4.1609, 4.0538, 4.6940, 4.4327, 4.1906], device='cuda:4'), covar=tensor([0.1367, 0.1476, 0.1478, 0.1897, 0.2720, 0.1007, 0.1100, 0.2468], device='cuda:4'), in_proj_covar=tensor([0.0262, 0.0363, 0.0340, 0.0307, 0.0420, 0.0371, 0.0293, 0.0414], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:23:40,610 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7232, 2.6885, 1.7445, 2.7221, 2.1139, 2.8237, 1.8430, 2.3559], device='cuda:4'), covar=tensor([0.0164, 0.0361, 0.1267, 0.0094, 0.0679, 0.0579, 0.1294, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0106, 0.0148, 0.0170, 0.0084, 0.0158, 0.0179, 0.0182, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 02:23:55,378 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33069.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:24:20,490 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5915, 2.4549, 2.4065, 4.1556, 1.8967, 3.7081, 2.3245, 2.4186], device='cuda:4'), covar=tensor([0.0465, 0.1149, 0.0636, 0.0252, 0.2173, 0.0420, 0.1199, 0.1698], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0275, 0.0229, 0.0289, 0.0339, 0.0257, 0.0253, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:24:39,065 INFO [train.py:904] (4/8) Epoch 4, batch 2650, loss[loss=0.2138, simple_loss=0.3042, pruned_loss=0.06168, over 17135.00 frames. ], tot_loss[loss=0.232, simple_loss=0.3067, pruned_loss=0.07862, over 3329638.44 frames. ], batch size: 49, lr: 1.68e-02, grad_scale: 8.0 2023-04-28 02:25:19,519 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33130.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:25:38,560 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.050e+02 3.285e+02 4.083e+02 4.956e+02 1.136e+03, threshold=8.166e+02, percent-clipped=3.0 2023-04-28 02:25:46,693 INFO [train.py:904] (4/8) Epoch 4, batch 2700, loss[loss=0.1871, simple_loss=0.2769, pruned_loss=0.04863, over 16823.00 frames. ], tot_loss[loss=0.231, simple_loss=0.3065, pruned_loss=0.07773, over 3325597.09 frames. ], batch size: 39, lr: 1.68e-02, grad_scale: 8.0 2023-04-28 02:25:54,241 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33155.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:26:38,350 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-28 02:26:56,821 INFO [train.py:904] (4/8) Epoch 4, batch 2750, loss[loss=0.2384, simple_loss=0.3037, pruned_loss=0.08659, over 16936.00 frames. ], tot_loss[loss=0.2305, simple_loss=0.3063, pruned_loss=0.07735, over 3326809.28 frames. ], batch size: 116, lr: 1.68e-02, grad_scale: 8.0 2023-04-28 02:27:16,365 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33216.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:27:47,416 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1875, 4.5944, 4.8005, 4.8287, 4.7559, 4.5255, 3.9732, 4.2700], device='cuda:4'), covar=tensor([0.0460, 0.0394, 0.0499, 0.0694, 0.0581, 0.0462, 0.1429, 0.0409], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0219, 0.0224, 0.0228, 0.0275, 0.0235, 0.0342, 0.0205], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 02:27:49,524 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-28 02:27:54,082 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.476e+02 3.299e+02 3.913e+02 4.982e+02 9.564e+02, threshold=7.826e+02, percent-clipped=3.0 2023-04-28 02:28:04,360 INFO [train.py:904] (4/8) Epoch 4, batch 2800, loss[loss=0.1903, simple_loss=0.2688, pruned_loss=0.05586, over 16978.00 frames. ], tot_loss[loss=0.2304, simple_loss=0.3064, pruned_loss=0.07724, over 3332424.56 frames. ], batch size: 41, lr: 1.68e-02, grad_scale: 8.0 2023-04-28 02:28:46,160 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7307, 4.7357, 5.2583, 5.3300, 5.2685, 4.8911, 4.8297, 4.6849], device='cuda:4'), covar=tensor([0.0238, 0.0326, 0.0281, 0.0386, 0.0392, 0.0263, 0.0709, 0.0282], device='cuda:4'), in_proj_covar=tensor([0.0227, 0.0225, 0.0230, 0.0234, 0.0282, 0.0240, 0.0351, 0.0210], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 02:28:51,524 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-04-28 02:29:10,594 INFO [train.py:904] (4/8) Epoch 4, batch 2850, loss[loss=0.2319, simple_loss=0.2984, pruned_loss=0.08273, over 16385.00 frames. ], tot_loss[loss=0.2297, simple_loss=0.3052, pruned_loss=0.07713, over 3332013.19 frames. ], batch size: 146, lr: 1.68e-02, grad_scale: 8.0 2023-04-28 02:29:39,595 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=33322.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:30:09,811 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.138e+02 3.527e+02 4.279e+02 5.315e+02 1.559e+03, threshold=8.559e+02, percent-clipped=4.0 2023-04-28 02:30:20,484 INFO [train.py:904] (4/8) Epoch 4, batch 2900, loss[loss=0.2119, simple_loss=0.2813, pruned_loss=0.0712, over 16803.00 frames. ], tot_loss[loss=0.2302, simple_loss=0.3043, pruned_loss=0.07811, over 3326143.59 frames. ], batch size: 96, lr: 1.68e-02, grad_scale: 8.0 2023-04-28 02:30:53,486 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-28 02:31:29,005 INFO [train.py:904] (4/8) Epoch 4, batch 2950, loss[loss=0.1995, simple_loss=0.2736, pruned_loss=0.06271, over 15934.00 frames. ], tot_loss[loss=0.2295, simple_loss=0.3031, pruned_loss=0.07794, over 3332252.63 frames. ], batch size: 35, lr: 1.68e-02, grad_scale: 8.0 2023-04-28 02:32:00,724 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-28 02:32:01,321 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=33425.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:32:24,388 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-28 02:32:27,159 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.112e+02 3.646e+02 4.580e+02 6.032e+02 1.054e+03, threshold=9.160e+02, percent-clipped=6.0 2023-04-28 02:32:28,137 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-28 02:32:35,894 INFO [train.py:904] (4/8) Epoch 4, batch 3000, loss[loss=0.1856, simple_loss=0.2687, pruned_loss=0.05124, over 15785.00 frames. ], tot_loss[loss=0.2299, simple_loss=0.3032, pruned_loss=0.07832, over 3325648.13 frames. ], batch size: 35, lr: 1.68e-02, grad_scale: 8.0 2023-04-28 02:32:35,894 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 02:32:45,553 INFO [train.py:938] (4/8) Epoch 4, validation: loss=0.1627, simple_loss=0.2694, pruned_loss=0.02796, over 944034.00 frames. 2023-04-28 02:32:45,554 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 02:32:50,975 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7043, 1.5456, 2.0897, 2.5161, 2.6397, 2.5551, 1.4986, 2.7114], device='cuda:4'), covar=tensor([0.0044, 0.0190, 0.0127, 0.0086, 0.0064, 0.0083, 0.0188, 0.0042], device='cuda:4'), in_proj_covar=tensor([0.0109, 0.0134, 0.0122, 0.0115, 0.0113, 0.0083, 0.0129, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 02:32:52,303 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-28 02:33:12,876 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33472.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:33:54,600 INFO [train.py:904] (4/8) Epoch 4, batch 3050, loss[loss=0.246, simple_loss=0.3063, pruned_loss=0.09284, over 16843.00 frames. ], tot_loss[loss=0.2307, simple_loss=0.3039, pruned_loss=0.07874, over 3319145.92 frames. ], batch size: 116, lr: 1.67e-02, grad_scale: 4.0 2023-04-28 02:34:07,857 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=33511.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:34:38,083 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33533.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:34:41,697 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2787, 4.5860, 4.2758, 4.3876, 4.0326, 4.0409, 4.1364, 4.5616], device='cuda:4'), covar=tensor([0.0645, 0.0626, 0.0887, 0.0421, 0.0626, 0.1006, 0.0650, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0457, 0.0392, 0.0287, 0.0291, 0.0286, 0.0362, 0.0321], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:34:44,003 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-04-28 02:34:44,928 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-28 02:34:54,166 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.952e+02 3.585e+02 4.279e+02 5.215e+02 1.682e+03, threshold=8.559e+02, percent-clipped=1.0 2023-04-28 02:35:02,724 INFO [train.py:904] (4/8) Epoch 4, batch 3100, loss[loss=0.2665, simple_loss=0.3252, pruned_loss=0.1039, over 12054.00 frames. ], tot_loss[loss=0.2311, simple_loss=0.3041, pruned_loss=0.07903, over 3310754.87 frames. ], batch size: 246, lr: 1.67e-02, grad_scale: 4.0 2023-04-28 02:36:10,677 INFO [train.py:904] (4/8) Epoch 4, batch 3150, loss[loss=0.26, simple_loss=0.3169, pruned_loss=0.1016, over 16467.00 frames. ], tot_loss[loss=0.2309, simple_loss=0.3033, pruned_loss=0.07925, over 3300939.88 frames. ], batch size: 145, lr: 1.67e-02, grad_scale: 4.0 2023-04-28 02:36:39,443 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=33622.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:36:50,126 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8033, 4.6982, 4.6633, 4.0271, 4.6635, 2.1314, 4.4325, 4.6501], device='cuda:4'), covar=tensor([0.0060, 0.0054, 0.0077, 0.0299, 0.0059, 0.1258, 0.0082, 0.0117], device='cuda:4'), in_proj_covar=tensor([0.0091, 0.0081, 0.0122, 0.0131, 0.0090, 0.0131, 0.0108, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:37:11,328 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.300e+02 3.211e+02 3.811e+02 4.570e+02 1.076e+03, threshold=7.622e+02, percent-clipped=2.0 2023-04-28 02:37:18,476 INFO [train.py:904] (4/8) Epoch 4, batch 3200, loss[loss=0.1958, simple_loss=0.2761, pruned_loss=0.05776, over 16976.00 frames. ], tot_loss[loss=0.2297, simple_loss=0.3025, pruned_loss=0.07841, over 3311737.21 frames. ], batch size: 41, lr: 1.67e-02, grad_scale: 8.0 2023-04-28 02:37:43,888 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0591, 3.9163, 3.1310, 5.2650, 5.1010, 4.6061, 1.9950, 3.6693], device='cuda:4'), covar=tensor([0.1173, 0.0396, 0.0843, 0.0077, 0.0231, 0.0274, 0.1141, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0139, 0.0166, 0.0081, 0.0173, 0.0159, 0.0156, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 02:37:44,850 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=33670.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:37:54,542 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33677.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:38:25,346 INFO [train.py:904] (4/8) Epoch 4, batch 3250, loss[loss=0.2039, simple_loss=0.2905, pruned_loss=0.05864, over 17186.00 frames. ], tot_loss[loss=0.2288, simple_loss=0.302, pruned_loss=0.07778, over 3313942.74 frames. ], batch size: 46, lr: 1.67e-02, grad_scale: 8.0 2023-04-28 02:38:58,275 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=33725.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:39:16,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33738.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:39:26,580 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.718e+02 3.624e+02 4.331e+02 5.289e+02 1.384e+03, threshold=8.662e+02, percent-clipped=5.0 2023-04-28 02:39:36,477 INFO [train.py:904] (4/8) Epoch 4, batch 3300, loss[loss=0.2452, simple_loss=0.3066, pruned_loss=0.09188, over 16761.00 frames. ], tot_loss[loss=0.2309, simple_loss=0.3039, pruned_loss=0.07893, over 3318690.62 frames. ], batch size: 124, lr: 1.67e-02, grad_scale: 8.0 2023-04-28 02:40:07,409 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=33773.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:40:25,735 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4747, 3.3990, 3.8084, 2.7581, 3.6664, 3.9033, 3.7734, 1.8486], device='cuda:4'), covar=tensor([0.0256, 0.0088, 0.0023, 0.0182, 0.0031, 0.0027, 0.0028, 0.0274], device='cuda:4'), in_proj_covar=tensor([0.0109, 0.0053, 0.0056, 0.0107, 0.0057, 0.0063, 0.0058, 0.0102], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:40:33,916 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-28 02:40:45,511 INFO [train.py:904] (4/8) Epoch 4, batch 3350, loss[loss=0.2198, simple_loss=0.3076, pruned_loss=0.06595, over 17083.00 frames. ], tot_loss[loss=0.2294, simple_loss=0.3032, pruned_loss=0.07785, over 3322037.46 frames. ], batch size: 53, lr: 1.67e-02, grad_scale: 8.0 2023-04-28 02:40:58,947 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=33811.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:41:22,107 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=33828.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:41:44,267 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.269e+02 3.302e+02 3.975e+02 4.780e+02 1.300e+03, threshold=7.950e+02, percent-clipped=1.0 2023-04-28 02:41:52,910 INFO [train.py:904] (4/8) Epoch 4, batch 3400, loss[loss=0.2345, simple_loss=0.2942, pruned_loss=0.08739, over 16753.00 frames. ], tot_loss[loss=0.2296, simple_loss=0.3032, pruned_loss=0.078, over 3323294.81 frames. ], batch size: 134, lr: 1.67e-02, grad_scale: 8.0 2023-04-28 02:41:53,290 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2373, 4.1696, 4.6879, 4.6671, 4.7010, 4.3229, 4.3441, 4.2102], device='cuda:4'), covar=tensor([0.0233, 0.0350, 0.0259, 0.0393, 0.0321, 0.0312, 0.0677, 0.0436], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0229, 0.0235, 0.0241, 0.0287, 0.0250, 0.0361, 0.0215], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 02:42:04,789 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=33859.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:43:03,518 INFO [train.py:904] (4/8) Epoch 4, batch 3450, loss[loss=0.191, simple_loss=0.2699, pruned_loss=0.05604, over 15884.00 frames. ], tot_loss[loss=0.2287, simple_loss=0.3024, pruned_loss=0.07746, over 3314029.77 frames. ], batch size: 35, lr: 1.66e-02, grad_scale: 8.0 2023-04-28 02:43:58,360 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33939.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:44:05,160 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.980e+02 3.338e+02 4.153e+02 5.120e+02 1.104e+03, threshold=8.306e+02, percent-clipped=4.0 2023-04-28 02:44:13,344 INFO [train.py:904] (4/8) Epoch 4, batch 3500, loss[loss=0.2263, simple_loss=0.3082, pruned_loss=0.07222, over 16713.00 frames. ], tot_loss[loss=0.2275, simple_loss=0.3016, pruned_loss=0.07665, over 3308341.73 frames. ], batch size: 57, lr: 1.66e-02, grad_scale: 8.0 2023-04-28 02:45:02,426 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5905, 4.5486, 3.9327, 2.1053, 3.0028, 2.6607, 3.8850, 4.0542], device='cuda:4'), covar=tensor([0.0242, 0.0391, 0.0398, 0.1339, 0.0677, 0.0791, 0.0612, 0.0739], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0132, 0.0154, 0.0141, 0.0133, 0.0127, 0.0141, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 02:45:26,922 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34000.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:45:27,574 INFO [train.py:904] (4/8) Epoch 4, batch 3550, loss[loss=0.2318, simple_loss=0.3162, pruned_loss=0.07363, over 17021.00 frames. ], tot_loss[loss=0.2261, simple_loss=0.3004, pruned_loss=0.07593, over 3311947.12 frames. ], batch size: 55, lr: 1.66e-02, grad_scale: 8.0 2023-04-28 02:45:29,031 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8920, 4.7552, 4.7243, 4.1463, 4.6468, 2.1513, 4.5508, 4.7417], device='cuda:4'), covar=tensor([0.0055, 0.0061, 0.0078, 0.0279, 0.0067, 0.1346, 0.0081, 0.0114], device='cuda:4'), in_proj_covar=tensor([0.0092, 0.0080, 0.0122, 0.0130, 0.0091, 0.0130, 0.0107, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:46:11,062 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34033.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 02:46:27,243 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34044.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:46:27,925 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.116e+02 3.324e+02 4.237e+02 5.025e+02 1.251e+03, threshold=8.474e+02, percent-clipped=3.0 2023-04-28 02:46:35,603 INFO [train.py:904] (4/8) Epoch 4, batch 3600, loss[loss=0.2235, simple_loss=0.2879, pruned_loss=0.0795, over 16733.00 frames. ], tot_loss[loss=0.2243, simple_loss=0.298, pruned_loss=0.07536, over 3316308.95 frames. ], batch size: 134, lr: 1.66e-02, grad_scale: 8.0 2023-04-28 02:47:12,100 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-28 02:47:21,392 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2095, 3.4196, 2.7536, 4.7551, 4.5314, 4.4423, 1.6625, 3.3980], device='cuda:4'), covar=tensor([0.0978, 0.0344, 0.0889, 0.0062, 0.0219, 0.0202, 0.1038, 0.0496], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0136, 0.0160, 0.0079, 0.0170, 0.0157, 0.0152, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 02:47:48,231 INFO [train.py:904] (4/8) Epoch 4, batch 3650, loss[loss=0.2379, simple_loss=0.282, pruned_loss=0.09688, over 16449.00 frames. ], tot_loss[loss=0.2245, simple_loss=0.2972, pruned_loss=0.07586, over 3322558.42 frames. ], batch size: 75, lr: 1.66e-02, grad_scale: 8.0 2023-04-28 02:47:54,675 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34105.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:48:29,167 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34128.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:48:53,739 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.336e+02 3.266e+02 3.802e+02 4.883e+02 8.817e+02, threshold=7.604e+02, percent-clipped=1.0 2023-04-28 02:48:54,320 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34145.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:49:02,642 INFO [train.py:904] (4/8) Epoch 4, batch 3700, loss[loss=0.2081, simple_loss=0.2808, pruned_loss=0.06769, over 16544.00 frames. ], tot_loss[loss=0.2255, simple_loss=0.2957, pruned_loss=0.07764, over 3301081.90 frames. ], batch size: 68, lr: 1.66e-02, grad_scale: 8.0 2023-04-28 02:49:12,322 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-28 02:49:41,422 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=34176.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:49:51,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1823, 4.0517, 4.1487, 1.8697, 4.2515, 4.2822, 3.1451, 3.0246], device='cuda:4'), covar=tensor([0.0819, 0.0075, 0.0104, 0.1091, 0.0046, 0.0048, 0.0305, 0.0375], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0081, 0.0080, 0.0140, 0.0072, 0.0075, 0.0113, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 02:50:17,552 INFO [train.py:904] (4/8) Epoch 4, batch 3750, loss[loss=0.2304, simple_loss=0.292, pruned_loss=0.08438, over 16821.00 frames. ], tot_loss[loss=0.2275, simple_loss=0.2964, pruned_loss=0.0793, over 3288409.77 frames. ], batch size: 83, lr: 1.66e-02, grad_scale: 8.0 2023-04-28 02:50:25,292 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34206.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:51:21,419 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.239e+02 3.358e+02 4.060e+02 5.073e+02 1.417e+03, threshold=8.120e+02, percent-clipped=4.0 2023-04-28 02:51:27,842 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9292, 5.8075, 5.7467, 5.6004, 5.6869, 6.1784, 5.9244, 5.5437], device='cuda:4'), covar=tensor([0.0646, 0.1162, 0.0897, 0.1497, 0.1788, 0.0737, 0.0819, 0.1664], device='cuda:4'), in_proj_covar=tensor([0.0260, 0.0367, 0.0344, 0.0312, 0.0409, 0.0368, 0.0292, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:51:29,935 INFO [train.py:904] (4/8) Epoch 4, batch 3800, loss[loss=0.2501, simple_loss=0.3107, pruned_loss=0.09474, over 16380.00 frames. ], tot_loss[loss=0.2296, simple_loss=0.2974, pruned_loss=0.08083, over 3287348.90 frames. ], batch size: 165, lr: 1.66e-02, grad_scale: 8.0 2023-04-28 02:52:36,355 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34295.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:52:44,482 INFO [train.py:904] (4/8) Epoch 4, batch 3850, loss[loss=0.2376, simple_loss=0.2984, pruned_loss=0.08841, over 16416.00 frames. ], tot_loss[loss=0.2291, simple_loss=0.2968, pruned_loss=0.08073, over 3293064.58 frames. ], batch size: 146, lr: 1.65e-02, grad_scale: 8.0 2023-04-28 02:53:31,120 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34333.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:53:39,195 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 02:53:49,430 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.155e+02 3.189e+02 3.757e+02 4.739e+02 9.284e+02, threshold=7.515e+02, percent-clipped=3.0 2023-04-28 02:53:57,007 INFO [train.py:904] (4/8) Epoch 4, batch 3900, loss[loss=0.2321, simple_loss=0.3004, pruned_loss=0.08189, over 16487.00 frames. ], tot_loss[loss=0.2288, simple_loss=0.2963, pruned_loss=0.0807, over 3292046.78 frames. ], batch size: 75, lr: 1.65e-02, grad_scale: 4.0 2023-04-28 02:54:08,054 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-28 02:54:40,913 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=34381.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:55:09,051 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34400.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:55:10,261 INFO [train.py:904] (4/8) Epoch 4, batch 3950, loss[loss=0.2387, simple_loss=0.2945, pruned_loss=0.09143, over 16429.00 frames. ], tot_loss[loss=0.2285, simple_loss=0.2954, pruned_loss=0.08086, over 3286349.70 frames. ], batch size: 146, lr: 1.65e-02, grad_scale: 4.0 2023-04-28 02:56:16,757 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.978e+02 3.579e+02 4.210e+02 5.438e+02 1.110e+03, threshold=8.420e+02, percent-clipped=9.0 2023-04-28 02:56:24,481 INFO [train.py:904] (4/8) Epoch 4, batch 4000, loss[loss=0.1991, simple_loss=0.2773, pruned_loss=0.06041, over 16703.00 frames. ], tot_loss[loss=0.2281, simple_loss=0.2948, pruned_loss=0.08071, over 3284471.57 frames. ], batch size: 89, lr: 1.65e-02, grad_scale: 8.0 2023-04-28 02:56:25,457 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 02:57:17,902 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1493, 1.6737, 2.3846, 3.0441, 2.8962, 3.3761, 1.5262, 3.1388], device='cuda:4'), covar=tensor([0.0050, 0.0205, 0.0116, 0.0090, 0.0070, 0.0039, 0.0223, 0.0026], device='cuda:4'), in_proj_covar=tensor([0.0105, 0.0135, 0.0121, 0.0116, 0.0115, 0.0087, 0.0130, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 02:57:33,359 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1491, 4.5708, 1.7337, 4.6261, 2.7892, 4.6725, 2.2140, 2.8090], device='cuda:4'), covar=tensor([0.0076, 0.0096, 0.1759, 0.0021, 0.0763, 0.0144, 0.1323, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0104, 0.0149, 0.0174, 0.0084, 0.0159, 0.0177, 0.0181, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 02:57:36,430 INFO [train.py:904] (4/8) Epoch 4, batch 4050, loss[loss=0.1831, simple_loss=0.2567, pruned_loss=0.05472, over 16510.00 frames. ], tot_loss[loss=0.2245, simple_loss=0.293, pruned_loss=0.07803, over 3278888.61 frames. ], batch size: 35, lr: 1.65e-02, grad_scale: 8.0 2023-04-28 02:57:36,750 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34501.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:58:41,514 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.628e+02 2.624e+02 2.984e+02 3.642e+02 7.677e+02, threshold=5.968e+02, percent-clipped=0.0 2023-04-28 02:58:48,880 INFO [train.py:904] (4/8) Epoch 4, batch 4100, loss[loss=0.285, simple_loss=0.3507, pruned_loss=0.1096, over 15518.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.293, pruned_loss=0.07601, over 3276044.48 frames. ], batch size: 191, lr: 1.65e-02, grad_scale: 8.0 2023-04-28 02:59:06,627 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-28 02:59:54,014 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34593.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:59:56,627 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34595.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:00:06,229 INFO [train.py:904] (4/8) Epoch 4, batch 4150, loss[loss=0.2683, simple_loss=0.3387, pruned_loss=0.09893, over 17052.00 frames. ], tot_loss[loss=0.2315, simple_loss=0.3021, pruned_loss=0.08042, over 3227357.23 frames. ], batch size: 55, lr: 1.65e-02, grad_scale: 8.0 2023-04-28 03:00:12,056 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.63 vs. limit=5.0 2023-04-28 03:00:38,832 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34622.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:00:44,941 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34626.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:00:50,054 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 03:01:10,642 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=34643.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:01:15,081 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.947e+02 3.447e+02 4.221e+02 5.425e+02 1.080e+03, threshold=8.441e+02, percent-clipped=15.0 2023-04-28 03:01:22,989 INFO [train.py:904] (4/8) Epoch 4, batch 4200, loss[loss=0.2415, simple_loss=0.3261, pruned_loss=0.07846, over 17120.00 frames. ], tot_loss[loss=0.2384, simple_loss=0.3106, pruned_loss=0.0831, over 3222910.05 frames. ], batch size: 48, lr: 1.65e-02, grad_scale: 8.0 2023-04-28 03:01:28,563 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34654.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:01:56,137 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-28 03:02:13,185 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34683.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:02:19,276 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34687.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:02:28,442 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7554, 2.9725, 2.4439, 4.2656, 3.7806, 3.8336, 1.4242, 2.9416], device='cuda:4'), covar=tensor([0.1174, 0.0426, 0.1042, 0.0054, 0.0176, 0.0304, 0.1288, 0.0662], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0132, 0.0157, 0.0073, 0.0150, 0.0149, 0.0148, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 03:02:39,595 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34700.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:02:40,350 INFO [train.py:904] (4/8) Epoch 4, batch 4250, loss[loss=0.2185, simple_loss=0.3054, pruned_loss=0.06582, over 16883.00 frames. ], tot_loss[loss=0.2393, simple_loss=0.313, pruned_loss=0.08282, over 3206029.21 frames. ], batch size: 116, lr: 1.65e-02, grad_scale: 8.0 2023-04-28 03:02:53,653 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34710.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:03:43,195 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0227, 4.1925, 1.8101, 4.4280, 2.7726, 4.3829, 2.1584, 2.7942], device='cuda:4'), covar=tensor([0.0107, 0.0151, 0.1802, 0.0025, 0.0710, 0.0228, 0.1441, 0.0712], device='cuda:4'), in_proj_covar=tensor([0.0106, 0.0146, 0.0175, 0.0082, 0.0159, 0.0172, 0.0182, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 03:03:47,976 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.098e+02 2.926e+02 3.520e+02 4.399e+02 1.246e+03, threshold=7.039e+02, percent-clipped=2.0 2023-04-28 03:03:52,078 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=34748.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:03:56,214 INFO [train.py:904] (4/8) Epoch 4, batch 4300, loss[loss=0.2969, simple_loss=0.355, pruned_loss=0.1194, over 11656.00 frames. ], tot_loss[loss=0.2391, simple_loss=0.3145, pruned_loss=0.08181, over 3211478.78 frames. ], batch size: 247, lr: 1.64e-02, grad_scale: 8.0 2023-04-28 03:04:16,704 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-28 03:04:27,340 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34771.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 03:05:10,251 INFO [train.py:904] (4/8) Epoch 4, batch 4350, loss[loss=0.2297, simple_loss=0.3176, pruned_loss=0.0709, over 17090.00 frames. ], tot_loss[loss=0.2439, simple_loss=0.3193, pruned_loss=0.08425, over 3196248.39 frames. ], batch size: 47, lr: 1.64e-02, grad_scale: 8.0 2023-04-28 03:05:10,646 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34801.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:05:10,994 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 03:05:37,150 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8717, 4.0451, 4.3802, 4.3088, 4.3622, 3.9551, 4.0028, 3.8778], device='cuda:4'), covar=tensor([0.0266, 0.0252, 0.0234, 0.0347, 0.0368, 0.0287, 0.0689, 0.0353], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0199, 0.0205, 0.0211, 0.0251, 0.0221, 0.0314, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 03:05:41,860 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-28 03:06:15,520 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.361e+02 3.334e+02 4.019e+02 4.887e+02 8.589e+02, threshold=8.038e+02, percent-clipped=2.0 2023-04-28 03:06:20,217 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=34849.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:06:22,168 INFO [train.py:904] (4/8) Epoch 4, batch 4400, loss[loss=0.2444, simple_loss=0.3201, pruned_loss=0.08433, over 16771.00 frames. ], tot_loss[loss=0.2461, simple_loss=0.3216, pruned_loss=0.0853, over 3183395.90 frames. ], batch size: 124, lr: 1.64e-02, grad_scale: 8.0 2023-04-28 03:07:11,819 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9970, 5.0068, 4.6510, 4.0854, 4.8399, 1.7502, 4.6282, 4.7635], device='cuda:4'), covar=tensor([0.0039, 0.0028, 0.0063, 0.0282, 0.0039, 0.1504, 0.0045, 0.0071], device='cuda:4'), in_proj_covar=tensor([0.0083, 0.0074, 0.0110, 0.0121, 0.0082, 0.0126, 0.0098, 0.0110], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 03:07:32,093 INFO [train.py:904] (4/8) Epoch 4, batch 4450, loss[loss=0.2383, simple_loss=0.3209, pruned_loss=0.07787, over 16559.00 frames. ], tot_loss[loss=0.247, simple_loss=0.3235, pruned_loss=0.08523, over 3200092.10 frames. ], batch size: 68, lr: 1.64e-02, grad_scale: 8.0 2023-04-28 03:08:36,347 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.991e+02 2.921e+02 3.471e+02 4.143e+02 7.527e+02, threshold=6.942e+02, percent-clipped=0.0 2023-04-28 03:08:40,976 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34949.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:08:43,009 INFO [train.py:904] (4/8) Epoch 4, batch 4500, loss[loss=0.2224, simple_loss=0.3037, pruned_loss=0.07055, over 16697.00 frames. ], tot_loss[loss=0.2463, simple_loss=0.3231, pruned_loss=0.0848, over 3200274.78 frames. ], batch size: 124, lr: 1.64e-02, grad_scale: 8.0 2023-04-28 03:09:23,015 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34978.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 03:09:28,650 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34982.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:09:54,763 INFO [train.py:904] (4/8) Epoch 4, batch 4550, loss[loss=0.2609, simple_loss=0.3372, pruned_loss=0.09229, over 16252.00 frames. ], tot_loss[loss=0.2465, simple_loss=0.3236, pruned_loss=0.08468, over 3223017.45 frames. ], batch size: 35, lr: 1.64e-02, grad_scale: 8.0 2023-04-28 03:10:35,873 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9788, 2.7981, 2.7603, 4.7895, 2.1815, 4.0299, 2.7005, 2.8163], device='cuda:4'), covar=tensor([0.0416, 0.1202, 0.0676, 0.0194, 0.2408, 0.0418, 0.1187, 0.1814], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0280, 0.0232, 0.0291, 0.0349, 0.0254, 0.0253, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:10:57,618 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.276e+02 2.863e+02 3.417e+02 4.239e+02 9.930e+02, threshold=6.834e+02, percent-clipped=3.0 2023-04-28 03:11:04,051 INFO [train.py:904] (4/8) Epoch 4, batch 4600, loss[loss=0.2738, simple_loss=0.3328, pruned_loss=0.1074, over 11219.00 frames. ], tot_loss[loss=0.2464, simple_loss=0.3239, pruned_loss=0.08447, over 3212687.56 frames. ], batch size: 246, lr: 1.64e-02, grad_scale: 8.0 2023-04-28 03:11:25,504 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35066.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:12:01,907 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35092.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:12:15,705 INFO [train.py:904] (4/8) Epoch 4, batch 4650, loss[loss=0.2176, simple_loss=0.2923, pruned_loss=0.07139, over 17153.00 frames. ], tot_loss[loss=0.2447, simple_loss=0.3222, pruned_loss=0.08356, over 3217729.17 frames. ], batch size: 46, lr: 1.64e-02, grad_scale: 8.0 2023-04-28 03:13:02,001 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5491, 1.3722, 1.9789, 2.4987, 2.4070, 2.7303, 1.6138, 2.7189], device='cuda:4'), covar=tensor([0.0066, 0.0206, 0.0123, 0.0092, 0.0097, 0.0058, 0.0190, 0.0041], device='cuda:4'), in_proj_covar=tensor([0.0103, 0.0130, 0.0115, 0.0112, 0.0113, 0.0079, 0.0129, 0.0071], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 03:13:13,848 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6540, 4.5705, 5.0548, 5.1324, 5.1754, 4.6301, 4.6616, 4.3048], device='cuda:4'), covar=tensor([0.0168, 0.0229, 0.0222, 0.0262, 0.0237, 0.0220, 0.0652, 0.0312], device='cuda:4'), in_proj_covar=tensor([0.0203, 0.0200, 0.0204, 0.0208, 0.0250, 0.0216, 0.0313, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 03:13:20,502 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.572e+02 2.712e+02 3.466e+02 4.003e+02 6.879e+02, threshold=6.932e+02, percent-clipped=1.0 2023-04-28 03:13:24,763 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.69 vs. limit=5.0 2023-04-28 03:13:28,256 INFO [train.py:904] (4/8) Epoch 4, batch 4700, loss[loss=0.2282, simple_loss=0.3053, pruned_loss=0.07551, over 16925.00 frames. ], tot_loss[loss=0.2428, simple_loss=0.3203, pruned_loss=0.08267, over 3218649.63 frames. ], batch size: 109, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:13:32,537 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35153.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:14:20,964 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8979, 2.6147, 2.4532, 4.6757, 1.8942, 3.8507, 2.5052, 2.6173], device='cuda:4'), covar=tensor([0.0399, 0.1218, 0.0706, 0.0180, 0.2490, 0.0395, 0.1234, 0.1681], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0279, 0.0233, 0.0294, 0.0350, 0.0252, 0.0255, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:14:41,628 INFO [train.py:904] (4/8) Epoch 4, batch 4750, loss[loss=0.2249, simple_loss=0.3101, pruned_loss=0.06989, over 16376.00 frames. ], tot_loss[loss=0.239, simple_loss=0.3164, pruned_loss=0.08077, over 3223124.58 frames. ], batch size: 146, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:15:45,876 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.092e+02 2.820e+02 3.474e+02 4.196e+02 7.562e+02, threshold=6.948e+02, percent-clipped=1.0 2023-04-28 03:15:52,034 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35249.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:15:53,750 INFO [train.py:904] (4/8) Epoch 4, batch 4800, loss[loss=0.3005, simple_loss=0.3533, pruned_loss=0.1239, over 12022.00 frames. ], tot_loss[loss=0.2361, simple_loss=0.3137, pruned_loss=0.07925, over 3221371.91 frames. ], batch size: 248, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:16:14,826 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-28 03:16:15,721 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6574, 4.4795, 4.5365, 4.9132, 4.9951, 4.4789, 5.0408, 4.9499], device='cuda:4'), covar=tensor([0.0561, 0.0568, 0.1074, 0.0392, 0.0414, 0.0444, 0.0377, 0.0411], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0388, 0.0520, 0.0403, 0.0303, 0.0291, 0.0316, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:16:24,896 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 03:16:33,438 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35278.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:16:38,640 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35282.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:16:59,981 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=35297.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:17:05,714 INFO [train.py:904] (4/8) Epoch 4, batch 4850, loss[loss=0.222, simple_loss=0.3008, pruned_loss=0.07161, over 16626.00 frames. ], tot_loss[loss=0.2363, simple_loss=0.3147, pruned_loss=0.07889, over 3202826.78 frames. ], batch size: 57, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:17:07,227 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9673, 3.1400, 3.5030, 3.5051, 3.4673, 3.2133, 3.2696, 3.2860], device='cuda:4'), covar=tensor([0.0312, 0.0554, 0.0348, 0.0413, 0.0426, 0.0350, 0.0688, 0.0384], device='cuda:4'), in_proj_covar=tensor([0.0205, 0.0202, 0.0208, 0.0208, 0.0250, 0.0218, 0.0313, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 03:17:41,749 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=35326.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:17:48,532 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=35330.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:18:11,872 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35345.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:18:12,516 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.805e+02 3.003e+02 3.470e+02 4.055e+02 1.018e+03, threshold=6.940e+02, percent-clipped=2.0 2023-04-28 03:18:18,445 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0926, 3.3932, 3.4076, 1.5861, 3.6531, 3.6394, 2.8084, 2.6141], device='cuda:4'), covar=tensor([0.0939, 0.0115, 0.0118, 0.1348, 0.0052, 0.0058, 0.0357, 0.0535], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0085, 0.0080, 0.0144, 0.0070, 0.0075, 0.0115, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:18:19,084 INFO [train.py:904] (4/8) Epoch 4, batch 4900, loss[loss=0.2266, simple_loss=0.3098, pruned_loss=0.0717, over 16326.00 frames. ], tot_loss[loss=0.2349, simple_loss=0.3137, pruned_loss=0.07801, over 3187596.87 frames. ], batch size: 146, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:18:42,141 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35366.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 03:18:53,774 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35374.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:19:32,910 INFO [train.py:904] (4/8) Epoch 4, batch 4950, loss[loss=0.2789, simple_loss=0.3588, pruned_loss=0.09949, over 15223.00 frames. ], tot_loss[loss=0.2346, simple_loss=0.3136, pruned_loss=0.07778, over 3200591.12 frames. ], batch size: 190, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:19:41,893 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35406.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 03:19:53,435 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=35414.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:20:22,782 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35435.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:20:37,430 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.040e+02 3.403e+02 3.998e+02 5.018e+02 8.259e+02, threshold=7.996e+02, percent-clipped=7.0 2023-04-28 03:20:41,762 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35448.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:20:45,630 INFO [train.py:904] (4/8) Epoch 4, batch 5000, loss[loss=0.2223, simple_loss=0.3092, pruned_loss=0.06769, over 16324.00 frames. ], tot_loss[loss=0.236, simple_loss=0.3154, pruned_loss=0.07826, over 3199618.44 frames. ], batch size: 165, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:21:57,693 INFO [train.py:904] (4/8) Epoch 4, batch 5050, loss[loss=0.2208, simple_loss=0.3065, pruned_loss=0.06756, over 17256.00 frames. ], tot_loss[loss=0.2359, simple_loss=0.3156, pruned_loss=0.0781, over 3194477.55 frames. ], batch size: 52, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:22:39,988 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35530.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:23:03,493 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.171e+02 3.027e+02 3.520e+02 4.379e+02 8.983e+02, threshold=7.040e+02, percent-clipped=1.0 2023-04-28 03:23:10,553 INFO [train.py:904] (4/8) Epoch 4, batch 5100, loss[loss=0.2128, simple_loss=0.2977, pruned_loss=0.06396, over 16902.00 frames. ], tot_loss[loss=0.2329, simple_loss=0.3129, pruned_loss=0.07645, over 3206210.18 frames. ], batch size: 90, lr: 1.63e-02, grad_scale: 8.0 2023-04-28 03:24:08,125 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35591.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:24:22,823 INFO [train.py:904] (4/8) Epoch 4, batch 5150, loss[loss=0.2494, simple_loss=0.3376, pruned_loss=0.0806, over 16897.00 frames. ], tot_loss[loss=0.2326, simple_loss=0.3136, pruned_loss=0.07581, over 3195946.30 frames. ], batch size: 109, lr: 1.62e-02, grad_scale: 8.0 2023-04-28 03:25:29,063 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.054e+02 2.805e+02 3.383e+02 4.164e+02 1.002e+03, threshold=6.766e+02, percent-clipped=7.0 2023-04-28 03:25:36,102 INFO [train.py:904] (4/8) Epoch 4, batch 5200, loss[loss=0.211, simple_loss=0.2878, pruned_loss=0.06709, over 16638.00 frames. ], tot_loss[loss=0.2307, simple_loss=0.3116, pruned_loss=0.07486, over 3204560.93 frames. ], batch size: 62, lr: 1.62e-02, grad_scale: 8.0 2023-04-28 03:26:03,067 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3896, 4.5044, 4.6207, 4.5012, 4.4279, 5.0397, 4.7415, 4.3623], device='cuda:4'), covar=tensor([0.0880, 0.1195, 0.1007, 0.1234, 0.2196, 0.0888, 0.0921, 0.1802], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0338, 0.0319, 0.0297, 0.0393, 0.0350, 0.0274, 0.0394], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 03:26:37,349 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2787, 5.5776, 5.1046, 5.3013, 4.9082, 4.6453, 4.9971, 5.5776], device='cuda:4'), covar=tensor([0.0473, 0.0547, 0.0870, 0.0367, 0.0506, 0.0482, 0.0479, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0439, 0.0385, 0.0286, 0.0281, 0.0284, 0.0350, 0.0307], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:26:46,332 INFO [train.py:904] (4/8) Epoch 4, batch 5250, loss[loss=0.2316, simple_loss=0.3166, pruned_loss=0.07329, over 16266.00 frames. ], tot_loss[loss=0.2293, simple_loss=0.3093, pruned_loss=0.07464, over 3208861.28 frames. ], batch size: 165, lr: 1.62e-02, grad_scale: 8.0 2023-04-28 03:26:47,351 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35701.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 03:26:57,958 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0986, 2.1795, 1.6410, 1.9755, 2.7246, 2.5460, 3.0626, 2.9689], device='cuda:4'), covar=tensor([0.0023, 0.0179, 0.0228, 0.0208, 0.0084, 0.0132, 0.0044, 0.0072], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0143, 0.0143, 0.0142, 0.0133, 0.0148, 0.0102, 0.0121], device='cuda:4'), out_proj_covar=tensor([9.0215e-05, 1.8888e-04, 1.8191e-04, 1.8474e-04, 1.7771e-04, 1.9696e-04, 1.3244e-04, 1.6140e-04], device='cuda:4') 2023-04-28 03:27:28,554 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35730.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:27:52,068 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.847e+02 2.886e+02 3.406e+02 4.102e+02 7.546e+02, threshold=6.813e+02, percent-clipped=1.0 2023-04-28 03:27:55,683 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35748.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:28:00,137 INFO [train.py:904] (4/8) Epoch 4, batch 5300, loss[loss=0.1903, simple_loss=0.2758, pruned_loss=0.0524, over 16803.00 frames. ], tot_loss[loss=0.2259, simple_loss=0.3053, pruned_loss=0.07323, over 3201033.93 frames. ], batch size: 102, lr: 1.62e-02, grad_scale: 8.0 2023-04-28 03:28:17,142 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3945, 4.2438, 3.7647, 2.0725, 2.9843, 2.6291, 3.8329, 3.9133], device='cuda:4'), covar=tensor([0.0228, 0.0413, 0.0452, 0.1494, 0.0712, 0.0852, 0.0571, 0.0512], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0123, 0.0153, 0.0142, 0.0136, 0.0128, 0.0142, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 03:28:28,847 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35771.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:29:02,750 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=35796.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:29:11,333 INFO [train.py:904] (4/8) Epoch 4, batch 5350, loss[loss=0.2302, simple_loss=0.315, pruned_loss=0.07266, over 16802.00 frames. ], tot_loss[loss=0.2241, simple_loss=0.3033, pruned_loss=0.07247, over 3206805.85 frames. ], batch size: 116, lr: 1.62e-02, grad_scale: 4.0 2023-04-28 03:29:22,546 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0538, 1.4269, 2.1727, 2.9545, 2.5331, 3.2969, 1.5138, 3.0741], device='cuda:4'), covar=tensor([0.0050, 0.0233, 0.0143, 0.0089, 0.0096, 0.0044, 0.0239, 0.0042], device='cuda:4'), in_proj_covar=tensor([0.0100, 0.0128, 0.0119, 0.0111, 0.0114, 0.0080, 0.0130, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 03:29:24,935 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35811.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:29:44,291 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-28 03:29:56,913 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35832.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:30:17,960 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.047e+02 3.081e+02 3.703e+02 4.509e+02 8.729e+02, threshold=7.406e+02, percent-clipped=5.0 2023-04-28 03:30:23,336 INFO [train.py:904] (4/8) Epoch 4, batch 5400, loss[loss=0.3163, simple_loss=0.3681, pruned_loss=0.1323, over 12235.00 frames. ], tot_loss[loss=0.2282, simple_loss=0.3072, pruned_loss=0.07464, over 3183982.00 frames. ], batch size: 248, lr: 1.62e-02, grad_scale: 4.0 2023-04-28 03:30:40,331 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-28 03:30:53,687 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35872.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:31:06,144 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6223, 1.9491, 1.4951, 1.6354, 2.3389, 2.1614, 2.5207, 2.4922], device='cuda:4'), covar=tensor([0.0027, 0.0177, 0.0225, 0.0229, 0.0091, 0.0153, 0.0047, 0.0073], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0144, 0.0144, 0.0141, 0.0136, 0.0149, 0.0101, 0.0120], device='cuda:4'), out_proj_covar=tensor([9.0794e-05, 1.8986e-04, 1.8346e-04, 1.8211e-04, 1.8120e-04, 1.9765e-04, 1.3149e-04, 1.6081e-04], device='cuda:4') 2023-04-28 03:31:13,807 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35886.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:31:25,536 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9021, 2.7595, 2.6919, 1.7460, 2.8905, 2.8736, 2.4932, 2.3019], device='cuda:4'), covar=tensor([0.0736, 0.0111, 0.0147, 0.0905, 0.0072, 0.0085, 0.0302, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0082, 0.0077, 0.0137, 0.0067, 0.0072, 0.0110, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:31:38,572 INFO [train.py:904] (4/8) Epoch 4, batch 5450, loss[loss=0.2815, simple_loss=0.3472, pruned_loss=0.1079, over 11972.00 frames. ], tot_loss[loss=0.2318, simple_loss=0.3107, pruned_loss=0.07644, over 3185370.47 frames. ], batch size: 246, lr: 1.62e-02, grad_scale: 4.0 2023-04-28 03:31:55,040 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-28 03:32:03,244 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.28 vs. limit=5.0 2023-04-28 03:32:50,253 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.590e+02 3.796e+02 5.362e+02 7.541e+02 3.116e+03, threshold=1.072e+03, percent-clipped=25.0 2023-04-28 03:32:56,433 INFO [train.py:904] (4/8) Epoch 4, batch 5500, loss[loss=0.3421, simple_loss=0.3833, pruned_loss=0.1504, over 11686.00 frames. ], tot_loss[loss=0.2459, simple_loss=0.3217, pruned_loss=0.08509, over 3141076.87 frames. ], batch size: 247, lr: 1.62e-02, grad_scale: 4.0 2023-04-28 03:34:16,678 INFO [train.py:904] (4/8) Epoch 4, batch 5550, loss[loss=0.2359, simple_loss=0.3266, pruned_loss=0.0726, over 16847.00 frames. ], tot_loss[loss=0.2569, simple_loss=0.3304, pruned_loss=0.09173, over 3126351.35 frames. ], batch size: 102, lr: 1.62e-02, grad_scale: 4.0 2023-04-28 03:34:17,784 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36001.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 03:34:35,520 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-28 03:35:02,199 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36030.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:35:29,646 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.087e+02 4.484e+02 5.536e+02 7.020e+02 1.224e+03, threshold=1.107e+03, percent-clipped=3.0 2023-04-28 03:35:33,005 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=36049.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:35:35,571 INFO [train.py:904] (4/8) Epoch 4, batch 5600, loss[loss=0.3552, simple_loss=0.3862, pruned_loss=0.1621, over 11117.00 frames. ], tot_loss[loss=0.2679, simple_loss=0.338, pruned_loss=0.09893, over 3077068.65 frames. ], batch size: 248, lr: 1.61e-02, grad_scale: 8.0 2023-04-28 03:36:08,398 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-28 03:36:20,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=36078.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:36:57,486 INFO [train.py:904] (4/8) Epoch 4, batch 5650, loss[loss=0.2513, simple_loss=0.324, pruned_loss=0.08928, over 16408.00 frames. ], tot_loss[loss=0.2751, simple_loss=0.3434, pruned_loss=0.1034, over 3061484.99 frames. ], batch size: 68, lr: 1.61e-02, grad_scale: 8.0 2023-04-28 03:37:41,228 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36127.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:37:41,360 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4567, 3.4175, 3.3399, 2.9179, 3.3674, 2.1703, 3.1761, 3.0835], device='cuda:4'), covar=tensor([0.0076, 0.0067, 0.0100, 0.0224, 0.0056, 0.1188, 0.0086, 0.0119], device='cuda:4'), in_proj_covar=tensor([0.0079, 0.0070, 0.0106, 0.0117, 0.0079, 0.0126, 0.0093, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:38:09,921 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.856e+02 5.092e+02 6.431e+02 7.893e+02 1.649e+03, threshold=1.286e+03, percent-clipped=5.0 2023-04-28 03:38:17,751 INFO [train.py:904] (4/8) Epoch 4, batch 5700, loss[loss=0.2618, simple_loss=0.3449, pruned_loss=0.08933, over 16713.00 frames. ], tot_loss[loss=0.2769, simple_loss=0.3447, pruned_loss=0.1045, over 3071007.18 frames. ], batch size: 83, lr: 1.61e-02, grad_scale: 8.0 2023-04-28 03:38:42,661 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36167.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:39:05,572 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36181.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:39:13,392 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36186.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:39:36,186 INFO [train.py:904] (4/8) Epoch 4, batch 5750, loss[loss=0.2773, simple_loss=0.3501, pruned_loss=0.1022, over 16208.00 frames. ], tot_loss[loss=0.2793, simple_loss=0.3475, pruned_loss=0.1056, over 3071092.62 frames. ], batch size: 165, lr: 1.61e-02, grad_scale: 8.0 2023-04-28 03:40:29,639 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=36234.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:40:42,451 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36242.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:40:49,464 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.564e+02 4.215e+02 5.340e+02 6.535e+02 1.180e+03, threshold=1.068e+03, percent-clipped=0.0 2023-04-28 03:40:55,811 INFO [train.py:904] (4/8) Epoch 4, batch 5800, loss[loss=0.2847, simple_loss=0.3525, pruned_loss=0.1084, over 16822.00 frames. ], tot_loss[loss=0.2788, simple_loss=0.3475, pruned_loss=0.1051, over 3059269.47 frames. ], batch size: 42, lr: 1.61e-02, grad_scale: 8.0 2023-04-28 03:40:58,245 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9382, 2.6814, 2.6457, 1.7700, 2.7816, 2.8664, 2.4296, 2.3541], device='cuda:4'), covar=tensor([0.0692, 0.0116, 0.0176, 0.0895, 0.0083, 0.0090, 0.0330, 0.0387], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0082, 0.0078, 0.0139, 0.0068, 0.0072, 0.0112, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:41:33,373 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-04-28 03:42:12,358 INFO [train.py:904] (4/8) Epoch 4, batch 5850, loss[loss=0.2615, simple_loss=0.3388, pruned_loss=0.09208, over 16100.00 frames. ], tot_loss[loss=0.2751, simple_loss=0.3449, pruned_loss=0.1027, over 3071355.82 frames. ], batch size: 165, lr: 1.61e-02, grad_scale: 8.0 2023-04-28 03:42:24,273 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1842, 4.2109, 4.3207, 4.3181, 4.2308, 4.7529, 4.4427, 4.2509], device='cuda:4'), covar=tensor([0.1204, 0.1502, 0.1013, 0.1753, 0.2379, 0.0889, 0.0973, 0.1972], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0348, 0.0331, 0.0309, 0.0403, 0.0364, 0.0283, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 03:43:29,064 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.296e+02 4.315e+02 5.493e+02 6.918e+02 1.909e+03, threshold=1.099e+03, percent-clipped=3.0 2023-04-28 03:43:33,995 INFO [train.py:904] (4/8) Epoch 4, batch 5900, loss[loss=0.2508, simple_loss=0.3319, pruned_loss=0.08484, over 17117.00 frames. ], tot_loss[loss=0.2736, simple_loss=0.344, pruned_loss=0.1015, over 3092650.44 frames. ], batch size: 48, lr: 1.61e-02, grad_scale: 4.0 2023-04-28 03:44:56,223 INFO [train.py:904] (4/8) Epoch 4, batch 5950, loss[loss=0.2891, simple_loss=0.348, pruned_loss=0.1151, over 11730.00 frames. ], tot_loss[loss=0.2706, simple_loss=0.3436, pruned_loss=0.09884, over 3104618.61 frames. ], batch size: 248, lr: 1.61e-02, grad_scale: 4.0 2023-04-28 03:45:10,532 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9231, 4.7948, 4.7368, 4.7153, 4.2925, 4.7036, 4.7518, 4.5119], device='cuda:4'), covar=tensor([0.0364, 0.0244, 0.0173, 0.0134, 0.0754, 0.0280, 0.0190, 0.0341], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0151, 0.0188, 0.0156, 0.0215, 0.0178, 0.0139, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:45:25,664 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-28 03:45:37,480 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36427.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:46:10,128 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.426e+02 4.337e+02 5.242e+02 6.578e+02 1.706e+03, threshold=1.048e+03, percent-clipped=3.0 2023-04-28 03:46:14,570 INFO [train.py:904] (4/8) Epoch 4, batch 6000, loss[loss=0.2478, simple_loss=0.3253, pruned_loss=0.08512, over 16501.00 frames. ], tot_loss[loss=0.2698, simple_loss=0.3426, pruned_loss=0.09851, over 3101761.28 frames. ], batch size: 68, lr: 1.61e-02, grad_scale: 8.0 2023-04-28 03:46:14,570 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 03:46:25,217 INFO [train.py:938] (4/8) Epoch 4, validation: loss=0.1966, simple_loss=0.3069, pruned_loss=0.04309, over 944034.00 frames. 2023-04-28 03:46:25,222 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 03:46:49,465 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36467.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:47:01,318 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=36475.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:47:02,803 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4148, 3.4433, 2.7374, 2.1478, 2.5317, 2.0532, 3.4231, 3.6765], device='cuda:4'), covar=tensor([0.2181, 0.0730, 0.1225, 0.1441, 0.1907, 0.1369, 0.0447, 0.0526], device='cuda:4'), in_proj_covar=tensor([0.0275, 0.0245, 0.0263, 0.0230, 0.0305, 0.0196, 0.0229, 0.0232], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:47:44,054 INFO [train.py:904] (4/8) Epoch 4, batch 6050, loss[loss=0.2729, simple_loss=0.3485, pruned_loss=0.09867, over 16332.00 frames. ], tot_loss[loss=0.2673, simple_loss=0.3402, pruned_loss=0.09722, over 3106106.36 frames. ], batch size: 165, lr: 1.61e-02, grad_scale: 8.0 2023-04-28 03:48:06,437 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=36515.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:48:24,380 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4685, 3.3392, 2.7108, 2.1949, 2.4867, 2.1187, 3.3690, 3.5894], device='cuda:4'), covar=tensor([0.1886, 0.0652, 0.1091, 0.1306, 0.1687, 0.1202, 0.0407, 0.0504], device='cuda:4'), in_proj_covar=tensor([0.0276, 0.0245, 0.0263, 0.0231, 0.0306, 0.0197, 0.0229, 0.0234], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:48:42,913 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36537.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:48:58,310 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.180e+02 4.133e+02 5.686e+02 7.938e+02 1.625e+03, threshold=1.137e+03, percent-clipped=9.0 2023-04-28 03:49:03,039 INFO [train.py:904] (4/8) Epoch 4, batch 6100, loss[loss=0.2664, simple_loss=0.3419, pruned_loss=0.09541, over 16606.00 frames. ], tot_loss[loss=0.266, simple_loss=0.339, pruned_loss=0.09644, over 3089440.38 frames. ], batch size: 57, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:50:22,075 INFO [train.py:904] (4/8) Epoch 4, batch 6150, loss[loss=0.2362, simple_loss=0.3131, pruned_loss=0.07959, over 16722.00 frames. ], tot_loss[loss=0.2632, simple_loss=0.3366, pruned_loss=0.09494, over 3117250.89 frames. ], batch size: 124, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:50:29,581 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36605.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:50:35,050 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36608.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:50:41,748 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36612.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:51:38,876 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.478e+02 3.823e+02 4.755e+02 6.026e+02 1.290e+03, threshold=9.509e+02, percent-clipped=1.0 2023-04-28 03:51:44,318 INFO [train.py:904] (4/8) Epoch 4, batch 6200, loss[loss=0.2563, simple_loss=0.3282, pruned_loss=0.09217, over 16512.00 frames. ], tot_loss[loss=0.2611, simple_loss=0.3346, pruned_loss=0.09375, over 3129683.33 frames. ], batch size: 68, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:52:05,822 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8005, 4.1146, 3.8496, 3.8955, 3.5357, 3.6380, 3.7704, 4.0057], device='cuda:4'), covar=tensor([0.0636, 0.0649, 0.0937, 0.0495, 0.0633, 0.1175, 0.0631, 0.0914], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0434, 0.0385, 0.0286, 0.0279, 0.0288, 0.0352, 0.0305], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:52:07,196 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36666.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:52:11,583 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36669.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:52:17,417 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36673.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:52:57,880 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36700.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:52:59,109 INFO [train.py:904] (4/8) Epoch 4, batch 6250, loss[loss=0.2537, simple_loss=0.3403, pruned_loss=0.08355, over 16825.00 frames. ], tot_loss[loss=0.2619, simple_loss=0.3352, pruned_loss=0.09432, over 3120960.94 frames. ], batch size: 102, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:54:06,743 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6277, 4.7218, 4.6281, 3.1735, 4.3338, 4.6264, 4.3993, 2.6890], device='cuda:4'), covar=tensor([0.0288, 0.0017, 0.0028, 0.0197, 0.0028, 0.0043, 0.0022, 0.0235], device='cuda:4'), in_proj_covar=tensor([0.0112, 0.0050, 0.0055, 0.0110, 0.0056, 0.0064, 0.0058, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 03:54:09,135 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.662e+02 4.271e+02 5.056e+02 6.233e+02 1.402e+03, threshold=1.011e+03, percent-clipped=5.0 2023-04-28 03:54:14,280 INFO [train.py:904] (4/8) Epoch 4, batch 6300, loss[loss=0.2799, simple_loss=0.3394, pruned_loss=0.1102, over 11998.00 frames. ], tot_loss[loss=0.2606, simple_loss=0.3345, pruned_loss=0.09337, over 3124468.45 frames. ], batch size: 247, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:54:31,229 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36761.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:55:05,997 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.17 vs. limit=5.0 2023-04-28 03:55:32,004 INFO [train.py:904] (4/8) Epoch 4, batch 6350, loss[loss=0.3349, simple_loss=0.3742, pruned_loss=0.1478, over 11174.00 frames. ], tot_loss[loss=0.2638, simple_loss=0.3364, pruned_loss=0.09561, over 3116703.64 frames. ], batch size: 248, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:55:38,768 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36805.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:56:27,999 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36837.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:56:43,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.626e+02 4.373e+02 5.288e+02 7.014e+02 2.020e+03, threshold=1.058e+03, percent-clipped=6.0 2023-04-28 03:56:48,100 INFO [train.py:904] (4/8) Epoch 4, batch 6400, loss[loss=0.3238, simple_loss=0.3732, pruned_loss=0.1372, over 11253.00 frames. ], tot_loss[loss=0.2658, simple_loss=0.3374, pruned_loss=0.09713, over 3093130.80 frames. ], batch size: 247, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:57:10,867 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36866.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:57:37,188 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1273, 3.9580, 4.0298, 4.2998, 4.3895, 3.9159, 4.3634, 4.3538], device='cuda:4'), covar=tensor([0.0737, 0.0633, 0.1086, 0.0450, 0.0398, 0.0866, 0.0493, 0.0405], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0421, 0.0537, 0.0422, 0.0317, 0.0309, 0.0344, 0.0347], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:57:39,967 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=36885.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:58:04,368 INFO [train.py:904] (4/8) Epoch 4, batch 6450, loss[loss=0.2382, simple_loss=0.3226, pruned_loss=0.07696, over 16845.00 frames. ], tot_loss[loss=0.2643, simple_loss=0.3366, pruned_loss=0.09604, over 3094417.71 frames. ], batch size: 102, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:58:20,084 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36911.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:58:52,154 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2997, 1.9001, 1.6046, 1.6319, 2.2777, 2.1305, 2.4615, 2.5190], device='cuda:4'), covar=tensor([0.0030, 0.0148, 0.0201, 0.0198, 0.0081, 0.0141, 0.0063, 0.0103], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0137, 0.0142, 0.0138, 0.0131, 0.0143, 0.0102, 0.0120], device='cuda:4'), out_proj_covar=tensor([8.7034e-05, 1.7916e-04, 1.8021e-04, 1.7598e-04, 1.7285e-04, 1.8722e-04, 1.2948e-04, 1.5817e-04], device='cuda:4') 2023-04-28 03:59:06,349 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 03:59:18,063 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.485e+02 3.772e+02 4.534e+02 6.144e+02 1.455e+03, threshold=9.067e+02, percent-clipped=1.0 2023-04-28 03:59:23,779 INFO [train.py:904] (4/8) Epoch 4, batch 6500, loss[loss=0.2788, simple_loss=0.3455, pruned_loss=0.106, over 16684.00 frames. ], tot_loss[loss=0.2627, simple_loss=0.3345, pruned_loss=0.09544, over 3084363.51 frames. ], batch size: 134, lr: 1.60e-02, grad_scale: 8.0 2023-04-28 03:59:35,233 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36958.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:59:39,313 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36961.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:59:44,099 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36964.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:59:49,631 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36968.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:59:56,672 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36972.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:00:44,409 INFO [train.py:904] (4/8) Epoch 4, batch 6550, loss[loss=0.2624, simple_loss=0.3516, pruned_loss=0.08656, over 16939.00 frames. ], tot_loss[loss=0.2653, simple_loss=0.3373, pruned_loss=0.09665, over 3070860.82 frames. ], batch size: 96, lr: 1.59e-02, grad_scale: 8.0 2023-04-28 04:00:50,594 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=37004.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:01:13,104 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=37019.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:01:56,563 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.603e+02 4.116e+02 5.104e+02 6.571e+02 1.425e+03, threshold=1.021e+03, percent-clipped=7.0 2023-04-28 04:02:01,481 INFO [train.py:904] (4/8) Epoch 4, batch 6600, loss[loss=0.2953, simple_loss=0.3605, pruned_loss=0.115, over 15190.00 frames. ], tot_loss[loss=0.2676, simple_loss=0.3402, pruned_loss=0.09748, over 3095116.38 frames. ], batch size: 191, lr: 1.59e-02, grad_scale: 8.0 2023-04-28 04:02:09,608 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37056.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 04:02:18,252 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5323, 3.2638, 3.2085, 2.0836, 3.1560, 3.1350, 3.2030, 1.6907], device='cuda:4'), covar=tensor([0.0394, 0.0023, 0.0036, 0.0249, 0.0034, 0.0058, 0.0029, 0.0304], device='cuda:4'), in_proj_covar=tensor([0.0113, 0.0051, 0.0055, 0.0113, 0.0057, 0.0064, 0.0058, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:02:22,621 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=37065.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:03:19,954 INFO [train.py:904] (4/8) Epoch 4, batch 6650, loss[loss=0.2291, simple_loss=0.3056, pruned_loss=0.07635, over 16518.00 frames. ], tot_loss[loss=0.27, simple_loss=0.3415, pruned_loss=0.09924, over 3078829.00 frames. ], batch size: 62, lr: 1.59e-02, grad_scale: 8.0 2023-04-28 04:03:26,777 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 04:03:51,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2253, 4.2118, 4.2960, 1.7462, 4.5805, 4.6483, 3.0865, 3.3260], device='cuda:4'), covar=tensor([0.0992, 0.0097, 0.0197, 0.1386, 0.0043, 0.0038, 0.0383, 0.0426], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0085, 0.0082, 0.0144, 0.0069, 0.0073, 0.0115, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 04:04:33,022 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.678e+02 4.226e+02 5.482e+02 6.973e+02 1.276e+03, threshold=1.096e+03, percent-clipped=4.0 2023-04-28 04:04:37,774 INFO [train.py:904] (4/8) Epoch 4, batch 6700, loss[loss=0.3241, simple_loss=0.3658, pruned_loss=0.1412, over 11297.00 frames. ], tot_loss[loss=0.2679, simple_loss=0.339, pruned_loss=0.09844, over 3066349.62 frames. ], batch size: 247, lr: 1.59e-02, grad_scale: 8.0 2023-04-28 04:04:53,075 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37161.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 04:05:17,991 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0157, 4.7418, 4.9928, 5.2919, 5.4055, 4.6155, 5.2984, 5.2614], device='cuda:4'), covar=tensor([0.0887, 0.0761, 0.1115, 0.0403, 0.0308, 0.0500, 0.0414, 0.0402], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0414, 0.0532, 0.0413, 0.0312, 0.0305, 0.0343, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:05:46,574 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2782, 5.1763, 5.0949, 5.0391, 4.5729, 5.1859, 5.0313, 4.7748], device='cuda:4'), covar=tensor([0.0368, 0.0195, 0.0168, 0.0129, 0.0836, 0.0183, 0.0161, 0.0392], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0156, 0.0194, 0.0161, 0.0222, 0.0184, 0.0146, 0.0208], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:05:53,495 INFO [train.py:904] (4/8) Epoch 4, batch 6750, loss[loss=0.2451, simple_loss=0.3257, pruned_loss=0.08222, over 16853.00 frames. ], tot_loss[loss=0.2676, simple_loss=0.3381, pruned_loss=0.09858, over 3070477.46 frames. ], batch size: 116, lr: 1.59e-02, grad_scale: 8.0 2023-04-28 04:06:08,583 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.04 vs. limit=2.0 2023-04-28 04:06:14,417 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5196, 3.5018, 2.9028, 2.2942, 2.5872, 2.0747, 3.6191, 3.6868], device='cuda:4'), covar=tensor([0.2260, 0.0739, 0.1272, 0.1417, 0.2142, 0.1409, 0.0456, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0248, 0.0262, 0.0233, 0.0312, 0.0198, 0.0233, 0.0240], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:06:55,751 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6012, 3.5315, 2.7206, 2.2979, 2.6383, 1.9985, 3.6351, 3.7803], device='cuda:4'), covar=tensor([0.1940, 0.0650, 0.1199, 0.1337, 0.1782, 0.1403, 0.0449, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0246, 0.0261, 0.0232, 0.0310, 0.0197, 0.0230, 0.0239], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:07:06,102 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.556e+02 4.104e+02 5.227e+02 6.545e+02 1.583e+03, threshold=1.045e+03, percent-clipped=5.0 2023-04-28 04:07:10,550 INFO [train.py:904] (4/8) Epoch 4, batch 6800, loss[loss=0.2903, simple_loss=0.3604, pruned_loss=0.1101, over 15367.00 frames. ], tot_loss[loss=0.2666, simple_loss=0.3377, pruned_loss=0.09771, over 3089205.05 frames. ], batch size: 191, lr: 1.59e-02, grad_scale: 8.0 2023-04-28 04:07:27,147 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37261.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:07:31,571 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37264.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:07:35,279 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37267.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:07:37,277 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37268.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:07:37,295 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=37268.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:08:02,039 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7230, 4.9114, 4.3163, 4.8189, 4.4652, 4.2503, 4.6696, 4.9493], device='cuda:4'), covar=tensor([0.0946, 0.0991, 0.2001, 0.0671, 0.0867, 0.1147, 0.0827, 0.1018], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0426, 0.0374, 0.0281, 0.0274, 0.0281, 0.0348, 0.0302], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:08:27,076 INFO [train.py:904] (4/8) Epoch 4, batch 6850, loss[loss=0.2676, simple_loss=0.3435, pruned_loss=0.09588, over 16690.00 frames. ], tot_loss[loss=0.269, simple_loss=0.3399, pruned_loss=0.09903, over 3081747.21 frames. ], batch size: 134, lr: 1.59e-02, grad_scale: 4.0 2023-04-28 04:08:38,764 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37309.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:08:43,640 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37312.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:08:46,917 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5659, 5.8739, 5.5330, 5.6463, 5.1500, 4.7775, 5.3582, 5.9513], device='cuda:4'), covar=tensor([0.0531, 0.0568, 0.0881, 0.0374, 0.0576, 0.0527, 0.0495, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0425, 0.0373, 0.0280, 0.0272, 0.0281, 0.0349, 0.0301], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:08:46,921 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37314.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:08:49,156 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37316.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:09:08,839 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=37329.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:09:36,159 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.986e+02 4.056e+02 4.810e+02 6.733e+02 1.178e+03, threshold=9.620e+02, percent-clipped=3.0 2023-04-28 04:09:39,829 INFO [train.py:904] (4/8) Epoch 4, batch 6900, loss[loss=0.2762, simple_loss=0.3493, pruned_loss=0.1016, over 16132.00 frames. ], tot_loss[loss=0.2685, simple_loss=0.3418, pruned_loss=0.09758, over 3109818.04 frames. ], batch size: 165, lr: 1.59e-02, grad_scale: 4.0 2023-04-28 04:09:48,206 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37356.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 04:09:50,507 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5935, 2.6065, 1.6134, 2.7097, 2.1171, 2.7326, 1.8973, 2.3626], device='cuda:4'), covar=tensor([0.0168, 0.0345, 0.1249, 0.0069, 0.0659, 0.0502, 0.1129, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0109, 0.0145, 0.0174, 0.0076, 0.0157, 0.0174, 0.0183, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 04:09:53,464 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37360.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:09:58,608 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7112, 4.6564, 4.5927, 3.7408, 4.5797, 1.9335, 4.3752, 4.6080], device='cuda:4'), covar=tensor([0.0072, 0.0062, 0.0082, 0.0367, 0.0059, 0.1486, 0.0094, 0.0140], device='cuda:4'), in_proj_covar=tensor([0.0080, 0.0068, 0.0107, 0.0116, 0.0077, 0.0129, 0.0093, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:09:59,877 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=37364.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:10:41,408 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.56 vs. limit=5.0 2023-04-28 04:10:44,225 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1304, 2.0516, 2.0877, 3.4112, 1.6178, 2.8554, 1.9522, 1.8952], device='cuda:4'), covar=tensor([0.0664, 0.1782, 0.0953, 0.0463, 0.3343, 0.0812, 0.1859, 0.2485], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0282, 0.0235, 0.0296, 0.0351, 0.0261, 0.0259, 0.0349], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:10:55,352 INFO [train.py:904] (4/8) Epoch 4, batch 6950, loss[loss=0.2637, simple_loss=0.3488, pruned_loss=0.08935, over 16450.00 frames. ], tot_loss[loss=0.2738, simple_loss=0.3454, pruned_loss=0.1011, over 3094697.14 frames. ], batch size: 68, lr: 1.59e-02, grad_scale: 4.0 2023-04-28 04:11:00,663 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37404.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 04:11:23,916 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3198, 4.1756, 3.6585, 1.9982, 2.9088, 2.4252, 3.6868, 3.8075], device='cuda:4'), covar=tensor([0.0220, 0.0406, 0.0476, 0.1594, 0.0780, 0.0929, 0.0544, 0.0614], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0121, 0.0152, 0.0141, 0.0135, 0.0126, 0.0143, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 04:11:32,951 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=37425.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:12:10,356 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.259e+02 4.534e+02 5.752e+02 7.141e+02 1.491e+03, threshold=1.150e+03, percent-clipped=11.0 2023-04-28 04:12:11,785 INFO [train.py:904] (4/8) Epoch 4, batch 7000, loss[loss=0.2788, simple_loss=0.3606, pruned_loss=0.0985, over 16932.00 frames. ], tot_loss[loss=0.2722, simple_loss=0.345, pruned_loss=0.09974, over 3105461.27 frames. ], batch size: 109, lr: 1.58e-02, grad_scale: 2.0 2023-04-28 04:12:18,915 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5758, 4.4759, 4.3912, 4.2890, 3.9351, 4.4299, 4.3887, 4.1513], device='cuda:4'), covar=tensor([0.0404, 0.0290, 0.0182, 0.0152, 0.0712, 0.0291, 0.0242, 0.0394], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0154, 0.0191, 0.0158, 0.0222, 0.0186, 0.0144, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:12:28,098 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37461.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 04:13:29,622 INFO [train.py:904] (4/8) Epoch 4, batch 7050, loss[loss=0.2263, simple_loss=0.3123, pruned_loss=0.07016, over 16707.00 frames. ], tot_loss[loss=0.2718, simple_loss=0.3449, pruned_loss=0.09936, over 3076559.18 frames. ], batch size: 57, lr: 1.58e-02, grad_scale: 2.0 2023-04-28 04:13:43,149 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37509.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 04:13:50,620 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 04:14:01,469 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7521, 1.9833, 1.4415, 1.9043, 2.5379, 2.3711, 3.0580, 2.8152], device='cuda:4'), covar=tensor([0.0038, 0.0190, 0.0276, 0.0212, 0.0108, 0.0156, 0.0051, 0.0096], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0137, 0.0143, 0.0139, 0.0131, 0.0143, 0.0102, 0.0119], device='cuda:4'), out_proj_covar=tensor([8.5800e-05, 1.7853e-04, 1.8198e-04, 1.7742e-04, 1.7131e-04, 1.8704e-04, 1.2809e-04, 1.5628e-04], device='cuda:4') 2023-04-28 04:14:24,274 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3687, 4.6276, 4.2969, 4.3140, 4.0229, 4.0024, 4.2138, 4.5878], device='cuda:4'), covar=tensor([0.0555, 0.0663, 0.0974, 0.0504, 0.0613, 0.0907, 0.0556, 0.0744], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0429, 0.0378, 0.0284, 0.0276, 0.0286, 0.0351, 0.0306], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:14:45,395 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 3.107e+02 4.211e+02 5.076e+02 6.594e+02 1.566e+03, threshold=1.015e+03, percent-clipped=2.0 2023-04-28 04:14:46,712 INFO [train.py:904] (4/8) Epoch 4, batch 7100, loss[loss=0.2169, simple_loss=0.2973, pruned_loss=0.06827, over 17182.00 frames. ], tot_loss[loss=0.2711, simple_loss=0.3433, pruned_loss=0.09945, over 3056260.95 frames. ], batch size: 46, lr: 1.58e-02, grad_scale: 2.0 2023-04-28 04:15:13,033 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37567.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:15:44,180 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3294, 3.2216, 3.3493, 3.5049, 3.4998, 3.1891, 3.4587, 3.5449], device='cuda:4'), covar=tensor([0.0651, 0.0567, 0.0871, 0.0409, 0.0451, 0.1577, 0.0594, 0.0418], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0411, 0.0527, 0.0419, 0.0314, 0.0309, 0.0336, 0.0343], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:16:02,654 INFO [train.py:904] (4/8) Epoch 4, batch 7150, loss[loss=0.2551, simple_loss=0.3392, pruned_loss=0.0855, over 16280.00 frames. ], tot_loss[loss=0.2702, simple_loss=0.3415, pruned_loss=0.09949, over 3038275.83 frames. ], batch size: 35, lr: 1.58e-02, grad_scale: 2.0 2023-04-28 04:16:23,243 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37614.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:16:24,830 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37615.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:16:38,303 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37624.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:16:50,234 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3030, 2.0706, 2.0903, 3.6931, 1.5571, 2.9103, 1.9768, 1.8305], device='cuda:4'), covar=tensor([0.0622, 0.1933, 0.1022, 0.0359, 0.3601, 0.0830, 0.2089, 0.2782], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0282, 0.0235, 0.0296, 0.0351, 0.0262, 0.0260, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:17:17,121 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.158e+02 3.757e+02 4.855e+02 6.210e+02 1.451e+03, threshold=9.710e+02, percent-clipped=1.0 2023-04-28 04:17:18,898 INFO [train.py:904] (4/8) Epoch 4, batch 7200, loss[loss=0.2194, simple_loss=0.3046, pruned_loss=0.06712, over 16320.00 frames. ], tot_loss[loss=0.2674, simple_loss=0.3391, pruned_loss=0.09781, over 3032384.04 frames. ], batch size: 146, lr: 1.58e-02, grad_scale: 4.0 2023-04-28 04:17:24,932 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.04 vs. limit=2.0 2023-04-28 04:17:33,223 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37660.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:17:35,455 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37662.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:18:40,008 INFO [train.py:904] (4/8) Epoch 4, batch 7250, loss[loss=0.2324, simple_loss=0.31, pruned_loss=0.07742, over 16222.00 frames. ], tot_loss[loss=0.264, simple_loss=0.3361, pruned_loss=0.0959, over 3041083.89 frames. ], batch size: 165, lr: 1.58e-02, grad_scale: 4.0 2023-04-28 04:18:52,049 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37708.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:19:10,333 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37720.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:19:55,472 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.396e+02 3.869e+02 4.810e+02 5.847e+02 1.256e+03, threshold=9.620e+02, percent-clipped=3.0 2023-04-28 04:19:57,415 INFO [train.py:904] (4/8) Epoch 4, batch 7300, loss[loss=0.313, simple_loss=0.3622, pruned_loss=0.1319, over 11634.00 frames. ], tot_loss[loss=0.2625, simple_loss=0.3348, pruned_loss=0.09507, over 3046888.39 frames. ], batch size: 247, lr: 1.58e-02, grad_scale: 4.0 2023-04-28 04:21:14,972 INFO [train.py:904] (4/8) Epoch 4, batch 7350, loss[loss=0.2297, simple_loss=0.308, pruned_loss=0.07568, over 17262.00 frames. ], tot_loss[loss=0.2625, simple_loss=0.3347, pruned_loss=0.09518, over 3037812.81 frames. ], batch size: 52, lr: 1.58e-02, grad_scale: 4.0 2023-04-28 04:22:24,028 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4231, 2.8381, 2.5211, 2.3726, 2.2508, 2.0346, 2.8687, 2.9985], device='cuda:4'), covar=tensor([0.1512, 0.0573, 0.1028, 0.1214, 0.2084, 0.1289, 0.0399, 0.0562], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0253, 0.0269, 0.0240, 0.0315, 0.0199, 0.0233, 0.0241], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:22:29,891 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.483e+02 3.907e+02 4.910e+02 6.467e+02 1.073e+03, threshold=9.820e+02, percent-clipped=2.0 2023-04-28 04:22:31,804 INFO [train.py:904] (4/8) Epoch 4, batch 7400, loss[loss=0.3134, simple_loss=0.3643, pruned_loss=0.1312, over 11277.00 frames. ], tot_loss[loss=0.2648, simple_loss=0.3366, pruned_loss=0.0965, over 3034417.22 frames. ], batch size: 250, lr: 1.58e-02, grad_scale: 4.0 2023-04-28 04:23:49,002 INFO [train.py:904] (4/8) Epoch 4, batch 7450, loss[loss=0.3456, simple_loss=0.384, pruned_loss=0.1536, over 11683.00 frames. ], tot_loss[loss=0.2657, simple_loss=0.3375, pruned_loss=0.097, over 3052908.38 frames. ], batch size: 247, lr: 1.58e-02, grad_scale: 4.0 2023-04-28 04:24:07,672 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8380, 3.6844, 3.8392, 2.8071, 3.7594, 1.5392, 3.5968, 3.6683], device='cuda:4'), covar=tensor([0.0106, 0.0099, 0.0108, 0.0503, 0.0093, 0.1978, 0.0121, 0.0203], device='cuda:4'), in_proj_covar=tensor([0.0081, 0.0070, 0.0106, 0.0116, 0.0079, 0.0130, 0.0094, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:24:10,279 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.64 vs. limit=5.0 2023-04-28 04:24:26,698 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37924.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:25:06,400 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.829e+02 4.545e+02 5.686e+02 6.937e+02 1.455e+03, threshold=1.137e+03, percent-clipped=4.0 2023-04-28 04:25:07,721 INFO [train.py:904] (4/8) Epoch 4, batch 7500, loss[loss=0.2716, simple_loss=0.3422, pruned_loss=0.1005, over 16432.00 frames. ], tot_loss[loss=0.2649, simple_loss=0.3376, pruned_loss=0.09609, over 3067143.65 frames. ], batch size: 146, lr: 1.57e-02, grad_scale: 4.0 2023-04-28 04:25:40,563 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=37972.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:25:49,066 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1736, 4.1582, 4.2750, 2.7349, 3.9120, 4.1796, 4.0232, 2.0642], device='cuda:4'), covar=tensor([0.0325, 0.0015, 0.0016, 0.0199, 0.0028, 0.0037, 0.0024, 0.0279], device='cuda:4'), in_proj_covar=tensor([0.0113, 0.0051, 0.0056, 0.0110, 0.0058, 0.0065, 0.0060, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:25:50,234 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=37979.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:26:27,335 INFO [train.py:904] (4/8) Epoch 4, batch 7550, loss[loss=0.2886, simple_loss=0.3474, pruned_loss=0.1149, over 11177.00 frames. ], tot_loss[loss=0.2634, simple_loss=0.3361, pruned_loss=0.09537, over 3083305.89 frames. ], batch size: 246, lr: 1.57e-02, grad_scale: 4.0 2023-04-28 04:26:51,771 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.42 vs. limit=5.0 2023-04-28 04:26:55,009 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=38020.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:27:24,033 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=38040.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:27:38,139 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.661e+02 3.786e+02 4.899e+02 6.266e+02 1.464e+03, threshold=9.797e+02, percent-clipped=2.0 2023-04-28 04:27:40,071 INFO [train.py:904] (4/8) Epoch 4, batch 7600, loss[loss=0.235, simple_loss=0.3122, pruned_loss=0.07893, over 16909.00 frames. ], tot_loss[loss=0.2638, simple_loss=0.3357, pruned_loss=0.09595, over 3079201.01 frames. ], batch size: 109, lr: 1.57e-02, grad_scale: 8.0 2023-04-28 04:28:03,680 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9637, 2.1062, 1.6521, 1.8465, 2.5868, 2.3843, 3.0327, 2.9629], device='cuda:4'), covar=tensor([0.0032, 0.0194, 0.0243, 0.0225, 0.0110, 0.0164, 0.0066, 0.0083], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0139, 0.0143, 0.0139, 0.0130, 0.0144, 0.0103, 0.0119], device='cuda:4'), out_proj_covar=tensor([8.5969e-05, 1.7909e-04, 1.8068e-04, 1.7635e-04, 1.6880e-04, 1.8688e-04, 1.2932e-04, 1.5571e-04], device='cuda:4') 2023-04-28 04:28:05,545 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=38068.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:28:55,058 INFO [train.py:904] (4/8) Epoch 4, batch 7650, loss[loss=0.3145, simple_loss=0.3568, pruned_loss=0.1361, over 11417.00 frames. ], tot_loss[loss=0.2656, simple_loss=0.3371, pruned_loss=0.09709, over 3086123.62 frames. ], batch size: 248, lr: 1.57e-02, grad_scale: 8.0 2023-04-28 04:29:12,128 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-28 04:30:08,813 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.664e+02 4.200e+02 5.231e+02 7.341e+02 1.286e+03, threshold=1.046e+03, percent-clipped=4.0 2023-04-28 04:30:09,971 INFO [train.py:904] (4/8) Epoch 4, batch 7700, loss[loss=0.2684, simple_loss=0.346, pruned_loss=0.09542, over 16782.00 frames. ], tot_loss[loss=0.2662, simple_loss=0.3371, pruned_loss=0.09769, over 3084417.17 frames. ], batch size: 39, lr: 1.57e-02, grad_scale: 8.0 2023-04-28 04:30:49,242 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0976, 3.0210, 2.6233, 2.0686, 2.5728, 2.1308, 2.7042, 2.9316], device='cuda:4'), covar=tensor([0.0243, 0.0357, 0.0367, 0.1176, 0.0576, 0.0790, 0.0470, 0.0444], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0122, 0.0155, 0.0141, 0.0134, 0.0127, 0.0141, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 04:31:14,387 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-28 04:31:26,736 INFO [train.py:904] (4/8) Epoch 4, batch 7750, loss[loss=0.2369, simple_loss=0.32, pruned_loss=0.07688, over 16715.00 frames. ], tot_loss[loss=0.2659, simple_loss=0.3369, pruned_loss=0.09743, over 3068796.05 frames. ], batch size: 134, lr: 1.57e-02, grad_scale: 8.0 2023-04-28 04:31:51,811 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7643, 3.6349, 2.9318, 2.6219, 3.0241, 2.2685, 3.9229, 3.9534], device='cuda:4'), covar=tensor([0.2070, 0.0898, 0.1373, 0.1333, 0.1734, 0.1377, 0.0460, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0248, 0.0264, 0.0236, 0.0308, 0.0198, 0.0234, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:32:17,646 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.60 vs. limit=2.0 2023-04-28 04:32:40,401 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.200e+02 4.651e+02 5.629e+02 7.358e+02 1.215e+03, threshold=1.126e+03, percent-clipped=5.0 2023-04-28 04:32:42,173 INFO [train.py:904] (4/8) Epoch 4, batch 7800, loss[loss=0.3044, simple_loss=0.3646, pruned_loss=0.1222, over 15225.00 frames. ], tot_loss[loss=0.2676, simple_loss=0.3381, pruned_loss=0.09854, over 3063755.29 frames. ], batch size: 190, lr: 1.57e-02, grad_scale: 8.0 2023-04-28 04:32:52,675 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7820, 2.2445, 1.6681, 1.8645, 2.6926, 2.4479, 3.0364, 2.9175], device='cuda:4'), covar=tensor([0.0037, 0.0163, 0.0238, 0.0210, 0.0092, 0.0152, 0.0069, 0.0093], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0140, 0.0145, 0.0140, 0.0132, 0.0144, 0.0104, 0.0120], device='cuda:4'), out_proj_covar=tensor([8.8676e-05, 1.8028e-04, 1.8257e-04, 1.7694e-04, 1.7094e-04, 1.8613e-04, 1.3069e-04, 1.5619e-04], device='cuda:4') 2023-04-28 04:33:15,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3272, 2.0647, 2.0803, 3.8589, 1.8080, 3.0048, 2.1791, 2.0380], device='cuda:4'), covar=tensor([0.0536, 0.1630, 0.0939, 0.0263, 0.2695, 0.0711, 0.1600, 0.2088], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0285, 0.0236, 0.0296, 0.0354, 0.0261, 0.0258, 0.0348], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:33:58,979 INFO [train.py:904] (4/8) Epoch 4, batch 7850, loss[loss=0.2765, simple_loss=0.3536, pruned_loss=0.09966, over 17119.00 frames. ], tot_loss[loss=0.2674, simple_loss=0.3388, pruned_loss=0.098, over 3069491.59 frames. ], batch size: 48, lr: 1.57e-02, grad_scale: 8.0 2023-04-28 04:34:03,266 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8011, 3.8305, 3.0976, 2.3384, 2.8321, 2.4113, 3.8533, 4.0721], device='cuda:4'), covar=tensor([0.1816, 0.0524, 0.1085, 0.1394, 0.1847, 0.1174, 0.0370, 0.0404], device='cuda:4'), in_proj_covar=tensor([0.0277, 0.0245, 0.0262, 0.0232, 0.0305, 0.0196, 0.0231, 0.0234], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:34:50,563 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=38335.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:35:00,171 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2552, 2.4996, 1.8333, 2.1235, 2.7596, 2.7373, 3.3369, 3.1525], device='cuda:4'), covar=tensor([0.0024, 0.0140, 0.0210, 0.0183, 0.0090, 0.0123, 0.0054, 0.0076], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0139, 0.0144, 0.0139, 0.0131, 0.0142, 0.0103, 0.0120], device='cuda:4'), out_proj_covar=tensor([8.7155e-05, 1.7952e-04, 1.8159e-04, 1.7532e-04, 1.6949e-04, 1.8451e-04, 1.2947e-04, 1.5611e-04], device='cuda:4') 2023-04-28 04:35:12,442 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.779e+02 3.909e+02 4.709e+02 5.998e+02 1.164e+03, threshold=9.418e+02, percent-clipped=1.0 2023-04-28 04:35:14,725 INFO [train.py:904] (4/8) Epoch 4, batch 7900, loss[loss=0.2428, simple_loss=0.3153, pruned_loss=0.08516, over 17004.00 frames. ], tot_loss[loss=0.2656, simple_loss=0.3374, pruned_loss=0.09695, over 3071321.22 frames. ], batch size: 55, lr: 1.57e-02, grad_scale: 8.0 2023-04-28 04:35:28,535 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=38360.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:35:41,936 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1254, 4.1841, 4.2157, 4.3140, 4.2811, 4.7427, 4.4158, 4.1658], device='cuda:4'), covar=tensor([0.1449, 0.1423, 0.1329, 0.1596, 0.2225, 0.0869, 0.1018, 0.2068], device='cuda:4'), in_proj_covar=tensor([0.0252, 0.0353, 0.0340, 0.0307, 0.0408, 0.0372, 0.0284, 0.0415], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:36:19,435 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1404, 4.0269, 4.2054, 4.3839, 4.4757, 4.0177, 4.4522, 4.4883], device='cuda:4'), covar=tensor([0.0894, 0.0695, 0.1055, 0.0448, 0.0410, 0.0820, 0.0436, 0.0385], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0417, 0.0538, 0.0431, 0.0323, 0.0309, 0.0348, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:36:34,208 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.67 vs. limit=2.0 2023-04-28 04:36:34,670 INFO [train.py:904] (4/8) Epoch 4, batch 7950, loss[loss=0.2577, simple_loss=0.3401, pruned_loss=0.0877, over 16778.00 frames. ], tot_loss[loss=0.2672, simple_loss=0.3384, pruned_loss=0.09799, over 3073682.29 frames. ], batch size: 89, lr: 1.57e-02, grad_scale: 8.0 2023-04-28 04:37:05,116 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=38421.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:37:15,804 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5403, 3.4818, 3.9642, 3.9573, 3.9438, 3.5918, 3.7096, 3.6413], device='cuda:4'), covar=tensor([0.0283, 0.0429, 0.0346, 0.0463, 0.0454, 0.0339, 0.0818, 0.0432], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0207, 0.0215, 0.0214, 0.0257, 0.0220, 0.0324, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 04:37:16,077 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.26 vs. limit=5.0 2023-04-28 04:37:49,209 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.619e+02 4.198e+02 4.896e+02 6.041e+02 1.381e+03, threshold=9.792e+02, percent-clipped=5.0 2023-04-28 04:37:50,954 INFO [train.py:904] (4/8) Epoch 4, batch 8000, loss[loss=0.2443, simple_loss=0.3219, pruned_loss=0.08335, over 16605.00 frames. ], tot_loss[loss=0.268, simple_loss=0.3387, pruned_loss=0.0986, over 3078948.47 frames. ], batch size: 57, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:38:17,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8579, 3.2106, 3.1829, 1.6245, 3.3322, 3.3544, 2.6347, 2.6220], device='cuda:4'), covar=tensor([0.0831, 0.0105, 0.0124, 0.1087, 0.0072, 0.0060, 0.0336, 0.0364], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0085, 0.0084, 0.0143, 0.0072, 0.0076, 0.0116, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 04:38:45,062 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 04:39:04,863 INFO [train.py:904] (4/8) Epoch 4, batch 8050, loss[loss=0.2382, simple_loss=0.3186, pruned_loss=0.07885, over 16545.00 frames. ], tot_loss[loss=0.2655, simple_loss=0.3371, pruned_loss=0.09692, over 3091047.65 frames. ], batch size: 68, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:39:10,175 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3739, 2.1682, 2.0041, 3.9826, 1.8138, 3.2445, 2.2332, 2.0981], device='cuda:4'), covar=tensor([0.0478, 0.1567, 0.0912, 0.0253, 0.2576, 0.0581, 0.1503, 0.2111], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0284, 0.0236, 0.0297, 0.0356, 0.0260, 0.0261, 0.0349], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:39:57,770 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 04:40:07,732 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4747, 2.3619, 2.2828, 4.2260, 1.8643, 3.3897, 2.3589, 2.2033], device='cuda:4'), covar=tensor([0.0511, 0.1473, 0.0826, 0.0229, 0.2632, 0.0523, 0.1388, 0.1985], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0284, 0.0235, 0.0294, 0.0354, 0.0259, 0.0259, 0.0348], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:40:21,959 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.546e+02 4.083e+02 4.841e+02 6.395e+02 1.721e+03, threshold=9.682e+02, percent-clipped=6.0 2023-04-28 04:40:23,266 INFO [train.py:904] (4/8) Epoch 4, batch 8100, loss[loss=0.2334, simple_loss=0.3131, pruned_loss=0.07688, over 17021.00 frames. ], tot_loss[loss=0.264, simple_loss=0.3362, pruned_loss=0.09586, over 3090743.29 frames. ], batch size: 55, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:41:41,522 INFO [train.py:904] (4/8) Epoch 4, batch 8150, loss[loss=0.2665, simple_loss=0.3304, pruned_loss=0.1013, over 16295.00 frames. ], tot_loss[loss=0.2604, simple_loss=0.3332, pruned_loss=0.09384, over 3117150.50 frames. ], batch size: 146, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:42:34,076 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=38635.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:42:44,574 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.01 vs. limit=2.0 2023-04-28 04:42:57,100 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.420e+02 4.459e+02 5.761e+02 7.393e+02 1.688e+03, threshold=1.152e+03, percent-clipped=6.0 2023-04-28 04:42:59,072 INFO [train.py:904] (4/8) Epoch 4, batch 8200, loss[loss=0.2661, simple_loss=0.3447, pruned_loss=0.09377, over 16194.00 frames. ], tot_loss[loss=0.2582, simple_loss=0.3305, pruned_loss=0.09298, over 3109912.62 frames. ], batch size: 165, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:43:01,533 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2092, 3.2943, 1.4651, 3.3151, 2.3148, 3.4626, 1.7662, 2.7083], device='cuda:4'), covar=tensor([0.0113, 0.0242, 0.1547, 0.0056, 0.0673, 0.0419, 0.1284, 0.0471], device='cuda:4'), in_proj_covar=tensor([0.0110, 0.0145, 0.0177, 0.0079, 0.0159, 0.0177, 0.0184, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 04:43:53,409 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=38683.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:44:13,276 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5583, 3.3903, 3.0851, 1.8790, 2.6505, 2.2001, 2.8503, 3.3580], device='cuda:4'), covar=tensor([0.0359, 0.0469, 0.0427, 0.1464, 0.0674, 0.0869, 0.0855, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0119, 0.0153, 0.0137, 0.0131, 0.0125, 0.0139, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 04:44:23,316 INFO [train.py:904] (4/8) Epoch 4, batch 8250, loss[loss=0.2095, simple_loss=0.2868, pruned_loss=0.06616, over 12262.00 frames. ], tot_loss[loss=0.2564, simple_loss=0.3302, pruned_loss=0.09131, over 3096563.56 frames. ], batch size: 247, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:44:49,018 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=38716.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:45:15,129 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1391, 4.0514, 3.9826, 3.4190, 3.8906, 1.6311, 3.8217, 3.9109], device='cuda:4'), covar=tensor([0.0061, 0.0057, 0.0081, 0.0227, 0.0067, 0.1555, 0.0078, 0.0119], device='cuda:4'), in_proj_covar=tensor([0.0080, 0.0069, 0.0106, 0.0113, 0.0080, 0.0129, 0.0094, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:45:43,965 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.138e+02 3.522e+02 4.130e+02 5.467e+02 1.123e+03, threshold=8.260e+02, percent-clipped=0.0 2023-04-28 04:45:45,841 INFO [train.py:904] (4/8) Epoch 4, batch 8300, loss[loss=0.2033, simple_loss=0.2987, pruned_loss=0.054, over 16710.00 frames. ], tot_loss[loss=0.2497, simple_loss=0.3257, pruned_loss=0.08688, over 3093838.09 frames. ], batch size: 89, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:47:07,345 INFO [train.py:904] (4/8) Epoch 4, batch 8350, loss[loss=0.2456, simple_loss=0.3318, pruned_loss=0.07967, over 16638.00 frames. ], tot_loss[loss=0.2463, simple_loss=0.3245, pruned_loss=0.08402, over 3100435.07 frames. ], batch size: 134, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:47:44,518 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=38823.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:48:13,576 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8163, 2.6297, 2.6858, 2.0150, 2.6224, 2.5562, 2.5905, 1.7492], device='cuda:4'), covar=tensor([0.0263, 0.0025, 0.0035, 0.0185, 0.0037, 0.0053, 0.0039, 0.0293], device='cuda:4'), in_proj_covar=tensor([0.0112, 0.0049, 0.0056, 0.0110, 0.0056, 0.0065, 0.0059, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:48:29,398 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.235e+02 3.176e+02 3.884e+02 4.976e+02 1.172e+03, threshold=7.768e+02, percent-clipped=5.0 2023-04-28 04:48:30,621 INFO [train.py:904] (4/8) Epoch 4, batch 8400, loss[loss=0.2196, simple_loss=0.3052, pruned_loss=0.067, over 15318.00 frames. ], tot_loss[loss=0.2416, simple_loss=0.321, pruned_loss=0.08113, over 3079670.45 frames. ], batch size: 191, lr: 1.56e-02, grad_scale: 8.0 2023-04-28 04:49:26,891 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=38884.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:49:53,140 INFO [train.py:904] (4/8) Epoch 4, batch 8450, loss[loss=0.2377, simple_loss=0.3169, pruned_loss=0.07923, over 15194.00 frames. ], tot_loss[loss=0.2375, simple_loss=0.3182, pruned_loss=0.07834, over 3091141.13 frames. ], batch size: 190, lr: 1.56e-02, grad_scale: 4.0 2023-04-28 04:51:13,821 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.235e+02 3.132e+02 4.045e+02 5.414e+02 1.265e+03, threshold=8.090e+02, percent-clipped=8.0 2023-04-28 04:51:13,836 INFO [train.py:904] (4/8) Epoch 4, batch 8500, loss[loss=0.1995, simple_loss=0.291, pruned_loss=0.05401, over 16732.00 frames. ], tot_loss[loss=0.232, simple_loss=0.3134, pruned_loss=0.07529, over 3083354.90 frames. ], batch size: 83, lr: 1.55e-02, grad_scale: 4.0 2023-04-28 04:52:20,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8530, 1.4630, 1.8189, 2.5440, 2.4853, 2.8787, 1.7230, 2.9150], device='cuda:4'), covar=tensor([0.0059, 0.0214, 0.0179, 0.0106, 0.0105, 0.0089, 0.0208, 0.0058], device='cuda:4'), in_proj_covar=tensor([0.0103, 0.0130, 0.0118, 0.0112, 0.0116, 0.0081, 0.0129, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 04:52:39,866 INFO [train.py:904] (4/8) Epoch 4, batch 8550, loss[loss=0.2101, simple_loss=0.2856, pruned_loss=0.06731, over 11885.00 frames. ], tot_loss[loss=0.2298, simple_loss=0.3111, pruned_loss=0.07432, over 3055783.34 frames. ], batch size: 248, lr: 1.55e-02, grad_scale: 4.0 2023-04-28 04:52:53,425 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4874, 5.8863, 5.5258, 5.6247, 5.1249, 4.8847, 5.4026, 5.9305], device='cuda:4'), covar=tensor([0.0516, 0.0610, 0.0961, 0.0374, 0.0557, 0.0542, 0.0527, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0304, 0.0410, 0.0357, 0.0266, 0.0265, 0.0281, 0.0332, 0.0293], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:53:09,942 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=39016.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:53:26,412 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0246, 1.3623, 1.5336, 2.0113, 2.1084, 2.1473, 1.4468, 2.1945], device='cuda:4'), covar=tensor([0.0082, 0.0209, 0.0135, 0.0129, 0.0097, 0.0085, 0.0203, 0.0053], device='cuda:4'), in_proj_covar=tensor([0.0103, 0.0130, 0.0117, 0.0112, 0.0117, 0.0080, 0.0130, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 04:53:29,728 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.98 vs. limit=5.0 2023-04-28 04:54:11,250 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5874, 4.9033, 4.6086, 4.6000, 4.3512, 4.2737, 4.4835, 4.9395], device='cuda:4'), covar=tensor([0.0587, 0.0715, 0.0990, 0.0492, 0.0677, 0.0798, 0.0590, 0.0722], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0415, 0.0361, 0.0270, 0.0269, 0.0284, 0.0336, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:54:21,289 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.792e+02 3.391e+02 3.909e+02 5.367e+02 1.114e+03, threshold=7.819e+02, percent-clipped=4.0 2023-04-28 04:54:21,304 INFO [train.py:904] (4/8) Epoch 4, batch 8600, loss[loss=0.2301, simple_loss=0.3196, pruned_loss=0.07035, over 16224.00 frames. ], tot_loss[loss=0.229, simple_loss=0.3113, pruned_loss=0.07338, over 3046031.08 frames. ], batch size: 165, lr: 1.55e-02, grad_scale: 4.0 2023-04-28 04:54:45,520 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9255, 4.1943, 3.9513, 3.9898, 3.6457, 3.7199, 3.9393, 4.1138], device='cuda:4'), covar=tensor([0.0633, 0.0698, 0.0873, 0.0487, 0.0675, 0.1210, 0.0542, 0.0867], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0412, 0.0359, 0.0269, 0.0267, 0.0282, 0.0333, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:54:50,961 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=39064.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:55:04,687 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2231, 2.0442, 2.1114, 3.5552, 1.7442, 3.0073, 2.1575, 1.8708], device='cuda:4'), covar=tensor([0.0483, 0.1703, 0.0937, 0.0287, 0.2944, 0.0595, 0.1595, 0.2363], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0280, 0.0231, 0.0284, 0.0346, 0.0254, 0.0255, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:55:58,609 INFO [train.py:904] (4/8) Epoch 4, batch 8650, loss[loss=0.2073, simple_loss=0.2846, pruned_loss=0.06503, over 11897.00 frames. ], tot_loss[loss=0.2264, simple_loss=0.3092, pruned_loss=0.07179, over 3032167.01 frames. ], batch size: 248, lr: 1.55e-02, grad_scale: 4.0 2023-04-28 04:57:21,584 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-28 04:57:27,191 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9239, 1.7774, 1.6006, 1.4815, 1.8924, 1.7333, 1.8793, 1.9434], device='cuda:4'), covar=tensor([0.0025, 0.0108, 0.0138, 0.0148, 0.0084, 0.0105, 0.0062, 0.0098], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0136, 0.0138, 0.0136, 0.0130, 0.0139, 0.0098, 0.0116], device='cuda:4'), out_proj_covar=tensor([8.3059e-05, 1.7568e-04, 1.7259e-04, 1.6965e-04, 1.6855e-04, 1.7834e-04, 1.2198e-04, 1.4897e-04], device='cuda:4') 2023-04-28 04:57:44,932 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.040e+02 3.067e+02 3.859e+02 4.894e+02 7.251e+02, threshold=7.718e+02, percent-clipped=0.0 2023-04-28 04:57:44,948 INFO [train.py:904] (4/8) Epoch 4, batch 8700, loss[loss=0.1868, simple_loss=0.2712, pruned_loss=0.05115, over 12240.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.3054, pruned_loss=0.06988, over 3033350.90 frames. ], batch size: 248, lr: 1.55e-02, grad_scale: 4.0 2023-04-28 04:58:36,850 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39179.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:58:50,414 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39185.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:59:20,254 INFO [train.py:904] (4/8) Epoch 4, batch 8750, loss[loss=0.1987, simple_loss=0.2918, pruned_loss=0.05283, over 12197.00 frames. ], tot_loss[loss=0.2212, simple_loss=0.3045, pruned_loss=0.06888, over 3031081.46 frames. ], batch size: 248, lr: 1.55e-02, grad_scale: 2.0 2023-04-28 04:59:35,800 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5371, 4.5706, 4.7379, 4.6989, 4.6988, 5.1413, 4.8609, 4.4992], device='cuda:4'), covar=tensor([0.0748, 0.1562, 0.1039, 0.1488, 0.1963, 0.0790, 0.0894, 0.2058], device='cuda:4'), in_proj_covar=tensor([0.0238, 0.0332, 0.0321, 0.0290, 0.0380, 0.0352, 0.0276, 0.0387], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:00:28,637 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.67 vs. limit=2.0 2023-04-28 05:01:04,972 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39246.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:01:14,441 INFO [train.py:904] (4/8) Epoch 4, batch 8800, loss[loss=0.2112, simple_loss=0.3034, pruned_loss=0.05952, over 16377.00 frames. ], tot_loss[loss=0.2187, simple_loss=0.3026, pruned_loss=0.06737, over 3051185.96 frames. ], batch size: 146, lr: 1.55e-02, grad_scale: 4.0 2023-04-28 05:01:15,986 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.302e+02 3.583e+02 4.376e+02 5.208e+02 1.205e+03, threshold=8.753e+02, percent-clipped=6.0 2023-04-28 05:02:23,546 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39284.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:02:55,596 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5438, 4.2605, 4.5311, 4.7059, 4.7964, 4.2298, 4.8826, 4.8184], device='cuda:4'), covar=tensor([0.0686, 0.0690, 0.0925, 0.0418, 0.0418, 0.0643, 0.0253, 0.0315], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0396, 0.0496, 0.0401, 0.0305, 0.0294, 0.0323, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:02:58,288 INFO [train.py:904] (4/8) Epoch 4, batch 8850, loss[loss=0.2429, simple_loss=0.3254, pruned_loss=0.0802, over 15273.00 frames. ], tot_loss[loss=0.219, simple_loss=0.305, pruned_loss=0.06651, over 3043072.05 frames. ], batch size: 190, lr: 1.55e-02, grad_scale: 4.0 2023-04-28 05:04:32,351 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39345.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:04:44,178 INFO [train.py:904] (4/8) Epoch 4, batch 8900, loss[loss=0.2067, simple_loss=0.3024, pruned_loss=0.05546, over 16201.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.3058, pruned_loss=0.0656, over 3069229.08 frames. ], batch size: 35, lr: 1.55e-02, grad_scale: 2.0 2023-04-28 05:04:49,516 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.906e+02 3.441e+02 4.038e+02 4.851e+02 9.886e+02, threshold=8.076e+02, percent-clipped=2.0 2023-04-28 05:06:47,799 INFO [train.py:904] (4/8) Epoch 4, batch 8950, loss[loss=0.1841, simple_loss=0.2745, pruned_loss=0.04685, over 16829.00 frames. ], tot_loss[loss=0.2176, simple_loss=0.3047, pruned_loss=0.06524, over 3082616.29 frames. ], batch size: 76, lr: 1.55e-02, grad_scale: 2.0 2023-04-28 05:06:51,498 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8884, 4.1672, 3.8948, 4.0269, 3.6677, 3.6677, 3.8446, 4.0704], device='cuda:4'), covar=tensor([0.0759, 0.0813, 0.1163, 0.0430, 0.0764, 0.1242, 0.0699, 0.1030], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0425, 0.0367, 0.0276, 0.0275, 0.0289, 0.0344, 0.0309], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:07:58,398 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.61 vs. limit=5.0 2023-04-28 05:08:35,741 INFO [train.py:904] (4/8) Epoch 4, batch 9000, loss[loss=0.1997, simple_loss=0.2889, pruned_loss=0.05523, over 15460.00 frames. ], tot_loss[loss=0.2146, simple_loss=0.3016, pruned_loss=0.06381, over 3082608.89 frames. ], batch size: 194, lr: 1.54e-02, grad_scale: 2.0 2023-04-28 05:08:35,742 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 05:08:45,782 INFO [train.py:938] (4/8) Epoch 4, validation: loss=0.1802, simple_loss=0.283, pruned_loss=0.0387, over 944034.00 frames. 2023-04-28 05:08:45,783 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 05:08:49,857 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.055e+02 3.044e+02 3.757e+02 4.852e+02 9.817e+02, threshold=7.514e+02, percent-clipped=1.0 2023-04-28 05:09:11,764 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39463.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:09:26,658 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-04-28 05:09:43,537 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=39479.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:10:29,876 INFO [train.py:904] (4/8) Epoch 4, batch 9050, loss[loss=0.237, simple_loss=0.316, pruned_loss=0.07899, over 16680.00 frames. ], tot_loss[loss=0.2154, simple_loss=0.3021, pruned_loss=0.0643, over 3081206.33 frames. ], batch size: 134, lr: 1.54e-02, grad_scale: 2.0 2023-04-28 05:11:07,330 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 05:11:17,843 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39524.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:11:21,897 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=39527.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:11:29,405 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0694, 3.1474, 3.1228, 1.5684, 3.3144, 3.3842, 2.9255, 2.7184], device='cuda:4'), covar=tensor([0.0721, 0.0111, 0.0106, 0.1088, 0.0062, 0.0058, 0.0244, 0.0327], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0083, 0.0078, 0.0140, 0.0068, 0.0072, 0.0110, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:11:54,712 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39541.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:12:14,618 INFO [train.py:904] (4/8) Epoch 4, batch 9100, loss[loss=0.217, simple_loss=0.3063, pruned_loss=0.06387, over 16815.00 frames. ], tot_loss[loss=0.2154, simple_loss=0.3012, pruned_loss=0.0648, over 3076756.98 frames. ], batch size: 124, lr: 1.54e-02, grad_scale: 2.0 2023-04-28 05:12:18,751 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.564e+02 3.376e+02 4.042e+02 5.182e+02 1.418e+03, threshold=8.084e+02, percent-clipped=5.0 2023-04-28 05:14:15,379 INFO [train.py:904] (4/8) Epoch 4, batch 9150, loss[loss=0.1987, simple_loss=0.289, pruned_loss=0.05415, over 16198.00 frames. ], tot_loss[loss=0.2153, simple_loss=0.3014, pruned_loss=0.06464, over 3063539.61 frames. ], batch size: 166, lr: 1.54e-02, grad_scale: 2.0 2023-04-28 05:15:17,307 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9799, 2.7596, 2.7037, 1.9255, 2.4540, 2.5945, 2.6522, 1.7860], device='cuda:4'), covar=tensor([0.0218, 0.0022, 0.0032, 0.0171, 0.0040, 0.0037, 0.0032, 0.0244], device='cuda:4'), in_proj_covar=tensor([0.0109, 0.0049, 0.0055, 0.0107, 0.0055, 0.0061, 0.0058, 0.0102], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 05:15:41,220 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39639.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 05:15:43,181 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39640.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:16:00,927 INFO [train.py:904] (4/8) Epoch 4, batch 9200, loss[loss=0.1981, simple_loss=0.278, pruned_loss=0.05909, over 17094.00 frames. ], tot_loss[loss=0.2117, simple_loss=0.2965, pruned_loss=0.0634, over 3064821.99 frames. ], batch size: 53, lr: 1.54e-02, grad_scale: 4.0 2023-04-28 05:16:01,988 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-28 05:16:04,311 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.381e+02 3.414e+02 4.371e+02 6.281e+02 1.474e+03, threshold=8.741e+02, percent-clipped=12.0 2023-04-28 05:17:35,020 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39700.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 05:17:35,637 INFO [train.py:904] (4/8) Epoch 4, batch 9250, loss[loss=0.2094, simple_loss=0.2804, pruned_loss=0.0692, over 12295.00 frames. ], tot_loss[loss=0.2119, simple_loss=0.2963, pruned_loss=0.06378, over 3054728.09 frames. ], batch size: 248, lr: 1.54e-02, grad_scale: 4.0 2023-04-28 05:19:26,136 INFO [train.py:904] (4/8) Epoch 4, batch 9300, loss[loss=0.2222, simple_loss=0.3012, pruned_loss=0.07158, over 12577.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2946, pruned_loss=0.06298, over 3060582.02 frames. ], batch size: 247, lr: 1.54e-02, grad_scale: 4.0 2023-04-28 05:19:30,018 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.165e+02 3.034e+02 3.554e+02 4.321e+02 7.893e+02, threshold=7.107e+02, percent-clipped=0.0 2023-04-28 05:19:30,978 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4905, 4.8113, 4.5253, 4.5756, 4.1471, 4.2797, 4.2733, 4.8042], device='cuda:4'), covar=tensor([0.0511, 0.0579, 0.0842, 0.0388, 0.0545, 0.0796, 0.0595, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0410, 0.0342, 0.0260, 0.0260, 0.0275, 0.0326, 0.0296], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:20:06,176 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39767.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:21:11,784 INFO [train.py:904] (4/8) Epoch 4, batch 9350, loss[loss=0.2293, simple_loss=0.3073, pruned_loss=0.07565, over 16986.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2939, pruned_loss=0.06257, over 3055126.59 frames. ], batch size: 109, lr: 1.54e-02, grad_scale: 2.0 2023-04-28 05:21:41,203 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39815.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:21:48,581 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39819.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:22:02,602 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-28 05:22:05,662 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39828.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:22:30,255 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=39841.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:22:51,019 INFO [train.py:904] (4/8) Epoch 4, batch 9400, loss[loss=0.1812, simple_loss=0.2689, pruned_loss=0.04673, over 12362.00 frames. ], tot_loss[loss=0.2085, simple_loss=0.2936, pruned_loss=0.06165, over 3055942.31 frames. ], batch size: 248, lr: 1.54e-02, grad_scale: 2.0 2023-04-28 05:22:57,591 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.054e+02 3.057e+02 3.854e+02 4.791e+02 1.161e+03, threshold=7.709e+02, percent-clipped=3.0 2023-04-28 05:23:41,194 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39876.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:24:09,957 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=39889.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:24:32,546 INFO [train.py:904] (4/8) Epoch 4, batch 9450, loss[loss=0.2043, simple_loss=0.2911, pruned_loss=0.05877, over 16668.00 frames. ], tot_loss[loss=0.2112, simple_loss=0.2966, pruned_loss=0.06287, over 3054709.63 frames. ], batch size: 134, lr: 1.54e-02, grad_scale: 2.0 2023-04-28 05:24:38,180 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-28 05:25:54,123 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=39940.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:26:06,634 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7854, 2.7630, 2.2510, 3.7634, 3.4139, 3.7192, 1.5701, 2.6651], device='cuda:4'), covar=tensor([0.1431, 0.0552, 0.1232, 0.0086, 0.0242, 0.0336, 0.1429, 0.0885], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0137, 0.0164, 0.0075, 0.0143, 0.0161, 0.0158, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 05:26:13,854 INFO [train.py:904] (4/8) Epoch 4, batch 9500, loss[loss=0.218, simple_loss=0.3031, pruned_loss=0.06643, over 16965.00 frames. ], tot_loss[loss=0.2101, simple_loss=0.2954, pruned_loss=0.06236, over 3052539.55 frames. ], batch size: 109, lr: 1.54e-02, grad_scale: 2.0 2023-04-28 05:26:21,219 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.115e+02 3.608e+02 4.489e+02 5.705e+02 9.491e+02, threshold=8.979e+02, percent-clipped=6.0 2023-04-28 05:27:13,804 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-28 05:27:21,870 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0826, 2.6453, 2.3712, 3.3914, 3.1797, 3.3142, 1.7910, 2.8573], device='cuda:4'), covar=tensor([0.0950, 0.0352, 0.0840, 0.0077, 0.0193, 0.0384, 0.0996, 0.0518], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0136, 0.0166, 0.0075, 0.0145, 0.0163, 0.0158, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 05:27:31,424 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=39988.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:27:31,468 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7836, 5.0180, 5.0722, 5.1237, 5.0043, 5.5703, 5.3255, 5.0060], device='cuda:4'), covar=tensor([0.0664, 0.1440, 0.1477, 0.1463, 0.2308, 0.0862, 0.0991, 0.1994], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0342, 0.0333, 0.0300, 0.0389, 0.0358, 0.0280, 0.0395], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:27:48,820 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39995.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 05:28:04,403 INFO [train.py:904] (4/8) Epoch 4, batch 9550, loss[loss=0.2106, simple_loss=0.3047, pruned_loss=0.05821, over 16872.00 frames. ], tot_loss[loss=0.2104, simple_loss=0.2953, pruned_loss=0.06272, over 3044007.00 frames. ], batch size: 96, lr: 1.53e-02, grad_scale: 2.0 2023-04-28 05:28:50,741 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-28 05:29:46,541 INFO [train.py:904] (4/8) Epoch 4, batch 9600, loss[loss=0.2404, simple_loss=0.3303, pruned_loss=0.07522, over 15344.00 frames. ], tot_loss[loss=0.2125, simple_loss=0.2975, pruned_loss=0.06372, over 3054313.57 frames. ], batch size: 191, lr: 1.53e-02, grad_scale: 4.0 2023-04-28 05:29:52,055 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.278e+02 3.502e+02 4.489e+02 5.494e+02 1.109e+03, threshold=8.977e+02, percent-clipped=3.0 2023-04-28 05:31:15,868 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3132, 3.1667, 2.3113, 2.2067, 2.3748, 1.9421, 3.1661, 3.2380], device='cuda:4'), covar=tensor([0.2296, 0.0904, 0.1529, 0.1440, 0.1677, 0.1545, 0.0569, 0.0602], device='cuda:4'), in_proj_covar=tensor([0.0267, 0.0237, 0.0254, 0.0229, 0.0239, 0.0189, 0.0221, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:31:21,408 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8622, 2.0548, 2.2175, 3.2054, 1.9541, 2.6775, 2.2117, 1.9050], device='cuda:4'), covar=tensor([0.0492, 0.1569, 0.0711, 0.0303, 0.2338, 0.0768, 0.1495, 0.2013], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0283, 0.0232, 0.0285, 0.0343, 0.0259, 0.0257, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:31:33,068 INFO [train.py:904] (4/8) Epoch 4, batch 9650, loss[loss=0.2139, simple_loss=0.2931, pruned_loss=0.06732, over 17016.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2994, pruned_loss=0.0642, over 3057108.78 frames. ], batch size: 109, lr: 1.53e-02, grad_scale: 4.0 2023-04-28 05:32:17,670 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40119.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:32:25,203 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40123.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:32:33,488 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3048, 1.2513, 1.7129, 2.1147, 2.1519, 2.2615, 1.4355, 2.2730], device='cuda:4'), covar=tensor([0.0071, 0.0242, 0.0155, 0.0135, 0.0113, 0.0086, 0.0227, 0.0053], device='cuda:4'), in_proj_covar=tensor([0.0104, 0.0131, 0.0118, 0.0114, 0.0114, 0.0080, 0.0131, 0.0072], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 05:33:21,181 INFO [train.py:904] (4/8) Epoch 4, batch 9700, loss[loss=0.2028, simple_loss=0.2933, pruned_loss=0.05621, over 16169.00 frames. ], tot_loss[loss=0.2131, simple_loss=0.2984, pruned_loss=0.0639, over 3063917.06 frames. ], batch size: 165, lr: 1.53e-02, grad_scale: 4.0 2023-04-28 05:33:26,546 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.999e+02 3.154e+02 3.870e+02 5.322e+02 1.510e+03, threshold=7.740e+02, percent-clipped=2.0 2023-04-28 05:33:53,098 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=40167.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:34:01,368 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40171.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:34:54,781 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40195.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:35:03,567 INFO [train.py:904] (4/8) Epoch 4, batch 9750, loss[loss=0.2074, simple_loss=0.2962, pruned_loss=0.05935, over 16817.00 frames. ], tot_loss[loss=0.2117, simple_loss=0.2966, pruned_loss=0.06341, over 3067543.62 frames. ], batch size: 124, lr: 1.53e-02, grad_scale: 4.0 2023-04-28 05:35:38,847 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40218.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:36:45,116 INFO [train.py:904] (4/8) Epoch 4, batch 9800, loss[loss=0.2202, simple_loss=0.3131, pruned_loss=0.06358, over 16792.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2965, pruned_loss=0.06202, over 3077502.41 frames. ], batch size: 124, lr: 1.53e-02, grad_scale: 4.0 2023-04-28 05:36:51,067 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.012e+02 3.170e+02 3.844e+02 4.650e+02 7.827e+02, threshold=7.689e+02, percent-clipped=1.0 2023-04-28 05:36:56,651 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40256.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:37:24,833 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40272.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:37:37,847 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40279.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:37:58,369 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2663, 3.2485, 2.6366, 2.1409, 2.2174, 2.0171, 3.3142, 3.2077], device='cuda:4'), covar=tensor([0.2114, 0.0747, 0.1189, 0.1496, 0.1706, 0.1381, 0.0386, 0.0630], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0241, 0.0257, 0.0232, 0.0241, 0.0193, 0.0224, 0.0223], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:38:17,457 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40295.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 05:38:22,459 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4932, 3.8767, 3.2157, 2.3309, 2.7830, 2.3108, 4.0998, 3.8414], device='cuda:4'), covar=tensor([0.2136, 0.0588, 0.0963, 0.1445, 0.1754, 0.1259, 0.0285, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0239, 0.0255, 0.0231, 0.0240, 0.0192, 0.0222, 0.0221], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:38:25,165 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.87 vs. limit=5.0 2023-04-28 05:38:29,681 INFO [train.py:904] (4/8) Epoch 4, batch 9850, loss[loss=0.1999, simple_loss=0.2875, pruned_loss=0.05618, over 16689.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2976, pruned_loss=0.06147, over 3090334.80 frames. ], batch size: 134, lr: 1.53e-02, grad_scale: 4.0 2023-04-28 05:39:31,333 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40330.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:39:39,147 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40333.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:40:05,815 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=40343.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 05:40:21,694 INFO [train.py:904] (4/8) Epoch 4, batch 9900, loss[loss=0.2093, simple_loss=0.3047, pruned_loss=0.057, over 15322.00 frames. ], tot_loss[loss=0.2107, simple_loss=0.2982, pruned_loss=0.06155, over 3082339.95 frames. ], batch size: 191, lr: 1.53e-02, grad_scale: 4.0 2023-04-28 05:40:27,910 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.084e+02 3.210e+02 3.891e+02 5.069e+02 8.589e+02, threshold=7.781e+02, percent-clipped=1.0 2023-04-28 05:40:47,770 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9138, 1.7054, 1.5086, 1.3157, 1.8695, 1.5642, 1.7949, 1.8844], device='cuda:4'), covar=tensor([0.0022, 0.0144, 0.0185, 0.0201, 0.0094, 0.0150, 0.0065, 0.0081], device='cuda:4'), in_proj_covar=tensor([0.0063, 0.0140, 0.0140, 0.0140, 0.0132, 0.0141, 0.0098, 0.0115], device='cuda:4'), out_proj_covar=tensor([7.6982e-05, 1.7859e-04, 1.7365e-04, 1.7497e-04, 1.6902e-04, 1.7972e-04, 1.1848e-04, 1.4584e-04], device='cuda:4') 2023-04-28 05:41:56,813 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40391.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:42:18,075 INFO [train.py:904] (4/8) Epoch 4, batch 9950, loss[loss=0.2153, simple_loss=0.3042, pruned_loss=0.06325, over 16658.00 frames. ], tot_loss[loss=0.2124, simple_loss=0.3007, pruned_loss=0.06205, over 3088173.42 frames. ], batch size: 134, lr: 1.53e-02, grad_scale: 4.0 2023-04-28 05:43:13,943 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40423.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:44:20,284 INFO [train.py:904] (4/8) Epoch 4, batch 10000, loss[loss=0.2124, simple_loss=0.3037, pruned_loss=0.06051, over 15355.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2983, pruned_loss=0.06085, over 3113387.15 frames. ], batch size: 191, lr: 1.53e-02, grad_scale: 8.0 2023-04-28 05:44:26,641 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.009e+02 3.138e+02 3.849e+02 4.764e+02 1.083e+03, threshold=7.697e+02, percent-clipped=5.0 2023-04-28 05:44:30,318 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3029, 3.6719, 1.6579, 3.7887, 2.3414, 3.7385, 1.7567, 2.6250], device='cuda:4'), covar=tensor([0.0136, 0.0234, 0.1686, 0.0050, 0.0949, 0.0394, 0.1703, 0.0714], device='cuda:4'), in_proj_covar=tensor([0.0108, 0.0135, 0.0170, 0.0077, 0.0153, 0.0163, 0.0181, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 05:44:35,316 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1780, 4.2883, 4.2583, 4.2878, 4.3238, 4.7351, 4.4890, 4.1776], device='cuda:4'), covar=tensor([0.1115, 0.1350, 0.1058, 0.1732, 0.2331, 0.0921, 0.0832, 0.1949], device='cuda:4'), in_proj_covar=tensor([0.0233, 0.0334, 0.0327, 0.0294, 0.0388, 0.0355, 0.0269, 0.0388], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:45:01,303 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=40471.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:45:01,425 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40471.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:45:47,344 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40492.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:46:02,920 INFO [train.py:904] (4/8) Epoch 4, batch 10050, loss[loss=0.2238, simple_loss=0.3148, pruned_loss=0.06644, over 16173.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2987, pruned_loss=0.06059, over 3119570.60 frames. ], batch size: 165, lr: 1.53e-02, grad_scale: 8.0 2023-04-28 05:46:38,831 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=40519.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:47:38,803 INFO [train.py:904] (4/8) Epoch 4, batch 10100, loss[loss=0.1885, simple_loss=0.2838, pruned_loss=0.04666, over 17062.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.2994, pruned_loss=0.06164, over 3088042.81 frames. ], batch size: 53, lr: 1.52e-02, grad_scale: 8.0 2023-04-28 05:47:39,353 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40551.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:47:42,212 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40553.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:47:42,813 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.191e+02 3.527e+02 4.401e+02 5.596e+02 1.013e+03, threshold=8.802e+02, percent-clipped=4.0 2023-04-28 05:47:49,726 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6309, 1.3029, 1.9026, 2.4484, 2.3654, 2.4219, 1.5896, 2.5363], device='cuda:4'), covar=tensor([0.0056, 0.0245, 0.0146, 0.0104, 0.0097, 0.0086, 0.0218, 0.0055], device='cuda:4'), in_proj_covar=tensor([0.0100, 0.0130, 0.0116, 0.0110, 0.0111, 0.0079, 0.0129, 0.0069], device='cuda:4'), out_proj_covar=tensor([1.4368e-04, 1.8461e-04, 1.6909e-04, 1.5893e-04, 1.5929e-04, 1.0932e-04, 1.8311e-04, 9.7455e-05], device='cuda:4') 2023-04-28 05:48:27,648 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40574.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:48:30,655 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1705, 3.1054, 3.1388, 1.6079, 3.3481, 3.3832, 2.7212, 2.6946], device='cuda:4'), covar=tensor([0.0727, 0.0142, 0.0130, 0.1138, 0.0069, 0.0065, 0.0378, 0.0386], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0083, 0.0074, 0.0139, 0.0067, 0.0071, 0.0108, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:48:40,906 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4785, 3.5243, 3.2081, 3.2047, 3.0525, 3.3830, 3.2027, 3.2497], device='cuda:4'), covar=tensor([0.0429, 0.0314, 0.0231, 0.0196, 0.0654, 0.0319, 0.0980, 0.0367], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0144, 0.0183, 0.0152, 0.0199, 0.0172, 0.0130, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:49:25,172 INFO [train.py:904] (4/8) Epoch 5, batch 0, loss[loss=0.3867, simple_loss=0.3956, pruned_loss=0.1889, over 16471.00 frames. ], tot_loss[loss=0.3867, simple_loss=0.3956, pruned_loss=0.1889, over 16471.00 frames. ], batch size: 75, lr: 1.42e-02, grad_scale: 8.0 2023-04-28 05:49:25,173 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 05:49:32,549 INFO [train.py:938] (4/8) Epoch 5, validation: loss=0.1789, simple_loss=0.2817, pruned_loss=0.03802, over 944034.00 frames. 2023-04-28 05:49:32,550 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 05:49:49,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0227, 4.3425, 3.4510, 2.6056, 3.2254, 2.5486, 4.4964, 4.1396], device='cuda:4'), covar=tensor([0.1728, 0.0545, 0.0979, 0.1242, 0.1799, 0.1256, 0.0282, 0.0541], device='cuda:4'), in_proj_covar=tensor([0.0259, 0.0238, 0.0253, 0.0227, 0.0234, 0.0190, 0.0220, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:50:11,884 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40628.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:50:42,807 INFO [train.py:904] (4/8) Epoch 5, batch 50, loss[loss=0.2692, simple_loss=0.3316, pruned_loss=0.1034, over 16271.00 frames. ], tot_loss[loss=0.2567, simple_loss=0.321, pruned_loss=0.09626, over 754546.49 frames. ], batch size: 165, lr: 1.42e-02, grad_scale: 1.0 2023-04-28 05:50:43,197 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40651.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:50:49,836 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.398e+02 3.847e+02 4.852e+02 6.011e+02 1.299e+03, threshold=9.705e+02, percent-clipped=3.0 2023-04-28 05:50:52,586 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-28 05:51:29,203 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40686.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:51:50,915 INFO [train.py:904] (4/8) Epoch 5, batch 100, loss[loss=0.2068, simple_loss=0.2989, pruned_loss=0.05737, over 17065.00 frames. ], tot_loss[loss=0.2433, simple_loss=0.312, pruned_loss=0.08729, over 1332030.83 frames. ], batch size: 50, lr: 1.42e-02, grad_scale: 1.0 2023-04-28 05:52:07,458 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40712.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:52:26,426 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 05:52:59,300 INFO [train.py:904] (4/8) Epoch 5, batch 150, loss[loss=0.2532, simple_loss=0.3325, pruned_loss=0.08691, over 16721.00 frames. ], tot_loss[loss=0.2376, simple_loss=0.308, pruned_loss=0.08364, over 1784273.32 frames. ], batch size: 62, lr: 1.42e-02, grad_scale: 1.0 2023-04-28 05:53:08,164 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.263e+02 3.680e+02 4.214e+02 5.175e+02 1.144e+03, threshold=8.427e+02, percent-clipped=1.0 2023-04-28 05:53:50,300 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-28 05:53:53,725 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-28 05:54:09,222 INFO [train.py:904] (4/8) Epoch 5, batch 200, loss[loss=0.2432, simple_loss=0.3212, pruned_loss=0.08261, over 17193.00 frames. ], tot_loss[loss=0.2367, simple_loss=0.3078, pruned_loss=0.08282, over 2128122.25 frames. ], batch size: 46, lr: 1.42e-02, grad_scale: 1.0 2023-04-28 05:54:38,357 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4777, 4.1275, 4.3820, 4.6761, 4.7559, 4.2281, 4.5849, 4.7244], device='cuda:4'), covar=tensor([0.0781, 0.0750, 0.1189, 0.0454, 0.0382, 0.0766, 0.0751, 0.0402], device='cuda:4'), in_proj_covar=tensor([0.0371, 0.0458, 0.0581, 0.0455, 0.0340, 0.0331, 0.0365, 0.0380], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:54:52,196 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2403, 4.2993, 4.3187, 4.4057, 4.2843, 4.8685, 4.5454, 4.2281], device='cuda:4'), covar=tensor([0.1292, 0.1679, 0.1413, 0.1530, 0.2638, 0.1033, 0.1051, 0.2136], device='cuda:4'), in_proj_covar=tensor([0.0259, 0.0381, 0.0364, 0.0328, 0.0431, 0.0385, 0.0297, 0.0437], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 05:55:13,920 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40848.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:55:17,642 INFO [train.py:904] (4/8) Epoch 5, batch 250, loss[loss=0.2415, simple_loss=0.307, pruned_loss=0.088, over 16767.00 frames. ], tot_loss[loss=0.235, simple_loss=0.3063, pruned_loss=0.0818, over 2390586.38 frames. ], batch size: 134, lr: 1.41e-02, grad_scale: 1.0 2023-04-28 05:55:18,003 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40851.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:55:25,557 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.087e+02 3.707e+02 4.511e+02 5.174e+02 8.449e+02, threshold=9.022e+02, percent-clipped=1.0 2023-04-28 05:55:27,171 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3213, 5.1820, 5.0719, 4.8929, 4.5135, 5.0556, 5.0953, 4.7168], device='cuda:4'), covar=tensor([0.0398, 0.0219, 0.0175, 0.0160, 0.0934, 0.0258, 0.0188, 0.0466], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0167, 0.0207, 0.0175, 0.0235, 0.0198, 0.0150, 0.0224], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 05:55:38,036 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40865.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:55:50,366 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40874.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:56:24,969 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=40899.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:56:26,389 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40900.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:56:27,683 INFO [train.py:904] (4/8) Epoch 5, batch 300, loss[loss=0.2286, simple_loss=0.2927, pruned_loss=0.08227, over 16295.00 frames. ], tot_loss[loss=0.2308, simple_loss=0.3028, pruned_loss=0.07942, over 2600568.37 frames. ], batch size: 165, lr: 1.41e-02, grad_scale: 1.0 2023-04-28 05:56:37,534 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9470, 4.2625, 3.5636, 2.5758, 3.1305, 2.5584, 4.5626, 4.4135], device='cuda:4'), covar=tensor([0.2181, 0.0663, 0.1072, 0.1463, 0.2340, 0.1322, 0.0313, 0.0525], device='cuda:4'), in_proj_covar=tensor([0.0278, 0.0253, 0.0270, 0.0242, 0.0282, 0.0203, 0.0234, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:56:55,551 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-28 05:56:58,126 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=40922.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:57:05,047 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40926.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:57:07,874 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40928.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:57:30,694 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0723, 2.5864, 2.0462, 2.2459, 2.8064, 2.6873, 3.1913, 3.0717], device='cuda:4'), covar=tensor([0.0033, 0.0132, 0.0195, 0.0169, 0.0089, 0.0140, 0.0084, 0.0082], device='cuda:4'), in_proj_covar=tensor([0.0072, 0.0146, 0.0147, 0.0144, 0.0139, 0.0147, 0.0110, 0.0124], device='cuda:4'), out_proj_covar=tensor([8.7354e-05, 1.8540e-04, 1.8304e-04, 1.7800e-04, 1.7663e-04, 1.8749e-04, 1.3454e-04, 1.5751e-04], device='cuda:4') 2023-04-28 05:57:39,676 INFO [train.py:904] (4/8) Epoch 5, batch 350, loss[loss=0.2767, simple_loss=0.3234, pruned_loss=0.115, over 16895.00 frames. ], tot_loss[loss=0.2269, simple_loss=0.2996, pruned_loss=0.07715, over 2759657.88 frames. ], batch size: 109, lr: 1.41e-02, grad_scale: 1.0 2023-04-28 05:57:48,102 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.380e+02 3.200e+02 3.923e+02 5.275e+02 9.514e+02, threshold=7.845e+02, percent-clipped=1.0 2023-04-28 05:57:54,097 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40961.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 05:58:14,879 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=40976.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:58:28,024 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40986.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:58:44,173 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40997.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:58:48,861 INFO [train.py:904] (4/8) Epoch 5, batch 400, loss[loss=0.2121, simple_loss=0.2815, pruned_loss=0.07131, over 12116.00 frames. ], tot_loss[loss=0.2261, simple_loss=0.2983, pruned_loss=0.07695, over 2874210.88 frames. ], batch size: 246, lr: 1.41e-02, grad_scale: 2.0 2023-04-28 05:58:59,023 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41007.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:59:36,590 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=41034.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:59:51,072 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41045.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:00:00,530 INFO [train.py:904] (4/8) Epoch 5, batch 450, loss[loss=0.2163, simple_loss=0.3076, pruned_loss=0.06246, over 16760.00 frames. ], tot_loss[loss=0.225, simple_loss=0.2972, pruned_loss=0.07643, over 2977591.91 frames. ], batch size: 57, lr: 1.41e-02, grad_scale: 2.0 2023-04-28 06:00:01,518 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5909, 4.6166, 5.2596, 5.2549, 5.1824, 4.6987, 4.8280, 4.4682], device='cuda:4'), covar=tensor([0.0227, 0.0311, 0.0298, 0.0359, 0.0408, 0.0288, 0.0655, 0.0381], device='cuda:4'), in_proj_covar=tensor([0.0226, 0.0220, 0.0226, 0.0231, 0.0265, 0.0238, 0.0339, 0.0203], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 06:00:09,409 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.105e+02 3.241e+02 3.884e+02 4.938e+02 1.053e+03, threshold=7.768e+02, percent-clipped=3.0 2023-04-28 06:00:11,813 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41058.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:01:10,604 INFO [train.py:904] (4/8) Epoch 5, batch 500, loss[loss=0.225, simple_loss=0.2848, pruned_loss=0.08267, over 16388.00 frames. ], tot_loss[loss=0.2227, simple_loss=0.2953, pruned_loss=0.07505, over 3047238.27 frames. ], batch size: 146, lr: 1.41e-02, grad_scale: 2.0 2023-04-28 06:01:12,252 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41102.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:01:16,741 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41106.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:01:55,464 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6569, 3.6983, 2.7864, 2.2437, 2.5724, 2.2546, 3.6107, 3.7502], device='cuda:4'), covar=tensor([0.1955, 0.0622, 0.1219, 0.1579, 0.2094, 0.1374, 0.0436, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0252, 0.0270, 0.0242, 0.0292, 0.0204, 0.0238, 0.0245], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:02:14,398 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41148.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:02:17,713 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2265, 2.4407, 1.9430, 2.1887, 2.7269, 2.5537, 3.3060, 3.0009], device='cuda:4'), covar=tensor([0.0030, 0.0165, 0.0227, 0.0204, 0.0124, 0.0176, 0.0098, 0.0116], device='cuda:4'), in_proj_covar=tensor([0.0076, 0.0150, 0.0149, 0.0147, 0.0143, 0.0150, 0.0114, 0.0127], device='cuda:4'), out_proj_covar=tensor([9.2207e-05, 1.8919e-04, 1.8466e-04, 1.8164e-04, 1.8185e-04, 1.9076e-04, 1.3951e-04, 1.6065e-04], device='cuda:4') 2023-04-28 06:02:19,017 INFO [train.py:904] (4/8) Epoch 5, batch 550, loss[loss=0.2109, simple_loss=0.2907, pruned_loss=0.0655, over 16713.00 frames. ], tot_loss[loss=0.223, simple_loss=0.2948, pruned_loss=0.07559, over 3098423.21 frames. ], batch size: 57, lr: 1.41e-02, grad_scale: 2.0 2023-04-28 06:02:27,241 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.046e+02 3.287e+02 3.927e+02 4.639e+02 9.245e+02, threshold=7.855e+02, percent-clipped=2.0 2023-04-28 06:02:35,971 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41163.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:03:06,256 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-28 06:03:22,055 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=41196.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:03:27,689 INFO [train.py:904] (4/8) Epoch 5, batch 600, loss[loss=0.2263, simple_loss=0.2878, pruned_loss=0.08243, over 16434.00 frames. ], tot_loss[loss=0.2215, simple_loss=0.293, pruned_loss=0.075, over 3147966.04 frames. ], batch size: 75, lr: 1.41e-02, grad_scale: 2.0 2023-04-28 06:03:54,840 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41221.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:04:33,941 INFO [train.py:904] (4/8) Epoch 5, batch 650, loss[loss=0.2133, simple_loss=0.2806, pruned_loss=0.073, over 16094.00 frames. ], tot_loss[loss=0.219, simple_loss=0.2909, pruned_loss=0.07356, over 3183744.20 frames. ], batch size: 165, lr: 1.41e-02, grad_scale: 2.0 2023-04-28 06:04:40,910 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41256.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 06:04:42,445 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.330e+02 3.128e+02 3.726e+02 4.799e+02 9.270e+02, threshold=7.452e+02, percent-clipped=1.0 2023-04-28 06:05:39,909 INFO [train.py:904] (4/8) Epoch 5, batch 700, loss[loss=0.2633, simple_loss=0.3224, pruned_loss=0.1021, over 15604.00 frames. ], tot_loss[loss=0.2175, simple_loss=0.29, pruned_loss=0.07249, over 3211721.64 frames. ], batch size: 190, lr: 1.41e-02, grad_scale: 2.0 2023-04-28 06:05:46,906 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-28 06:05:47,985 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0318, 4.2379, 3.5343, 2.4799, 3.1850, 2.3946, 4.4999, 4.4586], device='cuda:4'), covar=tensor([0.1936, 0.0635, 0.1071, 0.1530, 0.2370, 0.1461, 0.0358, 0.0532], device='cuda:4'), in_proj_covar=tensor([0.0276, 0.0252, 0.0266, 0.0239, 0.0292, 0.0202, 0.0236, 0.0248], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:05:48,984 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41307.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:05:57,685 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6303, 2.6773, 2.0844, 2.1522, 3.1483, 2.9124, 3.6668, 3.3279], device='cuda:4'), covar=tensor([0.0024, 0.0176, 0.0255, 0.0252, 0.0113, 0.0177, 0.0093, 0.0109], device='cuda:4'), in_proj_covar=tensor([0.0077, 0.0149, 0.0150, 0.0146, 0.0144, 0.0149, 0.0117, 0.0128], device='cuda:4'), out_proj_covar=tensor([9.4006e-05, 1.8831e-04, 1.8489e-04, 1.7942e-04, 1.8336e-04, 1.8894e-04, 1.4347e-04, 1.6208e-04], device='cuda:4') 2023-04-28 06:06:49,378 INFO [train.py:904] (4/8) Epoch 5, batch 750, loss[loss=0.1888, simple_loss=0.2706, pruned_loss=0.05354, over 16816.00 frames. ], tot_loss[loss=0.2169, simple_loss=0.2904, pruned_loss=0.07167, over 3245189.18 frames. ], batch size: 42, lr: 1.41e-02, grad_scale: 2.0 2023-04-28 06:06:52,744 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41353.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:06:55,089 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=41355.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:06:57,955 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.848e+02 3.021e+02 3.672e+02 4.220e+02 6.723e+02, threshold=7.344e+02, percent-clipped=0.0 2023-04-28 06:07:10,958 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-28 06:07:29,588 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4606, 2.6314, 2.1650, 2.2906, 2.9765, 2.7795, 3.5131, 3.2278], device='cuda:4'), covar=tensor([0.0028, 0.0184, 0.0221, 0.0226, 0.0107, 0.0187, 0.0099, 0.0107], device='cuda:4'), in_proj_covar=tensor([0.0079, 0.0152, 0.0151, 0.0148, 0.0146, 0.0152, 0.0120, 0.0131], device='cuda:4'), out_proj_covar=tensor([9.6717e-05, 1.9169e-04, 1.8688e-04, 1.8233e-04, 1.8569e-04, 1.9289e-04, 1.4701e-04, 1.6552e-04], device='cuda:4') 2023-04-28 06:07:58,820 INFO [train.py:904] (4/8) Epoch 5, batch 800, loss[loss=0.2259, simple_loss=0.2867, pruned_loss=0.08255, over 16447.00 frames. ], tot_loss[loss=0.2146, simple_loss=0.2889, pruned_loss=0.0702, over 3268074.59 frames. ], batch size: 146, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:07:59,098 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41401.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:08:40,221 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2335, 3.2656, 3.9506, 2.4633, 3.6539, 3.8897, 3.8064, 2.0095], device='cuda:4'), covar=tensor([0.0307, 0.0146, 0.0024, 0.0220, 0.0039, 0.0046, 0.0032, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0113, 0.0059, 0.0059, 0.0112, 0.0059, 0.0066, 0.0061, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:09:08,622 INFO [train.py:904] (4/8) Epoch 5, batch 850, loss[loss=0.1983, simple_loss=0.2679, pruned_loss=0.06436, over 16681.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.2884, pruned_loss=0.07029, over 3278287.21 frames. ], batch size: 134, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:09:16,402 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.580e+02 3.257e+02 3.799e+02 4.682e+02 9.659e+02, threshold=7.597e+02, percent-clipped=4.0 2023-04-28 06:09:17,869 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41458.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:10:16,011 INFO [train.py:904] (4/8) Epoch 5, batch 900, loss[loss=0.2405, simple_loss=0.314, pruned_loss=0.08354, over 17012.00 frames. ], tot_loss[loss=0.2126, simple_loss=0.2872, pruned_loss=0.06898, over 3297394.06 frames. ], batch size: 53, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:10:44,986 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41521.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:11:27,561 INFO [train.py:904] (4/8) Epoch 5, batch 950, loss[loss=0.1917, simple_loss=0.279, pruned_loss=0.05225, over 17108.00 frames. ], tot_loss[loss=0.213, simple_loss=0.2871, pruned_loss=0.06948, over 3298878.99 frames. ], batch size: 47, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:11:34,463 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41556.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 06:11:35,282 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.063e+02 2.970e+02 3.801e+02 5.319e+02 1.620e+03, threshold=7.602e+02, percent-clipped=9.0 2023-04-28 06:11:51,950 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3681, 3.5063, 1.6404, 3.5541, 2.4604, 3.5123, 1.9681, 2.6788], device='cuda:4'), covar=tensor([0.0112, 0.0243, 0.1661, 0.0104, 0.0787, 0.0454, 0.1279, 0.0602], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0153, 0.0177, 0.0087, 0.0161, 0.0184, 0.0186, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 06:11:53,682 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=41569.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:12:10,935 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2710, 1.9045, 1.4938, 1.7707, 2.1887, 2.0518, 2.1368, 2.2977], device='cuda:4'), covar=tensor([0.0046, 0.0173, 0.0217, 0.0191, 0.0100, 0.0147, 0.0101, 0.0109], device='cuda:4'), in_proj_covar=tensor([0.0080, 0.0155, 0.0151, 0.0150, 0.0147, 0.0152, 0.0122, 0.0133], device='cuda:4'), out_proj_covar=tensor([9.7994e-05, 1.9555e-04, 1.8667e-04, 1.8463e-04, 1.8720e-04, 1.9274e-04, 1.4989e-04, 1.6779e-04], device='cuda:4') 2023-04-28 06:12:37,216 INFO [train.py:904] (4/8) Epoch 5, batch 1000, loss[loss=0.2019, simple_loss=0.2892, pruned_loss=0.05727, over 17123.00 frames. ], tot_loss[loss=0.213, simple_loss=0.2868, pruned_loss=0.0696, over 3310723.67 frames. ], batch size: 47, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:12:41,514 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=41604.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:13:02,246 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-28 06:13:35,581 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5669, 3.7682, 4.2649, 2.9936, 3.8057, 4.1940, 3.9988, 2.5009], device='cuda:4'), covar=tensor([0.0283, 0.0035, 0.0021, 0.0185, 0.0043, 0.0037, 0.0027, 0.0226], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0061, 0.0060, 0.0115, 0.0061, 0.0068, 0.0063, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:13:36,782 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7852, 3.4586, 2.8796, 1.8974, 2.5093, 2.0695, 3.1965, 3.3327], device='cuda:4'), covar=tensor([0.0274, 0.0503, 0.0617, 0.1543, 0.0826, 0.1017, 0.0615, 0.0667], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0132, 0.0155, 0.0143, 0.0135, 0.0126, 0.0141, 0.0137], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 06:13:45,696 INFO [train.py:904] (4/8) Epoch 5, batch 1050, loss[loss=0.2317, simple_loss=0.3102, pruned_loss=0.07664, over 17018.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.2874, pruned_loss=0.06959, over 3316934.39 frames. ], batch size: 50, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:13:49,024 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41653.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:13:50,304 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41654.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:13:54,766 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.676e+02 2.880e+02 3.783e+02 4.446e+02 1.011e+03, threshold=7.566e+02, percent-clipped=5.0 2023-04-28 06:14:19,860 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0640, 2.3057, 2.3412, 4.6306, 1.7911, 3.4654, 2.3926, 2.3896], device='cuda:4'), covar=tensor([0.0464, 0.1938, 0.1014, 0.0256, 0.3091, 0.0772, 0.1787, 0.2632], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0307, 0.0249, 0.0309, 0.0358, 0.0291, 0.0272, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:14:56,088 INFO [train.py:904] (4/8) Epoch 5, batch 1100, loss[loss=0.1945, simple_loss=0.2665, pruned_loss=0.06125, over 16863.00 frames. ], tot_loss[loss=0.2112, simple_loss=0.2858, pruned_loss=0.06823, over 3322641.31 frames. ], batch size: 116, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:14:56,430 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=41701.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:14:56,483 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41701.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:15:15,771 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41715.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 06:16:02,694 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=41749.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:16:05,297 INFO [train.py:904] (4/8) Epoch 5, batch 1150, loss[loss=0.2376, simple_loss=0.2967, pruned_loss=0.08925, over 16709.00 frames. ], tot_loss[loss=0.2101, simple_loss=0.2851, pruned_loss=0.06761, over 3321210.76 frames. ], batch size: 134, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:16:12,830 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.896e+02 2.850e+02 3.603e+02 4.605e+02 7.939e+02, threshold=7.207e+02, percent-clipped=2.0 2023-04-28 06:16:13,919 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3155, 3.9239, 4.0796, 1.8502, 4.2061, 4.1705, 3.4035, 3.1105], device='cuda:4'), covar=tensor([0.0906, 0.0097, 0.0150, 0.1292, 0.0055, 0.0080, 0.0302, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0087, 0.0085, 0.0143, 0.0071, 0.0080, 0.0115, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:16:14,946 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41758.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:17:14,375 INFO [train.py:904] (4/8) Epoch 5, batch 1200, loss[loss=0.2091, simple_loss=0.2728, pruned_loss=0.07266, over 16752.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.2845, pruned_loss=0.0671, over 3320847.49 frames. ], batch size: 102, lr: 1.40e-02, grad_scale: 8.0 2023-04-28 06:17:16,632 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0260, 3.8986, 3.9713, 1.4772, 4.2071, 4.2448, 3.1426, 3.0479], device='cuda:4'), covar=tensor([0.1313, 0.0148, 0.0266, 0.1652, 0.0091, 0.0084, 0.0405, 0.0641], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0089, 0.0087, 0.0145, 0.0072, 0.0081, 0.0116, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:17:21,141 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=41806.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:17:33,463 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.54 vs. limit=5.0 2023-04-28 06:17:42,648 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-28 06:18:11,228 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6813, 4.3320, 3.9804, 1.7947, 3.1018, 2.3451, 3.8498, 3.8674], device='cuda:4'), covar=tensor([0.0223, 0.0428, 0.0404, 0.1632, 0.0620, 0.0974, 0.0586, 0.0823], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0135, 0.0157, 0.0145, 0.0137, 0.0127, 0.0144, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 06:18:23,575 INFO [train.py:904] (4/8) Epoch 5, batch 1250, loss[loss=0.229, simple_loss=0.3013, pruned_loss=0.07836, over 16627.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2849, pruned_loss=0.06758, over 3327450.38 frames. ], batch size: 62, lr: 1.40e-02, grad_scale: 8.0 2023-04-28 06:18:31,520 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.030e+02 3.380e+02 4.131e+02 4.913e+02 1.055e+03, threshold=8.263e+02, percent-clipped=6.0 2023-04-28 06:19:30,147 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7320, 3.9804, 3.3531, 2.3944, 3.0770, 2.3656, 4.2795, 4.2352], device='cuda:4'), covar=tensor([0.2185, 0.0648, 0.1045, 0.1416, 0.2163, 0.1359, 0.0316, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0276, 0.0253, 0.0269, 0.0242, 0.0297, 0.0202, 0.0237, 0.0259], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:19:30,797 INFO [train.py:904] (4/8) Epoch 5, batch 1300, loss[loss=0.2031, simple_loss=0.2795, pruned_loss=0.0634, over 16686.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2851, pruned_loss=0.0679, over 3320300.35 frames. ], batch size: 62, lr: 1.40e-02, grad_scale: 8.0 2023-04-28 06:19:38,932 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41906.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:20:18,986 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 06:20:37,973 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41948.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 06:20:42,274 INFO [train.py:904] (4/8) Epoch 5, batch 1350, loss[loss=0.202, simple_loss=0.2824, pruned_loss=0.06077, over 16720.00 frames. ], tot_loss[loss=0.2104, simple_loss=0.2856, pruned_loss=0.06762, over 3325321.47 frames. ], batch size: 62, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:20:51,204 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.161e+02 3.407e+02 4.000e+02 4.882e+02 1.065e+03, threshold=8.000e+02, percent-clipped=1.0 2023-04-28 06:20:52,818 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3301, 3.1667, 3.4773, 2.5089, 3.2145, 3.4843, 3.3563, 1.9191], device='cuda:4'), covar=tensor([0.0274, 0.0063, 0.0029, 0.0194, 0.0040, 0.0043, 0.0035, 0.0275], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0060, 0.0060, 0.0114, 0.0061, 0.0068, 0.0062, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:21:05,613 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41967.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:21:55,629 INFO [train.py:904] (4/8) Epoch 5, batch 1400, loss[loss=0.1996, simple_loss=0.2766, pruned_loss=0.0613, over 16827.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2853, pruned_loss=0.06763, over 3322885.54 frames. ], batch size: 42, lr: 1.40e-02, grad_scale: 4.0 2023-04-28 06:22:08,292 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42009.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 06:22:09,944 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42010.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 06:22:15,247 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1474, 5.6663, 5.8014, 5.7082, 5.5895, 6.1672, 5.8293, 5.5453], device='cuda:4'), covar=tensor([0.0553, 0.1461, 0.1231, 0.1329, 0.2301, 0.0756, 0.0876, 0.1846], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0399, 0.0380, 0.0340, 0.0455, 0.0410, 0.0313, 0.0455], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:22:21,487 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8620, 3.7752, 3.0146, 2.4369, 2.7136, 2.2092, 3.5576, 3.7304], device='cuda:4'), covar=tensor([0.1744, 0.0458, 0.1021, 0.1479, 0.2293, 0.1450, 0.0411, 0.0696], device='cuda:4'), in_proj_covar=tensor([0.0277, 0.0255, 0.0271, 0.0243, 0.0300, 0.0203, 0.0240, 0.0261], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:23:05,349 INFO [train.py:904] (4/8) Epoch 5, batch 1450, loss[loss=0.1962, simple_loss=0.2801, pruned_loss=0.05618, over 17169.00 frames. ], tot_loss[loss=0.2096, simple_loss=0.2838, pruned_loss=0.06767, over 3315584.24 frames. ], batch size: 46, lr: 1.39e-02, grad_scale: 4.0 2023-04-28 06:23:15,557 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.035e+02 3.059e+02 3.831e+02 4.678e+02 9.639e+02, threshold=7.661e+02, percent-clipped=1.0 2023-04-28 06:24:05,682 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5503, 4.1340, 4.3249, 2.0647, 4.6310, 4.4657, 3.2493, 3.2832], device='cuda:4'), covar=tensor([0.0763, 0.0116, 0.0197, 0.1142, 0.0050, 0.0069, 0.0311, 0.0459], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0085, 0.0084, 0.0141, 0.0069, 0.0079, 0.0113, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:24:14,179 INFO [train.py:904] (4/8) Epoch 5, batch 1500, loss[loss=0.1995, simple_loss=0.2849, pruned_loss=0.0571, over 17101.00 frames. ], tot_loss[loss=0.2101, simple_loss=0.284, pruned_loss=0.06807, over 3311898.19 frames. ], batch size: 49, lr: 1.39e-02, grad_scale: 4.0 2023-04-28 06:25:21,155 INFO [train.py:904] (4/8) Epoch 5, batch 1550, loss[loss=0.2016, simple_loss=0.2868, pruned_loss=0.0582, over 17139.00 frames. ], tot_loss[loss=0.2127, simple_loss=0.2861, pruned_loss=0.06965, over 3307615.56 frames. ], batch size: 48, lr: 1.39e-02, grad_scale: 4.0 2023-04-28 06:25:32,954 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.189e+02 3.587e+02 4.011e+02 4.553e+02 8.694e+02, threshold=8.021e+02, percent-clipped=2.0 2023-04-28 06:25:50,991 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9036, 4.2849, 4.6034, 2.0974, 4.9083, 4.8217, 3.4473, 3.8658], device='cuda:4'), covar=tensor([0.0660, 0.0124, 0.0164, 0.1060, 0.0048, 0.0047, 0.0285, 0.0311], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0086, 0.0083, 0.0141, 0.0070, 0.0078, 0.0113, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:26:09,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9430, 3.0125, 3.3982, 2.2734, 3.0699, 3.3890, 3.2717, 1.8719], device='cuda:4'), covar=tensor([0.0306, 0.0075, 0.0034, 0.0215, 0.0044, 0.0036, 0.0036, 0.0269], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0060, 0.0060, 0.0113, 0.0060, 0.0066, 0.0062, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:26:15,639 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4656, 3.9126, 4.1052, 2.0888, 4.2403, 4.1665, 3.2796, 3.2015], device='cuda:4'), covar=tensor([0.0750, 0.0104, 0.0101, 0.1042, 0.0054, 0.0063, 0.0269, 0.0387], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0086, 0.0083, 0.0141, 0.0070, 0.0078, 0.0112, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:26:32,629 INFO [train.py:904] (4/8) Epoch 5, batch 1600, loss[loss=0.2102, simple_loss=0.2886, pruned_loss=0.06593, over 16087.00 frames. ], tot_loss[loss=0.2143, simple_loss=0.288, pruned_loss=0.0703, over 3312236.10 frames. ], batch size: 35, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:27:31,591 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2318, 2.4256, 2.4205, 4.7888, 1.9041, 3.7227, 2.3586, 2.5650], device='cuda:4'), covar=tensor([0.0441, 0.1830, 0.0990, 0.0221, 0.3039, 0.0709, 0.1809, 0.2456], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0311, 0.0253, 0.0313, 0.0362, 0.0299, 0.0277, 0.0384], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:27:39,756 INFO [train.py:904] (4/8) Epoch 5, batch 1650, loss[loss=0.2313, simple_loss=0.3136, pruned_loss=0.0745, over 17097.00 frames. ], tot_loss[loss=0.2165, simple_loss=0.2905, pruned_loss=0.07122, over 3315088.99 frames. ], batch size: 55, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:27:49,835 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.983e+02 3.262e+02 3.932e+02 4.837e+02 9.371e+02, threshold=7.863e+02, percent-clipped=1.0 2023-04-28 06:27:56,371 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42262.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:27:58,791 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9627, 1.5234, 2.1787, 2.7950, 2.8259, 2.7320, 1.5364, 3.0077], device='cuda:4'), covar=tensor([0.0064, 0.0210, 0.0169, 0.0088, 0.0079, 0.0100, 0.0235, 0.0040], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0138, 0.0127, 0.0125, 0.0123, 0.0089, 0.0138, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 06:28:49,287 INFO [train.py:904] (4/8) Epoch 5, batch 1700, loss[loss=0.2338, simple_loss=0.3023, pruned_loss=0.08264, over 16463.00 frames. ], tot_loss[loss=0.2176, simple_loss=0.2924, pruned_loss=0.07141, over 3321213.74 frames. ], batch size: 146, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:28:54,198 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42304.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 06:29:01,253 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=42310.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 06:29:53,186 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 06:29:56,498 INFO [train.py:904] (4/8) Epoch 5, batch 1750, loss[loss=0.192, simple_loss=0.2734, pruned_loss=0.05525, over 16859.00 frames. ], tot_loss[loss=0.2186, simple_loss=0.2937, pruned_loss=0.07179, over 3318048.49 frames. ], batch size: 42, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:30:05,725 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.019e+02 3.418e+02 4.322e+02 5.636e+02 1.741e+03, threshold=8.644e+02, percent-clipped=7.0 2023-04-28 06:30:05,950 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=42358.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:30:35,449 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42379.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:31:06,096 INFO [train.py:904] (4/8) Epoch 5, batch 1800, loss[loss=0.2004, simple_loss=0.298, pruned_loss=0.05144, over 17127.00 frames. ], tot_loss[loss=0.2184, simple_loss=0.2943, pruned_loss=0.07124, over 3324186.66 frames. ], batch size: 48, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:32:01,415 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42440.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:32:16,841 INFO [train.py:904] (4/8) Epoch 5, batch 1850, loss[loss=0.2245, simple_loss=0.309, pruned_loss=0.07, over 17071.00 frames. ], tot_loss[loss=0.2178, simple_loss=0.2943, pruned_loss=0.07065, over 3332722.79 frames. ], batch size: 53, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:32:26,219 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.852e+02 3.144e+02 3.803e+02 4.355e+02 7.438e+02, threshold=7.606e+02, percent-clipped=0.0 2023-04-28 06:32:50,126 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9220, 4.0551, 1.8244, 4.4062, 2.5769, 4.2946, 1.9893, 3.0123], device='cuda:4'), covar=tensor([0.0104, 0.0234, 0.1677, 0.0041, 0.0698, 0.0262, 0.1470, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0153, 0.0174, 0.0086, 0.0159, 0.0184, 0.0183, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 06:33:25,573 INFO [train.py:904] (4/8) Epoch 5, batch 1900, loss[loss=0.2148, simple_loss=0.2895, pruned_loss=0.07006, over 16420.00 frames. ], tot_loss[loss=0.2174, simple_loss=0.2938, pruned_loss=0.07054, over 3328854.82 frames. ], batch size: 75, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:33:32,591 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42506.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:34:02,220 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-28 06:34:35,496 INFO [train.py:904] (4/8) Epoch 5, batch 1950, loss[loss=0.201, simple_loss=0.2724, pruned_loss=0.06485, over 16278.00 frames. ], tot_loss[loss=0.2156, simple_loss=0.293, pruned_loss=0.06912, over 3330116.79 frames. ], batch size: 36, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:34:47,175 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.870e+02 3.199e+02 3.788e+02 4.472e+02 9.515e+02, threshold=7.576e+02, percent-clipped=2.0 2023-04-28 06:34:52,088 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=42562.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:34:59,418 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42567.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:35:48,039 INFO [train.py:904] (4/8) Epoch 5, batch 2000, loss[loss=0.2741, simple_loss=0.3214, pruned_loss=0.1134, over 16806.00 frames. ], tot_loss[loss=0.2161, simple_loss=0.2925, pruned_loss=0.06984, over 3330206.16 frames. ], batch size: 124, lr: 1.39e-02, grad_scale: 8.0 2023-04-28 06:35:51,774 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=42604.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 06:36:00,882 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=42610.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:36:55,039 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42649.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:36:55,155 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3973, 3.9919, 3.1910, 1.8637, 2.6512, 2.1388, 3.6342, 3.7833], device='cuda:4'), covar=tensor([0.0189, 0.0378, 0.0573, 0.1655, 0.0861, 0.1040, 0.0483, 0.0632], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0133, 0.0152, 0.0142, 0.0134, 0.0125, 0.0141, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 06:36:57,536 INFO [train.py:904] (4/8) Epoch 5, batch 2050, loss[loss=0.2513, simple_loss=0.3129, pruned_loss=0.0948, over 16437.00 frames. ], tot_loss[loss=0.2178, simple_loss=0.2932, pruned_loss=0.07121, over 3329126.62 frames. ], batch size: 146, lr: 1.38e-02, grad_scale: 8.0 2023-04-28 06:36:59,031 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=42652.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 06:37:06,886 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.791e+02 3.076e+02 3.576e+02 4.219e+02 7.856e+02, threshold=7.152e+02, percent-clipped=1.0 2023-04-28 06:37:41,987 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-28 06:37:47,352 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0793, 4.7507, 5.0163, 5.2759, 5.4539, 4.4886, 5.4116, 5.3457], device='cuda:4'), covar=tensor([0.0690, 0.0786, 0.1186, 0.0454, 0.0311, 0.0641, 0.0316, 0.0379], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0494, 0.0635, 0.0516, 0.0384, 0.0376, 0.0401, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:37:51,049 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0988, 4.7642, 5.0524, 5.2973, 5.4871, 4.4817, 5.4311, 5.3581], device='cuda:4'), covar=tensor([0.0667, 0.0812, 0.1222, 0.0447, 0.0314, 0.0677, 0.0309, 0.0380], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0494, 0.0636, 0.0517, 0.0384, 0.0377, 0.0401, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:38:05,313 INFO [train.py:904] (4/8) Epoch 5, batch 2100, loss[loss=0.2606, simple_loss=0.3301, pruned_loss=0.09551, over 16777.00 frames. ], tot_loss[loss=0.2197, simple_loss=0.2947, pruned_loss=0.07234, over 3325252.95 frames. ], batch size: 83, lr: 1.38e-02, grad_scale: 8.0 2023-04-28 06:38:18,847 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42710.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:38:52,937 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42735.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:39:15,225 INFO [train.py:904] (4/8) Epoch 5, batch 2150, loss[loss=0.1952, simple_loss=0.2765, pruned_loss=0.05694, over 17262.00 frames. ], tot_loss[loss=0.2199, simple_loss=0.295, pruned_loss=0.07241, over 3312349.79 frames. ], batch size: 45, lr: 1.38e-02, grad_scale: 8.0 2023-04-28 06:39:24,099 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.183e+02 3.251e+02 3.882e+02 4.566e+02 8.930e+02, threshold=7.764e+02, percent-clipped=4.0 2023-04-28 06:40:23,566 INFO [train.py:904] (4/8) Epoch 5, batch 2200, loss[loss=0.253, simple_loss=0.3248, pruned_loss=0.09058, over 16556.00 frames. ], tot_loss[loss=0.2196, simple_loss=0.2951, pruned_loss=0.07206, over 3320150.89 frames. ], batch size: 68, lr: 1.38e-02, grad_scale: 8.0 2023-04-28 06:41:18,093 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1205, 4.9676, 4.8848, 4.2798, 4.8017, 2.1191, 4.7143, 4.9816], device='cuda:4'), covar=tensor([0.0059, 0.0059, 0.0078, 0.0336, 0.0075, 0.1473, 0.0079, 0.0108], device='cuda:4'), in_proj_covar=tensor([0.0097, 0.0082, 0.0127, 0.0133, 0.0097, 0.0138, 0.0111, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:41:34,694 INFO [train.py:904] (4/8) Epoch 5, batch 2250, loss[loss=0.2042, simple_loss=0.2798, pruned_loss=0.0643, over 16846.00 frames. ], tot_loss[loss=0.2206, simple_loss=0.296, pruned_loss=0.0726, over 3321721.10 frames. ], batch size: 39, lr: 1.38e-02, grad_scale: 8.0 2023-04-28 06:41:43,705 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.043e+02 3.207e+02 3.862e+02 4.939e+02 1.226e+03, threshold=7.724e+02, percent-clipped=4.0 2023-04-28 06:41:46,014 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42859.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:41:49,460 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42862.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:42:14,359 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42879.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:42:44,587 INFO [train.py:904] (4/8) Epoch 5, batch 2300, loss[loss=0.2157, simple_loss=0.2817, pruned_loss=0.0749, over 16876.00 frames. ], tot_loss[loss=0.22, simple_loss=0.2958, pruned_loss=0.07208, over 3327057.72 frames. ], batch size: 102, lr: 1.38e-02, grad_scale: 4.0 2023-04-28 06:43:11,573 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42920.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:43:38,249 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42940.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:43:42,363 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0836, 1.8903, 1.5515, 1.7165, 2.2112, 2.0812, 2.1725, 2.3050], device='cuda:4'), covar=tensor([0.0046, 0.0153, 0.0195, 0.0192, 0.0086, 0.0152, 0.0099, 0.0103], device='cuda:4'), in_proj_covar=tensor([0.0084, 0.0151, 0.0150, 0.0147, 0.0147, 0.0155, 0.0126, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:43:53,244 INFO [train.py:904] (4/8) Epoch 5, batch 2350, loss[loss=0.2331, simple_loss=0.3164, pruned_loss=0.07488, over 17121.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.2971, pruned_loss=0.07335, over 3310570.50 frames. ], batch size: 49, lr: 1.38e-02, grad_scale: 4.0 2023-04-28 06:44:03,299 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.280e+02 3.235e+02 3.889e+02 4.889e+02 8.777e+02, threshold=7.778e+02, percent-clipped=2.0 2023-04-28 06:44:24,564 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2249, 4.2582, 4.1783, 3.7105, 4.1850, 1.7620, 3.9995, 4.0137], device='cuda:4'), covar=tensor([0.0071, 0.0053, 0.0082, 0.0238, 0.0059, 0.1456, 0.0080, 0.0122], device='cuda:4'), in_proj_covar=tensor([0.0096, 0.0083, 0.0127, 0.0134, 0.0097, 0.0138, 0.0111, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:45:02,225 INFO [train.py:904] (4/8) Epoch 5, batch 2400, loss[loss=0.2935, simple_loss=0.3595, pruned_loss=0.1137, over 11905.00 frames. ], tot_loss[loss=0.2228, simple_loss=0.2983, pruned_loss=0.07362, over 3307503.07 frames. ], batch size: 246, lr: 1.38e-02, grad_scale: 8.0 2023-04-28 06:45:08,765 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=43005.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:45:27,059 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8816, 3.3606, 2.7185, 4.7898, 4.2192, 4.0318, 1.6576, 3.0994], device='cuda:4'), covar=tensor([0.1344, 0.0461, 0.1028, 0.0066, 0.0313, 0.0357, 0.1376, 0.0688], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0140, 0.0163, 0.0084, 0.0177, 0.0169, 0.0157, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 06:45:52,152 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43035.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:46:13,639 INFO [train.py:904] (4/8) Epoch 5, batch 2450, loss[loss=0.2059, simple_loss=0.2778, pruned_loss=0.06702, over 16802.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.2987, pruned_loss=0.07315, over 3303758.84 frames. ], batch size: 102, lr: 1.38e-02, grad_scale: 4.0 2023-04-28 06:46:19,357 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-28 06:46:26,010 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.576e+02 3.112e+02 3.702e+02 4.495e+02 8.852e+02, threshold=7.404e+02, percent-clipped=2.0 2023-04-28 06:46:57,983 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=43083.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:47:23,966 INFO [train.py:904] (4/8) Epoch 5, batch 2500, loss[loss=0.2046, simple_loss=0.2926, pruned_loss=0.05827, over 16433.00 frames. ], tot_loss[loss=0.2206, simple_loss=0.2974, pruned_loss=0.07192, over 3318436.85 frames. ], batch size: 36, lr: 1.38e-02, grad_scale: 4.0 2023-04-28 06:47:43,724 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0131, 4.1622, 4.3260, 1.8287, 4.5818, 4.6693, 3.3851, 3.6088], device='cuda:4'), covar=tensor([0.0560, 0.0116, 0.0150, 0.1152, 0.0047, 0.0051, 0.0264, 0.0321], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0086, 0.0086, 0.0142, 0.0072, 0.0079, 0.0116, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:48:16,920 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4145, 2.3865, 2.0004, 2.1709, 2.7260, 2.5714, 3.5773, 3.0992], device='cuda:4'), covar=tensor([0.0029, 0.0175, 0.0191, 0.0192, 0.0111, 0.0168, 0.0064, 0.0095], device='cuda:4'), in_proj_covar=tensor([0.0084, 0.0151, 0.0151, 0.0147, 0.0148, 0.0155, 0.0128, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:48:33,485 INFO [train.py:904] (4/8) Epoch 5, batch 2550, loss[loss=0.2257, simple_loss=0.2963, pruned_loss=0.07753, over 16883.00 frames. ], tot_loss[loss=0.222, simple_loss=0.2987, pruned_loss=0.07261, over 3309525.58 frames. ], batch size: 96, lr: 1.38e-02, grad_scale: 4.0 2023-04-28 06:48:42,413 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0361, 2.3834, 2.3931, 4.7182, 1.9247, 3.5498, 2.3196, 2.4922], device='cuda:4'), covar=tensor([0.0483, 0.2044, 0.1029, 0.0226, 0.3065, 0.0830, 0.1967, 0.2513], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0313, 0.0252, 0.0308, 0.0358, 0.0301, 0.0279, 0.0385], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:48:45,560 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.843e+02 3.229e+02 3.908e+02 4.674e+02 8.443e+02, threshold=7.816e+02, percent-clipped=2.0 2023-04-28 06:48:48,888 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43162.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:49:22,564 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3469, 5.3290, 5.1223, 4.9554, 4.4535, 5.2275, 5.1986, 4.7062], device='cuda:4'), covar=tensor([0.0422, 0.0212, 0.0194, 0.0181, 0.0989, 0.0223, 0.0166, 0.0524], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0197, 0.0231, 0.0201, 0.0267, 0.0223, 0.0170, 0.0257], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:49:43,647 INFO [train.py:904] (4/8) Epoch 5, batch 2600, loss[loss=0.2014, simple_loss=0.2761, pruned_loss=0.06338, over 16790.00 frames. ], tot_loss[loss=0.2197, simple_loss=0.2968, pruned_loss=0.07131, over 3321707.66 frames. ], batch size: 102, lr: 1.38e-02, grad_scale: 4.0 2023-04-28 06:49:55,551 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=43210.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:50:01,749 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=43215.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:50:04,870 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-28 06:50:31,301 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=43235.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:50:38,832 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7789, 3.5909, 2.7417, 5.1469, 4.6479, 4.4558, 1.8552, 3.3301], device='cuda:4'), covar=tensor([0.1344, 0.0483, 0.1047, 0.0062, 0.0264, 0.0295, 0.1230, 0.0658], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0140, 0.0162, 0.0084, 0.0179, 0.0169, 0.0156, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 06:50:53,870 INFO [train.py:904] (4/8) Epoch 5, batch 2650, loss[loss=0.2372, simple_loss=0.3065, pruned_loss=0.08392, over 16671.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.297, pruned_loss=0.07001, over 3329294.62 frames. ], batch size: 134, lr: 1.37e-02, grad_scale: 4.0 2023-04-28 06:51:05,579 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.996e+02 3.400e+02 3.860e+02 4.730e+02 9.799e+02, threshold=7.720e+02, percent-clipped=2.0 2023-04-28 06:51:25,047 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 06:51:39,962 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.76 vs. limit=5.0 2023-04-28 06:51:55,587 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-28 06:52:02,500 INFO [train.py:904] (4/8) Epoch 5, batch 2700, loss[loss=0.2838, simple_loss=0.3403, pruned_loss=0.1137, over 16875.00 frames. ], tot_loss[loss=0.2187, simple_loss=0.2975, pruned_loss=0.06997, over 3323560.88 frames. ], batch size: 116, lr: 1.37e-02, grad_scale: 4.0 2023-04-28 06:52:09,032 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43305.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:53:12,571 INFO [train.py:904] (4/8) Epoch 5, batch 2750, loss[loss=0.2134, simple_loss=0.2989, pruned_loss=0.064, over 17028.00 frames. ], tot_loss[loss=0.2184, simple_loss=0.2973, pruned_loss=0.06979, over 3320533.78 frames. ], batch size: 55, lr: 1.37e-02, grad_scale: 4.0 2023-04-28 06:53:15,876 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=43353.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:53:25,426 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.808e+02 2.783e+02 3.325e+02 4.453e+02 8.515e+02, threshold=6.649e+02, percent-clipped=2.0 2023-04-28 06:54:22,964 INFO [train.py:904] (4/8) Epoch 5, batch 2800, loss[loss=0.2078, simple_loss=0.3036, pruned_loss=0.05605, over 17121.00 frames. ], tot_loss[loss=0.219, simple_loss=0.2977, pruned_loss=0.0702, over 3321367.72 frames. ], batch size: 48, lr: 1.37e-02, grad_scale: 8.0 2023-04-28 06:54:33,470 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3186, 5.1213, 5.0954, 4.4257, 5.0733, 2.0938, 4.8805, 5.2294], device='cuda:4'), covar=tensor([0.0063, 0.0058, 0.0080, 0.0338, 0.0062, 0.1413, 0.0081, 0.0121], device='cuda:4'), in_proj_covar=tensor([0.0096, 0.0084, 0.0127, 0.0135, 0.0097, 0.0139, 0.0113, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:55:27,773 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4562, 4.2451, 4.3866, 4.7216, 4.7524, 4.2494, 4.6031, 4.7454], device='cuda:4'), covar=tensor([0.0865, 0.0723, 0.1263, 0.0427, 0.0444, 0.0824, 0.0825, 0.0394], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0501, 0.0652, 0.0517, 0.0384, 0.0386, 0.0401, 0.0426], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:55:30,666 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-28 06:55:33,337 INFO [train.py:904] (4/8) Epoch 5, batch 2850, loss[loss=0.2264, simple_loss=0.2896, pruned_loss=0.0816, over 16457.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.2962, pruned_loss=0.06985, over 3324129.48 frames. ], batch size: 146, lr: 1.37e-02, grad_scale: 8.0 2023-04-28 06:55:42,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2906, 4.2642, 3.3605, 5.4584, 5.0486, 4.7272, 2.3342, 3.8597], device='cuda:4'), covar=tensor([0.1068, 0.0339, 0.0754, 0.0068, 0.0210, 0.0278, 0.1014, 0.0469], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0141, 0.0164, 0.0086, 0.0181, 0.0170, 0.0157, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 06:55:45,525 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.398e+02 3.090e+02 3.971e+02 4.824e+02 1.597e+03, threshold=7.942e+02, percent-clipped=16.0 2023-04-28 06:56:41,596 INFO [train.py:904] (4/8) Epoch 5, batch 2900, loss[loss=0.2134, simple_loss=0.2736, pruned_loss=0.07662, over 16716.00 frames. ], tot_loss[loss=0.2178, simple_loss=0.2956, pruned_loss=0.07002, over 3320549.11 frames. ], batch size: 83, lr: 1.37e-02, grad_scale: 8.0 2023-04-28 06:57:00,339 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=43514.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:57:01,379 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43515.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:57:28,488 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43535.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:57:49,458 INFO [train.py:904] (4/8) Epoch 5, batch 2950, loss[loss=0.2012, simple_loss=0.2737, pruned_loss=0.06433, over 17197.00 frames. ], tot_loss[loss=0.2206, simple_loss=0.2963, pruned_loss=0.07247, over 3313245.66 frames. ], batch size: 46, lr: 1.37e-02, grad_scale: 8.0 2023-04-28 06:58:02,027 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.717e+02 3.523e+02 4.392e+02 5.474e+02 1.022e+03, threshold=8.784e+02, percent-clipped=2.0 2023-04-28 06:58:05,994 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=43563.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:58:06,195 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9099, 4.5005, 4.6179, 3.1413, 4.1815, 4.6153, 4.2472, 3.0840], device='cuda:4'), covar=tensor([0.0246, 0.0018, 0.0018, 0.0187, 0.0028, 0.0027, 0.0022, 0.0201], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0060, 0.0059, 0.0111, 0.0061, 0.0068, 0.0065, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:58:24,097 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=43575.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:58:34,162 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=43583.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:58:59,871 INFO [train.py:904] (4/8) Epoch 5, batch 3000, loss[loss=0.2219, simple_loss=0.2935, pruned_loss=0.0752, over 16728.00 frames. ], tot_loss[loss=0.2208, simple_loss=0.2963, pruned_loss=0.07268, over 3314389.04 frames. ], batch size: 124, lr: 1.37e-02, grad_scale: 8.0 2023-04-28 06:58:59,871 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 06:59:06,037 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8866, 4.7679, 5.0069, 3.3385, 5.1937, 5.3231, 4.3206, 4.6856], device='cuda:4'), covar=tensor([0.0418, 0.0064, 0.0106, 0.0625, 0.0033, 0.0032, 0.0150, 0.0164], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0083, 0.0084, 0.0137, 0.0071, 0.0078, 0.0113, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:59:08,833 INFO [train.py:938] (4/8) Epoch 5, validation: loss=0.1564, simple_loss=0.2632, pruned_loss=0.02477, over 944034.00 frames. 2023-04-28 06:59:08,833 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 06:59:13,928 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-28 06:59:15,143 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 06:59:36,396 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-28 07:00:01,292 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.19 vs. limit=5.0 2023-04-28 07:00:14,822 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9458, 2.7717, 2.6555, 1.7222, 2.8654, 2.8810, 2.4366, 2.3936], device='cuda:4'), covar=tensor([0.0726, 0.0138, 0.0189, 0.0889, 0.0072, 0.0088, 0.0376, 0.0410], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0084, 0.0085, 0.0140, 0.0071, 0.0079, 0.0115, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 07:00:18,524 INFO [train.py:904] (4/8) Epoch 5, batch 3050, loss[loss=0.2275, simple_loss=0.3096, pruned_loss=0.07276, over 17047.00 frames. ], tot_loss[loss=0.2202, simple_loss=0.2958, pruned_loss=0.07224, over 3317145.05 frames. ], batch size: 53, lr: 1.37e-02, grad_scale: 4.0 2023-04-28 07:00:31,425 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.472e+02 3.368e+02 3.840e+02 5.233e+02 1.219e+03, threshold=7.679e+02, percent-clipped=3.0 2023-04-28 07:01:18,787 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3461, 1.8332, 2.6354, 3.1832, 2.9676, 3.4492, 2.1214, 3.3438], device='cuda:4'), covar=tensor([0.0060, 0.0215, 0.0125, 0.0109, 0.0095, 0.0082, 0.0187, 0.0057], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0142, 0.0127, 0.0131, 0.0127, 0.0093, 0.0138, 0.0081], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 07:01:25,947 INFO [train.py:904] (4/8) Epoch 5, batch 3100, loss[loss=0.2348, simple_loss=0.3045, pruned_loss=0.08258, over 15471.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.2951, pruned_loss=0.07254, over 3318745.24 frames. ], batch size: 191, lr: 1.37e-02, grad_scale: 4.0 2023-04-28 07:01:27,472 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7992, 4.7695, 5.2662, 5.2775, 5.2777, 4.8495, 4.8481, 4.5797], device='cuda:4'), covar=tensor([0.0216, 0.0326, 0.0309, 0.0346, 0.0346, 0.0244, 0.0679, 0.0344], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0242, 0.0246, 0.0248, 0.0297, 0.0259, 0.0375, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 07:01:27,600 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3383, 2.4851, 1.9393, 2.1732, 2.8644, 2.6670, 3.5572, 3.2425], device='cuda:4'), covar=tensor([0.0035, 0.0165, 0.0220, 0.0197, 0.0119, 0.0162, 0.0087, 0.0091], device='cuda:4'), in_proj_covar=tensor([0.0084, 0.0150, 0.0151, 0.0147, 0.0148, 0.0155, 0.0130, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 07:02:01,738 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7499, 2.6438, 1.8969, 2.1624, 3.0277, 2.7196, 3.6423, 3.3095], device='cuda:4'), covar=tensor([0.0027, 0.0170, 0.0248, 0.0220, 0.0107, 0.0178, 0.0095, 0.0107], device='cuda:4'), in_proj_covar=tensor([0.0084, 0.0150, 0.0150, 0.0147, 0.0148, 0.0155, 0.0130, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 07:02:33,549 INFO [train.py:904] (4/8) Epoch 5, batch 3150, loss[loss=0.2138, simple_loss=0.2856, pruned_loss=0.07099, over 16685.00 frames. ], tot_loss[loss=0.2193, simple_loss=0.2941, pruned_loss=0.07223, over 3321532.01 frames. ], batch size: 134, lr: 1.37e-02, grad_scale: 4.0 2023-04-28 07:02:46,880 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.292e+02 2.988e+02 3.776e+02 4.574e+02 1.068e+03, threshold=7.553e+02, percent-clipped=4.0 2023-04-28 07:02:56,560 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-28 07:03:22,289 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=43787.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:03:42,785 INFO [train.py:904] (4/8) Epoch 5, batch 3200, loss[loss=0.2197, simple_loss=0.3068, pruned_loss=0.06631, over 17246.00 frames. ], tot_loss[loss=0.2191, simple_loss=0.2939, pruned_loss=0.07214, over 3324441.47 frames. ], batch size: 52, lr: 1.37e-02, grad_scale: 8.0 2023-04-28 07:04:23,070 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8524, 4.5623, 4.7977, 5.0759, 5.1999, 4.5142, 5.1775, 5.1838], device='cuda:4'), covar=tensor([0.0850, 0.0779, 0.1232, 0.0446, 0.0382, 0.0616, 0.0430, 0.0378], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0499, 0.0646, 0.0519, 0.0383, 0.0384, 0.0397, 0.0427], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 07:04:48,214 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=43848.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:04:52,114 INFO [train.py:904] (4/8) Epoch 5, batch 3250, loss[loss=0.2211, simple_loss=0.3106, pruned_loss=0.06574, over 17123.00 frames. ], tot_loss[loss=0.2195, simple_loss=0.2943, pruned_loss=0.07241, over 3320352.63 frames. ], batch size: 48, lr: 1.37e-02, grad_scale: 8.0 2023-04-28 07:05:06,682 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.826e+02 2.952e+02 3.819e+02 4.745e+02 7.662e+02, threshold=7.638e+02, percent-clipped=2.0 2023-04-28 07:05:19,447 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=43870.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:05:28,410 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2996, 5.7092, 5.4534, 5.5825, 5.0664, 5.0150, 5.2037, 5.8427], device='cuda:4'), covar=tensor([0.0725, 0.0767, 0.0937, 0.0441, 0.0622, 0.0589, 0.0654, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0367, 0.0506, 0.0426, 0.0326, 0.0315, 0.0325, 0.0403, 0.0354], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 07:06:01,738 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-28 07:06:03,319 INFO [train.py:904] (4/8) Epoch 5, batch 3300, loss[loss=0.1961, simple_loss=0.2852, pruned_loss=0.0535, over 17122.00 frames. ], tot_loss[loss=0.22, simple_loss=0.2954, pruned_loss=0.07234, over 3321923.62 frames. ], batch size: 49, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:06:14,741 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=43909.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:06:50,181 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8527, 4.9208, 5.4413, 5.3989, 5.3800, 4.9705, 4.9571, 4.7178], device='cuda:4'), covar=tensor([0.0248, 0.0284, 0.0273, 0.0351, 0.0343, 0.0257, 0.0694, 0.0332], device='cuda:4'), in_proj_covar=tensor([0.0255, 0.0248, 0.0252, 0.0255, 0.0303, 0.0266, 0.0382, 0.0224], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 07:07:12,748 INFO [train.py:904] (4/8) Epoch 5, batch 3350, loss[loss=0.1846, simple_loss=0.2752, pruned_loss=0.04697, over 17250.00 frames. ], tot_loss[loss=0.2189, simple_loss=0.295, pruned_loss=0.07145, over 3325714.51 frames. ], batch size: 52, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:07:24,285 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=43959.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:07:26,939 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.942e+02 2.864e+02 3.521e+02 4.539e+02 9.461e+02, threshold=7.042e+02, percent-clipped=3.0 2023-04-28 07:07:38,033 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2793, 4.4426, 4.5551, 2.0225, 4.8356, 4.8388, 3.5418, 4.0535], device='cuda:4'), covar=tensor([0.0491, 0.0084, 0.0120, 0.0977, 0.0034, 0.0047, 0.0262, 0.0218], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0087, 0.0087, 0.0142, 0.0074, 0.0081, 0.0117, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 07:07:39,258 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=43970.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:08:22,241 INFO [train.py:904] (4/8) Epoch 5, batch 3400, loss[loss=0.251, simple_loss=0.3144, pruned_loss=0.09379, over 16861.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.2941, pruned_loss=0.07082, over 3317064.63 frames. ], batch size: 96, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:08:26,225 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9370, 5.0099, 5.5269, 5.5016, 5.5489, 5.1528, 4.7872, 4.9382], device='cuda:4'), covar=tensor([0.0391, 0.0521, 0.0400, 0.0513, 0.0475, 0.0397, 0.1382, 0.0354], device='cuda:4'), in_proj_covar=tensor([0.0255, 0.0250, 0.0256, 0.0256, 0.0303, 0.0267, 0.0384, 0.0225], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 07:08:28,711 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9551, 4.3984, 3.3994, 2.4095, 3.3384, 2.5468, 4.6516, 4.5071], device='cuda:4'), covar=tensor([0.2118, 0.0632, 0.1188, 0.1722, 0.2216, 0.1456, 0.0301, 0.0582], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0256, 0.0271, 0.0245, 0.0309, 0.0205, 0.0240, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:08:47,579 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=44020.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:08:49,358 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0021, 5.4876, 5.5705, 5.4292, 5.4341, 5.9847, 5.7230, 5.4139], device='cuda:4'), covar=tensor([0.0760, 0.1405, 0.1218, 0.1632, 0.2688, 0.0900, 0.0979, 0.2101], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0410, 0.0394, 0.0349, 0.0467, 0.0423, 0.0328, 0.0471], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:09:31,564 INFO [train.py:904] (4/8) Epoch 5, batch 3450, loss[loss=0.193, simple_loss=0.2712, pruned_loss=0.05742, over 15793.00 frames. ], tot_loss[loss=0.2155, simple_loss=0.2921, pruned_loss=0.06949, over 3312627.31 frames. ], batch size: 35, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:09:33,938 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9329, 3.7425, 3.6445, 5.3086, 4.9670, 4.8545, 1.4530, 4.3795], device='cuda:4'), covar=tensor([0.1296, 0.0401, 0.0650, 0.0062, 0.0200, 0.0223, 0.1428, 0.0325], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0139, 0.0163, 0.0086, 0.0183, 0.0171, 0.0157, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 07:09:44,810 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.795e+02 3.034e+02 3.683e+02 4.468e+02 1.074e+03, threshold=7.367e+02, percent-clipped=2.0 2023-04-28 07:10:24,802 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5349, 4.3384, 3.9624, 2.0052, 3.0507, 2.4663, 3.8399, 4.0447], device='cuda:4'), covar=tensor([0.0332, 0.0480, 0.0404, 0.1529, 0.0669, 0.0848, 0.0626, 0.0818], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0135, 0.0153, 0.0140, 0.0131, 0.0124, 0.0141, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 07:10:39,128 INFO [train.py:904] (4/8) Epoch 5, batch 3500, loss[loss=0.2237, simple_loss=0.2941, pruned_loss=0.07664, over 16809.00 frames. ], tot_loss[loss=0.213, simple_loss=0.2895, pruned_loss=0.06828, over 3309660.28 frames. ], batch size: 102, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:11:37,791 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44143.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:11:49,615 INFO [train.py:904] (4/8) Epoch 5, batch 3550, loss[loss=0.2146, simple_loss=0.2894, pruned_loss=0.06988, over 16677.00 frames. ], tot_loss[loss=0.2127, simple_loss=0.2892, pruned_loss=0.06808, over 3313942.39 frames. ], batch size: 89, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:12:03,020 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.484e+02 3.064e+02 3.621e+02 4.319e+02 6.916e+02, threshold=7.242e+02, percent-clipped=0.0 2023-04-28 07:12:17,180 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=44170.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:12:59,202 INFO [train.py:904] (4/8) Epoch 5, batch 3600, loss[loss=0.2246, simple_loss=0.2894, pruned_loss=0.07992, over 11472.00 frames. ], tot_loss[loss=0.2122, simple_loss=0.2884, pruned_loss=0.06805, over 3301961.64 frames. ], batch size: 247, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:13:23,277 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=44218.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:13:26,070 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-28 07:14:09,668 INFO [train.py:904] (4/8) Epoch 5, batch 3650, loss[loss=0.1817, simple_loss=0.2678, pruned_loss=0.04775, over 17199.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2884, pruned_loss=0.06965, over 3298575.38 frames. ], batch size: 44, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:14:19,927 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-28 07:14:25,432 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.873e+02 3.015e+02 3.671e+02 4.591e+02 9.976e+02, threshold=7.342e+02, percent-clipped=3.0 2023-04-28 07:14:30,580 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44265.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:15:22,468 INFO [train.py:904] (4/8) Epoch 5, batch 3700, loss[loss=0.2005, simple_loss=0.2694, pruned_loss=0.06583, over 16782.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.287, pruned_loss=0.071, over 3288951.54 frames. ], batch size: 102, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:15:44,546 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44315.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:16:06,298 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5351, 4.5200, 4.4510, 4.3713, 4.0852, 4.4903, 4.4101, 4.2646], device='cuda:4'), covar=tensor([0.0478, 0.0316, 0.0197, 0.0176, 0.0847, 0.0329, 0.0355, 0.0428], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0194, 0.0225, 0.0200, 0.0261, 0.0222, 0.0164, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:16:36,542 INFO [train.py:904] (4/8) Epoch 5, batch 3750, loss[loss=0.2194, simple_loss=0.2811, pruned_loss=0.0789, over 16725.00 frames. ], tot_loss[loss=0.2164, simple_loss=0.2878, pruned_loss=0.07255, over 3267724.38 frames. ], batch size: 124, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:16:44,959 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0502, 4.9822, 4.8091, 4.1969, 4.9256, 1.9621, 4.6669, 4.7964], device='cuda:4'), covar=tensor([0.0051, 0.0045, 0.0095, 0.0285, 0.0041, 0.1536, 0.0085, 0.0091], device='cuda:4'), in_proj_covar=tensor([0.0097, 0.0086, 0.0128, 0.0136, 0.0099, 0.0139, 0.0114, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:16:51,455 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-04-28 07:16:52,717 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.995e+02 2.955e+02 3.644e+02 4.475e+02 7.424e+02, threshold=7.289e+02, percent-clipped=1.0 2023-04-28 07:16:56,349 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2922, 3.6631, 3.7041, 1.7859, 3.8540, 3.9141, 3.0970, 3.0174], device='cuda:4'), covar=tensor([0.0751, 0.0120, 0.0132, 0.1114, 0.0075, 0.0059, 0.0336, 0.0352], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0087, 0.0083, 0.0141, 0.0072, 0.0079, 0.0117, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 07:17:49,965 INFO [train.py:904] (4/8) Epoch 5, batch 3800, loss[loss=0.2192, simple_loss=0.2976, pruned_loss=0.07045, over 17019.00 frames. ], tot_loss[loss=0.2175, simple_loss=0.2879, pruned_loss=0.07353, over 3283621.38 frames. ], batch size: 55, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:18:50,508 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=44443.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:19:01,035 INFO [train.py:904] (4/8) Epoch 5, batch 3850, loss[loss=0.2188, simple_loss=0.2773, pruned_loss=0.08014, over 16911.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.2878, pruned_loss=0.07401, over 3280069.27 frames. ], batch size: 109, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:19:16,579 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.887e+02 2.978e+02 3.515e+02 4.161e+02 7.942e+02, threshold=7.030e+02, percent-clipped=2.0 2023-04-28 07:19:57,671 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=44491.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:20:11,834 INFO [train.py:904] (4/8) Epoch 5, batch 3900, loss[loss=0.2178, simple_loss=0.2881, pruned_loss=0.07372, over 16805.00 frames. ], tot_loss[loss=0.218, simple_loss=0.2876, pruned_loss=0.0742, over 3283934.05 frames. ], batch size: 102, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:20:16,590 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7361, 4.6618, 4.5761, 3.9523, 4.6063, 1.7968, 4.4450, 4.5116], device='cuda:4'), covar=tensor([0.0059, 0.0055, 0.0081, 0.0297, 0.0059, 0.1666, 0.0077, 0.0126], device='cuda:4'), in_proj_covar=tensor([0.0095, 0.0085, 0.0126, 0.0134, 0.0097, 0.0137, 0.0113, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:20:36,056 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-28 07:20:49,843 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=44527.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:21:10,954 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6454, 3.6864, 3.9731, 2.8893, 3.6357, 3.8532, 3.8784, 2.5131], device='cuda:4'), covar=tensor([0.0253, 0.0095, 0.0024, 0.0178, 0.0033, 0.0037, 0.0025, 0.0242], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0059, 0.0060, 0.0114, 0.0061, 0.0069, 0.0064, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:21:22,511 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=44550.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:21:23,256 INFO [train.py:904] (4/8) Epoch 5, batch 3950, loss[loss=0.2316, simple_loss=0.286, pruned_loss=0.08862, over 16847.00 frames. ], tot_loss[loss=0.2172, simple_loss=0.2861, pruned_loss=0.07419, over 3285722.06 frames. ], batch size: 109, lr: 1.36e-02, grad_scale: 8.0 2023-04-28 07:21:23,616 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8049, 5.1171, 4.7941, 4.8326, 4.5002, 4.4402, 4.5982, 5.1546], device='cuda:4'), covar=tensor([0.0647, 0.0728, 0.0916, 0.0494, 0.0655, 0.0910, 0.0702, 0.0731], device='cuda:4'), in_proj_covar=tensor([0.0367, 0.0495, 0.0419, 0.0323, 0.0314, 0.0322, 0.0400, 0.0355], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 07:21:37,717 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.214e+02 3.071e+02 3.714e+02 4.451e+02 1.184e+03, threshold=7.429e+02, percent-clipped=3.0 2023-04-28 07:21:44,193 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=44565.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:22:16,018 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=44588.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:22:34,879 INFO [train.py:904] (4/8) Epoch 5, batch 4000, loss[loss=0.2337, simple_loss=0.3045, pruned_loss=0.0815, over 16679.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.2862, pruned_loss=0.07475, over 3290400.24 frames. ], batch size: 134, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:22:49,510 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=44611.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:22:51,756 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=44613.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:22:54,936 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=44615.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:23:45,848 INFO [train.py:904] (4/8) Epoch 5, batch 4050, loss[loss=0.1833, simple_loss=0.2679, pruned_loss=0.04937, over 16773.00 frames. ], tot_loss[loss=0.216, simple_loss=0.2858, pruned_loss=0.07307, over 3281523.76 frames. ], batch size: 83, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:24:02,948 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 2.760e+02 3.326e+02 4.025e+02 7.816e+02, threshold=6.651e+02, percent-clipped=1.0 2023-04-28 07:24:04,434 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=44663.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:24:49,321 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5321, 3.6922, 2.8599, 2.2878, 2.7406, 2.1928, 3.6331, 3.8649], device='cuda:4'), covar=tensor([0.2112, 0.0565, 0.1190, 0.1450, 0.2010, 0.1445, 0.0400, 0.0502], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0253, 0.0272, 0.0250, 0.0317, 0.0209, 0.0241, 0.0268], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:25:01,114 INFO [train.py:904] (4/8) Epoch 5, batch 4100, loss[loss=0.2058, simple_loss=0.2915, pruned_loss=0.06002, over 16400.00 frames. ], tot_loss[loss=0.215, simple_loss=0.2866, pruned_loss=0.07166, over 3267347.26 frames. ], batch size: 35, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:26:16,728 INFO [train.py:904] (4/8) Epoch 5, batch 4150, loss[loss=0.2146, simple_loss=0.303, pruned_loss=0.06313, over 16829.00 frames. ], tot_loss[loss=0.2233, simple_loss=0.2954, pruned_loss=0.07562, over 3238408.19 frames. ], batch size: 96, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:26:34,908 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.630e+02 2.818e+02 3.479e+02 4.379e+02 9.731e+02, threshold=6.958e+02, percent-clipped=3.0 2023-04-28 07:26:42,222 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.51 vs. limit=5.0 2023-04-28 07:27:08,407 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5646, 3.3782, 3.4648, 2.9550, 3.4485, 2.0384, 3.2980, 3.1289], device='cuda:4'), covar=tensor([0.0078, 0.0069, 0.0084, 0.0193, 0.0055, 0.1345, 0.0088, 0.0128], device='cuda:4'), in_proj_covar=tensor([0.0093, 0.0083, 0.0121, 0.0130, 0.0093, 0.0135, 0.0108, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 07:27:24,488 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-28 07:27:35,372 INFO [train.py:904] (4/8) Epoch 5, batch 4200, loss[loss=0.2728, simple_loss=0.3455, pruned_loss=0.1, over 16450.00 frames. ], tot_loss[loss=0.2301, simple_loss=0.3034, pruned_loss=0.07841, over 3201289.42 frames. ], batch size: 146, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:27:36,195 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 07:28:14,231 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1860, 3.2738, 1.7305, 3.4547, 2.2921, 3.4803, 1.9325, 2.5648], device='cuda:4'), covar=tensor([0.0194, 0.0349, 0.1502, 0.0062, 0.0827, 0.0405, 0.1351, 0.0669], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0156, 0.0177, 0.0084, 0.0162, 0.0185, 0.0187, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 07:28:28,379 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 07:28:51,232 INFO [train.py:904] (4/8) Epoch 5, batch 4250, loss[loss=0.2657, simple_loss=0.3247, pruned_loss=0.1033, over 12452.00 frames. ], tot_loss[loss=0.2318, simple_loss=0.3069, pruned_loss=0.0784, over 3192154.25 frames. ], batch size: 248, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:29:02,274 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7502, 4.5126, 4.1735, 2.0279, 3.3035, 2.8113, 3.8765, 4.2429], device='cuda:4'), covar=tensor([0.0210, 0.0359, 0.0409, 0.1505, 0.0598, 0.0765, 0.0555, 0.0592], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0128, 0.0151, 0.0137, 0.0130, 0.0123, 0.0138, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 07:29:07,183 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 3.072e+02 3.670e+02 4.674e+02 1.163e+03, threshold=7.340e+02, percent-clipped=6.0 2023-04-28 07:29:38,859 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44883.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:30:06,000 INFO [train.py:904] (4/8) Epoch 5, batch 4300, loss[loss=0.2491, simple_loss=0.3387, pruned_loss=0.07976, over 16253.00 frames. ], tot_loss[loss=0.231, simple_loss=0.3079, pruned_loss=0.07702, over 3192354.47 frames. ], batch size: 165, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:30:13,088 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44906.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:30:56,806 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-28 07:30:57,963 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.96 vs. limit=5.0 2023-04-28 07:31:19,637 INFO [train.py:904] (4/8) Epoch 5, batch 4350, loss[loss=0.2335, simple_loss=0.3177, pruned_loss=0.07465, over 17115.00 frames. ], tot_loss[loss=0.2347, simple_loss=0.3119, pruned_loss=0.0787, over 3191844.99 frames. ], batch size: 47, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:31:36,570 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.041e+02 3.143e+02 3.858e+02 4.477e+02 9.466e+02, threshold=7.715e+02, percent-clipped=1.0 2023-04-28 07:32:35,451 INFO [train.py:904] (4/8) Epoch 5, batch 4400, loss[loss=0.2282, simple_loss=0.306, pruned_loss=0.07523, over 17223.00 frames. ], tot_loss[loss=0.2353, simple_loss=0.313, pruned_loss=0.07878, over 3203716.04 frames. ], batch size: 44, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:33:05,479 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6866, 2.5858, 2.0036, 2.4742, 3.0570, 2.7464, 3.7739, 3.3788], device='cuda:4'), covar=tensor([0.0016, 0.0144, 0.0211, 0.0162, 0.0085, 0.0157, 0.0039, 0.0073], device='cuda:4'), in_proj_covar=tensor([0.0079, 0.0147, 0.0151, 0.0146, 0.0145, 0.0154, 0.0122, 0.0134], device='cuda:4'), out_proj_covar=tensor([9.9095e-05, 1.8078e-04, 1.8227e-04, 1.7670e-04, 1.8205e-04, 1.9165e-04, 1.4821e-04, 1.6674e-04], device='cuda:4') 2023-04-28 07:33:22,160 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0512, 1.7143, 2.2624, 2.9894, 2.8155, 3.3965, 1.8891, 3.2951], device='cuda:4'), covar=tensor([0.0057, 0.0206, 0.0134, 0.0091, 0.0077, 0.0036, 0.0199, 0.0038], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0140, 0.0125, 0.0125, 0.0124, 0.0090, 0.0138, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 07:33:49,609 INFO [train.py:904] (4/8) Epoch 5, batch 4450, loss[loss=0.2394, simple_loss=0.3249, pruned_loss=0.07692, over 16787.00 frames. ], tot_loss[loss=0.2369, simple_loss=0.3156, pruned_loss=0.07916, over 3209238.97 frames. ], batch size: 102, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:33:56,570 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45056.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:34:02,440 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0140, 1.6044, 2.2972, 2.9486, 2.7630, 3.1325, 1.8253, 3.1539], device='cuda:4'), covar=tensor([0.0061, 0.0229, 0.0140, 0.0091, 0.0087, 0.0058, 0.0217, 0.0039], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0143, 0.0127, 0.0126, 0.0126, 0.0091, 0.0140, 0.0081], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 07:34:06,116 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.921e+02 2.841e+02 3.317e+02 4.350e+02 8.803e+02, threshold=6.635e+02, percent-clipped=2.0 2023-04-28 07:34:13,950 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1481, 4.1417, 4.5618, 4.5308, 4.5396, 4.0792, 4.1043, 3.9404], device='cuda:4'), covar=tensor([0.0187, 0.0267, 0.0237, 0.0284, 0.0302, 0.0266, 0.0838, 0.0439], device='cuda:4'), in_proj_covar=tensor([0.0229, 0.0220, 0.0224, 0.0228, 0.0274, 0.0242, 0.0347, 0.0203], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 07:34:33,622 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45081.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:35:01,844 INFO [train.py:904] (4/8) Epoch 5, batch 4500, loss[loss=0.2266, simple_loss=0.3066, pruned_loss=0.07326, over 16937.00 frames. ], tot_loss[loss=0.2356, simple_loss=0.3147, pruned_loss=0.07827, over 3219539.21 frames. ], batch size: 41, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:35:26,552 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45117.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 07:36:02,308 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45142.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:36:03,687 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8816, 3.2639, 2.7741, 4.7318, 3.9126, 4.0591, 1.5906, 3.2375], device='cuda:4'), covar=tensor([0.1328, 0.0505, 0.1023, 0.0043, 0.0219, 0.0294, 0.1399, 0.0662], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0139, 0.0164, 0.0082, 0.0172, 0.0167, 0.0158, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 07:36:13,663 INFO [train.py:904] (4/8) Epoch 5, batch 4550, loss[loss=0.2924, simple_loss=0.3582, pruned_loss=0.1133, over 15377.00 frames. ], tot_loss[loss=0.2367, simple_loss=0.3156, pruned_loss=0.07887, over 3235987.97 frames. ], batch size: 190, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:36:30,358 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.972e+02 2.621e+02 3.061e+02 3.788e+02 6.051e+02, threshold=6.121e+02, percent-clipped=0.0 2023-04-28 07:37:00,738 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=45183.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:37:27,614 INFO [train.py:904] (4/8) Epoch 5, batch 4600, loss[loss=0.2543, simple_loss=0.3206, pruned_loss=0.094, over 11430.00 frames. ], tot_loss[loss=0.2361, simple_loss=0.3157, pruned_loss=0.07825, over 3241081.45 frames. ], batch size: 246, lr: 1.35e-02, grad_scale: 8.0 2023-04-28 07:37:35,959 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=45206.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:37:45,260 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8493, 2.3295, 2.4118, 4.4611, 2.0568, 3.1565, 2.4855, 2.4315], device='cuda:4'), covar=tensor([0.0488, 0.1812, 0.0927, 0.0225, 0.2753, 0.0898, 0.1622, 0.2222], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0316, 0.0254, 0.0303, 0.0364, 0.0302, 0.0280, 0.0385], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 07:38:12,280 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=45231.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:38:29,522 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2430, 3.1885, 3.2183, 3.4174, 3.4431, 3.1497, 3.3857, 3.4981], device='cuda:4'), covar=tensor([0.0717, 0.0701, 0.1087, 0.0503, 0.0526, 0.2221, 0.0843, 0.0478], device='cuda:4'), in_proj_covar=tensor([0.0352, 0.0440, 0.0554, 0.0448, 0.0335, 0.0337, 0.0348, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 07:38:40,742 INFO [train.py:904] (4/8) Epoch 5, batch 4650, loss[loss=0.211, simple_loss=0.2948, pruned_loss=0.06362, over 16545.00 frames. ], tot_loss[loss=0.2351, simple_loss=0.3143, pruned_loss=0.07793, over 3220659.80 frames. ], batch size: 68, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:38:45,218 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=45254.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:38:57,177 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.863e+02 2.661e+02 3.124e+02 3.951e+02 6.864e+02, threshold=6.249e+02, percent-clipped=3.0 2023-04-28 07:39:43,095 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-28 07:39:54,500 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0841, 3.3615, 3.5400, 3.4972, 3.4538, 3.2475, 3.2335, 3.2712], device='cuda:4'), covar=tensor([0.0317, 0.0366, 0.0285, 0.0385, 0.0419, 0.0351, 0.0812, 0.0489], device='cuda:4'), in_proj_covar=tensor([0.0230, 0.0219, 0.0225, 0.0228, 0.0272, 0.0240, 0.0344, 0.0204], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 07:39:55,266 INFO [train.py:904] (4/8) Epoch 5, batch 4700, loss[loss=0.2286, simple_loss=0.2993, pruned_loss=0.07898, over 11596.00 frames. ], tot_loss[loss=0.2318, simple_loss=0.3111, pruned_loss=0.07624, over 3230051.42 frames. ], batch size: 246, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:41:03,107 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45348.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 07:41:06,800 INFO [train.py:904] (4/8) Epoch 5, batch 4750, loss[loss=0.2196, simple_loss=0.2973, pruned_loss=0.07096, over 16271.00 frames. ], tot_loss[loss=0.229, simple_loss=0.308, pruned_loss=0.07498, over 3210105.09 frames. ], batch size: 35, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:41:22,808 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.785e+02 2.710e+02 3.330e+02 4.145e+02 7.902e+02, threshold=6.661e+02, percent-clipped=5.0 2023-04-28 07:42:17,261 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8720, 1.3439, 1.9730, 2.5581, 2.5047, 2.9276, 1.2710, 2.7474], device='cuda:4'), covar=tensor([0.0050, 0.0227, 0.0134, 0.0094, 0.0085, 0.0055, 0.0280, 0.0035], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0141, 0.0127, 0.0125, 0.0125, 0.0089, 0.0138, 0.0081], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 07:42:20,533 INFO [train.py:904] (4/8) Epoch 5, batch 4800, loss[loss=0.2224, simple_loss=0.3036, pruned_loss=0.07057, over 17078.00 frames. ], tot_loss[loss=0.2255, simple_loss=0.3048, pruned_loss=0.0731, over 3204915.03 frames. ], batch size: 55, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:42:33,316 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45409.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 07:42:38,678 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=45412.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 07:43:14,510 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=45437.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:43:34,387 INFO [train.py:904] (4/8) Epoch 5, batch 4850, loss[loss=0.2171, simple_loss=0.3051, pruned_loss=0.06456, over 16419.00 frames. ], tot_loss[loss=0.2268, simple_loss=0.3065, pruned_loss=0.07356, over 3173100.37 frames. ], batch size: 146, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:43:50,673 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 2.510e+02 3.182e+02 3.913e+02 7.711e+02, threshold=6.364e+02, percent-clipped=1.0 2023-04-28 07:44:47,402 INFO [train.py:904] (4/8) Epoch 5, batch 4900, loss[loss=0.2131, simple_loss=0.2942, pruned_loss=0.06602, over 16849.00 frames. ], tot_loss[loss=0.2256, simple_loss=0.306, pruned_loss=0.0726, over 3182966.08 frames. ], batch size: 42, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:45:00,182 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7912, 1.9884, 1.3938, 1.9243, 2.6192, 2.3034, 3.0060, 2.8440], device='cuda:4'), covar=tensor([0.0037, 0.0191, 0.0299, 0.0220, 0.0094, 0.0185, 0.0055, 0.0091], device='cuda:4'), in_proj_covar=tensor([0.0078, 0.0150, 0.0152, 0.0147, 0.0146, 0.0154, 0.0121, 0.0133], device='cuda:4'), out_proj_covar=tensor([9.8378e-05, 1.8416e-04, 1.8283e-04, 1.7751e-04, 1.8268e-04, 1.9105e-04, 1.4469e-04, 1.6540e-04], device='cuda:4') 2023-04-28 07:45:57,846 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8298, 3.2412, 3.2066, 1.9721, 2.8076, 3.1690, 3.0448, 1.6398], device='cuda:4'), covar=tensor([0.0300, 0.0021, 0.0027, 0.0230, 0.0036, 0.0038, 0.0033, 0.0307], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0054, 0.0058, 0.0113, 0.0059, 0.0065, 0.0062, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:46:01,014 INFO [train.py:904] (4/8) Epoch 5, batch 4950, loss[loss=0.219, simple_loss=0.3032, pruned_loss=0.06737, over 17125.00 frames. ], tot_loss[loss=0.2257, simple_loss=0.3061, pruned_loss=0.07263, over 3177185.15 frames. ], batch size: 47, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:46:09,465 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3790, 3.5730, 1.5784, 3.9274, 2.4054, 3.7589, 1.7814, 2.6220], device='cuda:4'), covar=tensor([0.0141, 0.0240, 0.1606, 0.0031, 0.0731, 0.0292, 0.1336, 0.0648], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0148, 0.0173, 0.0078, 0.0159, 0.0178, 0.0185, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 07:46:13,615 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4984, 3.9091, 4.2219, 1.8759, 4.4221, 4.4718, 3.1837, 3.2607], device='cuda:4'), covar=tensor([0.0649, 0.0101, 0.0073, 0.1037, 0.0025, 0.0023, 0.0217, 0.0301], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0084, 0.0077, 0.0135, 0.0068, 0.0071, 0.0110, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 07:46:15,481 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.470e+02 3.018e+02 3.573e+02 4.428e+02 9.346e+02, threshold=7.147e+02, percent-clipped=9.0 2023-04-28 07:46:58,633 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8632, 2.7727, 2.7238, 1.8541, 2.4930, 2.6682, 2.6273, 1.7033], device='cuda:4'), covar=tensor([0.0245, 0.0025, 0.0030, 0.0219, 0.0044, 0.0038, 0.0037, 0.0267], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0054, 0.0058, 0.0115, 0.0059, 0.0066, 0.0063, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:47:13,801 INFO [train.py:904] (4/8) Epoch 5, batch 5000, loss[loss=0.2143, simple_loss=0.3045, pruned_loss=0.06198, over 16468.00 frames. ], tot_loss[loss=0.226, simple_loss=0.3073, pruned_loss=0.07236, over 3187796.58 frames. ], batch size: 75, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:47:49,026 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9124, 2.2049, 2.3077, 3.1212, 2.6792, 3.2021, 1.7424, 2.5884], device='cuda:4'), covar=tensor([0.1073, 0.0534, 0.0966, 0.0094, 0.0173, 0.0335, 0.1209, 0.0719], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0139, 0.0164, 0.0083, 0.0169, 0.0169, 0.0160, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 07:48:27,073 INFO [train.py:904] (4/8) Epoch 5, batch 5050, loss[loss=0.2187, simple_loss=0.3113, pruned_loss=0.06309, over 16708.00 frames. ], tot_loss[loss=0.2253, simple_loss=0.3071, pruned_loss=0.07171, over 3194388.57 frames. ], batch size: 89, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:48:43,127 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.866e+02 2.781e+02 3.302e+02 3.985e+02 6.386e+02, threshold=6.604e+02, percent-clipped=0.0 2023-04-28 07:49:39,971 INFO [train.py:904] (4/8) Epoch 5, batch 5100, loss[loss=0.2316, simple_loss=0.3061, pruned_loss=0.0786, over 15361.00 frames. ], tot_loss[loss=0.2227, simple_loss=0.3047, pruned_loss=0.07036, over 3207812.10 frames. ], batch size: 190, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:49:42,342 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45702.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:49:44,453 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=45704.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 07:49:46,524 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7007, 1.4141, 1.9291, 2.5361, 2.4944, 2.8089, 1.6035, 2.7838], device='cuda:4'), covar=tensor([0.0079, 0.0255, 0.0158, 0.0127, 0.0109, 0.0080, 0.0236, 0.0039], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0143, 0.0127, 0.0125, 0.0126, 0.0090, 0.0141, 0.0083], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 07:49:56,007 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=45712.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 07:50:33,731 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=45737.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:50:53,044 INFO [train.py:904] (4/8) Epoch 5, batch 5150, loss[loss=0.246, simple_loss=0.3326, pruned_loss=0.07976, over 15279.00 frames. ], tot_loss[loss=0.2221, simple_loss=0.3047, pruned_loss=0.06974, over 3187221.31 frames. ], batch size: 190, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:51:06,962 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=45760.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:51:08,846 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.833e+02 2.626e+02 3.241e+02 3.847e+02 8.905e+02, threshold=6.482e+02, percent-clipped=5.0 2023-04-28 07:51:10,589 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45763.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:51:26,058 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45774.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:51:28,100 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6393, 2.6974, 2.3655, 4.0228, 3.3396, 3.7711, 1.4505, 2.7999], device='cuda:4'), covar=tensor([0.1310, 0.0571, 0.1116, 0.0067, 0.0198, 0.0319, 0.1439, 0.0781], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0139, 0.0165, 0.0083, 0.0168, 0.0170, 0.0159, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 07:51:43,172 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=45785.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:52:05,634 INFO [train.py:904] (4/8) Epoch 5, batch 5200, loss[loss=0.1852, simple_loss=0.2737, pruned_loss=0.04838, over 16882.00 frames. ], tot_loss[loss=0.2217, simple_loss=0.304, pruned_loss=0.06965, over 3197785.10 frames. ], batch size: 102, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:52:54,187 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45835.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:53:16,334 INFO [train.py:904] (4/8) Epoch 5, batch 5250, loss[loss=0.2187, simple_loss=0.3011, pruned_loss=0.06819, over 16734.00 frames. ], tot_loss[loss=0.2205, simple_loss=0.3018, pruned_loss=0.06962, over 3190452.82 frames. ], batch size: 134, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:53:27,332 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45859.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:53:31,050 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.833e+02 2.734e+02 3.218e+02 3.993e+02 9.159e+02, threshold=6.436e+02, percent-clipped=4.0 2023-04-28 07:54:00,986 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6354, 4.4889, 4.4209, 2.8058, 3.8225, 4.3916, 4.1146, 2.4901], device='cuda:4'), covar=tensor([0.0291, 0.0011, 0.0022, 0.0236, 0.0046, 0.0045, 0.0027, 0.0292], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0055, 0.0060, 0.0115, 0.0060, 0.0068, 0.0064, 0.0108], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:54:26,350 INFO [train.py:904] (4/8) Epoch 5, batch 5300, loss[loss=0.1857, simple_loss=0.2731, pruned_loss=0.04911, over 16908.00 frames. ], tot_loss[loss=0.2172, simple_loss=0.298, pruned_loss=0.0682, over 3195148.57 frames. ], batch size: 96, lr: 1.34e-02, grad_scale: 8.0 2023-04-28 07:54:53,121 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45920.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:55:00,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9755, 4.9054, 4.7674, 4.6716, 4.3053, 4.8600, 4.7484, 4.5147], device='cuda:4'), covar=tensor([0.0363, 0.0325, 0.0196, 0.0177, 0.0808, 0.0289, 0.0193, 0.0396], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0190, 0.0219, 0.0189, 0.0247, 0.0216, 0.0155, 0.0235], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 07:55:24,548 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8773, 3.8748, 4.2977, 4.2577, 4.2456, 3.8595, 3.8807, 3.8186], device='cuda:4'), covar=tensor([0.0245, 0.0529, 0.0311, 0.0376, 0.0400, 0.0272, 0.0773, 0.0400], device='cuda:4'), in_proj_covar=tensor([0.0239, 0.0232, 0.0230, 0.0234, 0.0286, 0.0249, 0.0353, 0.0210], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 07:55:36,423 INFO [train.py:904] (4/8) Epoch 5, batch 5350, loss[loss=0.2227, simple_loss=0.3018, pruned_loss=0.07181, over 16143.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.2956, pruned_loss=0.06707, over 3183042.52 frames. ], batch size: 35, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 07:55:53,057 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.681e+02 2.936e+02 3.263e+02 4.063e+02 1.076e+03, threshold=6.526e+02, percent-clipped=3.0 2023-04-28 07:56:52,488 INFO [train.py:904] (4/8) Epoch 5, batch 5400, loss[loss=0.2545, simple_loss=0.3284, pruned_loss=0.09034, over 15396.00 frames. ], tot_loss[loss=0.217, simple_loss=0.2983, pruned_loss=0.06782, over 3174905.77 frames. ], batch size: 190, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 07:56:57,453 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46004.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 07:58:09,281 INFO [train.py:904] (4/8) Epoch 5, batch 5450, loss[loss=0.2855, simple_loss=0.3475, pruned_loss=0.1118, over 12078.00 frames. ], tot_loss[loss=0.2208, simple_loss=0.3019, pruned_loss=0.06983, over 3165822.41 frames. ], batch size: 246, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 07:58:11,188 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=46052.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 07:58:19,234 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46058.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 07:58:24,190 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0693, 3.2129, 3.5209, 3.4829, 3.4479, 3.1569, 3.2466, 3.3294], device='cuda:4'), covar=tensor([0.0334, 0.0575, 0.0336, 0.0427, 0.0503, 0.0404, 0.0820, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0237, 0.0229, 0.0227, 0.0232, 0.0287, 0.0249, 0.0349, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 07:58:24,984 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.677e+02 3.104e+02 3.655e+02 4.572e+02 1.003e+03, threshold=7.310e+02, percent-clipped=8.0 2023-04-28 07:59:22,159 INFO [train.py:904] (4/8) Epoch 5, batch 5500, loss[loss=0.2706, simple_loss=0.3445, pruned_loss=0.09832, over 16681.00 frames. ], tot_loss[loss=0.2327, simple_loss=0.3117, pruned_loss=0.07679, over 3160596.63 frames. ], batch size: 124, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 07:59:46,524 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-28 08:00:07,596 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46130.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:00:39,132 INFO [train.py:904] (4/8) Epoch 5, batch 5550, loss[loss=0.2654, simple_loss=0.3404, pruned_loss=0.0952, over 16848.00 frames. ], tot_loss[loss=0.2451, simple_loss=0.3214, pruned_loss=0.08437, over 3144020.79 frames. ], batch size: 102, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 08:00:56,874 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.500e+02 4.429e+02 5.470e+02 6.755e+02 1.488e+03, threshold=1.094e+03, percent-clipped=17.0 2023-04-28 08:01:58,727 INFO [train.py:904] (4/8) Epoch 5, batch 5600, loss[loss=0.2635, simple_loss=0.3383, pruned_loss=0.0943, over 16362.00 frames. ], tot_loss[loss=0.2557, simple_loss=0.329, pruned_loss=0.09117, over 3097476.76 frames. ], batch size: 146, lr: 1.33e-02, grad_scale: 16.0 2023-04-28 08:02:23,964 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46215.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:03:23,179 INFO [train.py:904] (4/8) Epoch 5, batch 5650, loss[loss=0.2591, simple_loss=0.3326, pruned_loss=0.09284, over 16560.00 frames. ], tot_loss[loss=0.2646, simple_loss=0.3351, pruned_loss=0.09705, over 3054695.79 frames. ], batch size: 68, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 08:03:42,413 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.653e+02 4.676e+02 6.240e+02 7.908e+02 1.367e+03, threshold=1.248e+03, percent-clipped=1.0 2023-04-28 08:04:04,191 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=46276.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:04:43,823 INFO [train.py:904] (4/8) Epoch 5, batch 5700, loss[loss=0.3482, simple_loss=0.3764, pruned_loss=0.16, over 11060.00 frames. ], tot_loss[loss=0.2655, simple_loss=0.3359, pruned_loss=0.09753, over 3078719.88 frames. ], batch size: 248, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 08:04:51,425 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 08:05:03,634 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5167, 2.5872, 2.2897, 3.8348, 2.9857, 3.6945, 1.4312, 2.7668], device='cuda:4'), covar=tensor([0.1787, 0.0689, 0.1386, 0.0144, 0.0354, 0.0377, 0.1873, 0.0887], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0140, 0.0167, 0.0083, 0.0174, 0.0174, 0.0162, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 08:05:11,516 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2268, 2.4092, 1.8228, 2.1290, 2.8282, 2.4979, 3.2125, 3.0451], device='cuda:4'), covar=tensor([0.0031, 0.0179, 0.0260, 0.0207, 0.0108, 0.0178, 0.0074, 0.0103], device='cuda:4'), in_proj_covar=tensor([0.0077, 0.0152, 0.0155, 0.0152, 0.0147, 0.0155, 0.0124, 0.0135], device='cuda:4'), out_proj_covar=tensor([9.6646e-05, 1.8666e-04, 1.8621e-04, 1.8289e-04, 1.8278e-04, 1.9084e-04, 1.4856e-04, 1.6678e-04], device='cuda:4') 2023-04-28 08:05:41,919 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=46337.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:06:04,707 INFO [train.py:904] (4/8) Epoch 5, batch 5750, loss[loss=0.2766, simple_loss=0.3462, pruned_loss=0.1035, over 16376.00 frames. ], tot_loss[loss=0.2689, simple_loss=0.339, pruned_loss=0.09939, over 3061954.57 frames. ], batch size: 146, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 08:06:16,559 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46358.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:06:23,495 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.622e+02 4.192e+02 4.948e+02 6.377e+02 1.174e+03, threshold=9.897e+02, percent-clipped=0.0 2023-04-28 08:06:26,622 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8259, 2.6777, 2.5838, 1.8576, 2.4333, 2.5739, 2.5617, 1.8502], device='cuda:4'), covar=tensor([0.0265, 0.0028, 0.0040, 0.0207, 0.0047, 0.0058, 0.0037, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0053, 0.0058, 0.0114, 0.0058, 0.0067, 0.0062, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:07:10,470 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9745, 2.7410, 2.6573, 1.7221, 2.7760, 2.8306, 2.4239, 2.3605], device='cuda:4'), covar=tensor([0.0696, 0.0150, 0.0182, 0.0953, 0.0095, 0.0112, 0.0323, 0.0399], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0086, 0.0079, 0.0138, 0.0070, 0.0075, 0.0115, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 08:07:15,863 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-28 08:07:26,288 INFO [train.py:904] (4/8) Epoch 5, batch 5800, loss[loss=0.3202, simple_loss=0.3618, pruned_loss=0.1393, over 11821.00 frames. ], tot_loss[loss=0.2689, simple_loss=0.3393, pruned_loss=0.09929, over 3023469.72 frames. ], batch size: 247, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 08:07:35,758 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=46406.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:07:59,622 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1719, 1.4129, 1.8040, 2.1203, 2.1871, 2.2292, 1.5213, 2.1021], device='cuda:4'), covar=tensor([0.0069, 0.0189, 0.0118, 0.0119, 0.0093, 0.0072, 0.0188, 0.0038], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0138, 0.0125, 0.0122, 0.0124, 0.0088, 0.0138, 0.0077], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 08:08:04,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9431, 3.1881, 1.4163, 3.3130, 2.1869, 3.1780, 1.6515, 2.4624], device='cuda:4'), covar=tensor([0.0175, 0.0250, 0.1648, 0.0052, 0.0722, 0.0420, 0.1455, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0148, 0.0178, 0.0079, 0.0162, 0.0181, 0.0186, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:08:13,781 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46430.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:08:46,734 INFO [train.py:904] (4/8) Epoch 5, batch 5850, loss[loss=0.2573, simple_loss=0.3346, pruned_loss=0.08995, over 16634.00 frames. ], tot_loss[loss=0.2641, simple_loss=0.3358, pruned_loss=0.09617, over 3045413.32 frames. ], batch size: 62, lr: 1.33e-02, grad_scale: 4.0 2023-04-28 08:09:06,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.458e+02 3.799e+02 4.852e+02 6.197e+02 1.255e+03, threshold=9.704e+02, percent-clipped=4.0 2023-04-28 08:09:29,289 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=46478.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:10:09,066 INFO [train.py:904] (4/8) Epoch 5, batch 5900, loss[loss=0.3242, simple_loss=0.3639, pruned_loss=0.1422, over 11971.00 frames. ], tot_loss[loss=0.2625, simple_loss=0.3345, pruned_loss=0.09522, over 3054879.12 frames. ], batch size: 246, lr: 1.33e-02, grad_scale: 4.0 2023-04-28 08:10:36,181 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46515.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:11:31,951 INFO [train.py:904] (4/8) Epoch 5, batch 5950, loss[loss=0.2611, simple_loss=0.3479, pruned_loss=0.0871, over 16181.00 frames. ], tot_loss[loss=0.2609, simple_loss=0.3349, pruned_loss=0.09347, over 3047036.88 frames. ], batch size: 35, lr: 1.33e-02, grad_scale: 4.0 2023-04-28 08:11:37,798 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6713, 2.6056, 2.5736, 1.8106, 2.4387, 2.4900, 2.5780, 1.7059], device='cuda:4'), covar=tensor([0.0286, 0.0036, 0.0042, 0.0217, 0.0055, 0.0067, 0.0043, 0.0272], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0054, 0.0058, 0.0115, 0.0058, 0.0069, 0.0063, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:11:52,723 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=46563.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:11:53,494 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.225e+02 3.478e+02 4.456e+02 5.303e+02 9.511e+02, threshold=8.911e+02, percent-clipped=0.0 2023-04-28 08:12:37,086 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2212, 3.5494, 3.6845, 1.5538, 3.8761, 3.9113, 2.8793, 2.7301], device='cuda:4'), covar=tensor([0.0789, 0.0122, 0.0134, 0.1202, 0.0049, 0.0051, 0.0315, 0.0444], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0086, 0.0079, 0.0138, 0.0070, 0.0075, 0.0114, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 08:12:43,413 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=46595.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:12:51,978 INFO [train.py:904] (4/8) Epoch 5, batch 6000, loss[loss=0.2143, simple_loss=0.2969, pruned_loss=0.06581, over 16653.00 frames. ], tot_loss[loss=0.2597, simple_loss=0.3338, pruned_loss=0.09276, over 3063485.27 frames. ], batch size: 62, lr: 1.33e-02, grad_scale: 8.0 2023-04-28 08:12:51,978 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 08:13:04,052 INFO [train.py:938] (4/8) Epoch 5, validation: loss=0.1879, simple_loss=0.2992, pruned_loss=0.03826, over 944034.00 frames. 2023-04-28 08:13:04,052 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 08:13:05,695 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=46602.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:13:51,746 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46632.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:14:27,390 INFO [train.py:904] (4/8) Epoch 5, batch 6050, loss[loss=0.2724, simple_loss=0.3504, pruned_loss=0.09722, over 17027.00 frames. ], tot_loss[loss=0.2594, simple_loss=0.3331, pruned_loss=0.09282, over 3061126.65 frames. ], batch size: 41, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:14:37,487 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=46656.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 08:14:48,362 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=46663.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:14:48,996 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.236e+02 4.108e+02 4.980e+02 6.205e+02 1.363e+03, threshold=9.960e+02, percent-clipped=7.0 2023-04-28 08:15:45,451 INFO [train.py:904] (4/8) Epoch 5, batch 6100, loss[loss=0.2189, simple_loss=0.3054, pruned_loss=0.06617, over 16403.00 frames. ], tot_loss[loss=0.2564, simple_loss=0.3315, pruned_loss=0.09063, over 3078339.17 frames. ], batch size: 146, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:16:39,306 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3983, 3.5784, 2.8211, 2.1987, 2.5806, 2.1344, 3.6300, 3.7384], device='cuda:4'), covar=tensor([0.2292, 0.0655, 0.1333, 0.1540, 0.1875, 0.1418, 0.0412, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0249, 0.0268, 0.0243, 0.0292, 0.0201, 0.0241, 0.0251], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:17:01,186 INFO [train.py:904] (4/8) Epoch 5, batch 6150, loss[loss=0.2893, simple_loss=0.3425, pruned_loss=0.118, over 11707.00 frames. ], tot_loss[loss=0.2542, simple_loss=0.3292, pruned_loss=0.08957, over 3098767.88 frames. ], batch size: 246, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:17:22,716 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.183e+02 3.664e+02 4.778e+02 7.050e+02 1.596e+03, threshold=9.557e+02, percent-clipped=5.0 2023-04-28 08:18:21,097 INFO [train.py:904] (4/8) Epoch 5, batch 6200, loss[loss=0.2352, simple_loss=0.315, pruned_loss=0.07767, over 16995.00 frames. ], tot_loss[loss=0.2504, simple_loss=0.3256, pruned_loss=0.08758, over 3103975.47 frames. ], batch size: 41, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:18:36,422 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=46810.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:19:09,740 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3439, 3.4302, 1.6477, 3.6809, 2.4414, 3.5563, 1.7840, 2.6206], device='cuda:4'), covar=tensor([0.0126, 0.0256, 0.1537, 0.0051, 0.0695, 0.0419, 0.1488, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0147, 0.0178, 0.0081, 0.0161, 0.0183, 0.0185, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:19:25,527 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.76 vs. limit=5.0 2023-04-28 08:19:37,501 INFO [train.py:904] (4/8) Epoch 5, batch 6250, loss[loss=0.2422, simple_loss=0.3249, pruned_loss=0.07979, over 16728.00 frames. ], tot_loss[loss=0.2495, simple_loss=0.3249, pruned_loss=0.08703, over 3121242.94 frames. ], batch size: 124, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:19:57,338 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.405e+02 3.628e+02 4.524e+02 5.799e+02 1.377e+03, threshold=9.049e+02, percent-clipped=5.0 2023-04-28 08:20:08,306 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2319, 1.8176, 1.9913, 3.5701, 1.7021, 2.5452, 1.9722, 1.8634], device='cuda:4'), covar=tensor([0.0688, 0.2569, 0.1310, 0.0439, 0.3580, 0.1297, 0.2357, 0.2948], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0311, 0.0258, 0.0305, 0.0366, 0.0299, 0.0281, 0.0374], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:20:09,430 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=46871.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:20:30,060 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6340, 2.2073, 1.6461, 1.9302, 2.6864, 2.3787, 2.9302, 2.8431], device='cuda:4'), covar=tensor([0.0043, 0.0174, 0.0235, 0.0217, 0.0099, 0.0172, 0.0071, 0.0101], device='cuda:4'), in_proj_covar=tensor([0.0076, 0.0151, 0.0153, 0.0150, 0.0146, 0.0154, 0.0126, 0.0135], device='cuda:4'), out_proj_covar=tensor([9.5265e-05, 1.8508e-04, 1.8329e-04, 1.8049e-04, 1.8139e-04, 1.8924e-04, 1.5053e-04, 1.6638e-04], device='cuda:4') 2023-04-28 08:20:31,483 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2416, 2.4295, 1.7669, 2.0703, 2.9164, 2.5524, 3.3529, 3.1287], device='cuda:4'), covar=tensor([0.0029, 0.0167, 0.0251, 0.0211, 0.0114, 0.0168, 0.0070, 0.0096], device='cuda:4'), in_proj_covar=tensor([0.0076, 0.0151, 0.0153, 0.0150, 0.0146, 0.0154, 0.0126, 0.0135], device='cuda:4'), out_proj_covar=tensor([9.5248e-05, 1.8506e-04, 1.8323e-04, 1.8056e-04, 1.8141e-04, 1.8925e-04, 1.5055e-04, 1.6643e-04], device='cuda:4') 2023-04-28 08:20:41,288 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7013, 2.9386, 2.4326, 4.2842, 3.4470, 3.9753, 1.5198, 3.0316], device='cuda:4'), covar=tensor([0.1371, 0.0535, 0.1228, 0.0066, 0.0285, 0.0359, 0.1500, 0.0711], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0144, 0.0169, 0.0085, 0.0179, 0.0179, 0.0164, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:20:55,067 INFO [train.py:904] (4/8) Epoch 5, batch 6300, loss[loss=0.3348, simple_loss=0.3717, pruned_loss=0.1489, over 11765.00 frames. ], tot_loss[loss=0.2492, simple_loss=0.3251, pruned_loss=0.08667, over 3124210.67 frames. ], batch size: 247, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:21:43,817 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46932.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:22:12,095 INFO [train.py:904] (4/8) Epoch 5, batch 6350, loss[loss=0.2166, simple_loss=0.2922, pruned_loss=0.07054, over 17140.00 frames. ], tot_loss[loss=0.2518, simple_loss=0.3266, pruned_loss=0.08844, over 3115293.15 frames. ], batch size: 48, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:22:13,436 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46951.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 08:22:22,791 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46958.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:22:31,677 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.349e+02 3.778e+02 4.701e+02 5.946e+02 1.466e+03, threshold=9.402e+02, percent-clipped=8.0 2023-04-28 08:22:35,627 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6294, 1.3462, 1.5159, 1.6205, 1.8751, 1.8367, 1.4684, 1.6517], device='cuda:4'), covar=tensor([0.0084, 0.0142, 0.0084, 0.0101, 0.0074, 0.0058, 0.0138, 0.0038], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0141, 0.0126, 0.0121, 0.0125, 0.0089, 0.0138, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 08:22:54,452 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7060, 3.3585, 3.0346, 5.0743, 4.3644, 4.3710, 1.9604, 3.0768], device='cuda:4'), covar=tensor([0.1412, 0.0529, 0.0920, 0.0064, 0.0274, 0.0298, 0.1254, 0.0801], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0144, 0.0169, 0.0086, 0.0180, 0.0180, 0.0164, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:22:57,126 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=46980.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:23:10,318 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2494, 4.1503, 4.0654, 3.9684, 3.7057, 4.1682, 3.9943, 3.8980], device='cuda:4'), covar=tensor([0.0422, 0.0279, 0.0214, 0.0184, 0.0760, 0.0259, 0.0433, 0.0459], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0182, 0.0206, 0.0177, 0.0234, 0.0206, 0.0149, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:23:28,012 INFO [train.py:904] (4/8) Epoch 5, batch 6400, loss[loss=0.2744, simple_loss=0.3571, pruned_loss=0.09588, over 16624.00 frames. ], tot_loss[loss=0.2539, simple_loss=0.3279, pruned_loss=0.08993, over 3099447.71 frames. ], batch size: 134, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:23:32,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2004, 1.9058, 2.1884, 3.5124, 1.9184, 2.6453, 2.1365, 1.9749], device='cuda:4'), covar=tensor([0.0619, 0.2006, 0.1076, 0.0392, 0.2908, 0.1070, 0.1902, 0.2339], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0313, 0.0258, 0.0306, 0.0370, 0.0302, 0.0282, 0.0376], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:23:41,716 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-28 08:24:40,867 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-28 08:24:42,540 INFO [train.py:904] (4/8) Epoch 5, batch 6450, loss[loss=0.2118, simple_loss=0.3002, pruned_loss=0.0617, over 16690.00 frames. ], tot_loss[loss=0.2518, simple_loss=0.3267, pruned_loss=0.08847, over 3101612.73 frames. ], batch size: 89, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:24:44,915 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-28 08:24:57,318 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7564, 3.3275, 3.2520, 2.0491, 3.0548, 3.2055, 3.2172, 1.6070], device='cuda:4'), covar=tensor([0.0339, 0.0022, 0.0031, 0.0248, 0.0048, 0.0054, 0.0034, 0.0323], device='cuda:4'), in_proj_covar=tensor([0.0113, 0.0052, 0.0057, 0.0113, 0.0058, 0.0067, 0.0062, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:25:02,354 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.185e+02 3.689e+02 4.668e+02 5.826e+02 1.210e+03, threshold=9.337e+02, percent-clipped=4.0 2023-04-28 08:25:59,690 INFO [train.py:904] (4/8) Epoch 5, batch 6500, loss[loss=0.2389, simple_loss=0.3175, pruned_loss=0.08012, over 16846.00 frames. ], tot_loss[loss=0.2492, simple_loss=0.324, pruned_loss=0.08722, over 3108099.82 frames. ], batch size: 96, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:26:12,801 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9561, 2.6591, 2.6109, 1.7358, 2.8276, 2.7939, 2.2856, 2.3345], device='cuda:4'), covar=tensor([0.0760, 0.0170, 0.0183, 0.0993, 0.0096, 0.0135, 0.0448, 0.0428], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0089, 0.0081, 0.0140, 0.0071, 0.0077, 0.0116, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 08:27:18,005 INFO [train.py:904] (4/8) Epoch 5, batch 6550, loss[loss=0.329, simple_loss=0.38, pruned_loss=0.139, over 11673.00 frames. ], tot_loss[loss=0.2529, simple_loss=0.3275, pruned_loss=0.08912, over 3108738.18 frames. ], batch size: 246, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:27:37,105 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.214e+02 3.842e+02 4.732e+02 5.920e+02 1.050e+03, threshold=9.464e+02, percent-clipped=2.0 2023-04-28 08:27:41,299 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=47166.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:27:50,898 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-28 08:28:33,655 INFO [train.py:904] (4/8) Epoch 5, batch 6600, loss[loss=0.2494, simple_loss=0.3243, pruned_loss=0.08729, over 16708.00 frames. ], tot_loss[loss=0.2547, simple_loss=0.33, pruned_loss=0.08972, over 3112343.62 frames. ], batch size: 89, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:29:08,196 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9856, 3.2801, 3.5594, 3.5262, 3.4986, 3.2346, 3.2675, 3.3271], device='cuda:4'), covar=tensor([0.0382, 0.0446, 0.0311, 0.0375, 0.0448, 0.0401, 0.0935, 0.0460], device='cuda:4'), in_proj_covar=tensor([0.0234, 0.0225, 0.0233, 0.0232, 0.0279, 0.0247, 0.0358, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 08:29:19,195 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8150, 2.6181, 2.6086, 1.8148, 2.5630, 2.5427, 2.6225, 1.8034], device='cuda:4'), covar=tensor([0.0254, 0.0035, 0.0045, 0.0197, 0.0052, 0.0064, 0.0039, 0.0239], device='cuda:4'), in_proj_covar=tensor([0.0113, 0.0052, 0.0057, 0.0112, 0.0058, 0.0068, 0.0061, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:29:35,100 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47241.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:29:51,328 INFO [train.py:904] (4/8) Epoch 5, batch 6650, loss[loss=0.2384, simple_loss=0.3124, pruned_loss=0.08215, over 16683.00 frames. ], tot_loss[loss=0.2565, simple_loss=0.3308, pruned_loss=0.09106, over 3114060.95 frames. ], batch size: 134, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:29:52,448 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=47251.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:30:02,606 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=47258.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:30:11,434 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.376e+02 3.750e+02 4.777e+02 6.524e+02 9.688e+02, threshold=9.554e+02, percent-clipped=1.0 2023-04-28 08:31:02,816 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-28 08:31:04,890 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=47299.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:31:06,905 INFO [train.py:904] (4/8) Epoch 5, batch 6700, loss[loss=0.3449, simple_loss=0.3763, pruned_loss=0.1568, over 11350.00 frames. ], tot_loss[loss=0.2559, simple_loss=0.3296, pruned_loss=0.09108, over 3111044.36 frames. ], batch size: 247, lr: 1.32e-02, grad_scale: 8.0 2023-04-28 08:31:09,261 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47302.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:31:16,566 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=47306.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:31:45,317 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7616, 3.5836, 3.0648, 1.7608, 2.6546, 2.1551, 3.2018, 3.3778], device='cuda:4'), covar=tensor([0.0301, 0.0442, 0.0557, 0.1633, 0.0774, 0.0924, 0.0653, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0127, 0.0154, 0.0141, 0.0132, 0.0124, 0.0140, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:32:25,987 INFO [train.py:904] (4/8) Epoch 5, batch 6750, loss[loss=0.272, simple_loss=0.3402, pruned_loss=0.1019, over 15339.00 frames. ], tot_loss[loss=0.2538, simple_loss=0.3275, pruned_loss=0.09008, over 3117474.01 frames. ], batch size: 190, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:32:45,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.669e+02 3.980e+02 4.802e+02 5.958e+02 9.394e+02, threshold=9.603e+02, percent-clipped=0.0 2023-04-28 08:33:09,115 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47379.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:33:40,654 INFO [train.py:904] (4/8) Epoch 5, batch 6800, loss[loss=0.2643, simple_loss=0.3461, pruned_loss=0.09122, over 16863.00 frames. ], tot_loss[loss=0.2542, simple_loss=0.3279, pruned_loss=0.09027, over 3099059.95 frames. ], batch size: 109, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:33:51,947 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5133, 5.4607, 5.3508, 5.2801, 4.7893, 5.4230, 5.2641, 5.0906], device='cuda:4'), covar=tensor([0.0431, 0.0178, 0.0165, 0.0133, 0.0915, 0.0200, 0.0142, 0.0409], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0183, 0.0206, 0.0179, 0.0236, 0.0207, 0.0151, 0.0230], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:34:13,719 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6643, 3.3926, 2.9356, 1.8045, 2.6753, 2.1454, 3.0555, 3.1767], device='cuda:4'), covar=tensor([0.0264, 0.0518, 0.0584, 0.1631, 0.0753, 0.0929, 0.0710, 0.0709], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0125, 0.0152, 0.0139, 0.0131, 0.0123, 0.0138, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:34:31,911 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-28 08:34:40,981 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47440.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:34:57,631 INFO [train.py:904] (4/8) Epoch 5, batch 6850, loss[loss=0.2338, simple_loss=0.3355, pruned_loss=0.0661, over 16424.00 frames. ], tot_loss[loss=0.2551, simple_loss=0.329, pruned_loss=0.09059, over 3087602.86 frames. ], batch size: 75, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:35:17,575 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.600e+02 3.467e+02 4.392e+02 5.504e+02 1.073e+03, threshold=8.784e+02, percent-clipped=2.0 2023-04-28 08:35:20,990 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=47466.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:36:08,010 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1578, 3.3510, 1.6684, 3.4225, 2.3525, 3.4046, 1.8062, 2.5559], device='cuda:4'), covar=tensor([0.0160, 0.0221, 0.1446, 0.0054, 0.0700, 0.0416, 0.1365, 0.0567], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0146, 0.0177, 0.0079, 0.0159, 0.0180, 0.0187, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:36:12,143 INFO [train.py:904] (4/8) Epoch 5, batch 6900, loss[loss=0.2662, simple_loss=0.3499, pruned_loss=0.09127, over 16813.00 frames. ], tot_loss[loss=0.2569, simple_loss=0.3321, pruned_loss=0.09091, over 3090763.10 frames. ], batch size: 83, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:36:33,892 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=47514.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:36:38,723 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7161, 4.7009, 4.5614, 3.8998, 4.6191, 1.8292, 4.3327, 4.5146], device='cuda:4'), covar=tensor([0.0051, 0.0048, 0.0079, 0.0281, 0.0046, 0.1607, 0.0081, 0.0113], device='cuda:4'), in_proj_covar=tensor([0.0089, 0.0077, 0.0118, 0.0123, 0.0088, 0.0140, 0.0104, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:37:16,224 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9607, 3.6701, 3.8058, 2.4802, 3.4820, 3.7074, 3.6914, 2.1474], device='cuda:4'), covar=tensor([0.0329, 0.0020, 0.0028, 0.0223, 0.0036, 0.0061, 0.0024, 0.0266], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0053, 0.0058, 0.0113, 0.0058, 0.0068, 0.0062, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:37:30,502 INFO [train.py:904] (4/8) Epoch 5, batch 6950, loss[loss=0.2377, simple_loss=0.3116, pruned_loss=0.08191, over 16753.00 frames. ], tot_loss[loss=0.259, simple_loss=0.3333, pruned_loss=0.09239, over 3096164.39 frames. ], batch size: 124, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:37:50,868 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.715e+02 3.906e+02 5.173e+02 7.099e+02 1.028e+03, threshold=1.035e+03, percent-clipped=6.0 2023-04-28 08:38:06,227 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2156, 3.8482, 3.9074, 2.7337, 3.5815, 3.7850, 3.8782, 2.2610], device='cuda:4'), covar=tensor([0.0344, 0.0019, 0.0028, 0.0215, 0.0042, 0.0086, 0.0021, 0.0278], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0052, 0.0058, 0.0114, 0.0058, 0.0068, 0.0062, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:38:17,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5291, 4.4151, 5.0617, 4.9725, 5.0043, 4.5230, 4.6618, 4.3961], device='cuda:4'), covar=tensor([0.0228, 0.0344, 0.0298, 0.0385, 0.0394, 0.0294, 0.0735, 0.0337], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0227, 0.0235, 0.0235, 0.0284, 0.0248, 0.0352, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 08:38:23,320 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7614, 4.6120, 5.2967, 5.2208, 5.2634, 4.7589, 4.8450, 4.5100], device='cuda:4'), covar=tensor([0.0222, 0.0355, 0.0291, 0.0352, 0.0342, 0.0250, 0.0759, 0.0351], device='cuda:4'), in_proj_covar=tensor([0.0234, 0.0225, 0.0234, 0.0233, 0.0282, 0.0247, 0.0349, 0.0208], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 08:38:42,869 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=47597.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:38:49,094 INFO [train.py:904] (4/8) Epoch 5, batch 7000, loss[loss=0.2537, simple_loss=0.3374, pruned_loss=0.08506, over 16682.00 frames. ], tot_loss[loss=0.2583, simple_loss=0.3333, pruned_loss=0.09164, over 3089590.86 frames. ], batch size: 134, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:39:41,240 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.65 vs. limit=5.0 2023-04-28 08:40:02,365 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8451, 5.3058, 5.4584, 5.2732, 5.4527, 5.9771, 5.4503, 5.3080], device='cuda:4'), covar=tensor([0.0791, 0.1535, 0.1468, 0.1459, 0.2075, 0.0839, 0.1116, 0.2106], device='cuda:4'), in_proj_covar=tensor([0.0275, 0.0376, 0.0391, 0.0336, 0.0439, 0.0404, 0.0311, 0.0457], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:40:06,953 INFO [train.py:904] (4/8) Epoch 5, batch 7050, loss[loss=0.2985, simple_loss=0.3435, pruned_loss=0.1268, over 11372.00 frames. ], tot_loss[loss=0.2595, simple_loss=0.3347, pruned_loss=0.09216, over 3090258.89 frames. ], batch size: 247, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:40:26,882 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.583e+02 3.724e+02 4.498e+02 5.703e+02 1.501e+03, threshold=8.996e+02, percent-clipped=4.0 2023-04-28 08:41:26,196 INFO [train.py:904] (4/8) Epoch 5, batch 7100, loss[loss=0.2201, simple_loss=0.3081, pruned_loss=0.06609, over 16391.00 frames. ], tot_loss[loss=0.257, simple_loss=0.3323, pruned_loss=0.09088, over 3099076.75 frames. ], batch size: 146, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:42:20,349 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=47735.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:42:24,162 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47738.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:42:44,668 INFO [train.py:904] (4/8) Epoch 5, batch 7150, loss[loss=0.225, simple_loss=0.302, pruned_loss=0.07401, over 16435.00 frames. ], tot_loss[loss=0.2552, simple_loss=0.3296, pruned_loss=0.09042, over 3094554.15 frames. ], batch size: 75, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:43:03,968 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.186e+02 3.709e+02 4.695e+02 5.943e+02 1.245e+03, threshold=9.389e+02, percent-clipped=5.0 2023-04-28 08:43:39,970 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47788.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:43:57,735 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47799.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:44:00,011 INFO [train.py:904] (4/8) Epoch 5, batch 7200, loss[loss=0.2362, simple_loss=0.3146, pruned_loss=0.07893, over 16732.00 frames. ], tot_loss[loss=0.2516, simple_loss=0.327, pruned_loss=0.0881, over 3098999.78 frames. ], batch size: 134, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:44:07,562 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 08:44:13,965 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47809.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 08:44:53,285 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8964, 2.4222, 2.4111, 4.6574, 2.0918, 3.2974, 2.5619, 2.6435], device='cuda:4'), covar=tensor([0.0560, 0.1974, 0.1139, 0.0260, 0.2987, 0.1040, 0.1756, 0.2322], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0318, 0.0260, 0.0306, 0.0372, 0.0308, 0.0285, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:45:22,076 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47849.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:45:24,156 INFO [train.py:904] (4/8) Epoch 5, batch 7250, loss[loss=0.2114, simple_loss=0.2878, pruned_loss=0.06748, over 16687.00 frames. ], tot_loss[loss=0.2489, simple_loss=0.3247, pruned_loss=0.0866, over 3110211.63 frames. ], batch size: 134, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:45:43,493 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.043e+02 3.368e+02 4.070e+02 5.805e+02 1.027e+03, threshold=8.141e+02, percent-clipped=1.0 2023-04-28 08:45:53,864 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47870.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 08:46:34,234 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=47897.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:46:40,682 INFO [train.py:904] (4/8) Epoch 5, batch 7300, loss[loss=0.2421, simple_loss=0.3236, pruned_loss=0.08029, over 16666.00 frames. ], tot_loss[loss=0.2485, simple_loss=0.3244, pruned_loss=0.08631, over 3115965.60 frames. ], batch size: 62, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:47:47,330 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2036, 1.8487, 1.9589, 3.7237, 1.7441, 2.7083, 2.0435, 1.9709], device='cuda:4'), covar=tensor([0.0645, 0.2189, 0.1255, 0.0343, 0.3112, 0.1089, 0.2018, 0.2544], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0315, 0.0258, 0.0304, 0.0371, 0.0306, 0.0284, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:47:49,459 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=47945.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:47:58,198 INFO [train.py:904] (4/8) Epoch 5, batch 7350, loss[loss=0.2379, simple_loss=0.3313, pruned_loss=0.07226, over 16493.00 frames. ], tot_loss[loss=0.2497, simple_loss=0.3252, pruned_loss=0.08709, over 3098870.64 frames. ], batch size: 75, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:48:17,667 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.639e+02 4.067e+02 5.121e+02 6.309e+02 1.579e+03, threshold=1.024e+03, percent-clipped=9.0 2023-04-28 08:48:23,054 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-28 08:48:50,243 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.50 vs. limit=5.0 2023-04-28 08:49:18,342 INFO [train.py:904] (4/8) Epoch 5, batch 7400, loss[loss=0.2333, simple_loss=0.3128, pruned_loss=0.07694, over 17019.00 frames. ], tot_loss[loss=0.2498, simple_loss=0.3255, pruned_loss=0.08702, over 3114071.73 frames. ], batch size: 55, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:50:13,353 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48035.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:50:38,165 INFO [train.py:904] (4/8) Epoch 5, batch 7450, loss[loss=0.2334, simple_loss=0.3116, pruned_loss=0.07767, over 16601.00 frames. ], tot_loss[loss=0.2515, simple_loss=0.3266, pruned_loss=0.0882, over 3112867.09 frames. ], batch size: 62, lr: 1.31e-02, grad_scale: 8.0 2023-04-28 08:50:53,385 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6045, 2.6197, 2.1915, 3.8679, 2.9142, 3.8210, 1.4089, 2.6977], device='cuda:4'), covar=tensor([0.1484, 0.0670, 0.1281, 0.0128, 0.0225, 0.0396, 0.1531, 0.0819], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0145, 0.0169, 0.0086, 0.0181, 0.0179, 0.0161, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 08:51:00,810 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.428e+02 4.070e+02 4.599e+02 6.099e+02 1.111e+03, threshold=9.198e+02, percent-clipped=2.0 2023-04-28 08:51:14,586 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48072.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:51:27,860 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9108, 2.3633, 2.4034, 3.2695, 2.4436, 3.2776, 1.7316, 2.7642], device='cuda:4'), covar=tensor([0.1194, 0.0500, 0.0908, 0.0098, 0.0181, 0.0408, 0.1273, 0.0661], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0144, 0.0168, 0.0085, 0.0180, 0.0178, 0.0160, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 08:51:31,925 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=48083.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:51:49,935 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48094.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:52:00,138 INFO [train.py:904] (4/8) Epoch 5, batch 7500, loss[loss=0.2186, simple_loss=0.2892, pruned_loss=0.07403, over 16999.00 frames. ], tot_loss[loss=0.2516, simple_loss=0.3271, pruned_loss=0.08809, over 3105559.86 frames. ], batch size: 55, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 08:52:25,767 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-28 08:52:47,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8332, 5.3935, 5.4740, 5.3983, 5.4521, 6.0007, 5.4966, 5.2818], device='cuda:4'), covar=tensor([0.0667, 0.1493, 0.1261, 0.1453, 0.2097, 0.0724, 0.1054, 0.1787], device='cuda:4'), in_proj_covar=tensor([0.0275, 0.0376, 0.0383, 0.0333, 0.0436, 0.0400, 0.0312, 0.0452], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:52:50,884 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48133.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:53:08,579 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48144.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:53:13,877 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6673, 1.5031, 2.0148, 2.5403, 2.5298, 2.8885, 1.5016, 2.6387], device='cuda:4'), covar=tensor([0.0067, 0.0279, 0.0161, 0.0109, 0.0111, 0.0065, 0.0244, 0.0048], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0143, 0.0125, 0.0122, 0.0126, 0.0091, 0.0139, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 08:53:18,029 INFO [train.py:904] (4/8) Epoch 5, batch 7550, loss[loss=0.2082, simple_loss=0.2939, pruned_loss=0.06128, over 16675.00 frames. ], tot_loss[loss=0.2522, simple_loss=0.3268, pruned_loss=0.08877, over 3103236.35 frames. ], batch size: 89, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 08:53:38,522 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.261e+02 3.568e+02 4.415e+02 5.679e+02 1.356e+03, threshold=8.829e+02, percent-clipped=4.0 2023-04-28 08:53:40,653 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48165.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 08:54:34,586 INFO [train.py:904] (4/8) Epoch 5, batch 7600, loss[loss=0.2661, simple_loss=0.3294, pruned_loss=0.1014, over 15318.00 frames. ], tot_loss[loss=0.2515, simple_loss=0.3258, pruned_loss=0.08865, over 3100513.36 frames. ], batch size: 190, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 08:54:52,542 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48212.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 08:55:45,624 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4982, 3.8576, 4.0880, 2.9593, 3.5690, 3.9921, 3.8864, 2.2761], device='cuda:4'), covar=tensor([0.0283, 0.0020, 0.0023, 0.0183, 0.0043, 0.0051, 0.0032, 0.0267], device='cuda:4'), in_proj_covar=tensor([0.0113, 0.0052, 0.0057, 0.0112, 0.0059, 0.0067, 0.0062, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 08:55:51,264 INFO [train.py:904] (4/8) Epoch 5, batch 7650, loss[loss=0.2599, simple_loss=0.3367, pruned_loss=0.09161, over 16510.00 frames. ], tot_loss[loss=0.2528, simple_loss=0.3268, pruned_loss=0.08933, over 3111834.38 frames. ], batch size: 68, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 08:56:12,527 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.451e+02 3.891e+02 4.972e+02 6.319e+02 1.654e+03, threshold=9.944e+02, percent-clipped=8.0 2023-04-28 08:56:26,033 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48273.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 08:57:08,646 INFO [train.py:904] (4/8) Epoch 5, batch 7700, loss[loss=0.2452, simple_loss=0.3203, pruned_loss=0.08503, over 16428.00 frames. ], tot_loss[loss=0.2537, simple_loss=0.3275, pruned_loss=0.08992, over 3105322.47 frames. ], batch size: 146, lr: 1.30e-02, grad_scale: 4.0 2023-04-28 08:57:44,821 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48324.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:58:06,658 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5558, 5.3503, 5.3358, 5.2178, 4.8772, 5.3087, 5.1926, 4.9463], device='cuda:4'), covar=tensor([0.0440, 0.0291, 0.0171, 0.0151, 0.0744, 0.0325, 0.0213, 0.0546], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0184, 0.0206, 0.0179, 0.0237, 0.0210, 0.0152, 0.0233], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 08:58:22,053 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-28 08:58:26,138 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-28 08:58:26,511 INFO [train.py:904] (4/8) Epoch 5, batch 7750, loss[loss=0.2408, simple_loss=0.3177, pruned_loss=0.0819, over 16623.00 frames. ], tot_loss[loss=0.2536, simple_loss=0.3275, pruned_loss=0.08981, over 3096247.06 frames. ], batch size: 62, lr: 1.30e-02, grad_scale: 4.0 2023-04-28 08:58:27,120 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7355, 3.5238, 2.8138, 1.5846, 2.4361, 1.8964, 3.1731, 3.4028], device='cuda:4'), covar=tensor([0.0345, 0.0448, 0.0751, 0.2049, 0.1063, 0.1202, 0.0747, 0.0738], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0127, 0.0154, 0.0140, 0.0133, 0.0125, 0.0140, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:58:40,211 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-28 08:58:47,829 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.847e+02 4.062e+02 4.544e+02 6.017e+02 9.038e+02, threshold=9.088e+02, percent-clipped=0.0 2023-04-28 08:58:58,096 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7317, 4.1954, 2.0089, 4.3602, 2.6425, 4.2243, 2.0518, 2.9613], device='cuda:4'), covar=tensor([0.0127, 0.0203, 0.1638, 0.0037, 0.0736, 0.0421, 0.1508, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0149, 0.0178, 0.0080, 0.0161, 0.0180, 0.0187, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 08:59:20,065 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48385.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:59:34,781 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48394.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:59:39,521 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48397.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 08:59:45,047 INFO [train.py:904] (4/8) Epoch 5, batch 7800, loss[loss=0.2383, simple_loss=0.3168, pruned_loss=0.07989, over 16447.00 frames. ], tot_loss[loss=0.2542, simple_loss=0.3278, pruned_loss=0.09026, over 3095779.77 frames. ], batch size: 68, lr: 1.30e-02, grad_scale: 4.0 2023-04-28 09:00:12,323 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8259, 3.1910, 2.4810, 4.6229, 3.6697, 4.2727, 1.5947, 2.9726], device='cuda:4'), covar=tensor([0.1380, 0.0556, 0.1211, 0.0105, 0.0369, 0.0307, 0.1461, 0.0776], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0142, 0.0167, 0.0085, 0.0182, 0.0177, 0.0158, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 09:00:21,164 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48423.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:00:28,781 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48428.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:00:49,711 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=48442.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:00:52,290 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48444.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:01:02,290 INFO [train.py:904] (4/8) Epoch 5, batch 7850, loss[loss=0.2573, simple_loss=0.3376, pruned_loss=0.08846, over 16255.00 frames. ], tot_loss[loss=0.2544, simple_loss=0.329, pruned_loss=0.08993, over 3098582.99 frames. ], batch size: 165, lr: 1.30e-02, grad_scale: 4.0 2023-04-28 09:01:13,149 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48458.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:01:23,602 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.309e+02 3.753e+02 4.508e+02 5.540e+02 1.129e+03, threshold=9.017e+02, percent-clipped=6.0 2023-04-28 09:01:23,968 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48465.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 09:01:25,436 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1613, 3.4704, 3.5805, 1.4722, 3.6521, 3.7010, 2.7994, 2.7603], device='cuda:4'), covar=tensor([0.0836, 0.0112, 0.0130, 0.1322, 0.0059, 0.0081, 0.0348, 0.0432], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0086, 0.0081, 0.0141, 0.0070, 0.0078, 0.0115, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 09:01:52,543 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48484.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:02:04,083 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=48492.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:02:16,621 INFO [train.py:904] (4/8) Epoch 5, batch 7900, loss[loss=0.2139, simple_loss=0.299, pruned_loss=0.06442, over 16602.00 frames. ], tot_loss[loss=0.2527, simple_loss=0.3272, pruned_loss=0.08915, over 3095233.94 frames. ], batch size: 57, lr: 1.30e-02, grad_scale: 4.0 2023-04-28 09:02:35,251 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=48513.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 09:03:36,605 INFO [train.py:904] (4/8) Epoch 5, batch 7950, loss[loss=0.2779, simple_loss=0.334, pruned_loss=0.1109, over 11652.00 frames. ], tot_loss[loss=0.2533, simple_loss=0.3273, pruned_loss=0.08967, over 3098814.43 frames. ], batch size: 246, lr: 1.30e-02, grad_scale: 4.0 2023-04-28 09:03:56,964 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.950e+02 3.717e+02 4.429e+02 5.526e+02 1.268e+03, threshold=8.857e+02, percent-clipped=1.0 2023-04-28 09:04:02,930 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48568.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 09:04:50,721 INFO [train.py:904] (4/8) Epoch 5, batch 8000, loss[loss=0.2645, simple_loss=0.3404, pruned_loss=0.09429, over 16847.00 frames. ], tot_loss[loss=0.2549, simple_loss=0.3285, pruned_loss=0.09061, over 3105317.89 frames. ], batch size: 102, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 09:04:59,627 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3833, 1.9191, 1.5480, 1.5713, 2.1187, 1.9863, 2.2918, 2.3101], device='cuda:4'), covar=tensor([0.0038, 0.0136, 0.0204, 0.0213, 0.0090, 0.0153, 0.0088, 0.0102], device='cuda:4'), in_proj_covar=tensor([0.0074, 0.0149, 0.0152, 0.0148, 0.0143, 0.0153, 0.0125, 0.0134], device='cuda:4'), out_proj_covar=tensor([9.0888e-05, 1.8033e-04, 1.8077e-04, 1.7663e-04, 1.7609e-04, 1.8580e-04, 1.4759e-04, 1.6326e-04], device='cuda:4') 2023-04-28 09:06:06,080 INFO [train.py:904] (4/8) Epoch 5, batch 8050, loss[loss=0.241, simple_loss=0.3154, pruned_loss=0.0833, over 16900.00 frames. ], tot_loss[loss=0.2543, simple_loss=0.328, pruned_loss=0.09026, over 3098437.72 frames. ], batch size: 109, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 09:06:26,963 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.628e+02 3.970e+02 4.648e+02 5.523e+02 1.617e+03, threshold=9.296e+02, percent-clipped=6.0 2023-04-28 09:06:49,713 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48680.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:07:13,924 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4604, 3.3230, 2.8114, 2.1985, 2.5769, 2.1696, 3.5035, 3.6326], device='cuda:4'), covar=tensor([0.2282, 0.0860, 0.1252, 0.1577, 0.1779, 0.1394, 0.0425, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0246, 0.0269, 0.0242, 0.0291, 0.0201, 0.0240, 0.0253], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:07:21,595 INFO [train.py:904] (4/8) Epoch 5, batch 8100, loss[loss=0.2411, simple_loss=0.3129, pruned_loss=0.0847, over 16450.00 frames. ], tot_loss[loss=0.2523, simple_loss=0.327, pruned_loss=0.08879, over 3098567.97 frames. ], batch size: 35, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 09:08:06,102 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48728.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:08:40,601 INFO [train.py:904] (4/8) Epoch 5, batch 8150, loss[loss=0.2674, simple_loss=0.3281, pruned_loss=0.1033, over 11796.00 frames. ], tot_loss[loss=0.2511, simple_loss=0.3254, pruned_loss=0.08837, over 3090529.53 frames. ], batch size: 247, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 09:08:44,612 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48753.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:09:01,945 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.147e+02 3.868e+02 4.692e+02 5.916e+02 1.218e+03, threshold=9.385e+02, percent-clipped=3.0 2023-04-28 09:09:20,846 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=48776.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:09:24,963 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48779.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:10:00,528 INFO [train.py:904] (4/8) Epoch 5, batch 8200, loss[loss=0.3018, simple_loss=0.351, pruned_loss=0.1262, over 11924.00 frames. ], tot_loss[loss=0.2497, simple_loss=0.3234, pruned_loss=0.08795, over 3074114.72 frames. ], batch size: 246, lr: 1.30e-02, grad_scale: 8.0 2023-04-28 09:10:32,519 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48821.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:10:50,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1410, 4.0929, 4.5302, 4.5098, 4.5070, 4.1683, 4.1702, 4.1128], device='cuda:4'), covar=tensor([0.0245, 0.0379, 0.0327, 0.0333, 0.0399, 0.0268, 0.0870, 0.0360], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0225, 0.0234, 0.0231, 0.0282, 0.0247, 0.0355, 0.0208], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 09:11:00,871 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48838.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:11:22,104 INFO [train.py:904] (4/8) Epoch 5, batch 8250, loss[loss=0.2402, simple_loss=0.3099, pruned_loss=0.08529, over 12191.00 frames. ], tot_loss[loss=0.2468, simple_loss=0.3227, pruned_loss=0.08547, over 3076920.03 frames. ], batch size: 247, lr: 1.29e-02, grad_scale: 8.0 2023-04-28 09:11:44,535 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.248e+02 3.760e+02 4.474e+02 5.646e+02 1.182e+03, threshold=8.947e+02, percent-clipped=2.0 2023-04-28 09:11:50,744 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48868.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 09:12:13,773 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48882.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:12:41,192 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48899.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:12:43,300 INFO [train.py:904] (4/8) Epoch 5, batch 8300, loss[loss=0.2054, simple_loss=0.2962, pruned_loss=0.05731, over 15246.00 frames. ], tot_loss[loss=0.242, simple_loss=0.3194, pruned_loss=0.08228, over 3047321.51 frames. ], batch size: 190, lr: 1.29e-02, grad_scale: 8.0 2023-04-28 09:12:54,716 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4985, 3.5919, 3.3365, 3.2451, 3.0938, 3.4441, 3.2555, 3.2590], device='cuda:4'), covar=tensor([0.0415, 0.0266, 0.0199, 0.0178, 0.0484, 0.0235, 0.0739, 0.0359], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0180, 0.0204, 0.0177, 0.0231, 0.0205, 0.0147, 0.0230], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:12:59,245 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6677, 2.6543, 2.4199, 3.6772, 2.7847, 3.7525, 1.4578, 2.8574], device='cuda:4'), covar=tensor([0.1446, 0.0523, 0.1043, 0.0079, 0.0174, 0.0321, 0.1489, 0.0693], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0139, 0.0165, 0.0084, 0.0178, 0.0174, 0.0159, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 09:13:08,659 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=48916.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 09:14:06,483 INFO [train.py:904] (4/8) Epoch 5, batch 8350, loss[loss=0.2182, simple_loss=0.3104, pruned_loss=0.06298, over 16700.00 frames. ], tot_loss[loss=0.238, simple_loss=0.3179, pruned_loss=0.07908, over 3065238.70 frames. ], batch size: 124, lr: 1.29e-02, grad_scale: 8.0 2023-04-28 09:14:30,517 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.953e+02 3.108e+02 3.875e+02 4.541e+02 1.583e+03, threshold=7.750e+02, percent-clipped=2.0 2023-04-28 09:14:54,817 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48980.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:14:57,266 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.63 vs. limit=2.0 2023-04-28 09:15:30,220 INFO [train.py:904] (4/8) Epoch 5, batch 8400, loss[loss=0.2016, simple_loss=0.2906, pruned_loss=0.05634, over 16530.00 frames. ], tot_loss[loss=0.2333, simple_loss=0.3142, pruned_loss=0.07626, over 3059480.31 frames. ], batch size: 62, lr: 1.29e-02, grad_scale: 8.0 2023-04-28 09:16:14,934 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=49028.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:16:28,533 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8438, 4.0967, 1.8796, 4.2909, 2.5235, 4.1703, 2.1482, 2.9361], device='cuda:4'), covar=tensor([0.0121, 0.0178, 0.1639, 0.0033, 0.0829, 0.0285, 0.1443, 0.0592], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0146, 0.0174, 0.0076, 0.0157, 0.0174, 0.0182, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 09:16:51,002 INFO [train.py:904] (4/8) Epoch 5, batch 8450, loss[loss=0.2184, simple_loss=0.2937, pruned_loss=0.0716, over 12333.00 frames. ], tot_loss[loss=0.2299, simple_loss=0.3117, pruned_loss=0.07406, over 3054811.42 frames. ], batch size: 247, lr: 1.29e-02, grad_scale: 8.0 2023-04-28 09:16:55,338 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49053.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:17:13,777 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.140e+02 2.885e+02 3.477e+02 4.145e+02 8.809e+02, threshold=6.955e+02, percent-clipped=1.0 2023-04-28 09:17:38,435 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49079.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:17:41,871 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6757, 3.4796, 2.8275, 2.2336, 2.4266, 2.1463, 3.5950, 3.5738], device='cuda:4'), covar=tensor([0.1935, 0.0640, 0.1208, 0.1687, 0.1907, 0.1500, 0.0358, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0234, 0.0256, 0.0234, 0.0262, 0.0194, 0.0229, 0.0236], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:18:13,180 INFO [train.py:904] (4/8) Epoch 5, batch 8500, loss[loss=0.1867, simple_loss=0.2666, pruned_loss=0.05341, over 12078.00 frames. ], tot_loss[loss=0.2243, simple_loss=0.307, pruned_loss=0.07086, over 3046699.77 frames. ], batch size: 246, lr: 1.29e-02, grad_scale: 8.0 2023-04-28 09:18:13,691 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=49101.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:18:57,905 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=49127.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:19:30,020 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7737, 3.6542, 3.8192, 3.9715, 3.9932, 3.5793, 4.0123, 3.9926], device='cuda:4'), covar=tensor([0.0852, 0.0651, 0.1015, 0.0494, 0.0479, 0.1620, 0.0508, 0.0473], device='cuda:4'), in_proj_covar=tensor([0.0356, 0.0447, 0.0555, 0.0468, 0.0350, 0.0344, 0.0367, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:19:38,542 INFO [train.py:904] (4/8) Epoch 5, batch 8550, loss[loss=0.2047, simple_loss=0.2907, pruned_loss=0.05937, over 16369.00 frames. ], tot_loss[loss=0.2213, simple_loss=0.3044, pruned_loss=0.0691, over 3058015.03 frames. ], batch size: 68, lr: 1.29e-02, grad_scale: 8.0 2023-04-28 09:20:03,570 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1195, 1.8286, 2.0011, 3.5510, 1.7388, 2.5963, 2.1070, 1.9170], device='cuda:4'), covar=tensor([0.0620, 0.2603, 0.1332, 0.0335, 0.3478, 0.1226, 0.2106, 0.2861], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0310, 0.0256, 0.0296, 0.0365, 0.0303, 0.0283, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:20:04,105 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.964e+02 3.233e+02 3.773e+02 4.691e+02 1.040e+03, threshold=7.547e+02, percent-clipped=6.0 2023-04-28 09:20:24,382 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6601, 3.6152, 4.0406, 4.0194, 4.0330, 3.7187, 3.7958, 3.7541], device='cuda:4'), covar=tensor([0.0257, 0.0387, 0.0356, 0.0396, 0.0386, 0.0312, 0.0698, 0.0308], device='cuda:4'), in_proj_covar=tensor([0.0221, 0.0213, 0.0220, 0.0216, 0.0265, 0.0233, 0.0325, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-28 09:20:29,025 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49177.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:21:04,488 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49194.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:21:17,317 INFO [train.py:904] (4/8) Epoch 5, batch 8600, loss[loss=0.2294, simple_loss=0.3161, pruned_loss=0.0714, over 16148.00 frames. ], tot_loss[loss=0.2205, simple_loss=0.3047, pruned_loss=0.06814, over 3063468.82 frames. ], batch size: 165, lr: 1.29e-02, grad_scale: 8.0 2023-04-28 09:21:25,292 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8716, 2.4255, 2.2935, 3.1910, 2.4136, 3.3617, 1.6614, 2.7197], device='cuda:4'), covar=tensor([0.1519, 0.0542, 0.1050, 0.0117, 0.0167, 0.0400, 0.1558, 0.0742], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0140, 0.0164, 0.0085, 0.0171, 0.0175, 0.0160, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 09:21:43,583 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7981, 3.5910, 3.5083, 3.9443, 4.0391, 3.7007, 4.0793, 4.0577], device='cuda:4'), covar=tensor([0.0976, 0.1000, 0.1999, 0.0861, 0.0773, 0.1273, 0.0638, 0.0696], device='cuda:4'), in_proj_covar=tensor([0.0353, 0.0444, 0.0553, 0.0465, 0.0346, 0.0342, 0.0363, 0.0384], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:22:06,324 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4077, 4.3823, 4.1947, 3.7254, 4.1837, 1.4249, 3.9724, 4.0513], device='cuda:4'), covar=tensor([0.0055, 0.0044, 0.0093, 0.0207, 0.0058, 0.1958, 0.0084, 0.0126], device='cuda:4'), in_proj_covar=tensor([0.0090, 0.0077, 0.0121, 0.0118, 0.0090, 0.0144, 0.0105, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:22:28,813 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=49236.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:22:58,656 INFO [train.py:904] (4/8) Epoch 5, batch 8650, loss[loss=0.1843, simple_loss=0.2833, pruned_loss=0.04268, over 16887.00 frames. ], tot_loss[loss=0.2164, simple_loss=0.3015, pruned_loss=0.06569, over 3062357.49 frames. ], batch size: 102, lr: 1.29e-02, grad_scale: 4.0 2023-04-28 09:23:33,946 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.003e+02 3.105e+02 3.609e+02 4.531e+02 9.141e+02, threshold=7.218e+02, percent-clipped=3.0 2023-04-28 09:24:39,678 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=49297.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:24:40,003 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-28 09:24:47,014 INFO [train.py:904] (4/8) Epoch 5, batch 8700, loss[loss=0.188, simple_loss=0.2844, pruned_loss=0.04578, over 16890.00 frames. ], tot_loss[loss=0.2128, simple_loss=0.298, pruned_loss=0.0638, over 3054024.99 frames. ], batch size: 102, lr: 1.29e-02, grad_scale: 4.0 2023-04-28 09:26:24,150 INFO [train.py:904] (4/8) Epoch 5, batch 8750, loss[loss=0.2188, simple_loss=0.3109, pruned_loss=0.06335, over 16926.00 frames. ], tot_loss[loss=0.2113, simple_loss=0.2971, pruned_loss=0.06271, over 3067215.06 frames. ], batch size: 116, lr: 1.29e-02, grad_scale: 4.0 2023-04-28 09:26:51,530 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6330, 4.1876, 4.2469, 3.0550, 4.0556, 4.3546, 3.9415, 2.4192], device='cuda:4'), covar=tensor([0.0289, 0.0011, 0.0015, 0.0207, 0.0026, 0.0015, 0.0024, 0.0271], device='cuda:4'), in_proj_covar=tensor([0.0112, 0.0052, 0.0057, 0.0111, 0.0058, 0.0063, 0.0061, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 09:27:05,486 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.641e+02 2.935e+02 3.673e+02 5.022e+02 7.467e+02, threshold=7.345e+02, percent-clipped=1.0 2023-04-28 09:28:17,602 INFO [train.py:904] (4/8) Epoch 5, batch 8800, loss[loss=0.2376, simple_loss=0.3298, pruned_loss=0.07268, over 16905.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2958, pruned_loss=0.06197, over 3049997.03 frames. ], batch size: 116, lr: 1.29e-02, grad_scale: 4.0 2023-04-28 09:28:38,026 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3683, 3.2754, 3.3857, 3.5236, 3.5406, 3.2270, 3.5184, 3.5475], device='cuda:4'), covar=tensor([0.0668, 0.0656, 0.1087, 0.0549, 0.0482, 0.1462, 0.0597, 0.0474], device='cuda:4'), in_proj_covar=tensor([0.0351, 0.0441, 0.0546, 0.0458, 0.0345, 0.0338, 0.0360, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:29:02,399 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=49422.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:29:24,411 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=49432.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 09:30:04,510 INFO [train.py:904] (4/8) Epoch 5, batch 8850, loss[loss=0.214, simple_loss=0.3123, pruned_loss=0.05789, over 16718.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2981, pruned_loss=0.06145, over 3047287.89 frames. ], batch size: 134, lr: 1.29e-02, grad_scale: 4.0 2023-04-28 09:30:20,575 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0275, 4.9432, 4.7959, 4.6787, 4.4351, 4.9336, 4.8075, 4.6397], device='cuda:4'), covar=tensor([0.0339, 0.0232, 0.0188, 0.0143, 0.0655, 0.0236, 0.0221, 0.0331], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0172, 0.0197, 0.0167, 0.0217, 0.0195, 0.0139, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:30:38,662 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.896e+02 3.290e+02 4.056e+02 5.025e+02 1.173e+03, threshold=8.111e+02, percent-clipped=7.0 2023-04-28 09:31:02,520 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49477.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:31:16,932 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=49483.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:31:17,131 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-28 09:31:23,976 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8782, 3.7056, 3.9056, 4.0748, 4.2020, 3.7268, 4.2126, 4.1790], device='cuda:4'), covar=tensor([0.0917, 0.0744, 0.1161, 0.0593, 0.0431, 0.0967, 0.0388, 0.0402], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0435, 0.0536, 0.0450, 0.0338, 0.0329, 0.0353, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:31:37,660 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=49493.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 09:31:39,661 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49494.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:31:54,235 INFO [train.py:904] (4/8) Epoch 5, batch 8900, loss[loss=0.2125, simple_loss=0.3027, pruned_loss=0.06116, over 16803.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.2984, pruned_loss=0.06054, over 3064673.03 frames. ], batch size: 124, lr: 1.29e-02, grad_scale: 4.0 2023-04-28 09:32:48,143 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=49525.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:32:57,637 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.71 vs. limit=2.0 2023-04-28 09:33:39,419 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=49542.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:33:56,805 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1939, 5.5687, 5.3082, 5.3447, 4.9598, 4.6684, 5.0801, 5.5990], device='cuda:4'), covar=tensor([0.0719, 0.0624, 0.0789, 0.0368, 0.0569, 0.0597, 0.0578, 0.0629], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0460, 0.0386, 0.0298, 0.0289, 0.0306, 0.0372, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:33:59,795 INFO [train.py:904] (4/8) Epoch 5, batch 8950, loss[loss=0.1888, simple_loss=0.2763, pruned_loss=0.05067, over 16957.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.299, pruned_loss=0.06185, over 3066294.61 frames. ], batch size: 109, lr: 1.29e-02, grad_scale: 4.0 2023-04-28 09:34:35,629 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.980e+02 2.796e+02 3.504e+02 4.655e+02 7.441e+02, threshold=7.007e+02, percent-clipped=0.0 2023-04-28 09:35:33,173 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49592.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:35:51,807 INFO [train.py:904] (4/8) Epoch 5, batch 9000, loss[loss=0.1828, simple_loss=0.27, pruned_loss=0.04778, over 16549.00 frames. ], tot_loss[loss=0.2087, simple_loss=0.2962, pruned_loss=0.06055, over 3066370.62 frames. ], batch size: 68, lr: 1.28e-02, grad_scale: 4.0 2023-04-28 09:35:51,807 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 09:36:02,102 INFO [train.py:938] (4/8) Epoch 5, validation: loss=0.1735, simple_loss=0.2766, pruned_loss=0.0352, over 944034.00 frames. 2023-04-28 09:36:02,103 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 09:36:19,368 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2530, 1.2852, 1.7877, 2.1824, 2.2719, 2.2855, 1.5138, 2.3366], device='cuda:4'), covar=tensor([0.0082, 0.0269, 0.0166, 0.0139, 0.0112, 0.0104, 0.0244, 0.0052], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0142, 0.0128, 0.0124, 0.0129, 0.0090, 0.0141, 0.0077], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 09:37:36,047 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.97 vs. limit=5.0 2023-04-28 09:37:44,739 INFO [train.py:904] (4/8) Epoch 5, batch 9050, loss[loss=0.1881, simple_loss=0.2696, pruned_loss=0.05333, over 16843.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2963, pruned_loss=0.06069, over 3065531.04 frames. ], batch size: 96, lr: 1.28e-02, grad_scale: 4.0 2023-04-28 09:38:18,678 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.067e+02 3.141e+02 3.840e+02 5.023e+02 8.628e+02, threshold=7.679e+02, percent-clipped=5.0 2023-04-28 09:38:32,134 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 09:39:29,265 INFO [train.py:904] (4/8) Epoch 5, batch 9100, loss[loss=0.2159, simple_loss=0.3102, pruned_loss=0.06083, over 15371.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2959, pruned_loss=0.0609, over 3069918.76 frames. ], batch size: 191, lr: 1.28e-02, grad_scale: 4.0 2023-04-28 09:41:26,339 INFO [train.py:904] (4/8) Epoch 5, batch 9150, loss[loss=0.1926, simple_loss=0.2851, pruned_loss=0.05007, over 15278.00 frames. ], tot_loss[loss=0.2085, simple_loss=0.2962, pruned_loss=0.06038, over 3078407.35 frames. ], batch size: 191, lr: 1.28e-02, grad_scale: 4.0 2023-04-28 09:42:00,617 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=49766.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:42:01,427 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.296e+02 3.040e+02 3.844e+02 4.947e+02 8.229e+02, threshold=7.688e+02, percent-clipped=4.0 2023-04-28 09:42:25,693 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49778.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:42:47,422 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49788.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 09:43:09,765 INFO [train.py:904] (4/8) Epoch 5, batch 9200, loss[loss=0.185, simple_loss=0.2615, pruned_loss=0.05423, over 12307.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.291, pruned_loss=0.059, over 3065760.56 frames. ], batch size: 248, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:43:28,079 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-28 09:43:58,048 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=49827.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:44:46,685 INFO [train.py:904] (4/8) Epoch 5, batch 9250, loss[loss=0.2079, simple_loss=0.2959, pruned_loss=0.05994, over 16446.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2903, pruned_loss=0.05894, over 3053337.12 frames. ], batch size: 146, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:45:18,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.407e+02 3.305e+02 4.013e+02 4.841e+02 9.780e+02, threshold=8.027e+02, percent-clipped=1.0 2023-04-28 09:45:40,216 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1569, 1.3700, 1.7173, 2.0918, 2.1364, 2.1344, 1.4972, 2.0877], device='cuda:4'), covar=tensor([0.0103, 0.0243, 0.0144, 0.0164, 0.0132, 0.0123, 0.0247, 0.0057], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0141, 0.0127, 0.0123, 0.0127, 0.0089, 0.0138, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 09:46:21,457 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49892.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:46:40,054 INFO [train.py:904] (4/8) Epoch 5, batch 9300, loss[loss=0.2133, simple_loss=0.2776, pruned_loss=0.07451, over 12266.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.2888, pruned_loss=0.05833, over 3048785.75 frames. ], batch size: 250, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:48:05,855 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=49940.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:48:26,053 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 09:48:26,418 INFO [train.py:904] (4/8) Epoch 5, batch 9350, loss[loss=0.2064, simple_loss=0.2827, pruned_loss=0.06503, over 12589.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2885, pruned_loss=0.05801, over 3065566.09 frames. ], batch size: 247, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:49:00,529 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.703e+02 3.133e+02 3.801e+02 4.449e+02 7.989e+02, threshold=7.602e+02, percent-clipped=0.0 2023-04-28 09:50:11,379 INFO [train.py:904] (4/8) Epoch 5, batch 9400, loss[loss=0.2288, simple_loss=0.3206, pruned_loss=0.06855, over 16970.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.288, pruned_loss=0.05748, over 3056413.15 frames. ], batch size: 109, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:51:21,850 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5055, 3.7053, 1.4733, 3.9567, 2.3156, 3.8627, 1.9677, 2.7760], device='cuda:4'), covar=tensor([0.0181, 0.0261, 0.1928, 0.0054, 0.0936, 0.0335, 0.1575, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0143, 0.0173, 0.0075, 0.0154, 0.0168, 0.0180, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 09:51:51,664 INFO [train.py:904] (4/8) Epoch 5, batch 9450, loss[loss=0.2153, simple_loss=0.2946, pruned_loss=0.06807, over 12516.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2906, pruned_loss=0.05831, over 3047175.72 frames. ], batch size: 247, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:52:21,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.776e+02 2.955e+02 3.683e+02 5.023e+02 1.227e+03, threshold=7.366e+02, percent-clipped=5.0 2023-04-28 09:52:41,536 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.66 vs. limit=5.0 2023-04-28 09:52:41,604 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 09:52:47,403 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50078.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:53:08,967 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50088.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 09:53:33,284 INFO [train.py:904] (4/8) Epoch 5, batch 9500, loss[loss=0.1925, simple_loss=0.2867, pruned_loss=0.04917, over 16162.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2896, pruned_loss=0.05781, over 3041846.85 frames. ], batch size: 165, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:54:02,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5139, 2.5417, 2.2511, 2.2266, 3.0871, 2.7962, 3.5200, 3.2503], device='cuda:4'), covar=tensor([0.0021, 0.0192, 0.0205, 0.0222, 0.0102, 0.0168, 0.0061, 0.0089], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0151, 0.0150, 0.0148, 0.0144, 0.0152, 0.0118, 0.0130], device='cuda:4'), out_proj_covar=tensor([8.5363e-05, 1.8219e-04, 1.7688e-04, 1.7521e-04, 1.7530e-04, 1.8261e-04, 1.3517e-04, 1.5624e-04], device='cuda:4') 2023-04-28 09:54:18,247 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=50122.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:54:25,574 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=50126.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:54:29,352 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50128.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:54:44,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=50136.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 09:55:18,505 INFO [train.py:904] (4/8) Epoch 5, batch 9550, loss[loss=0.2266, simple_loss=0.3142, pruned_loss=0.06949, over 16317.00 frames. ], tot_loss[loss=0.2032, simple_loss=0.2902, pruned_loss=0.05811, over 3064919.75 frames. ], batch size: 146, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:55:53,349 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.669e+02 2.772e+02 3.459e+02 4.281e+02 6.687e+02, threshold=6.919e+02, percent-clipped=0.0 2023-04-28 09:56:00,193 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8381, 3.6694, 3.9373, 4.0614, 4.1914, 3.7208, 4.1642, 4.1711], device='cuda:4'), covar=tensor([0.0960, 0.0797, 0.1005, 0.0537, 0.0426, 0.1047, 0.0472, 0.0425], device='cuda:4'), in_proj_covar=tensor([0.0348, 0.0434, 0.0541, 0.0449, 0.0336, 0.0332, 0.0352, 0.0371], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 09:56:16,647 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.50 vs. limit=5.0 2023-04-28 09:56:38,111 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1123, 3.3373, 3.3310, 2.3231, 3.1571, 3.2936, 3.2719, 1.8605], device='cuda:4'), covar=tensor([0.0296, 0.0021, 0.0030, 0.0232, 0.0043, 0.0040, 0.0034, 0.0309], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0053, 0.0058, 0.0114, 0.0058, 0.0064, 0.0062, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 09:56:38,164 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50189.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 09:56:52,851 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50197.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 09:56:59,037 INFO [train.py:904] (4/8) Epoch 5, batch 9600, loss[loss=0.2254, simple_loss=0.3163, pruned_loss=0.06726, over 15199.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2921, pruned_loss=0.05941, over 3064023.52 frames. ], batch size: 190, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:58:47,442 INFO [train.py:904] (4/8) Epoch 5, batch 9650, loss[loss=0.2072, simple_loss=0.2862, pruned_loss=0.06406, over 12491.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2932, pruned_loss=0.05946, over 3042949.81 frames. ], batch size: 248, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 09:59:05,792 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50258.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 09:59:27,461 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.258e+02 2.991e+02 3.657e+02 4.626e+02 9.582e+02, threshold=7.315e+02, percent-clipped=7.0 2023-04-28 10:00:35,786 INFO [train.py:904] (4/8) Epoch 5, batch 9700, loss[loss=0.1989, simple_loss=0.2852, pruned_loss=0.0563, over 15252.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2914, pruned_loss=0.05862, over 3045943.35 frames. ], batch size: 190, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 10:02:05,563 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4813, 3.5141, 3.2863, 3.1805, 3.0957, 3.4020, 3.2813, 3.2034], device='cuda:4'), covar=tensor([0.0441, 0.0336, 0.0195, 0.0171, 0.0566, 0.0263, 0.0662, 0.0369], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0177, 0.0200, 0.0170, 0.0221, 0.0203, 0.0140, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:02:18,556 INFO [train.py:904] (4/8) Epoch 5, batch 9750, loss[loss=0.2103, simple_loss=0.2828, pruned_loss=0.0689, over 12506.00 frames. ], tot_loss[loss=0.2035, simple_loss=0.2897, pruned_loss=0.05863, over 3033159.56 frames. ], batch size: 250, lr: 1.28e-02, grad_scale: 8.0 2023-04-28 10:02:35,257 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.37 vs. limit=5.0 2023-04-28 10:02:37,659 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6203, 4.9471, 4.7101, 4.6559, 4.2967, 4.2762, 4.4412, 4.9551], device='cuda:4'), covar=tensor([0.0697, 0.0689, 0.0739, 0.0446, 0.0684, 0.0925, 0.0716, 0.0691], device='cuda:4'), in_proj_covar=tensor([0.0352, 0.0467, 0.0388, 0.0304, 0.0300, 0.0310, 0.0380, 0.0338], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:02:50,575 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.871e+02 3.120e+02 3.818e+02 4.621e+02 8.897e+02, threshold=7.636e+02, percent-clipped=1.0 2023-04-28 10:03:29,108 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 10:03:35,870 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-28 10:03:56,548 INFO [train.py:904] (4/8) Epoch 5, batch 9800, loss[loss=0.2072, simple_loss=0.302, pruned_loss=0.05621, over 16841.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2898, pruned_loss=0.05745, over 3042312.93 frames. ], batch size: 124, lr: 1.27e-02, grad_scale: 8.0 2023-04-28 10:04:36,838 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50422.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:05:22,838 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 10:05:41,979 INFO [train.py:904] (4/8) Epoch 5, batch 9850, loss[loss=0.1938, simple_loss=0.2931, pruned_loss=0.04724, over 16663.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2909, pruned_loss=0.05699, over 3049012.23 frames. ], batch size: 89, lr: 1.27e-02, grad_scale: 8.0 2023-04-28 10:05:48,163 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.62 vs. limit=2.0 2023-04-28 10:06:14,680 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.848e+02 2.952e+02 3.639e+02 4.340e+02 9.232e+02, threshold=7.278e+02, percent-clipped=1.0 2023-04-28 10:06:21,890 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=50470.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:06:52,945 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=50484.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:07:31,481 INFO [train.py:904] (4/8) Epoch 5, batch 9900, loss[loss=0.2207, simple_loss=0.3124, pruned_loss=0.06447, over 16711.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.292, pruned_loss=0.05773, over 3034298.51 frames. ], batch size: 134, lr: 1.27e-02, grad_scale: 8.0 2023-04-28 10:08:04,995 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6122, 4.5594, 4.3976, 3.9995, 4.2872, 1.6266, 4.1093, 4.2436], device='cuda:4'), covar=tensor([0.0050, 0.0039, 0.0074, 0.0165, 0.0059, 0.1738, 0.0077, 0.0128], device='cuda:4'), in_proj_covar=tensor([0.0086, 0.0074, 0.0115, 0.0107, 0.0086, 0.0141, 0.0099, 0.0108], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:08:56,422 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0437, 4.4277, 2.0772, 4.6032, 2.9237, 4.4167, 2.1052, 2.9189], device='cuda:4'), covar=tensor([0.0107, 0.0089, 0.1280, 0.0020, 0.0615, 0.0216, 0.1299, 0.0575], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0144, 0.0172, 0.0074, 0.0155, 0.0166, 0.0181, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 10:09:17,068 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50546.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:09:27,670 INFO [train.py:904] (4/8) Epoch 5, batch 9950, loss[loss=0.1935, simple_loss=0.2852, pruned_loss=0.05088, over 16724.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2945, pruned_loss=0.05808, over 3061875.18 frames. ], batch size: 83, lr: 1.27e-02, grad_scale: 8.0 2023-04-28 10:09:33,669 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=50553.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 10:09:54,744 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4472, 3.9438, 3.2847, 5.5765, 5.1734, 5.1284, 2.3990, 3.7419], device='cuda:4'), covar=tensor([0.1098, 0.0401, 0.0853, 0.0043, 0.0135, 0.0184, 0.1034, 0.0545], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0140, 0.0164, 0.0082, 0.0148, 0.0171, 0.0160, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 10:10:04,528 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.956e+02 3.001e+02 3.652e+02 4.906e+02 2.314e+03, threshold=7.303e+02, percent-clipped=6.0 2023-04-28 10:10:48,512 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 10:11:29,335 INFO [train.py:904] (4/8) Epoch 5, batch 10000, loss[loss=0.1855, simple_loss=0.2799, pruned_loss=0.0456, over 16565.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2915, pruned_loss=0.05672, over 3070588.12 frames. ], batch size: 62, lr: 1.27e-02, grad_scale: 8.0 2023-04-28 10:11:41,396 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4110, 3.5594, 1.7064, 3.7274, 2.3402, 3.6654, 1.8949, 2.5456], device='cuda:4'), covar=tensor([0.0139, 0.0203, 0.1630, 0.0055, 0.0885, 0.0342, 0.1488, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0145, 0.0174, 0.0074, 0.0157, 0.0168, 0.0182, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 10:11:43,811 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50607.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:13:08,522 INFO [train.py:904] (4/8) Epoch 5, batch 10050, loss[loss=0.2073, simple_loss=0.2981, pruned_loss=0.05828, over 12216.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.291, pruned_loss=0.05608, over 3068850.05 frames. ], batch size: 248, lr: 1.27e-02, grad_scale: 8.0 2023-04-28 10:13:38,879 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.875e+02 2.547e+02 3.217e+02 3.917e+02 1.118e+03, threshold=6.434e+02, percent-clipped=1.0 2023-04-28 10:14:01,071 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5250, 3.4665, 3.4126, 3.0774, 3.3734, 1.8848, 3.2201, 2.9700], device='cuda:4'), covar=tensor([0.0083, 0.0071, 0.0099, 0.0172, 0.0075, 0.1621, 0.0104, 0.0160], device='cuda:4'), in_proj_covar=tensor([0.0088, 0.0076, 0.0119, 0.0110, 0.0088, 0.0144, 0.0101, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:14:22,763 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4059, 1.9353, 2.1071, 3.9113, 1.8698, 2.6476, 2.2351, 2.0549], device='cuda:4'), covar=tensor([0.0609, 0.2266, 0.1220, 0.0292, 0.3030, 0.1249, 0.1974, 0.2599], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0304, 0.0259, 0.0297, 0.0356, 0.0306, 0.0284, 0.0355], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:14:39,000 INFO [train.py:904] (4/8) Epoch 5, batch 10100, loss[loss=0.1941, simple_loss=0.2771, pruned_loss=0.05552, over 15360.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2913, pruned_loss=0.05658, over 3065147.54 frames. ], batch size: 190, lr: 1.27e-02, grad_scale: 8.0 2023-04-28 10:16:20,268 INFO [train.py:904] (4/8) Epoch 6, batch 0, loss[loss=0.2477, simple_loss=0.3176, pruned_loss=0.08891, over 17255.00 frames. ], tot_loss[loss=0.2477, simple_loss=0.3176, pruned_loss=0.08891, over 17255.00 frames. ], batch size: 44, lr: 1.19e-02, grad_scale: 8.0 2023-04-28 10:16:20,269 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 10:16:27,645 INFO [train.py:938] (4/8) Epoch 6, validation: loss=0.1727, simple_loss=0.2755, pruned_loss=0.03501, over 944034.00 frames. 2023-04-28 10:16:27,646 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 10:16:52,395 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.958e+02 3.546e+02 4.478e+02 5.747e+02 1.222e+03, threshold=8.956e+02, percent-clipped=19.0 2023-04-28 10:17:10,805 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50782.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:17:13,045 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50784.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:17:35,095 INFO [train.py:904] (4/8) Epoch 6, batch 50, loss[loss=0.2547, simple_loss=0.3381, pruned_loss=0.08566, over 17072.00 frames. ], tot_loss[loss=0.2509, simple_loss=0.319, pruned_loss=0.09136, over 748374.38 frames. ], batch size: 55, lr: 1.19e-02, grad_scale: 2.0 2023-04-28 10:17:39,060 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9175, 5.4638, 5.4717, 5.3611, 5.3291, 5.9133, 5.5596, 5.3511], device='cuda:4'), covar=tensor([0.0780, 0.1518, 0.1630, 0.1688, 0.2688, 0.1009, 0.1067, 0.2190], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0377, 0.0372, 0.0327, 0.0434, 0.0400, 0.0303, 0.0437], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:17:46,442 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50808.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 10:17:46,610 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9166, 2.0498, 2.3336, 4.5532, 1.9310, 3.0224, 2.2839, 2.2757], device='cuda:4'), covar=tensor([0.0526, 0.2329, 0.1174, 0.0266, 0.3112, 0.1145, 0.1978, 0.2669], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0310, 0.0262, 0.0303, 0.0363, 0.0311, 0.0289, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:17:58,711 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-28 10:18:20,194 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=50832.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:18:34,204 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50843.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:18:47,052 INFO [train.py:904] (4/8) Epoch 6, batch 100, loss[loss=0.2263, simple_loss=0.2979, pruned_loss=0.07736, over 16410.00 frames. ], tot_loss[loss=0.2364, simple_loss=0.3094, pruned_loss=0.08168, over 1326682.40 frames. ], batch size: 75, lr: 1.18e-02, grad_scale: 2.0 2023-04-28 10:18:49,526 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50853.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 10:19:11,789 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.233e+02 3.517e+02 4.205e+02 5.493e+02 1.012e+03, threshold=8.410e+02, percent-clipped=3.0 2023-04-28 10:19:12,214 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50869.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 10:19:49,068 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8181, 4.2571, 4.3334, 1.8597, 4.6438, 4.5494, 3.2368, 3.4087], device='cuda:4'), covar=tensor([0.0680, 0.0097, 0.0165, 0.1125, 0.0038, 0.0066, 0.0313, 0.0337], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0088, 0.0076, 0.0137, 0.0068, 0.0076, 0.0112, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 10:19:54,850 INFO [train.py:904] (4/8) Epoch 6, batch 150, loss[loss=0.2292, simple_loss=0.293, pruned_loss=0.08274, over 16896.00 frames. ], tot_loss[loss=0.2323, simple_loss=0.3062, pruned_loss=0.07923, over 1763388.91 frames. ], batch size: 109, lr: 1.18e-02, grad_scale: 2.0 2023-04-28 10:19:55,127 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=50901.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 10:19:56,135 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=50902.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:20:10,446 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-28 10:20:57,583 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50946.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:21:05,274 INFO [train.py:904] (4/8) Epoch 6, batch 200, loss[loss=0.2256, simple_loss=0.2909, pruned_loss=0.08015, over 16782.00 frames. ], tot_loss[loss=0.231, simple_loss=0.3048, pruned_loss=0.07863, over 2107173.93 frames. ], batch size: 134, lr: 1.18e-02, grad_scale: 2.0 2023-04-28 10:21:28,624 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.959e+02 3.434e+02 3.943e+02 4.796e+02 1.132e+03, threshold=7.886e+02, percent-clipped=3.0 2023-04-28 10:22:12,665 INFO [train.py:904] (4/8) Epoch 6, batch 250, loss[loss=0.2259, simple_loss=0.2933, pruned_loss=0.07921, over 12070.00 frames. ], tot_loss[loss=0.2273, simple_loss=0.3013, pruned_loss=0.07667, over 2371896.27 frames. ], batch size: 247, lr: 1.18e-02, grad_scale: 2.0 2023-04-28 10:22:20,813 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51007.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:22:24,908 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-28 10:22:35,846 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8333, 1.6566, 2.1303, 2.7169, 2.6830, 2.5724, 1.5524, 2.7931], device='cuda:4'), covar=tensor([0.0072, 0.0228, 0.0149, 0.0113, 0.0102, 0.0123, 0.0226, 0.0056], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0146, 0.0131, 0.0127, 0.0133, 0.0095, 0.0144, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 10:22:37,398 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-28 10:22:59,053 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-28 10:23:20,591 INFO [train.py:904] (4/8) Epoch 6, batch 300, loss[loss=0.2006, simple_loss=0.2864, pruned_loss=0.05735, over 17049.00 frames. ], tot_loss[loss=0.2218, simple_loss=0.2965, pruned_loss=0.07357, over 2582579.93 frames. ], batch size: 53, lr: 1.18e-02, grad_scale: 2.0 2023-04-28 10:23:37,427 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51062.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:23:45,699 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.679e+02 2.797e+02 3.691e+02 4.478e+02 8.058e+02, threshold=7.381e+02, percent-clipped=1.0 2023-04-28 10:24:27,717 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7944, 5.5965, 5.6341, 5.4828, 5.3623, 6.0126, 5.7411, 5.4383], device='cuda:4'), covar=tensor([0.0799, 0.1683, 0.1651, 0.1768, 0.3063, 0.1014, 0.1162, 0.2233], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0414, 0.0403, 0.0356, 0.0475, 0.0433, 0.0330, 0.0476], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 10:24:30,408 INFO [train.py:904] (4/8) Epoch 6, batch 350, loss[loss=0.2234, simple_loss=0.2868, pruned_loss=0.07998, over 16796.00 frames. ], tot_loss[loss=0.2188, simple_loss=0.2935, pruned_loss=0.07207, over 2750854.98 frames. ], batch size: 102, lr: 1.18e-02, grad_scale: 2.0 2023-04-28 10:24:49,888 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.47 vs. limit=5.0 2023-04-28 10:25:01,800 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51123.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:25:13,557 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2659, 3.0470, 2.5250, 1.8362, 2.4731, 1.9874, 2.8924, 2.8665], device='cuda:4'), covar=tensor([0.0309, 0.0471, 0.0648, 0.1670, 0.0781, 0.1157, 0.0572, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0122, 0.0149, 0.0138, 0.0127, 0.0125, 0.0135, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 10:25:19,515 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51138.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:25:37,100 INFO [train.py:904] (4/8) Epoch 6, batch 400, loss[loss=0.1833, simple_loss=0.2599, pruned_loss=0.05333, over 16807.00 frames. ], tot_loss[loss=0.2177, simple_loss=0.2917, pruned_loss=0.07184, over 2873844.65 frames. ], batch size: 39, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:25:55,421 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51164.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 10:26:01,706 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.109e+02 2.940e+02 3.610e+02 4.239e+02 7.005e+02, threshold=7.220e+02, percent-clipped=1.0 2023-04-28 10:26:05,631 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51172.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 10:26:20,480 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7307, 4.0637, 4.0000, 2.1838, 4.1887, 4.2159, 3.1748, 3.2709], device='cuda:4'), covar=tensor([0.0653, 0.0090, 0.0146, 0.0957, 0.0058, 0.0089, 0.0368, 0.0320], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0090, 0.0079, 0.0138, 0.0069, 0.0078, 0.0114, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 10:26:45,054 INFO [train.py:904] (4/8) Epoch 6, batch 450, loss[loss=0.1802, simple_loss=0.2598, pruned_loss=0.0503, over 16768.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2892, pruned_loss=0.06931, over 2985275.50 frames. ], batch size: 83, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:26:47,147 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51202.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:27:13,485 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.17 vs. limit=5.0 2023-04-28 10:27:29,539 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51233.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 10:27:52,964 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=51250.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:27:53,833 INFO [train.py:904] (4/8) Epoch 6, batch 500, loss[loss=0.1938, simple_loss=0.2675, pruned_loss=0.06, over 16562.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.2871, pruned_loss=0.0679, over 3064434.06 frames. ], batch size: 68, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:28:17,360 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.948e+02 3.090e+02 3.738e+02 4.449e+02 8.454e+02, threshold=7.475e+02, percent-clipped=2.0 2023-04-28 10:28:25,521 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 10:29:01,771 INFO [train.py:904] (4/8) Epoch 6, batch 550, loss[loss=0.205, simple_loss=0.2924, pruned_loss=0.05879, over 17192.00 frames. ], tot_loss[loss=0.2104, simple_loss=0.2864, pruned_loss=0.06723, over 3123532.05 frames. ], batch size: 46, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:29:03,273 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51302.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:29:05,983 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1794, 3.8693, 3.1064, 1.9274, 2.7167, 2.3293, 3.6603, 3.6094], device='cuda:4'), covar=tensor([0.0230, 0.0492, 0.0605, 0.1552, 0.0738, 0.0917, 0.0451, 0.0735], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0126, 0.0150, 0.0139, 0.0129, 0.0124, 0.0136, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 10:30:13,393 INFO [train.py:904] (4/8) Epoch 6, batch 600, loss[loss=0.1855, simple_loss=0.2697, pruned_loss=0.05063, over 17191.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2852, pruned_loss=0.06675, over 3165233.81 frames. ], batch size: 46, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:30:22,120 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5494, 2.0318, 2.2670, 4.1732, 2.0079, 2.8188, 2.2469, 2.1930], device='cuda:4'), covar=tensor([0.0597, 0.2239, 0.1203, 0.0305, 0.2723, 0.1215, 0.2146, 0.2246], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0324, 0.0271, 0.0316, 0.0372, 0.0334, 0.0298, 0.0393], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:30:38,608 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.911e+02 2.906e+02 3.488e+02 4.318e+02 1.008e+03, threshold=6.975e+02, percent-clipped=1.0 2023-04-28 10:30:55,664 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51382.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:31:10,894 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51392.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:31:23,783 INFO [train.py:904] (4/8) Epoch 6, batch 650, loss[loss=0.2088, simple_loss=0.2905, pruned_loss=0.06349, over 16988.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.284, pruned_loss=0.06618, over 3210730.24 frames. ], batch size: 41, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:31:47,279 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51418.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:32:14,534 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51438.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:32:21,295 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51443.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:32:31,562 INFO [train.py:904] (4/8) Epoch 6, batch 700, loss[loss=0.228, simple_loss=0.3078, pruned_loss=0.07405, over 17138.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2827, pruned_loss=0.06502, over 3232262.05 frames. ], batch size: 49, lr: 1.18e-02, grad_scale: 2.0 2023-04-28 10:32:35,167 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51453.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:32:49,175 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51464.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 10:32:57,145 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.941e+02 3.067e+02 3.781e+02 4.816e+02 1.338e+03, threshold=7.562e+02, percent-clipped=6.0 2023-04-28 10:33:19,513 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=51486.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:33:32,410 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51495.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:33:39,717 INFO [train.py:904] (4/8) Epoch 6, batch 750, loss[loss=0.1809, simple_loss=0.2628, pruned_loss=0.04945, over 16546.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2837, pruned_loss=0.06578, over 3254013.54 frames. ], batch size: 68, lr: 1.18e-02, grad_scale: 2.0 2023-04-28 10:33:56,286 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=51512.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 10:34:19,279 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51528.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 10:34:53,036 INFO [train.py:904] (4/8) Epoch 6, batch 800, loss[loss=0.1921, simple_loss=0.2709, pruned_loss=0.05669, over 17237.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2831, pruned_loss=0.06601, over 3259959.17 frames. ], batch size: 45, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:34:59,989 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51556.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:35:11,437 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7064, 3.6987, 2.7696, 2.3114, 2.5342, 2.1817, 3.6828, 3.5763], device='cuda:4'), covar=tensor([0.1982, 0.0497, 0.1305, 0.1703, 0.1902, 0.1521, 0.0466, 0.0791], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0253, 0.0275, 0.0249, 0.0280, 0.0206, 0.0250, 0.0262], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:35:19,942 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.998e+02 2.921e+02 3.386e+02 4.262e+02 8.289e+02, threshold=6.772e+02, percent-clipped=2.0 2023-04-28 10:36:01,900 INFO [train.py:904] (4/8) Epoch 6, batch 850, loss[loss=0.1857, simple_loss=0.2755, pruned_loss=0.048, over 17170.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.282, pruned_loss=0.06507, over 3280621.67 frames. ], batch size: 46, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:36:04,112 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51602.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:37:10,632 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=51650.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:37:11,527 INFO [train.py:904] (4/8) Epoch 6, batch 900, loss[loss=0.1688, simple_loss=0.2587, pruned_loss=0.03945, over 17255.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2803, pruned_loss=0.06395, over 3290298.77 frames. ], batch size: 45, lr: 1.18e-02, grad_scale: 4.0 2023-04-28 10:37:39,516 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.821e+02 2.905e+02 3.526e+02 4.406e+02 8.198e+02, threshold=7.052e+02, percent-clipped=7.0 2023-04-28 10:38:22,587 INFO [train.py:904] (4/8) Epoch 6, batch 950, loss[loss=0.1958, simple_loss=0.2838, pruned_loss=0.05387, over 17125.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2814, pruned_loss=0.06477, over 3298575.71 frames. ], batch size: 47, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:38:46,333 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51718.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:39:10,600 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-28 10:39:12,960 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51738.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:39:25,710 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51747.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:39:26,802 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51748.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:39:30,782 INFO [train.py:904] (4/8) Epoch 6, batch 1000, loss[loss=0.2106, simple_loss=0.2687, pruned_loss=0.07625, over 16840.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2801, pruned_loss=0.06383, over 3308291.55 frames. ], batch size: 102, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:39:51,036 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=51766.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:39:51,405 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-28 10:39:57,332 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.119e+02 2.827e+02 3.297e+02 4.150e+02 8.828e+02, threshold=6.593e+02, percent-clipped=4.0 2023-04-28 10:40:12,049 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7539, 3.8732, 4.2578, 2.9631, 3.7924, 4.1614, 4.0613, 2.5112], device='cuda:4'), covar=tensor([0.0304, 0.0039, 0.0024, 0.0224, 0.0049, 0.0043, 0.0032, 0.0276], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0064, 0.0061, 0.0117, 0.0064, 0.0071, 0.0064, 0.0110], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 10:40:27,234 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5367, 3.3286, 3.7982, 2.7628, 3.6479, 3.7486, 3.7447, 2.1104], device='cuda:4'), covar=tensor([0.0300, 0.0137, 0.0036, 0.0202, 0.0045, 0.0064, 0.0040, 0.0312], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0064, 0.0061, 0.0118, 0.0064, 0.0072, 0.0065, 0.0110], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 10:40:39,945 INFO [train.py:904] (4/8) Epoch 6, batch 1050, loss[loss=0.188, simple_loss=0.2659, pruned_loss=0.05499, over 17217.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2795, pruned_loss=0.06391, over 3313384.80 frames. ], batch size: 44, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:40:49,347 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51808.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:41:17,951 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51828.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 10:41:27,734 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7094, 3.1839, 2.5144, 4.3089, 3.8730, 4.1385, 1.5952, 3.1434], device='cuda:4'), covar=tensor([0.1287, 0.0431, 0.0982, 0.0085, 0.0236, 0.0316, 0.1175, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0143, 0.0168, 0.0093, 0.0182, 0.0185, 0.0161, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 10:41:32,846 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 10:41:49,451 INFO [train.py:904] (4/8) Epoch 6, batch 1100, loss[loss=0.1833, simple_loss=0.2694, pruned_loss=0.04859, over 17135.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2795, pruned_loss=0.0649, over 3317193.12 frames. ], batch size: 48, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:41:49,771 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51851.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:42:16,638 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.653e+02 2.702e+02 3.374e+02 4.258e+02 9.547e+02, threshold=6.748e+02, percent-clipped=3.0 2023-04-28 10:42:24,556 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=51876.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 10:42:59,133 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-28 10:42:59,412 INFO [train.py:904] (4/8) Epoch 6, batch 1150, loss[loss=0.2185, simple_loss=0.2972, pruned_loss=0.06993, over 16701.00 frames. ], tot_loss[loss=0.2035, simple_loss=0.2794, pruned_loss=0.06379, over 3323977.40 frames. ], batch size: 57, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:44:08,004 INFO [train.py:904] (4/8) Epoch 6, batch 1200, loss[loss=0.1758, simple_loss=0.2725, pruned_loss=0.0396, over 17067.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2779, pruned_loss=0.06334, over 3327630.16 frames. ], batch size: 50, lr: 1.17e-02, grad_scale: 8.0 2023-04-28 10:44:12,550 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4792, 3.4514, 4.0813, 2.6971, 3.7552, 3.9645, 3.8328, 2.4727], device='cuda:4'), covar=tensor([0.0321, 0.0197, 0.0023, 0.0221, 0.0047, 0.0052, 0.0047, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0064, 0.0061, 0.0116, 0.0063, 0.0070, 0.0065, 0.0108], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 10:44:33,628 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.991e+02 2.935e+02 3.395e+02 4.046e+02 1.078e+03, threshold=6.791e+02, percent-clipped=4.0 2023-04-28 10:44:51,295 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7957, 3.3051, 2.8138, 5.0039, 4.4394, 4.6132, 1.5931, 3.5348], device='cuda:4'), covar=tensor([0.1289, 0.0566, 0.1015, 0.0101, 0.0272, 0.0321, 0.1363, 0.0574], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0143, 0.0166, 0.0092, 0.0183, 0.0185, 0.0160, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 10:45:20,379 INFO [train.py:904] (4/8) Epoch 6, batch 1250, loss[loss=0.2189, simple_loss=0.2825, pruned_loss=0.07769, over 11936.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2774, pruned_loss=0.06364, over 3328781.48 frames. ], batch size: 248, lr: 1.17e-02, grad_scale: 8.0 2023-04-28 10:45:23,208 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 10:45:28,087 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9173, 4.1589, 1.9912, 4.5086, 2.7652, 4.5279, 2.1554, 3.1678], device='cuda:4'), covar=tensor([0.0141, 0.0255, 0.1479, 0.0065, 0.0692, 0.0288, 0.1358, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0161, 0.0179, 0.0086, 0.0161, 0.0192, 0.0185, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 10:46:13,378 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52038.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:46:15,830 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52040.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:46:26,760 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52048.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:46:30,468 INFO [train.py:904] (4/8) Epoch 6, batch 1300, loss[loss=0.2181, simple_loss=0.31, pruned_loss=0.06308, over 17257.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2775, pruned_loss=0.06346, over 3320818.77 frames. ], batch size: 52, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:46:58,326 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.022e+02 2.918e+02 3.490e+02 4.202e+02 7.834e+02, threshold=6.979e+02, percent-clipped=4.0 2023-04-28 10:47:03,023 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5251, 3.5799, 4.0375, 2.7954, 3.6505, 3.9160, 3.8380, 2.5178], device='cuda:4'), covar=tensor([0.0305, 0.0121, 0.0028, 0.0212, 0.0046, 0.0049, 0.0036, 0.0237], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0063, 0.0061, 0.0117, 0.0064, 0.0071, 0.0065, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 10:47:20,308 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=52086.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:47:34,137 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=52096.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:47:41,500 INFO [train.py:904] (4/8) Epoch 6, batch 1350, loss[loss=0.1972, simple_loss=0.2871, pruned_loss=0.0536, over 17179.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2779, pruned_loss=0.0635, over 3308951.39 frames. ], batch size: 46, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:47:42,035 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52101.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:47:44,068 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52103.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:48:12,625 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1388, 5.5691, 5.7001, 5.5369, 5.6157, 6.0962, 5.7507, 5.4622], device='cuda:4'), covar=tensor([0.0656, 0.1522, 0.1421, 0.1590, 0.2509, 0.0787, 0.1044, 0.2059], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0420, 0.0416, 0.0358, 0.0485, 0.0446, 0.0338, 0.0480], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 10:48:51,205 INFO [train.py:904] (4/8) Epoch 6, batch 1400, loss[loss=0.1915, simple_loss=0.273, pruned_loss=0.05501, over 17176.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2784, pruned_loss=0.06334, over 3314543.76 frames. ], batch size: 46, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:48:51,459 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52151.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:49:19,188 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.784e+02 2.727e+02 3.349e+02 4.009e+02 8.330e+02, threshold=6.698e+02, percent-clipped=2.0 2023-04-28 10:49:49,794 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52193.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:49:58,142 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=52199.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:49:58,275 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0239, 4.8072, 5.0352, 5.2963, 5.4288, 4.7087, 5.4150, 5.3682], device='cuda:4'), covar=tensor([0.0911, 0.0786, 0.1292, 0.0505, 0.0379, 0.0709, 0.0400, 0.0392], device='cuda:4'), in_proj_covar=tensor([0.0445, 0.0545, 0.0697, 0.0559, 0.0421, 0.0414, 0.0436, 0.0470], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:50:00,566 INFO [train.py:904] (4/8) Epoch 6, batch 1450, loss[loss=0.1904, simple_loss=0.2616, pruned_loss=0.05957, over 15418.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2771, pruned_loss=0.06294, over 3310429.51 frames. ], batch size: 190, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:51:10,829 INFO [train.py:904] (4/8) Epoch 6, batch 1500, loss[loss=0.1922, simple_loss=0.2565, pruned_loss=0.06397, over 16833.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.277, pruned_loss=0.06272, over 3317726.24 frames. ], batch size: 102, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:51:15,537 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52254.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:51:38,534 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.881e+02 3.021e+02 3.666e+02 4.541e+02 9.509e+02, threshold=7.332e+02, percent-clipped=3.0 2023-04-28 10:52:18,645 INFO [train.py:904] (4/8) Epoch 6, batch 1550, loss[loss=0.211, simple_loss=0.2944, pruned_loss=0.06378, over 16697.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2801, pruned_loss=0.06462, over 3321478.74 frames. ], batch size: 57, lr: 1.17e-02, grad_scale: 4.0 2023-04-28 10:52:33,867 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52312.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:53:18,371 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0162, 3.0270, 3.4345, 2.2219, 3.0566, 3.3091, 3.1987, 1.9645], device='cuda:4'), covar=tensor([0.0303, 0.0099, 0.0026, 0.0227, 0.0055, 0.0039, 0.0041, 0.0259], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0062, 0.0060, 0.0114, 0.0063, 0.0071, 0.0065, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 10:53:27,894 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.92 vs. limit=5.0 2023-04-28 10:53:28,025 INFO [train.py:904] (4/8) Epoch 6, batch 1600, loss[loss=0.1906, simple_loss=0.265, pruned_loss=0.05813, over 16844.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2823, pruned_loss=0.06517, over 3319186.79 frames. ], batch size: 42, lr: 1.17e-02, grad_scale: 8.0 2023-04-28 10:53:55,823 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.894e+02 2.932e+02 3.332e+02 3.992e+02 7.476e+02, threshold=6.664e+02, percent-clipped=1.0 2023-04-28 10:53:59,280 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52373.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:54:08,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5529, 5.9617, 5.6538, 5.6822, 5.2621, 4.9894, 5.3888, 6.0304], device='cuda:4'), covar=tensor([0.0838, 0.0615, 0.0794, 0.0496, 0.0659, 0.0592, 0.0692, 0.0670], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0552, 0.0451, 0.0353, 0.0338, 0.0348, 0.0446, 0.0386], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 10:54:30,441 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52396.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:54:36,616 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-28 10:54:37,117 INFO [train.py:904] (4/8) Epoch 6, batch 1650, loss[loss=0.1991, simple_loss=0.286, pruned_loss=0.05615, over 17052.00 frames. ], tot_loss[loss=0.2083, simple_loss=0.2842, pruned_loss=0.06619, over 3320584.24 frames. ], batch size: 50, lr: 1.17e-02, grad_scale: 8.0 2023-04-28 10:54:40,267 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52403.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:54:46,960 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.02 vs. limit=2.0 2023-04-28 10:55:25,809 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6507, 3.8142, 2.0587, 4.0843, 2.8287, 4.0829, 2.1620, 3.0405], device='cuda:4'), covar=tensor([0.0138, 0.0331, 0.1333, 0.0088, 0.0606, 0.0417, 0.1244, 0.0521], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0162, 0.0178, 0.0087, 0.0160, 0.0193, 0.0186, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 10:55:45,565 INFO [train.py:904] (4/8) Epoch 6, batch 1700, loss[loss=0.1859, simple_loss=0.2699, pruned_loss=0.05096, over 16835.00 frames. ], tot_loss[loss=0.2091, simple_loss=0.2855, pruned_loss=0.0663, over 3321558.38 frames. ], batch size: 42, lr: 1.17e-02, grad_scale: 8.0 2023-04-28 10:55:45,911 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=52451.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:56:14,196 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.976e+02 2.939e+02 3.544e+02 4.398e+02 9.866e+02, threshold=7.088e+02, percent-clipped=4.0 2023-04-28 10:56:57,691 INFO [train.py:904] (4/8) Epoch 6, batch 1750, loss[loss=0.23, simple_loss=0.295, pruned_loss=0.0825, over 16772.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2871, pruned_loss=0.06702, over 3316634.42 frames. ], batch size: 83, lr: 1.17e-02, grad_scale: 8.0 2023-04-28 10:58:05,487 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52549.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 10:58:07,617 INFO [train.py:904] (4/8) Epoch 6, batch 1800, loss[loss=0.2395, simple_loss=0.2991, pruned_loss=0.09001, over 16870.00 frames. ], tot_loss[loss=0.2118, simple_loss=0.2885, pruned_loss=0.06755, over 3316409.57 frames. ], batch size: 109, lr: 1.17e-02, grad_scale: 8.0 2023-04-28 10:58:13,014 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5956, 3.6081, 4.1140, 3.0144, 3.6744, 3.9049, 3.7360, 2.5114], device='cuda:4'), covar=tensor([0.0298, 0.0089, 0.0026, 0.0189, 0.0050, 0.0048, 0.0042, 0.0257], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0061, 0.0059, 0.0111, 0.0061, 0.0070, 0.0064, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 10:58:14,092 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52555.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 10:58:36,301 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.850e+02 3.092e+02 3.773e+02 4.797e+02 9.614e+02, threshold=7.547e+02, percent-clipped=5.0 2023-04-28 10:59:00,767 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7162, 3.8306, 3.0587, 5.1511, 4.6011, 4.7366, 1.5128, 3.4768], device='cuda:4'), covar=tensor([0.1415, 0.0476, 0.0929, 0.0084, 0.0226, 0.0309, 0.1490, 0.0626], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0144, 0.0167, 0.0095, 0.0186, 0.0186, 0.0161, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 10:59:17,730 INFO [train.py:904] (4/8) Epoch 6, batch 1850, loss[loss=0.1801, simple_loss=0.2548, pruned_loss=0.05274, over 16791.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.2889, pruned_loss=0.06708, over 3325427.94 frames. ], batch size: 39, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 10:59:38,401 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52616.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 11:00:27,089 INFO [train.py:904] (4/8) Epoch 6, batch 1900, loss[loss=0.1975, simple_loss=0.2765, pruned_loss=0.05925, over 16434.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2878, pruned_loss=0.06601, over 3333706.92 frames. ], batch size: 146, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:00:51,260 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52668.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:00:54,523 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.841e+02 2.667e+02 3.390e+02 4.181e+02 1.051e+03, threshold=6.780e+02, percent-clipped=5.0 2023-04-28 11:01:30,652 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52696.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:01:36,640 INFO [train.py:904] (4/8) Epoch 6, batch 1950, loss[loss=0.2299, simple_loss=0.3073, pruned_loss=0.07628, over 16668.00 frames. ], tot_loss[loss=0.2097, simple_loss=0.2878, pruned_loss=0.06578, over 3333404.87 frames. ], batch size: 62, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:01:38,156 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3113, 3.2357, 3.2999, 3.4496, 3.5218, 3.2274, 3.3206, 3.5380], device='cuda:4'), covar=tensor([0.0887, 0.0742, 0.1099, 0.0544, 0.0524, 0.2308, 0.1371, 0.0565], device='cuda:4'), in_proj_covar=tensor([0.0449, 0.0553, 0.0703, 0.0569, 0.0424, 0.0418, 0.0442, 0.0475], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:01:47,294 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4698, 4.7916, 4.5006, 4.5638, 4.2614, 4.1692, 4.3051, 4.7933], device='cuda:4'), covar=tensor([0.0753, 0.0695, 0.0973, 0.0488, 0.0721, 0.1142, 0.0733, 0.0735], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0559, 0.0459, 0.0358, 0.0343, 0.0349, 0.0450, 0.0395], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:01:56,005 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7580, 3.1861, 2.7488, 4.7656, 4.1763, 4.4477, 1.5494, 3.3084], device='cuda:4'), covar=tensor([0.1319, 0.0567, 0.1052, 0.0101, 0.0292, 0.0316, 0.1397, 0.0624], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0142, 0.0166, 0.0095, 0.0184, 0.0184, 0.0160, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 11:02:36,906 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=52744.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:02:47,589 INFO [train.py:904] (4/8) Epoch 6, batch 2000, loss[loss=0.2014, simple_loss=0.2956, pruned_loss=0.05356, over 17154.00 frames. ], tot_loss[loss=0.2097, simple_loss=0.2878, pruned_loss=0.06584, over 3312980.15 frames. ], batch size: 49, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:02:49,208 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52752.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:03:15,694 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.007e+02 2.923e+02 3.553e+02 4.132e+02 6.210e+02, threshold=7.105e+02, percent-clipped=0.0 2023-04-28 11:03:29,010 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7794, 2.6836, 2.6892, 1.8997, 2.4341, 2.5540, 2.6135, 1.7231], device='cuda:4'), covar=tensor([0.0287, 0.0062, 0.0037, 0.0226, 0.0080, 0.0055, 0.0054, 0.0273], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0062, 0.0060, 0.0112, 0.0062, 0.0071, 0.0064, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:03:33,176 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5273, 4.2846, 4.4763, 4.7046, 4.8062, 4.2920, 4.5496, 4.7458], device='cuda:4'), covar=tensor([0.0725, 0.0724, 0.1050, 0.0477, 0.0451, 0.0801, 0.1185, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0439, 0.0541, 0.0687, 0.0558, 0.0419, 0.0409, 0.0433, 0.0466], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:03:46,442 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6936, 4.6254, 5.1299, 5.2087, 5.2196, 4.7997, 4.7238, 4.4340], device='cuda:4'), covar=tensor([0.0265, 0.0417, 0.0389, 0.0413, 0.0374, 0.0290, 0.0764, 0.0366], device='cuda:4'), in_proj_covar=tensor([0.0265, 0.0265, 0.0263, 0.0262, 0.0318, 0.0279, 0.0393, 0.0228], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 11:03:57,170 INFO [train.py:904] (4/8) Epoch 6, batch 2050, loss[loss=0.2186, simple_loss=0.3021, pruned_loss=0.06758, over 17022.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2877, pruned_loss=0.06665, over 3315369.62 frames. ], batch size: 55, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:04:14,415 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52813.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:05:06,132 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52849.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:05:08,343 INFO [train.py:904] (4/8) Epoch 6, batch 2100, loss[loss=0.2097, simple_loss=0.2795, pruned_loss=0.06997, over 16781.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2872, pruned_loss=0.06627, over 3320899.90 frames. ], batch size: 96, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:05:36,283 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.586e+02 2.933e+02 3.490e+02 4.297e+02 7.985e+02, threshold=6.979e+02, percent-clipped=3.0 2023-04-28 11:06:12,953 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=52897.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:06:18,012 INFO [train.py:904] (4/8) Epoch 6, batch 2150, loss[loss=0.2362, simple_loss=0.3026, pruned_loss=0.08488, over 16757.00 frames. ], tot_loss[loss=0.2104, simple_loss=0.2879, pruned_loss=0.06646, over 3318698.27 frames. ], batch size: 124, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:06:33,075 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52911.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 11:07:29,309 INFO [train.py:904] (4/8) Epoch 6, batch 2200, loss[loss=0.222, simple_loss=0.2918, pruned_loss=0.0761, over 16785.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.289, pruned_loss=0.06683, over 3328123.29 frames. ], batch size: 124, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:07:52,373 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52968.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:07:56,705 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.110e+02 3.315e+02 3.926e+02 4.640e+02 8.894e+02, threshold=7.853e+02, percent-clipped=3.0 2023-04-28 11:08:38,583 INFO [train.py:904] (4/8) Epoch 6, batch 2250, loss[loss=0.2174, simple_loss=0.3133, pruned_loss=0.06077, over 16752.00 frames. ], tot_loss[loss=0.2111, simple_loss=0.2888, pruned_loss=0.06669, over 3332204.73 frames. ], batch size: 57, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:08:59,894 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=53016.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:09:13,412 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.64 vs. limit=5.0 2023-04-28 11:09:47,159 INFO [train.py:904] (4/8) Epoch 6, batch 2300, loss[loss=0.1912, simple_loss=0.2794, pruned_loss=0.05147, over 17046.00 frames. ], tot_loss[loss=0.2111, simple_loss=0.2891, pruned_loss=0.06658, over 3328113.32 frames. ], batch size: 50, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:09:53,225 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4807, 5.4196, 5.2387, 5.0691, 4.8253, 5.2731, 5.2567, 4.9518], device='cuda:4'), covar=tensor([0.0439, 0.0277, 0.0206, 0.0196, 0.0914, 0.0290, 0.0202, 0.0501], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0223, 0.0246, 0.0219, 0.0281, 0.0249, 0.0174, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:09:54,630 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53056.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:10:15,630 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.951e+02 3.084e+02 3.786e+02 4.874e+02 1.349e+03, threshold=7.572e+02, percent-clipped=4.0 2023-04-28 11:10:57,588 INFO [train.py:904] (4/8) Epoch 6, batch 2350, loss[loss=0.227, simple_loss=0.2965, pruned_loss=0.07872, over 16517.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.2895, pruned_loss=0.06665, over 3333214.43 frames. ], batch size: 68, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:11:08,086 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53108.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:11:20,211 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53117.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:11:27,276 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2867, 1.9125, 2.1389, 3.7497, 1.9005, 2.5886, 2.1317, 2.0638], device='cuda:4'), covar=tensor([0.0668, 0.2317, 0.1249, 0.0346, 0.2686, 0.1364, 0.2126, 0.2250], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0335, 0.0275, 0.0317, 0.0376, 0.0352, 0.0306, 0.0406], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:12:04,689 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8941, 4.7918, 4.7016, 4.4972, 4.2750, 4.7421, 4.6813, 4.3788], device='cuda:4'), covar=tensor([0.0482, 0.0335, 0.0227, 0.0239, 0.0874, 0.0313, 0.0386, 0.0609], device='cuda:4'), in_proj_covar=tensor([0.0207, 0.0225, 0.0246, 0.0218, 0.0279, 0.0248, 0.0174, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:12:08,724 INFO [train.py:904] (4/8) Epoch 6, batch 2400, loss[loss=0.1784, simple_loss=0.261, pruned_loss=0.04788, over 16789.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.29, pruned_loss=0.0666, over 3326744.55 frames. ], batch size: 39, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:12:18,564 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4975, 4.3888, 4.3586, 4.2114, 4.0311, 4.4057, 4.2452, 4.1334], device='cuda:4'), covar=tensor([0.0509, 0.0435, 0.0221, 0.0208, 0.0834, 0.0356, 0.0369, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0207, 0.0226, 0.0246, 0.0219, 0.0279, 0.0248, 0.0173, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:12:36,943 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.873e+02 2.840e+02 3.422e+02 4.151e+02 8.672e+02, threshold=6.844e+02, percent-clipped=2.0 2023-04-28 11:12:52,713 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9203, 4.9641, 5.4433, 5.4894, 5.4492, 5.0448, 5.0596, 4.6628], device='cuda:4'), covar=tensor([0.0246, 0.0350, 0.0236, 0.0298, 0.0333, 0.0235, 0.0697, 0.0331], device='cuda:4'), in_proj_covar=tensor([0.0261, 0.0263, 0.0256, 0.0257, 0.0306, 0.0272, 0.0383, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 11:13:19,222 INFO [train.py:904] (4/8) Epoch 6, batch 2450, loss[loss=0.2154, simple_loss=0.2901, pruned_loss=0.07039, over 16837.00 frames. ], tot_loss[loss=0.2124, simple_loss=0.2909, pruned_loss=0.06694, over 3315256.15 frames. ], batch size: 102, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:13:33,716 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53211.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 11:13:41,117 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3188, 5.6482, 5.3919, 5.5044, 5.0359, 4.7498, 5.1935, 5.7871], device='cuda:4'), covar=tensor([0.0778, 0.0781, 0.0897, 0.0449, 0.0700, 0.0672, 0.0673, 0.0725], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0548, 0.0450, 0.0352, 0.0335, 0.0344, 0.0441, 0.0384], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:13:45,214 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9051, 4.2628, 2.8539, 2.3884, 3.1380, 2.2513, 4.3402, 4.1960], device='cuda:4'), covar=tensor([0.2308, 0.0654, 0.1486, 0.1745, 0.2288, 0.1664, 0.0429, 0.0660], device='cuda:4'), in_proj_covar=tensor([0.0276, 0.0254, 0.0266, 0.0247, 0.0286, 0.0202, 0.0243, 0.0268], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:14:04,136 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.88 vs. limit=5.0 2023-04-28 11:14:16,001 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53241.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:14:29,175 INFO [train.py:904] (4/8) Epoch 6, batch 2500, loss[loss=0.1959, simple_loss=0.2812, pruned_loss=0.05526, over 17135.00 frames. ], tot_loss[loss=0.2124, simple_loss=0.2906, pruned_loss=0.06707, over 3321871.70 frames. ], batch size: 46, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:14:39,969 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=53259.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 11:14:57,075 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.925e+02 2.700e+02 3.263e+02 4.236e+02 1.034e+03, threshold=6.525e+02, percent-clipped=4.0 2023-04-28 11:15:38,466 INFO [train.py:904] (4/8) Epoch 6, batch 2550, loss[loss=0.1778, simple_loss=0.2682, pruned_loss=0.04369, over 17115.00 frames. ], tot_loss[loss=0.2115, simple_loss=0.29, pruned_loss=0.06651, over 3328672.23 frames. ], batch size: 49, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:15:40,036 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53302.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:15:43,231 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53304.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:15:46,564 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-28 11:15:58,243 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53315.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:16:39,684 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7326, 4.7876, 5.1978, 5.1672, 5.2358, 4.7174, 4.7818, 4.4474], device='cuda:4'), covar=tensor([0.0238, 0.0329, 0.0255, 0.0328, 0.0307, 0.0261, 0.0723, 0.0369], device='cuda:4'), in_proj_covar=tensor([0.0262, 0.0262, 0.0257, 0.0259, 0.0308, 0.0275, 0.0388, 0.0230], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 11:16:48,720 INFO [train.py:904] (4/8) Epoch 6, batch 2600, loss[loss=0.2017, simple_loss=0.2776, pruned_loss=0.06292, over 16486.00 frames. ], tot_loss[loss=0.2119, simple_loss=0.29, pruned_loss=0.06686, over 3331759.32 frames. ], batch size: 75, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:17:09,319 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53365.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:17:11,632 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53367.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:17:16,414 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.993e+02 2.796e+02 3.424e+02 4.100e+02 9.306e+02, threshold=6.847e+02, percent-clipped=3.0 2023-04-28 11:17:24,121 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53376.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:17:57,470 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-28 11:17:59,456 INFO [train.py:904] (4/8) Epoch 6, batch 2650, loss[loss=0.224, simple_loss=0.2993, pruned_loss=0.07433, over 16744.00 frames. ], tot_loss[loss=0.2124, simple_loss=0.291, pruned_loss=0.06685, over 3333260.25 frames. ], batch size: 83, lr: 1.16e-02, grad_scale: 8.0 2023-04-28 11:18:10,307 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53408.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:18:14,692 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53412.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:18:37,231 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53428.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 11:19:09,533 INFO [train.py:904] (4/8) Epoch 6, batch 2700, loss[loss=0.2104, simple_loss=0.2955, pruned_loss=0.06261, over 16473.00 frames. ], tot_loss[loss=0.2108, simple_loss=0.2903, pruned_loss=0.06563, over 3335560.70 frames. ], batch size: 68, lr: 1.16e-02, grad_scale: 4.0 2023-04-28 11:19:16,987 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=53456.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:19:38,362 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.962e+02 2.946e+02 3.581e+02 4.267e+02 8.371e+02, threshold=7.163e+02, percent-clipped=3.0 2023-04-28 11:20:19,909 INFO [train.py:904] (4/8) Epoch 6, batch 2750, loss[loss=0.1976, simple_loss=0.2886, pruned_loss=0.05331, over 16093.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2902, pruned_loss=0.06535, over 3328558.92 frames. ], batch size: 35, lr: 1.16e-02, grad_scale: 4.0 2023-04-28 11:21:15,765 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6834, 3.1482, 2.6033, 4.9216, 4.0212, 4.4744, 1.8096, 3.2661], device='cuda:4'), covar=tensor([0.1676, 0.0738, 0.1317, 0.0131, 0.0433, 0.0358, 0.1562, 0.0840], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0144, 0.0168, 0.0098, 0.0196, 0.0188, 0.0161, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 11:21:31,892 INFO [train.py:904] (4/8) Epoch 6, batch 2800, loss[loss=0.251, simple_loss=0.3209, pruned_loss=0.09055, over 12614.00 frames. ], tot_loss[loss=0.2097, simple_loss=0.2895, pruned_loss=0.06494, over 3327119.05 frames. ], batch size: 246, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:21:56,381 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-28 11:22:01,718 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.012e+02 2.830e+02 3.513e+02 4.760e+02 1.218e+03, threshold=7.026e+02, percent-clipped=5.0 2023-04-28 11:22:37,898 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53597.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:22:43,359 INFO [train.py:904] (4/8) Epoch 6, batch 2850, loss[loss=0.2332, simple_loss=0.315, pruned_loss=0.07575, over 16715.00 frames. ], tot_loss[loss=0.2083, simple_loss=0.2881, pruned_loss=0.06429, over 3323225.93 frames. ], batch size: 57, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:23:11,181 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0702, 2.9947, 3.2464, 2.2386, 2.9858, 3.2951, 3.0574, 1.6544], device='cuda:4'), covar=tensor([0.0290, 0.0078, 0.0046, 0.0233, 0.0066, 0.0067, 0.0067, 0.0353], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0062, 0.0062, 0.0114, 0.0063, 0.0072, 0.0065, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:23:29,508 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6692, 3.3251, 2.6576, 4.9880, 4.4812, 4.5775, 1.4484, 3.2231], device='cuda:4'), covar=tensor([0.1349, 0.0504, 0.1090, 0.0089, 0.0271, 0.0332, 0.1396, 0.0679], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0142, 0.0166, 0.0096, 0.0193, 0.0186, 0.0159, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 11:23:51,283 INFO [train.py:904] (4/8) Epoch 6, batch 2900, loss[loss=0.1827, simple_loss=0.2788, pruned_loss=0.04335, over 17129.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.2878, pruned_loss=0.06588, over 3318483.28 frames. ], batch size: 48, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:24:04,494 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53660.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:24:19,307 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53671.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:24:20,247 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.792e+02 2.863e+02 3.412e+02 4.656e+02 8.517e+02, threshold=6.823e+02, percent-clipped=1.0 2023-04-28 11:24:44,518 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6399, 3.7074, 2.6965, 2.2246, 2.6362, 2.1816, 3.5946, 3.6170], device='cuda:4'), covar=tensor([0.1896, 0.0505, 0.1199, 0.1744, 0.1859, 0.1458, 0.0476, 0.0734], device='cuda:4'), in_proj_covar=tensor([0.0278, 0.0255, 0.0269, 0.0249, 0.0289, 0.0205, 0.0246, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:24:46,765 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6930, 3.9143, 2.1262, 3.9450, 2.6743, 3.9630, 2.1349, 2.9141], device='cuda:4'), covar=tensor([0.0117, 0.0223, 0.1140, 0.0095, 0.0609, 0.0316, 0.1113, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0163, 0.0178, 0.0088, 0.0163, 0.0196, 0.0186, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 11:24:59,792 INFO [train.py:904] (4/8) Epoch 6, batch 2950, loss[loss=0.194, simple_loss=0.2732, pruned_loss=0.05739, over 16787.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2867, pruned_loss=0.0666, over 3315551.65 frames. ], batch size: 39, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:25:15,494 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53712.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:25:30,149 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53723.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 11:26:08,565 INFO [train.py:904] (4/8) Epoch 6, batch 3000, loss[loss=0.1757, simple_loss=0.2587, pruned_loss=0.04631, over 16810.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.287, pruned_loss=0.06592, over 3327216.46 frames. ], batch size: 42, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:26:08,566 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 11:26:17,402 INFO [train.py:938] (4/8) Epoch 6, validation: loss=0.1514, simple_loss=0.258, pruned_loss=0.02246, over 944034.00 frames. 2023-04-28 11:26:17,403 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 11:26:29,470 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=53760.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:26:45,982 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.538e+02 2.839e+02 3.365e+02 4.010e+02 8.699e+02, threshold=6.731e+02, percent-clipped=2.0 2023-04-28 11:26:58,866 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-28 11:27:17,180 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1563, 4.5318, 2.4353, 4.7982, 2.9617, 4.7537, 2.4038, 3.2929], device='cuda:4'), covar=tensor([0.0125, 0.0208, 0.1209, 0.0062, 0.0698, 0.0264, 0.1297, 0.0515], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0163, 0.0177, 0.0089, 0.0164, 0.0198, 0.0186, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 11:27:25,828 INFO [train.py:904] (4/8) Epoch 6, batch 3050, loss[loss=0.2175, simple_loss=0.3026, pruned_loss=0.06621, over 16726.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2874, pruned_loss=0.06626, over 3328309.49 frames. ], batch size: 57, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:27:52,251 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1063, 3.9097, 4.1170, 4.3269, 4.4303, 3.9938, 4.1092, 4.3435], device='cuda:4'), covar=tensor([0.1031, 0.0808, 0.1233, 0.0538, 0.0469, 0.1124, 0.1926, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0445, 0.0538, 0.0697, 0.0553, 0.0417, 0.0414, 0.0437, 0.0470], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:27:54,619 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6032, 2.4141, 1.9878, 2.3682, 2.8664, 2.7087, 3.6113, 3.1391], device='cuda:4'), covar=tensor([0.0036, 0.0202, 0.0265, 0.0219, 0.0129, 0.0185, 0.0090, 0.0116], device='cuda:4'), in_proj_covar=tensor([0.0095, 0.0164, 0.0164, 0.0161, 0.0162, 0.0167, 0.0155, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:28:25,084 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0766, 1.7840, 2.2997, 2.9657, 2.8103, 3.4310, 2.0910, 3.1791], device='cuda:4'), covar=tensor([0.0092, 0.0215, 0.0155, 0.0121, 0.0122, 0.0082, 0.0202, 0.0081], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0152, 0.0136, 0.0136, 0.0141, 0.0103, 0.0145, 0.0088], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 11:28:30,035 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-28 11:28:35,659 INFO [train.py:904] (4/8) Epoch 6, batch 3100, loss[loss=0.2188, simple_loss=0.3078, pruned_loss=0.06492, over 16794.00 frames. ], tot_loss[loss=0.211, simple_loss=0.2886, pruned_loss=0.06669, over 3327333.08 frames. ], batch size: 57, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:28:38,484 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3071, 5.2849, 5.1254, 4.9497, 4.7214, 5.2266, 5.2424, 4.7934], device='cuda:4'), covar=tensor([0.0404, 0.0247, 0.0184, 0.0173, 0.0933, 0.0232, 0.0188, 0.0530], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0228, 0.0249, 0.0220, 0.0281, 0.0249, 0.0175, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:29:05,910 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.546e+02 2.841e+02 3.355e+02 4.130e+02 6.611e+02, threshold=6.710e+02, percent-clipped=0.0 2023-04-28 11:29:40,275 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53897.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:29:45,136 INFO [train.py:904] (4/8) Epoch 6, batch 3150, loss[loss=0.2146, simple_loss=0.3076, pruned_loss=0.06085, over 16656.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.2877, pruned_loss=0.06597, over 3325476.81 frames. ], batch size: 57, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:30:33,087 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2105, 4.3268, 4.4913, 2.1031, 4.7949, 4.8073, 3.3067, 3.7231], device='cuda:4'), covar=tensor([0.0571, 0.0121, 0.0191, 0.1107, 0.0058, 0.0081, 0.0364, 0.0305], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0093, 0.0085, 0.0139, 0.0073, 0.0085, 0.0118, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 11:30:49,057 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=53945.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:30:56,670 INFO [train.py:904] (4/8) Epoch 6, batch 3200, loss[loss=0.2161, simple_loss=0.2866, pruned_loss=0.07278, over 16567.00 frames. ], tot_loss[loss=0.2087, simple_loss=0.2866, pruned_loss=0.06539, over 3325713.89 frames. ], batch size: 75, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:31:09,534 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53960.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:31:25,882 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53971.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:31:26,626 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.709e+02 2.936e+02 3.448e+02 4.117e+02 1.255e+03, threshold=6.896e+02, percent-clipped=5.0 2023-04-28 11:32:09,691 INFO [train.py:904] (4/8) Epoch 6, batch 3250, loss[loss=0.2124, simple_loss=0.288, pruned_loss=0.06839, over 16020.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2866, pruned_loss=0.06553, over 3316932.71 frames. ], batch size: 35, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:32:19,695 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=54008.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:32:35,405 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=54019.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:32:40,146 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=54023.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 11:33:17,039 INFO [train.py:904] (4/8) Epoch 6, batch 3300, loss[loss=0.249, simple_loss=0.3153, pruned_loss=0.09133, over 16683.00 frames. ], tot_loss[loss=0.2092, simple_loss=0.2871, pruned_loss=0.06567, over 3311984.63 frames. ], batch size: 134, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:33:45,159 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=54071.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:33:46,070 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.774e+02 3.091e+02 3.848e+02 4.464e+02 8.976e+02, threshold=7.696e+02, percent-clipped=3.0 2023-04-28 11:33:55,959 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7463, 4.0681, 2.0602, 4.2933, 2.7968, 4.4073, 2.3619, 3.0131], device='cuda:4'), covar=tensor([0.0160, 0.0251, 0.1392, 0.0081, 0.0707, 0.0290, 0.1229, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0165, 0.0177, 0.0090, 0.0164, 0.0202, 0.0187, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 11:34:26,275 INFO [train.py:904] (4/8) Epoch 6, batch 3350, loss[loss=0.1841, simple_loss=0.2712, pruned_loss=0.04855, over 17153.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.2881, pruned_loss=0.06651, over 3308067.08 frames. ], batch size: 47, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:35:12,755 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0099, 5.4868, 5.5395, 5.5367, 5.5171, 6.0369, 5.6753, 5.4352], device='cuda:4'), covar=tensor([0.0740, 0.1329, 0.1457, 0.1407, 0.2183, 0.0701, 0.1026, 0.1993], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0428, 0.0423, 0.0366, 0.0495, 0.0450, 0.0346, 0.0495], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:35:34,480 INFO [train.py:904] (4/8) Epoch 6, batch 3400, loss[loss=0.195, simple_loss=0.2846, pruned_loss=0.05273, over 17127.00 frames. ], tot_loss[loss=0.2109, simple_loss=0.2884, pruned_loss=0.06676, over 3310136.78 frames. ], batch size: 49, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:36:05,329 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.856e+02 2.871e+02 3.468e+02 4.191e+02 6.688e+02, threshold=6.936e+02, percent-clipped=0.0 2023-04-28 11:36:28,208 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 11:36:46,918 INFO [train.py:904] (4/8) Epoch 6, batch 3450, loss[loss=0.1681, simple_loss=0.2497, pruned_loss=0.04322, over 17244.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.287, pruned_loss=0.06584, over 3304366.16 frames. ], batch size: 44, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:36:48,933 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-28 11:37:58,612 INFO [train.py:904] (4/8) Epoch 6, batch 3500, loss[loss=0.2018, simple_loss=0.2672, pruned_loss=0.06815, over 16886.00 frames. ], tot_loss[loss=0.2067, simple_loss=0.2851, pruned_loss=0.0642, over 3313611.51 frames. ], batch size: 116, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:38:28,902 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.000e+02 2.700e+02 3.164e+02 3.995e+02 8.205e+02, threshold=6.329e+02, percent-clipped=2.0 2023-04-28 11:39:10,634 INFO [train.py:904] (4/8) Epoch 6, batch 3550, loss[loss=0.185, simple_loss=0.2812, pruned_loss=0.04446, over 17110.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.2836, pruned_loss=0.06341, over 3324240.82 frames. ], batch size: 47, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:39:17,722 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.44 vs. limit=5.0 2023-04-28 11:40:05,052 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4565, 3.3857, 3.4185, 2.9636, 3.2943, 2.1521, 3.1117, 2.8568], device='cuda:4'), covar=tensor([0.0083, 0.0073, 0.0117, 0.0246, 0.0066, 0.1558, 0.0106, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0106, 0.0095, 0.0146, 0.0142, 0.0109, 0.0151, 0.0126, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:40:21,754 INFO [train.py:904] (4/8) Epoch 6, batch 3600, loss[loss=0.1765, simple_loss=0.247, pruned_loss=0.05296, over 16873.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2828, pruned_loss=0.06341, over 3318368.79 frames. ], batch size: 96, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:40:51,130 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.869e+02 3.018e+02 3.592e+02 4.498e+02 7.311e+02, threshold=7.184e+02, percent-clipped=7.0 2023-04-28 11:40:55,248 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5343, 4.5480, 4.6351, 4.6510, 4.5764, 5.1929, 4.7510, 4.4506], device='cuda:4'), covar=tensor([0.1161, 0.1519, 0.1442, 0.1801, 0.2440, 0.1041, 0.1279, 0.2631], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0429, 0.0425, 0.0366, 0.0492, 0.0454, 0.0350, 0.0494], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:41:18,335 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-28 11:41:26,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8311, 4.9457, 5.3768, 5.3812, 5.3267, 4.9831, 5.0440, 4.6687], device='cuda:4'), covar=tensor([0.0231, 0.0390, 0.0323, 0.0330, 0.0383, 0.0231, 0.0621, 0.0322], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0272, 0.0266, 0.0266, 0.0327, 0.0283, 0.0397, 0.0236], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 11:41:33,847 INFO [train.py:904] (4/8) Epoch 6, batch 3650, loss[loss=0.2026, simple_loss=0.2658, pruned_loss=0.06975, over 16808.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2811, pruned_loss=0.06384, over 3309377.59 frames. ], batch size: 83, lr: 1.15e-02, grad_scale: 8.0 2023-04-28 11:42:46,526 INFO [train.py:904] (4/8) Epoch 6, batch 3700, loss[loss=0.2277, simple_loss=0.2916, pruned_loss=0.08187, over 11429.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2801, pruned_loss=0.06541, over 3293799.37 frames. ], batch size: 247, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 11:42:49,081 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9292, 3.5368, 3.0207, 1.8076, 2.5574, 2.1627, 3.3636, 3.3838], device='cuda:4'), covar=tensor([0.0237, 0.0503, 0.0534, 0.1575, 0.0749, 0.0869, 0.0525, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0135, 0.0153, 0.0138, 0.0131, 0.0122, 0.0137, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 11:43:17,519 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.788e+02 2.928e+02 3.594e+02 4.333e+02 7.725e+02, threshold=7.187e+02, percent-clipped=2.0 2023-04-28 11:43:59,568 INFO [train.py:904] (4/8) Epoch 6, batch 3750, loss[loss=0.1943, simple_loss=0.2648, pruned_loss=0.06195, over 16729.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2809, pruned_loss=0.06735, over 3269935.77 frames. ], batch size: 76, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 11:44:13,124 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6132, 5.6284, 5.4373, 4.8334, 5.4574, 2.4287, 5.3129, 5.4585], device='cuda:4'), covar=tensor([0.0034, 0.0026, 0.0069, 0.0224, 0.0042, 0.1503, 0.0059, 0.0072], device='cuda:4'), in_proj_covar=tensor([0.0104, 0.0093, 0.0143, 0.0138, 0.0108, 0.0149, 0.0124, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:45:08,603 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7231, 3.9374, 2.9449, 2.2847, 2.8122, 2.1910, 3.8533, 3.8275], device='cuda:4'), covar=tensor([0.2163, 0.0564, 0.1330, 0.1993, 0.2171, 0.1566, 0.0521, 0.0756], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0256, 0.0271, 0.0254, 0.0294, 0.0207, 0.0250, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:45:13,203 INFO [train.py:904] (4/8) Epoch 6, batch 3800, loss[loss=0.2416, simple_loss=0.3074, pruned_loss=0.08794, over 16901.00 frames. ], tot_loss[loss=0.2102, simple_loss=0.2825, pruned_loss=0.06892, over 3270666.36 frames. ], batch size: 96, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 11:45:46,020 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.146e+02 2.814e+02 3.370e+02 4.517e+02 7.386e+02, threshold=6.740e+02, percent-clipped=3.0 2023-04-28 11:45:50,510 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6101, 4.6996, 4.7682, 4.8967, 4.8353, 5.3374, 4.9609, 4.6843], device='cuda:4'), covar=tensor([0.1201, 0.1430, 0.1274, 0.1547, 0.2262, 0.0925, 0.1088, 0.2151], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0426, 0.0418, 0.0363, 0.0485, 0.0443, 0.0345, 0.0489], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:46:20,436 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2938, 4.2523, 4.1838, 4.1241, 3.9106, 4.2177, 3.9571, 4.0514], device='cuda:4'), covar=tensor([0.0470, 0.0362, 0.0229, 0.0190, 0.0791, 0.0353, 0.0577, 0.0442], device='cuda:4'), in_proj_covar=tensor([0.0206, 0.0224, 0.0245, 0.0217, 0.0277, 0.0245, 0.0172, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:46:28,644 INFO [train.py:904] (4/8) Epoch 6, batch 3850, loss[loss=0.1952, simple_loss=0.2723, pruned_loss=0.05908, over 16380.00 frames. ], tot_loss[loss=0.21, simple_loss=0.282, pruned_loss=0.069, over 3265239.03 frames. ], batch size: 68, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 11:46:31,595 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-28 11:47:40,658 INFO [train.py:904] (4/8) Epoch 6, batch 3900, loss[loss=0.2201, simple_loss=0.2865, pruned_loss=0.07687, over 16509.00 frames. ], tot_loss[loss=0.2102, simple_loss=0.2816, pruned_loss=0.0694, over 3263009.43 frames. ], batch size: 146, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 11:47:59,633 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=54664.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 11:48:10,861 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.870e+02 2.751e+02 3.451e+02 4.155e+02 1.044e+03, threshold=6.903e+02, percent-clipped=2.0 2023-04-28 11:48:29,792 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2433, 4.0288, 4.1862, 4.4196, 4.5106, 4.0767, 4.2587, 4.4819], device='cuda:4'), covar=tensor([0.0881, 0.0758, 0.1296, 0.0507, 0.0488, 0.0948, 0.1250, 0.0483], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0514, 0.0664, 0.0535, 0.0402, 0.0395, 0.0418, 0.0450], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:48:53,478 INFO [train.py:904] (4/8) Epoch 6, batch 3950, loss[loss=0.1866, simple_loss=0.257, pruned_loss=0.05805, over 16504.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2809, pruned_loss=0.06987, over 3269379.87 frames. ], batch size: 75, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 11:49:27,768 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=54725.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 11:50:04,041 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6441, 2.6985, 2.4376, 4.5960, 3.7326, 4.2694, 1.5856, 3.0652], device='cuda:4'), covar=tensor([0.1321, 0.0635, 0.1166, 0.0068, 0.0269, 0.0264, 0.1318, 0.0751], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0143, 0.0166, 0.0097, 0.0191, 0.0187, 0.0159, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 11:50:04,726 INFO [train.py:904] (4/8) Epoch 6, batch 4000, loss[loss=0.2123, simple_loss=0.2968, pruned_loss=0.06393, over 16603.00 frames. ], tot_loss[loss=0.2113, simple_loss=0.2818, pruned_loss=0.07043, over 3273930.21 frames. ], batch size: 76, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 11:50:36,654 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.045e+02 2.875e+02 3.507e+02 4.223e+02 8.153e+02, threshold=7.015e+02, percent-clipped=4.0 2023-04-28 11:50:45,071 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0817, 5.0923, 4.9393, 4.7953, 4.5605, 4.9773, 4.8631, 4.6569], device='cuda:4'), covar=tensor([0.0363, 0.0194, 0.0150, 0.0161, 0.0752, 0.0190, 0.0248, 0.0431], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0219, 0.0240, 0.0213, 0.0269, 0.0240, 0.0168, 0.0266], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:51:10,365 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 11:51:17,347 INFO [train.py:904] (4/8) Epoch 6, batch 4050, loss[loss=0.2084, simple_loss=0.2973, pruned_loss=0.05977, over 17118.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2815, pruned_loss=0.06881, over 3274684.55 frames. ], batch size: 47, lr: 1.14e-02, grad_scale: 4.0 2023-04-28 11:52:09,229 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2319, 4.4061, 1.8737, 4.7491, 2.7949, 4.6590, 2.0103, 3.0551], device='cuda:4'), covar=tensor([0.0096, 0.0196, 0.1815, 0.0031, 0.0663, 0.0177, 0.1704, 0.0581], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0163, 0.0180, 0.0087, 0.0165, 0.0196, 0.0188, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 11:52:22,775 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-28 11:52:28,288 INFO [train.py:904] (4/8) Epoch 6, batch 4100, loss[loss=0.2057, simple_loss=0.2824, pruned_loss=0.06456, over 16624.00 frames. ], tot_loss[loss=0.2085, simple_loss=0.282, pruned_loss=0.06753, over 3265742.42 frames. ], batch size: 57, lr: 1.14e-02, grad_scale: 4.0 2023-04-28 11:52:45,812 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=54862.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:52:57,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1261, 5.6260, 5.7774, 5.6300, 5.7126, 6.1479, 5.7808, 5.5649], device='cuda:4'), covar=tensor([0.0637, 0.1255, 0.1126, 0.1456, 0.1873, 0.0770, 0.0936, 0.1617], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0421, 0.0413, 0.0363, 0.0484, 0.0439, 0.0344, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 11:53:01,849 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.687e+02 2.498e+02 3.136e+02 3.951e+02 8.635e+02, threshold=6.272e+02, percent-clipped=2.0 2023-04-28 11:53:11,395 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9445, 4.9567, 4.8009, 3.7931, 4.9174, 1.6977, 4.6553, 4.7236], device='cuda:4'), covar=tensor([0.0059, 0.0043, 0.0094, 0.0399, 0.0049, 0.2195, 0.0076, 0.0133], device='cuda:4'), in_proj_covar=tensor([0.0103, 0.0091, 0.0140, 0.0138, 0.0106, 0.0148, 0.0122, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 11:53:43,637 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9380, 4.1004, 1.9610, 4.5423, 2.6530, 4.4454, 2.1890, 2.9263], device='cuda:4'), covar=tensor([0.0133, 0.0244, 0.1640, 0.0030, 0.0793, 0.0253, 0.1384, 0.0668], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0163, 0.0182, 0.0087, 0.0166, 0.0198, 0.0188, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 11:53:45,438 INFO [train.py:904] (4/8) Epoch 6, batch 4150, loss[loss=0.237, simple_loss=0.3129, pruned_loss=0.08054, over 16803.00 frames. ], tot_loss[loss=0.2147, simple_loss=0.2892, pruned_loss=0.0701, over 3260816.83 frames. ], batch size: 83, lr: 1.14e-02, grad_scale: 4.0 2023-04-28 11:54:11,006 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-28 11:54:19,558 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=54923.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 11:55:01,110 INFO [train.py:904] (4/8) Epoch 6, batch 4200, loss[loss=0.2778, simple_loss=0.3409, pruned_loss=0.1073, over 11554.00 frames. ], tot_loss[loss=0.2217, simple_loss=0.2973, pruned_loss=0.07298, over 3216195.75 frames. ], batch size: 246, lr: 1.14e-02, grad_scale: 4.0 2023-04-28 11:55:01,980 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.84 vs. limit=5.0 2023-04-28 11:55:12,467 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0854, 1.4307, 1.7733, 2.1320, 2.0906, 2.2902, 1.4546, 2.1041], device='cuda:4'), covar=tensor([0.0090, 0.0224, 0.0108, 0.0119, 0.0115, 0.0087, 0.0209, 0.0054], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0151, 0.0135, 0.0133, 0.0140, 0.0101, 0.0145, 0.0088], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 11:55:33,716 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.052e+02 3.214e+02 3.756e+02 4.661e+02 1.168e+03, threshold=7.512e+02, percent-clipped=6.0 2023-04-28 11:56:05,875 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-28 11:56:15,276 INFO [train.py:904] (4/8) Epoch 6, batch 4250, loss[loss=0.2168, simple_loss=0.2888, pruned_loss=0.07245, over 12153.00 frames. ], tot_loss[loss=0.2233, simple_loss=0.3005, pruned_loss=0.07303, over 3193878.63 frames. ], batch size: 247, lr: 1.14e-02, grad_scale: 4.0 2023-04-28 11:56:44,700 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=55020.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 11:57:29,393 INFO [train.py:904] (4/8) Epoch 6, batch 4300, loss[loss=0.2459, simple_loss=0.326, pruned_loss=0.08285, over 15413.00 frames. ], tot_loss[loss=0.2227, simple_loss=0.3016, pruned_loss=0.07189, over 3192704.81 frames. ], batch size: 191, lr: 1.14e-02, grad_scale: 4.0 2023-04-28 11:57:34,978 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4230, 1.4777, 1.9029, 2.4536, 2.4327, 2.7367, 1.4011, 2.4554], device='cuda:4'), covar=tensor([0.0091, 0.0256, 0.0163, 0.0138, 0.0124, 0.0066, 0.0258, 0.0058], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0150, 0.0134, 0.0133, 0.0139, 0.0100, 0.0145, 0.0087], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 11:58:01,801 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.697e+02 2.606e+02 3.247e+02 3.921e+02 6.703e+02, threshold=6.494e+02, percent-clipped=0.0 2023-04-28 11:58:43,132 INFO [train.py:904] (4/8) Epoch 6, batch 4350, loss[loss=0.2315, simple_loss=0.311, pruned_loss=0.07604, over 17116.00 frames. ], tot_loss[loss=0.2264, simple_loss=0.3058, pruned_loss=0.07356, over 3185744.69 frames. ], batch size: 49, lr: 1.14e-02, grad_scale: 4.0 2023-04-28 11:59:00,694 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6371, 2.9385, 2.6381, 4.7763, 3.9205, 4.3872, 1.7813, 2.9808], device='cuda:4'), covar=tensor([0.1421, 0.0661, 0.1119, 0.0120, 0.0347, 0.0260, 0.1470, 0.0822], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0144, 0.0168, 0.0095, 0.0192, 0.0188, 0.0162, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 11:59:46,839 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2054, 4.2623, 4.7009, 4.6475, 4.6747, 4.2821, 4.3418, 4.0289], device='cuda:4'), covar=tensor([0.0235, 0.0314, 0.0273, 0.0350, 0.0357, 0.0258, 0.0690, 0.0406], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0247, 0.0245, 0.0241, 0.0296, 0.0261, 0.0363, 0.0217], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 11:59:52,994 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=55148.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 11:59:56,586 INFO [train.py:904] (4/8) Epoch 6, batch 4400, loss[loss=0.2564, simple_loss=0.3365, pruned_loss=0.08819, over 16803.00 frames. ], tot_loss[loss=0.2284, simple_loss=0.3077, pruned_loss=0.07452, over 3180621.87 frames. ], batch size: 124, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 12:00:27,146 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.875e+02 2.770e+02 3.222e+02 4.108e+02 7.043e+02, threshold=6.445e+02, percent-clipped=2.0 2023-04-28 12:01:04,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3535, 4.0162, 4.0091, 2.8126, 3.4881, 3.9383, 3.7432, 2.3977], device='cuda:4'), covar=tensor([0.0329, 0.0012, 0.0019, 0.0208, 0.0037, 0.0053, 0.0024, 0.0233], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0057, 0.0059, 0.0113, 0.0062, 0.0069, 0.0064, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 12:01:06,617 INFO [train.py:904] (4/8) Epoch 6, batch 4450, loss[loss=0.2372, simple_loss=0.3276, pruned_loss=0.07339, over 16695.00 frames. ], tot_loss[loss=0.2299, simple_loss=0.3102, pruned_loss=0.07473, over 3192648.23 frames. ], batch size: 89, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 12:01:19,234 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=55209.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:01:21,602 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=55211.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:01:31,358 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=55218.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 12:01:55,293 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.98 vs. limit=5.0 2023-04-28 12:02:16,913 INFO [train.py:904] (4/8) Epoch 6, batch 4500, loss[loss=0.2167, simple_loss=0.2968, pruned_loss=0.06833, over 16385.00 frames. ], tot_loss[loss=0.2288, simple_loss=0.3095, pruned_loss=0.07405, over 3202919.72 frames. ], batch size: 68, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 12:02:44,583 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-28 12:02:47,712 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=55272.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:02:48,408 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.800e+02 2.322e+02 2.860e+02 3.458e+02 5.535e+02, threshold=5.719e+02, percent-clipped=0.0 2023-04-28 12:03:03,029 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-28 12:03:29,863 INFO [train.py:904] (4/8) Epoch 6, batch 4550, loss[loss=0.2217, simple_loss=0.3087, pruned_loss=0.06737, over 16747.00 frames. ], tot_loss[loss=0.2293, simple_loss=0.3098, pruned_loss=0.07441, over 3209935.94 frames. ], batch size: 124, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 12:03:49,033 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7107, 2.0696, 1.4684, 1.9825, 2.6189, 2.2529, 2.8485, 2.8299], device='cuda:4'), covar=tensor([0.0046, 0.0196, 0.0288, 0.0227, 0.0100, 0.0188, 0.0073, 0.0099], device='cuda:4'), in_proj_covar=tensor([0.0089, 0.0158, 0.0160, 0.0159, 0.0155, 0.0162, 0.0144, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:03:57,461 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=55320.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 12:04:41,225 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-28 12:04:41,598 INFO [train.py:904] (4/8) Epoch 6, batch 4600, loss[loss=0.2107, simple_loss=0.2947, pruned_loss=0.06338, over 16430.00 frames. ], tot_loss[loss=0.2291, simple_loss=0.3105, pruned_loss=0.07384, over 3230662.35 frames. ], batch size: 35, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 12:05:07,048 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=55368.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 12:05:13,322 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.543e+02 2.321e+02 2.807e+02 3.689e+02 1.502e+03, threshold=5.614e+02, percent-clipped=3.0 2023-04-28 12:05:42,091 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3789, 4.1255, 4.3797, 4.6150, 4.7294, 4.2007, 4.6839, 4.6812], device='cuda:4'), covar=tensor([0.0901, 0.0808, 0.1209, 0.0424, 0.0368, 0.0790, 0.0407, 0.0398], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0489, 0.0636, 0.0500, 0.0381, 0.0381, 0.0395, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:05:52,743 INFO [train.py:904] (4/8) Epoch 6, batch 4650, loss[loss=0.2468, simple_loss=0.3267, pruned_loss=0.08342, over 16722.00 frames. ], tot_loss[loss=0.2277, simple_loss=0.3088, pruned_loss=0.07333, over 3230194.59 frames. ], batch size: 134, lr: 1.14e-02, grad_scale: 8.0 2023-04-28 12:06:47,976 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9100, 3.9447, 4.3174, 4.2780, 4.3054, 3.9784, 4.0262, 3.8216], device='cuda:4'), covar=tensor([0.0252, 0.0332, 0.0325, 0.0406, 0.0317, 0.0286, 0.0718, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0243, 0.0245, 0.0242, 0.0296, 0.0261, 0.0363, 0.0217], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 12:06:54,966 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-28 12:07:03,016 INFO [train.py:904] (4/8) Epoch 6, batch 4700, loss[loss=0.2079, simple_loss=0.2927, pruned_loss=0.06154, over 16800.00 frames. ], tot_loss[loss=0.2249, simple_loss=0.3061, pruned_loss=0.07184, over 3230744.42 frames. ], batch size: 83, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:07:15,472 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-28 12:07:32,552 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 12:07:34,099 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.678e+02 2.601e+02 3.111e+02 3.691e+02 9.251e+02, threshold=6.221e+02, percent-clipped=2.0 2023-04-28 12:08:12,119 INFO [train.py:904] (4/8) Epoch 6, batch 4750, loss[loss=0.1863, simple_loss=0.2721, pruned_loss=0.05024, over 16859.00 frames. ], tot_loss[loss=0.2216, simple_loss=0.3029, pruned_loss=0.0702, over 3216500.24 frames. ], batch size: 96, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:08:17,374 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=55504.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:08:36,542 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=55518.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 12:08:46,649 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3517, 3.4899, 1.7390, 3.7295, 2.3917, 3.5978, 1.9506, 2.6684], device='cuda:4'), covar=tensor([0.0147, 0.0258, 0.1539, 0.0044, 0.0759, 0.0418, 0.1324, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0155, 0.0176, 0.0082, 0.0160, 0.0187, 0.0183, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 12:09:22,944 INFO [train.py:904] (4/8) Epoch 6, batch 4800, loss[loss=0.2147, simple_loss=0.3003, pruned_loss=0.06457, over 16680.00 frames. ], tot_loss[loss=0.2184, simple_loss=0.2999, pruned_loss=0.06841, over 3218704.96 frames. ], batch size: 134, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:09:45,467 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=55566.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:09:46,592 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=55567.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:09:53,990 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.426e+02 2.469e+02 2.878e+02 3.568e+02 5.612e+02, threshold=5.756e+02, percent-clipped=0.0 2023-04-28 12:10:16,369 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-28 12:10:35,250 INFO [train.py:904] (4/8) Epoch 6, batch 4850, loss[loss=0.2415, simple_loss=0.3216, pruned_loss=0.08067, over 16590.00 frames. ], tot_loss[loss=0.2196, simple_loss=0.3014, pruned_loss=0.06885, over 3198368.12 frames. ], batch size: 62, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:10:55,278 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 12:11:18,618 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6646, 2.2375, 2.3150, 4.3845, 1.9981, 2.9870, 2.3403, 2.4227], device='cuda:4'), covar=tensor([0.0590, 0.2146, 0.1222, 0.0246, 0.2914, 0.1109, 0.1997, 0.2054], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0329, 0.0272, 0.0303, 0.0373, 0.0342, 0.0297, 0.0393], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:11:37,190 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0915, 3.8570, 4.0812, 4.2639, 4.4418, 3.9865, 4.3480, 4.3922], device='cuda:4'), covar=tensor([0.0949, 0.0819, 0.1294, 0.0591, 0.0395, 0.1002, 0.0555, 0.0425], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0484, 0.0622, 0.0499, 0.0377, 0.0375, 0.0392, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:11:48,114 INFO [train.py:904] (4/8) Epoch 6, batch 4900, loss[loss=0.1982, simple_loss=0.2815, pruned_loss=0.05742, over 17120.00 frames. ], tot_loss[loss=0.2182, simple_loss=0.3008, pruned_loss=0.06782, over 3194998.94 frames. ], batch size: 49, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:11:54,580 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-28 12:12:11,398 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4504, 1.8681, 1.5338, 1.7315, 2.2923, 2.0395, 2.3249, 2.4115], device='cuda:4'), covar=tensor([0.0045, 0.0206, 0.0268, 0.0293, 0.0109, 0.0195, 0.0091, 0.0117], device='cuda:4'), in_proj_covar=tensor([0.0087, 0.0157, 0.0159, 0.0157, 0.0152, 0.0162, 0.0142, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:12:19,090 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.651e+02 2.308e+02 2.929e+02 3.604e+02 9.107e+02, threshold=5.857e+02, percent-clipped=2.0 2023-04-28 12:12:39,919 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-28 12:12:59,378 INFO [train.py:904] (4/8) Epoch 6, batch 4950, loss[loss=0.2573, simple_loss=0.33, pruned_loss=0.09232, over 11643.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.3008, pruned_loss=0.06746, over 3177410.69 frames. ], batch size: 246, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:14:05,627 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=55748.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:14:08,593 INFO [train.py:904] (4/8) Epoch 6, batch 5000, loss[loss=0.2078, simple_loss=0.2997, pruned_loss=0.05798, over 16821.00 frames. ], tot_loss[loss=0.2188, simple_loss=0.3025, pruned_loss=0.06758, over 3193855.64 frames. ], batch size: 102, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:14:38,593 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.944e+02 2.723e+02 3.199e+02 3.755e+02 7.199e+02, threshold=6.398e+02, percent-clipped=5.0 2023-04-28 12:14:53,882 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6989, 3.6171, 3.7442, 3.5420, 3.7215, 4.1681, 3.8997, 3.5408], device='cuda:4'), covar=tensor([0.1815, 0.1554, 0.1534, 0.2133, 0.2366, 0.1453, 0.1188, 0.2444], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0392, 0.0393, 0.0341, 0.0457, 0.0424, 0.0327, 0.0463], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 12:15:21,197 INFO [train.py:904] (4/8) Epoch 6, batch 5050, loss[loss=0.2022, simple_loss=0.2882, pruned_loss=0.05805, over 17196.00 frames. ], tot_loss[loss=0.2184, simple_loss=0.3024, pruned_loss=0.06722, over 3195138.85 frames. ], batch size: 45, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:15:25,527 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=55804.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:15:33,241 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=55809.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:15:37,931 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5191, 4.3203, 4.1437, 2.8232, 3.6463, 4.1791, 3.6457, 1.8651], device='cuda:4'), covar=tensor([0.0408, 0.0050, 0.0043, 0.0278, 0.0086, 0.0206, 0.0244, 0.0430], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0056, 0.0060, 0.0116, 0.0063, 0.0071, 0.0065, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 12:16:18,889 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.46 vs. limit=5.0 2023-04-28 12:16:18,930 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 12:16:28,576 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=55848.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:16:31,556 INFO [train.py:904] (4/8) Epoch 6, batch 5100, loss[loss=0.2044, simple_loss=0.2925, pruned_loss=0.0581, over 16470.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.3003, pruned_loss=0.06609, over 3205338.35 frames. ], batch size: 146, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:16:32,963 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=55852.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:16:54,906 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=55867.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:17:04,477 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.683e+02 3.138e+02 3.930e+02 6.713e+02, threshold=6.277e+02, percent-clipped=2.0 2023-04-28 12:17:45,242 INFO [train.py:904] (4/8) Epoch 6, batch 5150, loss[loss=0.1993, simple_loss=0.287, pruned_loss=0.05581, over 16589.00 frames. ], tot_loss[loss=0.2151, simple_loss=0.2997, pruned_loss=0.0653, over 3199925.13 frames. ], batch size: 57, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:17:57,067 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=55909.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:18:01,900 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-28 12:18:06,164 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=55915.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:18:27,082 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0137, 3.7873, 3.7299, 2.3863, 3.2924, 3.5412, 3.5140, 1.8248], device='cuda:4'), covar=tensor([0.0370, 0.0019, 0.0021, 0.0235, 0.0044, 0.0065, 0.0036, 0.0343], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0055, 0.0058, 0.0112, 0.0061, 0.0070, 0.0064, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 12:18:56,200 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.99 vs. limit=5.0 2023-04-28 12:18:58,946 INFO [train.py:904] (4/8) Epoch 6, batch 5200, loss[loss=0.2183, simple_loss=0.3034, pruned_loss=0.06661, over 15451.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2984, pruned_loss=0.06474, over 3193010.87 frames. ], batch size: 191, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:19:17,080 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.45 vs. limit=5.0 2023-04-28 12:19:30,989 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.992e+02 2.558e+02 3.011e+02 3.644e+02 8.071e+02, threshold=6.021e+02, percent-clipped=1.0 2023-04-28 12:20:15,818 INFO [train.py:904] (4/8) Epoch 6, batch 5250, loss[loss=0.2015, simple_loss=0.2834, pruned_loss=0.05975, over 16533.00 frames. ], tot_loss[loss=0.2124, simple_loss=0.2958, pruned_loss=0.06454, over 3208890.53 frames. ], batch size: 68, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:21:11,111 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-28 12:21:20,200 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8340, 1.9683, 2.2131, 3.2181, 1.9974, 2.4907, 2.2824, 2.0305], device='cuda:4'), covar=tensor([0.0745, 0.2120, 0.1168, 0.0393, 0.2691, 0.1285, 0.1879, 0.2223], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0332, 0.0275, 0.0310, 0.0376, 0.0345, 0.0299, 0.0392], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:21:28,014 INFO [train.py:904] (4/8) Epoch 6, batch 5300, loss[loss=0.1723, simple_loss=0.2565, pruned_loss=0.04404, over 16752.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2909, pruned_loss=0.06232, over 3216870.18 frames. ], batch size: 89, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:21:33,600 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6286, 3.9443, 4.1324, 1.9638, 4.4268, 4.4123, 3.1437, 3.2319], device='cuda:4'), covar=tensor([0.0723, 0.0142, 0.0138, 0.1153, 0.0029, 0.0032, 0.0335, 0.0370], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0097, 0.0083, 0.0144, 0.0073, 0.0080, 0.0120, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 12:21:59,801 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.725e+02 2.658e+02 3.030e+02 3.663e+02 6.971e+02, threshold=6.060e+02, percent-clipped=3.0 2023-04-28 12:22:23,273 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.04 vs. limit=2.0 2023-04-28 12:22:40,507 INFO [train.py:904] (4/8) Epoch 6, batch 5350, loss[loss=0.1828, simple_loss=0.2745, pruned_loss=0.04552, over 16562.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.289, pruned_loss=0.06201, over 3212080.59 frames. ], batch size: 68, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:22:45,576 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=56104.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:23:45,019 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56145.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:23:52,596 INFO [train.py:904] (4/8) Epoch 6, batch 5400, loss[loss=0.2167, simple_loss=0.3006, pruned_loss=0.06641, over 16936.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.292, pruned_loss=0.06276, over 3220587.73 frames. ], batch size: 109, lr: 1.13e-02, grad_scale: 4.0 2023-04-28 12:24:26,635 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.771e+02 2.610e+02 3.124e+02 3.605e+02 7.209e+02, threshold=6.248e+02, percent-clipped=3.0 2023-04-28 12:25:08,991 INFO [train.py:904] (4/8) Epoch 6, batch 5450, loss[loss=0.2389, simple_loss=0.3074, pruned_loss=0.0852, over 17061.00 frames. ], tot_loss[loss=0.2121, simple_loss=0.2952, pruned_loss=0.0645, over 3223555.08 frames. ], batch size: 53, lr: 1.13e-02, grad_scale: 4.0 2023-04-28 12:25:14,996 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=56204.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:25:17,606 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=56206.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:25:24,326 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9691, 1.7308, 1.4278, 1.5683, 1.9128, 1.6697, 1.8335, 1.9749], device='cuda:4'), covar=tensor([0.0041, 0.0112, 0.0167, 0.0134, 0.0080, 0.0118, 0.0081, 0.0074], device='cuda:4'), in_proj_covar=tensor([0.0090, 0.0163, 0.0164, 0.0164, 0.0158, 0.0168, 0.0147, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:25:37,922 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8337, 5.1317, 4.8347, 4.8575, 4.5633, 4.5272, 4.7080, 5.2165], device='cuda:4'), covar=tensor([0.0726, 0.0686, 0.0998, 0.0622, 0.0651, 0.0700, 0.0711, 0.0674], device='cuda:4'), in_proj_covar=tensor([0.0392, 0.0516, 0.0438, 0.0337, 0.0317, 0.0331, 0.0426, 0.0372], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:25:39,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1592, 2.9601, 2.6310, 1.9847, 2.5672, 2.1792, 2.6913, 2.9172], device='cuda:4'), covar=tensor([0.0272, 0.0441, 0.0426, 0.1267, 0.0575, 0.0798, 0.0512, 0.0437], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0124, 0.0150, 0.0137, 0.0128, 0.0122, 0.0135, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 12:26:27,337 INFO [train.py:904] (4/8) Epoch 6, batch 5500, loss[loss=0.336, simple_loss=0.3819, pruned_loss=0.145, over 11855.00 frames. ], tot_loss[loss=0.2244, simple_loss=0.3052, pruned_loss=0.07181, over 3175770.09 frames. ], batch size: 247, lr: 1.13e-02, grad_scale: 4.0 2023-04-28 12:26:37,081 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.4923, 2.7857, 3.0137, 4.4347, 3.7062, 4.1981, 1.2720, 3.4105], device='cuda:4'), covar=tensor([0.1432, 0.0589, 0.0847, 0.0108, 0.0223, 0.0293, 0.1575, 0.0593], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0146, 0.0173, 0.0096, 0.0194, 0.0190, 0.0166, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 12:27:01,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.186e+02 3.422e+02 4.497e+02 5.690e+02 1.317e+03, threshold=8.994e+02, percent-clipped=17.0 2023-04-28 12:27:46,056 INFO [train.py:904] (4/8) Epoch 6, batch 5550, loss[loss=0.364, simple_loss=0.3884, pruned_loss=0.1698, over 10871.00 frames. ], tot_loss[loss=0.2352, simple_loss=0.3137, pruned_loss=0.07833, over 3148728.01 frames. ], batch size: 247, lr: 1.13e-02, grad_scale: 4.0 2023-04-28 12:28:10,739 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9462, 3.1624, 3.3989, 3.3839, 3.3676, 3.1806, 3.0355, 3.2818], device='cuda:4'), covar=tensor([0.0484, 0.0629, 0.0459, 0.0551, 0.0602, 0.0514, 0.1220, 0.0509], device='cuda:4'), in_proj_covar=tensor([0.0255, 0.0247, 0.0252, 0.0251, 0.0303, 0.0273, 0.0375, 0.0222], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 12:29:07,882 INFO [train.py:904] (4/8) Epoch 6, batch 5600, loss[loss=0.2419, simple_loss=0.3246, pruned_loss=0.07954, over 16728.00 frames. ], tot_loss[loss=0.2439, simple_loss=0.3202, pruned_loss=0.08384, over 3122900.48 frames. ], batch size: 62, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:29:45,254 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.502e+02 4.330e+02 5.114e+02 6.903e+02 1.749e+03, threshold=1.023e+03, percent-clipped=9.0 2023-04-28 12:30:11,545 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6564, 2.7210, 2.4552, 4.1521, 3.1487, 3.9648, 1.4522, 2.7858], device='cuda:4'), covar=tensor([0.1308, 0.0550, 0.1111, 0.0084, 0.0230, 0.0338, 0.1437, 0.0788], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0144, 0.0170, 0.0095, 0.0192, 0.0189, 0.0162, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 12:30:29,972 INFO [train.py:904] (4/8) Epoch 6, batch 5650, loss[loss=0.3293, simple_loss=0.3709, pruned_loss=0.1439, over 11401.00 frames. ], tot_loss[loss=0.2565, simple_loss=0.3285, pruned_loss=0.09223, over 3042496.41 frames. ], batch size: 248, lr: 1.13e-02, grad_scale: 8.0 2023-04-28 12:30:34,928 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=56404.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:31:19,390 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 12:31:48,511 INFO [train.py:904] (4/8) Epoch 6, batch 5700, loss[loss=0.26, simple_loss=0.3324, pruned_loss=0.09376, over 16824.00 frames. ], tot_loss[loss=0.2612, simple_loss=0.3317, pruned_loss=0.09537, over 3021209.22 frames. ], batch size: 83, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:31:51,665 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=56452.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:32:25,510 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.692e+02 4.150e+02 5.062e+02 6.862e+02 1.724e+03, threshold=1.012e+03, percent-clipped=2.0 2023-04-28 12:32:41,731 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0319, 3.2733, 3.4809, 3.4815, 3.4665, 3.2484, 3.2949, 3.3005], device='cuda:4'), covar=tensor([0.0408, 0.0509, 0.0419, 0.0448, 0.0479, 0.0448, 0.0839, 0.0508], device='cuda:4'), in_proj_covar=tensor([0.0257, 0.0250, 0.0254, 0.0255, 0.0308, 0.0274, 0.0377, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 12:33:08,899 INFO [train.py:904] (4/8) Epoch 6, batch 5750, loss[loss=0.2648, simple_loss=0.32, pruned_loss=0.1048, over 11221.00 frames. ], tot_loss[loss=0.2638, simple_loss=0.3341, pruned_loss=0.09679, over 2994462.91 frames. ], batch size: 246, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:33:09,308 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=56501.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:33:13,968 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=56504.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:33:44,544 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56523.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 12:34:03,139 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.90 vs. limit=5.0 2023-04-28 12:34:30,321 INFO [train.py:904] (4/8) Epoch 6, batch 5800, loss[loss=0.2258, simple_loss=0.3094, pruned_loss=0.0711, over 16722.00 frames. ], tot_loss[loss=0.2609, simple_loss=0.3332, pruned_loss=0.09433, over 3010387.44 frames. ], batch size: 134, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:34:32,753 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=56552.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:35:05,967 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.365e+02 3.759e+02 4.783e+02 6.283e+02 1.634e+03, threshold=9.566e+02, percent-clipped=2.0 2023-04-28 12:35:23,433 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=56584.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 12:35:49,148 INFO [train.py:904] (4/8) Epoch 6, batch 5850, loss[loss=0.2587, simple_loss=0.3321, pruned_loss=0.09261, over 16889.00 frames. ], tot_loss[loss=0.2564, simple_loss=0.3301, pruned_loss=0.09136, over 3027668.92 frames. ], batch size: 116, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:37:11,643 INFO [train.py:904] (4/8) Epoch 6, batch 5900, loss[loss=0.2605, simple_loss=0.3352, pruned_loss=0.09295, over 17172.00 frames. ], tot_loss[loss=0.2545, simple_loss=0.3284, pruned_loss=0.09027, over 3031229.10 frames. ], batch size: 46, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:37:52,136 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.523e+02 3.504e+02 4.467e+02 5.504e+02 1.096e+03, threshold=8.934e+02, percent-clipped=2.0 2023-04-28 12:38:16,503 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-28 12:38:33,952 INFO [train.py:904] (4/8) Epoch 6, batch 5950, loss[loss=0.2481, simple_loss=0.3397, pruned_loss=0.07829, over 16781.00 frames. ], tot_loss[loss=0.2538, simple_loss=0.3298, pruned_loss=0.08885, over 3052165.16 frames. ], batch size: 124, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:39:31,523 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-28 12:39:52,035 INFO [train.py:904] (4/8) Epoch 6, batch 6000, loss[loss=0.3075, simple_loss=0.3497, pruned_loss=0.1326, over 11511.00 frames. ], tot_loss[loss=0.2533, simple_loss=0.3287, pruned_loss=0.08898, over 3041835.57 frames. ], batch size: 248, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:39:52,035 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 12:40:01,521 INFO [train.py:938] (4/8) Epoch 6, validation: loss=0.18, simple_loss=0.2922, pruned_loss=0.03386, over 944034.00 frames. 2023-04-28 12:40:01,521 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 12:40:36,509 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.218e+02 3.558e+02 4.306e+02 5.339e+02 1.394e+03, threshold=8.611e+02, percent-clipped=5.0 2023-04-28 12:41:19,686 INFO [train.py:904] (4/8) Epoch 6, batch 6050, loss[loss=0.2292, simple_loss=0.3155, pruned_loss=0.07147, over 16742.00 frames. ], tot_loss[loss=0.2517, simple_loss=0.327, pruned_loss=0.0882, over 3052484.72 frames. ], batch size: 124, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:41:20,598 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=56801.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:41:23,934 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56803.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:42:33,595 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=56849.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:42:36,667 INFO [train.py:904] (4/8) Epoch 6, batch 6100, loss[loss=0.2261, simple_loss=0.3085, pruned_loss=0.07191, over 16398.00 frames. ], tot_loss[loss=0.2487, simple_loss=0.325, pruned_loss=0.08613, over 3070396.49 frames. ], batch size: 68, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:42:58,752 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=56864.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:43:13,027 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.259e+02 3.305e+02 4.282e+02 5.206e+02 1.267e+03, threshold=8.564e+02, percent-clipped=3.0 2023-04-28 12:43:20,733 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=56879.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 12:43:56,259 INFO [train.py:904] (4/8) Epoch 6, batch 6150, loss[loss=0.2112, simple_loss=0.2977, pruned_loss=0.06238, over 16931.00 frames. ], tot_loss[loss=0.246, simple_loss=0.3226, pruned_loss=0.08467, over 3086398.80 frames. ], batch size: 109, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:44:08,672 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56908.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:45:17,271 INFO [train.py:904] (4/8) Epoch 6, batch 6200, loss[loss=0.2955, simple_loss=0.3432, pruned_loss=0.1238, over 11930.00 frames. ], tot_loss[loss=0.2448, simple_loss=0.3209, pruned_loss=0.08435, over 3083846.80 frames. ], batch size: 248, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:45:18,893 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4960, 3.5110, 3.2595, 3.1330, 3.0682, 3.4038, 3.2241, 3.1968], device='cuda:4'), covar=tensor([0.0467, 0.0349, 0.0202, 0.0179, 0.0557, 0.0291, 0.1058, 0.0415], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0211, 0.0223, 0.0196, 0.0251, 0.0226, 0.0158, 0.0254], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:45:46,284 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=56969.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:45:53,729 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.179e+02 3.479e+02 4.387e+02 5.634e+02 1.000e+03, threshold=8.774e+02, percent-clipped=2.0 2023-04-28 12:46:22,672 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56993.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:46:26,560 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56995.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:46:34,326 INFO [train.py:904] (4/8) Epoch 6, batch 6250, loss[loss=0.2428, simple_loss=0.318, pruned_loss=0.08377, over 16721.00 frames. ], tot_loss[loss=0.244, simple_loss=0.3201, pruned_loss=0.0839, over 3098123.49 frames. ], batch size: 124, lr: 1.12e-02, grad_scale: 4.0 2023-04-28 12:46:53,422 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3161, 4.6424, 5.0754, 5.0050, 4.9526, 4.6071, 4.1947, 4.1191], device='cuda:4'), covar=tensor([0.0534, 0.0542, 0.0361, 0.0470, 0.0473, 0.0401, 0.1143, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0257, 0.0250, 0.0254, 0.0249, 0.0303, 0.0271, 0.0372, 0.0223], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 12:47:15,299 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7559, 3.5875, 3.8050, 3.6980, 3.7623, 4.1532, 3.9095, 3.6509], device='cuda:4'), covar=tensor([0.1803, 0.2139, 0.1662, 0.2027, 0.2372, 0.1461, 0.1198, 0.2348], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0403, 0.0408, 0.0353, 0.0462, 0.0433, 0.0335, 0.0474], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 12:47:50,597 INFO [train.py:904] (4/8) Epoch 6, batch 6300, loss[loss=0.2357, simple_loss=0.3134, pruned_loss=0.07895, over 16160.00 frames. ], tot_loss[loss=0.2433, simple_loss=0.3198, pruned_loss=0.0834, over 3084478.84 frames. ], batch size: 35, lr: 1.12e-02, grad_scale: 4.0 2023-04-28 12:47:56,711 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57054.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:47:59,851 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57056.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:48:29,802 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.167e+02 3.660e+02 4.420e+02 5.766e+02 1.393e+03, threshold=8.841e+02, percent-clipped=5.0 2023-04-28 12:48:45,655 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7242, 2.5119, 1.9235, 2.3601, 3.1069, 2.8480, 3.5436, 3.4161], device='cuda:4'), covar=tensor([0.0024, 0.0216, 0.0297, 0.0239, 0.0119, 0.0161, 0.0098, 0.0090], device='cuda:4'), in_proj_covar=tensor([0.0086, 0.0160, 0.0164, 0.0162, 0.0156, 0.0165, 0.0148, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:49:09,391 INFO [train.py:904] (4/8) Epoch 6, batch 6350, loss[loss=0.2454, simple_loss=0.3172, pruned_loss=0.08682, over 17204.00 frames. ], tot_loss[loss=0.2436, simple_loss=0.3199, pruned_loss=0.08368, over 3113026.42 frames. ], batch size: 44, lr: 1.12e-02, grad_scale: 4.0 2023-04-28 12:50:26,695 INFO [train.py:904] (4/8) Epoch 6, batch 6400, loss[loss=0.2684, simple_loss=0.3379, pruned_loss=0.09941, over 16694.00 frames. ], tot_loss[loss=0.2447, simple_loss=0.3202, pruned_loss=0.08457, over 3120799.38 frames. ], batch size: 83, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:50:38,655 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57159.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:51:01,322 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.642e+02 4.267e+02 5.177e+02 6.433e+02 1.516e+03, threshold=1.035e+03, percent-clipped=6.0 2023-04-28 12:51:08,481 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57179.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 12:51:19,407 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57186.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:51:36,904 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57197.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:51:42,172 INFO [train.py:904] (4/8) Epoch 6, batch 6450, loss[loss=0.2297, simple_loss=0.2904, pruned_loss=0.08452, over 11533.00 frames. ], tot_loss[loss=0.2435, simple_loss=0.32, pruned_loss=0.08353, over 3119707.23 frames. ], batch size: 247, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:52:26,738 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=57227.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 12:52:52,980 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57244.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:52:57,212 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57247.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:53:03,828 INFO [train.py:904] (4/8) Epoch 6, batch 6500, loss[loss=0.2314, simple_loss=0.3022, pruned_loss=0.08026, over 16878.00 frames. ], tot_loss[loss=0.2409, simple_loss=0.317, pruned_loss=0.0824, over 3110441.13 frames. ], batch size: 116, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:53:15,515 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57258.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 12:53:25,365 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57264.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:53:41,776 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.959e+02 3.641e+02 4.251e+02 5.070e+02 1.044e+03, threshold=8.501e+02, percent-clipped=1.0 2023-04-28 12:54:07,419 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2248, 5.2102, 5.1264, 4.9351, 4.5434, 5.1833, 5.0597, 4.7958], device='cuda:4'), covar=tensor([0.0594, 0.0605, 0.0211, 0.0165, 0.0957, 0.0367, 0.0188, 0.0592], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0214, 0.0225, 0.0197, 0.0252, 0.0229, 0.0162, 0.0258], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:54:25,244 INFO [train.py:904] (4/8) Epoch 6, batch 6550, loss[loss=0.3373, simple_loss=0.3764, pruned_loss=0.1491, over 11208.00 frames. ], tot_loss[loss=0.2458, simple_loss=0.3219, pruned_loss=0.0849, over 3088204.19 frames. ], batch size: 248, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:54:29,281 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5401, 3.9078, 4.1826, 1.8812, 4.3666, 4.4254, 3.1966, 3.2758], device='cuda:4'), covar=tensor([0.0805, 0.0134, 0.0138, 0.1172, 0.0049, 0.0080, 0.0314, 0.0378], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0091, 0.0082, 0.0139, 0.0073, 0.0082, 0.0116, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 12:54:31,722 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57305.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:55:32,168 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1518, 5.1308, 4.8758, 4.3023, 4.9669, 1.6431, 4.7136, 4.9439], device='cuda:4'), covar=tensor([0.0052, 0.0051, 0.0083, 0.0296, 0.0051, 0.1889, 0.0075, 0.0113], device='cuda:4'), in_proj_covar=tensor([0.0098, 0.0086, 0.0131, 0.0129, 0.0098, 0.0149, 0.0114, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 12:55:40,094 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57349.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:55:42,171 INFO [train.py:904] (4/8) Epoch 6, batch 6600, loss[loss=0.2751, simple_loss=0.3558, pruned_loss=0.09721, over 15305.00 frames. ], tot_loss[loss=0.249, simple_loss=0.3253, pruned_loss=0.08635, over 3069694.75 frames. ], batch size: 190, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:55:43,278 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57351.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:56:03,092 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.76 vs. limit=5.0 2023-04-28 12:56:18,766 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.503e+02 3.780e+02 4.683e+02 5.878e+02 1.016e+03, threshold=9.365e+02, percent-clipped=5.0 2023-04-28 12:56:56,484 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57398.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:57:00,528 INFO [train.py:904] (4/8) Epoch 6, batch 6650, loss[loss=0.3058, simple_loss=0.3544, pruned_loss=0.1286, over 11319.00 frames. ], tot_loss[loss=0.2499, simple_loss=0.3258, pruned_loss=0.08703, over 3073729.22 frames. ], batch size: 247, lr: 1.12e-02, grad_scale: 8.0 2023-04-28 12:58:15,912 INFO [train.py:904] (4/8) Epoch 6, batch 6700, loss[loss=0.3183, simple_loss=0.3617, pruned_loss=0.1374, over 11293.00 frames. ], tot_loss[loss=0.2487, simple_loss=0.3243, pruned_loss=0.08651, over 3084130.49 frames. ], batch size: 248, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 12:58:29,463 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57459.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:58:29,620 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57459.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:58:34,670 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57462.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:58:54,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.152e+02 3.718e+02 4.462e+02 5.305e+02 7.989e+02, threshold=8.923e+02, percent-clipped=0.0 2023-04-28 12:58:55,520 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 12:59:24,779 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57495.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 12:59:34,116 INFO [train.py:904] (4/8) Epoch 6, batch 6750, loss[loss=0.2893, simple_loss=0.3477, pruned_loss=0.1155, over 11583.00 frames. ], tot_loss[loss=0.2485, simple_loss=0.3235, pruned_loss=0.08673, over 3092742.65 frames. ], batch size: 247, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 12:59:43,067 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=57507.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:00:07,997 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57523.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:00:36,643 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57542.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:00:49,718 INFO [train.py:904] (4/8) Epoch 6, batch 6800, loss[loss=0.2931, simple_loss=0.3525, pruned_loss=0.1168, over 11674.00 frames. ], tot_loss[loss=0.2476, simple_loss=0.323, pruned_loss=0.08614, over 3082260.58 frames. ], batch size: 248, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:00:54,409 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57553.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 13:00:58,740 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57556.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:01:08,621 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0730, 5.2831, 5.0417, 5.0233, 4.7009, 4.4581, 4.8403, 5.3484], device='cuda:4'), covar=tensor([0.0656, 0.0659, 0.0823, 0.0486, 0.0610, 0.0743, 0.0672, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0404, 0.0522, 0.0446, 0.0341, 0.0324, 0.0343, 0.0427, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:01:09,768 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57564.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:01:25,687 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-28 13:01:27,504 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.272e+02 3.624e+02 4.504e+02 5.898e+02 1.165e+03, threshold=9.008e+02, percent-clipped=3.0 2023-04-28 13:01:41,293 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-28 13:02:06,924 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57600.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:02:07,781 INFO [train.py:904] (4/8) Epoch 6, batch 6850, loss[loss=0.239, simple_loss=0.3343, pruned_loss=0.07191, over 17301.00 frames. ], tot_loss[loss=0.2474, simple_loss=0.3232, pruned_loss=0.08578, over 3103517.44 frames. ], batch size: 52, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:02:24,686 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=57612.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:03:22,271 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57649.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:03:22,412 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4714, 3.9644, 4.0956, 1.9932, 4.2743, 4.3082, 3.1829, 3.2165], device='cuda:4'), covar=tensor([0.0747, 0.0111, 0.0127, 0.1049, 0.0041, 0.0058, 0.0279, 0.0345], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0091, 0.0083, 0.0140, 0.0072, 0.0082, 0.0116, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 13:03:24,924 INFO [train.py:904] (4/8) Epoch 6, batch 6900, loss[loss=0.3392, simple_loss=0.3718, pruned_loss=0.1534, over 11285.00 frames. ], tot_loss[loss=0.2479, simple_loss=0.3255, pruned_loss=0.0851, over 3109205.57 frames. ], batch size: 247, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:03:25,254 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57651.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:04:02,708 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.992e+02 3.443e+02 4.296e+02 5.667e+02 1.150e+03, threshold=8.592e+02, percent-clipped=3.0 2023-04-28 13:04:37,887 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=57697.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:04:40,788 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=57699.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:04:43,531 INFO [train.py:904] (4/8) Epoch 6, batch 6950, loss[loss=0.2396, simple_loss=0.3222, pruned_loss=0.07857, over 16422.00 frames. ], tot_loss[loss=0.2494, simple_loss=0.3266, pruned_loss=0.08613, over 3115056.05 frames. ], batch size: 146, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:05:34,181 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6136, 2.6559, 1.7637, 2.7600, 2.0628, 2.7559, 1.9423, 2.3845], device='cuda:4'), covar=tensor([0.0200, 0.0405, 0.1192, 0.0094, 0.0669, 0.0454, 0.1169, 0.0561], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0154, 0.0177, 0.0083, 0.0161, 0.0189, 0.0187, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 13:05:34,184 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57733.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:06:01,323 INFO [train.py:904] (4/8) Epoch 6, batch 7000, loss[loss=0.2577, simple_loss=0.3432, pruned_loss=0.08612, over 16848.00 frames. ], tot_loss[loss=0.2495, simple_loss=0.3266, pruned_loss=0.08617, over 3093830.16 frames. ], batch size: 116, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:06:06,475 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57754.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:06:38,242 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.975e+02 3.231e+02 4.685e+02 6.535e+02 1.606e+03, threshold=9.370e+02, percent-clipped=8.0 2023-04-28 13:07:00,314 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-04-28 13:07:07,508 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57794.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:07:17,559 INFO [train.py:904] (4/8) Epoch 6, batch 7050, loss[loss=0.2508, simple_loss=0.3283, pruned_loss=0.08666, over 16924.00 frames. ], tot_loss[loss=0.2486, simple_loss=0.3268, pruned_loss=0.08515, over 3111487.45 frames. ], batch size: 116, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:07:18,031 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5943, 4.3685, 4.3466, 2.9899, 3.9348, 4.3380, 4.0589, 2.2368], device='cuda:4'), covar=tensor([0.0302, 0.0015, 0.0022, 0.0203, 0.0040, 0.0055, 0.0029, 0.0278], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0055, 0.0059, 0.0115, 0.0062, 0.0073, 0.0065, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 13:07:44,478 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57818.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:08:20,858 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57842.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:08:34,156 INFO [train.py:904] (4/8) Epoch 6, batch 7100, loss[loss=0.2191, simple_loss=0.3034, pruned_loss=0.06736, over 16689.00 frames. ], tot_loss[loss=0.2487, simple_loss=0.326, pruned_loss=0.0857, over 3084798.48 frames. ], batch size: 76, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:08:34,651 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57851.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:08:37,850 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57853.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 13:09:10,525 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.050e+02 3.825e+02 4.803e+02 6.106e+02 1.425e+03, threshold=9.606e+02, percent-clipped=4.0 2023-04-28 13:09:32,897 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=57890.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:09:48,257 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57900.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:09:49,598 INFO [train.py:904] (4/8) Epoch 6, batch 7150, loss[loss=0.2027, simple_loss=0.2857, pruned_loss=0.05987, over 16752.00 frames. ], tot_loss[loss=0.2474, simple_loss=0.3241, pruned_loss=0.08528, over 3088008.83 frames. ], batch size: 89, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:09:49,970 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=57901.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:10:55,523 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9334, 2.2804, 2.2933, 2.9914, 2.2754, 3.2272, 1.7010, 2.6514], device='cuda:4'), covar=tensor([0.1114, 0.0540, 0.0956, 0.0107, 0.0178, 0.0385, 0.1245, 0.0713], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0144, 0.0170, 0.0095, 0.0197, 0.0192, 0.0164, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 13:11:00,231 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=57948.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:11:00,505 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5328, 3.5830, 2.7966, 2.2155, 2.5956, 2.2388, 3.7493, 3.5912], device='cuda:4'), covar=tensor([0.2248, 0.0753, 0.1415, 0.1682, 0.1959, 0.1484, 0.0420, 0.0724], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0253, 0.0269, 0.0254, 0.0285, 0.0203, 0.0250, 0.0264], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:11:04,887 INFO [train.py:904] (4/8) Epoch 6, batch 7200, loss[loss=0.2449, simple_loss=0.3259, pruned_loss=0.08191, over 15334.00 frames. ], tot_loss[loss=0.2435, simple_loss=0.3215, pruned_loss=0.08276, over 3093435.88 frames. ], batch size: 190, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:11:10,451 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.73 vs. limit=5.0 2023-04-28 13:11:41,800 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.198e+02 3.384e+02 4.240e+02 5.535e+02 1.102e+03, threshold=8.480e+02, percent-clipped=1.0 2023-04-28 13:12:28,071 INFO [train.py:904] (4/8) Epoch 6, batch 7250, loss[loss=0.242, simple_loss=0.3138, pruned_loss=0.08509, over 16671.00 frames. ], tot_loss[loss=0.2415, simple_loss=0.3192, pruned_loss=0.08192, over 3087515.06 frames. ], batch size: 134, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:13:22,217 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5122, 2.0578, 2.1896, 4.2560, 1.9313, 2.8188, 2.2360, 2.2370], device='cuda:4'), covar=tensor([0.0718, 0.2584, 0.1528, 0.0311, 0.3418, 0.1530, 0.2302, 0.2805], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0340, 0.0280, 0.0313, 0.0386, 0.0353, 0.0305, 0.0401], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:13:44,368 INFO [train.py:904] (4/8) Epoch 6, batch 7300, loss[loss=0.2274, simple_loss=0.3221, pruned_loss=0.06633, over 17123.00 frames. ], tot_loss[loss=0.2414, simple_loss=0.3187, pruned_loss=0.08209, over 3084877.15 frames. ], batch size: 47, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:13:49,558 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58054.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:14:21,941 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.373e+02 3.804e+02 4.653e+02 5.753e+02 1.061e+03, threshold=9.306e+02, percent-clipped=1.0 2023-04-28 13:14:36,071 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9613, 1.6977, 2.1776, 2.9229, 2.7072, 3.5266, 1.5235, 3.0925], device='cuda:4'), covar=tensor([0.0093, 0.0261, 0.0178, 0.0106, 0.0114, 0.0046, 0.0278, 0.0059], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0146, 0.0129, 0.0128, 0.0133, 0.0098, 0.0144, 0.0085], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 13:14:43,348 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=58089.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:14:46,520 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7497, 4.7533, 5.2627, 5.2057, 5.2415, 4.7723, 4.8586, 4.3836], device='cuda:4'), covar=tensor([0.0231, 0.0288, 0.0284, 0.0342, 0.0357, 0.0309, 0.0726, 0.0412], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0243, 0.0249, 0.0246, 0.0290, 0.0267, 0.0364, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 13:15:02,126 INFO [train.py:904] (4/8) Epoch 6, batch 7350, loss[loss=0.2193, simple_loss=0.308, pruned_loss=0.06531, over 16724.00 frames. ], tot_loss[loss=0.2426, simple_loss=0.3191, pruned_loss=0.08306, over 3050654.19 frames. ], batch size: 83, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:15:03,734 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=58102.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:15:09,361 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8529, 4.1107, 3.8719, 3.9457, 3.5924, 3.7311, 3.8414, 4.0452], device='cuda:4'), covar=tensor([0.0734, 0.0718, 0.0940, 0.0518, 0.0684, 0.1110, 0.0610, 0.0812], device='cuda:4'), in_proj_covar=tensor([0.0403, 0.0525, 0.0450, 0.0340, 0.0326, 0.0348, 0.0425, 0.0376], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:15:28,189 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58118.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:15:38,229 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6367, 4.6274, 4.3859, 3.6821, 4.4538, 1.5854, 4.2133, 4.2694], device='cuda:4'), covar=tensor([0.0059, 0.0046, 0.0114, 0.0317, 0.0059, 0.1951, 0.0087, 0.0149], device='cuda:4'), in_proj_covar=tensor([0.0098, 0.0083, 0.0130, 0.0129, 0.0098, 0.0149, 0.0113, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:16:04,927 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58141.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:16:09,236 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58144.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:16:19,318 INFO [train.py:904] (4/8) Epoch 6, batch 7400, loss[loss=0.3169, simple_loss=0.3614, pruned_loss=0.1362, over 11074.00 frames. ], tot_loss[loss=0.2434, simple_loss=0.32, pruned_loss=0.08337, over 3063494.95 frames. ], batch size: 247, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:16:19,732 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58151.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:16:43,737 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=58166.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:16:57,163 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.315e+02 3.701e+02 4.363e+02 5.153e+02 1.099e+03, threshold=8.726e+02, percent-clipped=1.0 2023-04-28 13:17:01,684 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.36 vs. limit=5.0 2023-04-28 13:17:09,055 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0877, 3.7746, 3.5441, 1.7719, 2.9942, 2.4586, 3.3468, 3.6355], device='cuda:4'), covar=tensor([0.0257, 0.0479, 0.0475, 0.1771, 0.0672, 0.0855, 0.0635, 0.0726], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0128, 0.0153, 0.0141, 0.0133, 0.0126, 0.0138, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 13:17:27,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1566, 1.5495, 1.7765, 2.1542, 2.2319, 2.3233, 1.5321, 2.2271], device='cuda:4'), covar=tensor([0.0111, 0.0252, 0.0154, 0.0152, 0.0113, 0.0098, 0.0216, 0.0063], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0146, 0.0128, 0.0129, 0.0133, 0.0098, 0.0144, 0.0084], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 13:17:34,943 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=58199.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:17:37,750 INFO [train.py:904] (4/8) Epoch 6, batch 7450, loss[loss=0.2285, simple_loss=0.3194, pruned_loss=0.06877, over 16804.00 frames. ], tot_loss[loss=0.2466, simple_loss=0.3223, pruned_loss=0.08545, over 3050011.83 frames. ], batch size: 83, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:17:39,880 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58202.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:17:45,781 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58205.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:18:23,370 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-28 13:18:59,284 INFO [train.py:904] (4/8) Epoch 6, batch 7500, loss[loss=0.2131, simple_loss=0.2935, pruned_loss=0.06632, over 16725.00 frames. ], tot_loss[loss=0.2457, simple_loss=0.3224, pruned_loss=0.08447, over 3068202.44 frames. ], batch size: 57, lr: 1.11e-02, grad_scale: 4.0 2023-04-28 13:19:39,439 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.280e+02 3.913e+02 4.700e+02 5.850e+02 1.352e+03, threshold=9.401e+02, percent-clipped=2.0 2023-04-28 13:20:11,875 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 13:20:13,852 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 13:20:17,495 INFO [train.py:904] (4/8) Epoch 6, batch 7550, loss[loss=0.2598, simple_loss=0.3315, pruned_loss=0.09402, over 15480.00 frames. ], tot_loss[loss=0.2462, simple_loss=0.3219, pruned_loss=0.08525, over 3045048.22 frames. ], batch size: 190, lr: 1.11e-02, grad_scale: 4.0 2023-04-28 13:21:33,625 INFO [train.py:904] (4/8) Epoch 6, batch 7600, loss[loss=0.2829, simple_loss=0.3349, pruned_loss=0.1154, over 11105.00 frames. ], tot_loss[loss=0.2455, simple_loss=0.3206, pruned_loss=0.08519, over 3043182.85 frames. ], batch size: 248, lr: 1.11e-02, grad_scale: 8.0 2023-04-28 13:22:14,297 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.923e+02 3.777e+02 4.488e+02 5.757e+02 1.119e+03, threshold=8.975e+02, percent-clipped=2.0 2023-04-28 13:22:33,768 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58389.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:22:52,233 INFO [train.py:904] (4/8) Epoch 6, batch 7650, loss[loss=0.2393, simple_loss=0.3248, pruned_loss=0.07691, over 17001.00 frames. ], tot_loss[loss=0.2471, simple_loss=0.3216, pruned_loss=0.0863, over 3026965.55 frames. ], batch size: 50, lr: 1.11e-02, grad_scale: 4.0 2023-04-28 13:23:49,829 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=58437.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:24:11,081 INFO [train.py:904] (4/8) Epoch 6, batch 7700, loss[loss=0.2306, simple_loss=0.3168, pruned_loss=0.07219, over 16868.00 frames. ], tot_loss[loss=0.2464, simple_loss=0.3209, pruned_loss=0.08595, over 3047206.07 frames. ], batch size: 96, lr: 1.11e-02, grad_scale: 4.0 2023-04-28 13:24:51,994 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.098e+02 3.672e+02 4.554e+02 5.723e+02 8.873e+02, threshold=9.107e+02, percent-clipped=1.0 2023-04-28 13:25:22,819 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.02 vs. limit=2.0 2023-04-28 13:25:23,658 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=58497.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:25:28,762 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=58500.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:25:29,686 INFO [train.py:904] (4/8) Epoch 6, batch 7750, loss[loss=0.2213, simple_loss=0.3117, pruned_loss=0.06552, over 16794.00 frames. ], tot_loss[loss=0.2476, simple_loss=0.3221, pruned_loss=0.08657, over 3043296.34 frames. ], batch size: 102, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:25:56,689 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3118, 1.8550, 1.5586, 1.6536, 2.1755, 2.0266, 2.1546, 2.3730], device='cuda:4'), covar=tensor([0.0051, 0.0180, 0.0248, 0.0228, 0.0100, 0.0173, 0.0105, 0.0121], device='cuda:4'), in_proj_covar=tensor([0.0087, 0.0161, 0.0168, 0.0164, 0.0157, 0.0168, 0.0148, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:25:58,062 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-28 13:26:46,071 INFO [train.py:904] (4/8) Epoch 6, batch 7800, loss[loss=0.3027, simple_loss=0.3441, pruned_loss=0.1306, over 11245.00 frames. ], tot_loss[loss=0.2495, simple_loss=0.323, pruned_loss=0.08803, over 3026825.67 frames. ], batch size: 248, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:27:26,211 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.190e+02 3.955e+02 4.761e+02 5.914e+02 1.211e+03, threshold=9.523e+02, percent-clipped=2.0 2023-04-28 13:27:29,865 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0851, 1.8348, 2.0849, 3.6194, 1.9088, 2.4687, 2.0894, 1.9624], device='cuda:4'), covar=tensor([0.0709, 0.2681, 0.1462, 0.0353, 0.3182, 0.1455, 0.2311, 0.2579], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0342, 0.0280, 0.0316, 0.0388, 0.0353, 0.0304, 0.0399], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:28:01,703 INFO [train.py:904] (4/8) Epoch 6, batch 7850, loss[loss=0.2378, simple_loss=0.3187, pruned_loss=0.0784, over 16373.00 frames. ], tot_loss[loss=0.2505, simple_loss=0.3243, pruned_loss=0.0883, over 3020281.03 frames. ], batch size: 146, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:28:15,608 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.19 vs. limit=5.0 2023-04-28 13:28:17,195 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-28 13:28:30,878 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0767, 3.7858, 3.7514, 2.1991, 3.5087, 3.7384, 3.5021, 1.8249], device='cuda:4'), covar=tensor([0.0408, 0.0017, 0.0028, 0.0305, 0.0055, 0.0060, 0.0043, 0.0377], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0054, 0.0058, 0.0114, 0.0062, 0.0074, 0.0063, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 13:28:51,085 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 13:29:14,704 INFO [train.py:904] (4/8) Epoch 6, batch 7900, loss[loss=0.2701, simple_loss=0.3271, pruned_loss=0.1065, over 11804.00 frames. ], tot_loss[loss=0.2489, simple_loss=0.3232, pruned_loss=0.08727, over 3034690.22 frames. ], batch size: 247, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:29:18,062 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1529, 4.2538, 4.2699, 4.2644, 4.1893, 4.7333, 4.2741, 4.0563], device='cuda:4'), covar=tensor([0.1344, 0.1566, 0.1568, 0.1858, 0.2611, 0.0950, 0.1328, 0.2388], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0407, 0.0414, 0.0359, 0.0467, 0.0440, 0.0334, 0.0479], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 13:29:43,573 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-28 13:29:53,308 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.999e+02 3.363e+02 4.237e+02 5.568e+02 1.033e+03, threshold=8.474e+02, percent-clipped=1.0 2023-04-28 13:30:02,298 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58681.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:30:09,271 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8982, 4.1025, 1.9340, 4.5752, 2.4959, 4.4571, 2.2367, 2.8857], device='cuda:4'), covar=tensor([0.0141, 0.0290, 0.1662, 0.0048, 0.0783, 0.0435, 0.1359, 0.0635], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0153, 0.0178, 0.0086, 0.0161, 0.0189, 0.0189, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 13:30:31,227 INFO [train.py:904] (4/8) Epoch 6, batch 7950, loss[loss=0.225, simple_loss=0.3088, pruned_loss=0.07059, over 17110.00 frames. ], tot_loss[loss=0.2481, simple_loss=0.3228, pruned_loss=0.08676, over 3054391.08 frames. ], batch size: 47, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:31:33,484 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58742.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:31:46,443 INFO [train.py:904] (4/8) Epoch 6, batch 8000, loss[loss=0.2225, simple_loss=0.3115, pruned_loss=0.06673, over 16892.00 frames. ], tot_loss[loss=0.2498, simple_loss=0.3238, pruned_loss=0.08789, over 3040837.15 frames. ], batch size: 96, lr: 1.10e-02, grad_scale: 8.0 2023-04-28 13:32:25,054 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58775.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:32:27,145 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.294e+02 3.253e+02 3.971e+02 4.697e+02 8.095e+02, threshold=7.943e+02, percent-clipped=0.0 2023-04-28 13:32:56,777 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.95 vs. limit=5.0 2023-04-28 13:32:57,929 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58797.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:33:02,611 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58800.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:33:03,398 INFO [train.py:904] (4/8) Epoch 6, batch 8050, loss[loss=0.2365, simple_loss=0.3158, pruned_loss=0.07863, over 16946.00 frames. ], tot_loss[loss=0.2487, simple_loss=0.323, pruned_loss=0.08721, over 3037736.09 frames. ], batch size: 109, lr: 1.10e-02, grad_scale: 8.0 2023-04-28 13:33:58,014 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58836.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:34:12,440 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=58845.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:34:16,878 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=58848.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:34:20,748 INFO [train.py:904] (4/8) Epoch 6, batch 8100, loss[loss=0.2211, simple_loss=0.299, pruned_loss=0.07156, over 17205.00 frames. ], tot_loss[loss=0.2466, simple_loss=0.3218, pruned_loss=0.08573, over 3051696.78 frames. ], batch size: 52, lr: 1.10e-02, grad_scale: 2.0 2023-04-28 13:34:27,291 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2807, 4.2303, 4.1729, 3.4913, 4.1296, 1.4977, 3.9203, 3.9238], device='cuda:4'), covar=tensor([0.0070, 0.0057, 0.0095, 0.0290, 0.0069, 0.2067, 0.0099, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0097, 0.0084, 0.0130, 0.0128, 0.0099, 0.0150, 0.0113, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:35:03,715 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.302e+02 3.498e+02 4.445e+02 5.357e+02 1.535e+03, threshold=8.891e+02, percent-clipped=10.0 2023-04-28 13:35:10,473 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1576, 1.4032, 1.8339, 2.1857, 2.1790, 2.3632, 1.5407, 2.2380], device='cuda:4'), covar=tensor([0.0103, 0.0265, 0.0149, 0.0149, 0.0129, 0.0096, 0.0233, 0.0060], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0148, 0.0130, 0.0128, 0.0136, 0.0100, 0.0145, 0.0084], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 13:35:38,303 INFO [train.py:904] (4/8) Epoch 6, batch 8150, loss[loss=0.2142, simple_loss=0.2946, pruned_loss=0.06685, over 16499.00 frames. ], tot_loss[loss=0.2445, simple_loss=0.3198, pruned_loss=0.08455, over 3074366.26 frames. ], batch size: 75, lr: 1.10e-02, grad_scale: 2.0 2023-04-28 13:35:46,344 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58906.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:36:54,139 INFO [train.py:904] (4/8) Epoch 6, batch 8200, loss[loss=0.2672, simple_loss=0.3192, pruned_loss=0.1075, over 11440.00 frames. ], tot_loss[loss=0.2418, simple_loss=0.3168, pruned_loss=0.0834, over 3072441.71 frames. ], batch size: 248, lr: 1.10e-02, grad_scale: 2.0 2023-04-28 13:37:21,740 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58967.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:37:40,735 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.295e+02 3.597e+02 4.719e+02 5.797e+02 1.131e+03, threshold=9.438e+02, percent-clipped=6.0 2023-04-28 13:38:17,122 INFO [train.py:904] (4/8) Epoch 6, batch 8250, loss[loss=0.2016, simple_loss=0.2792, pruned_loss=0.06202, over 11868.00 frames. ], tot_loss[loss=0.2392, simple_loss=0.3155, pruned_loss=0.08142, over 3045686.60 frames. ], batch size: 246, lr: 1.10e-02, grad_scale: 2.0 2023-04-28 13:38:23,840 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.58 vs. limit=5.0 2023-04-28 13:39:16,124 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59037.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:39:37,601 INFO [train.py:904] (4/8) Epoch 6, batch 8300, loss[loss=0.2238, simple_loss=0.3137, pruned_loss=0.06691, over 15292.00 frames. ], tot_loss[loss=0.2333, simple_loss=0.3115, pruned_loss=0.07756, over 3034490.89 frames. ], batch size: 190, lr: 1.10e-02, grad_scale: 2.0 2023-04-28 13:40:22,347 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.568e+02 2.786e+02 3.488e+02 4.487e+02 8.597e+02, threshold=6.977e+02, percent-clipped=0.0 2023-04-28 13:40:59,217 INFO [train.py:904] (4/8) Epoch 6, batch 8350, loss[loss=0.2582, simple_loss=0.3198, pruned_loss=0.09827, over 12162.00 frames. ], tot_loss[loss=0.2302, simple_loss=0.3103, pruned_loss=0.07502, over 3034318.37 frames. ], batch size: 246, lr: 1.10e-02, grad_scale: 2.0 2023-04-28 13:41:27,544 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4535, 4.2714, 4.4629, 4.6559, 4.8041, 4.3709, 4.7841, 4.7597], device='cuda:4'), covar=tensor([0.1009, 0.0792, 0.1209, 0.0492, 0.0413, 0.0686, 0.0356, 0.0379], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0482, 0.0615, 0.0496, 0.0381, 0.0372, 0.0395, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:41:39,522 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8107, 3.5114, 3.3013, 1.9215, 2.8665, 2.2052, 3.2865, 3.3978], device='cuda:4'), covar=tensor([0.0258, 0.0516, 0.0446, 0.1611, 0.0682, 0.0947, 0.0656, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0125, 0.0149, 0.0137, 0.0129, 0.0122, 0.0135, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 13:41:49,263 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59131.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:42:21,175 INFO [train.py:904] (4/8) Epoch 6, batch 8400, loss[loss=0.2063, simple_loss=0.2883, pruned_loss=0.0621, over 11991.00 frames. ], tot_loss[loss=0.2256, simple_loss=0.3068, pruned_loss=0.0722, over 3030088.19 frames. ], batch size: 247, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:42:42,297 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=59164.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:42:53,088 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1027, 4.1588, 4.2934, 4.2601, 4.2647, 4.7362, 4.3773, 4.0782], device='cuda:4'), covar=tensor([0.1428, 0.1562, 0.1380, 0.1644, 0.2114, 0.0884, 0.1104, 0.1957], device='cuda:4'), in_proj_covar=tensor([0.0276, 0.0380, 0.0394, 0.0332, 0.0436, 0.0417, 0.0315, 0.0449], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:43:05,301 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.645e+02 2.850e+02 3.501e+02 4.655e+02 7.905e+02, threshold=7.001e+02, percent-clipped=2.0 2023-04-28 13:43:40,726 INFO [train.py:904] (4/8) Epoch 6, batch 8450, loss[loss=0.1906, simple_loss=0.2737, pruned_loss=0.05378, over 12528.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.3049, pruned_loss=0.07014, over 3029734.69 frames. ], batch size: 248, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:43:47,929 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3249, 4.2610, 4.1297, 3.6455, 4.1303, 1.5661, 3.9446, 3.9692], device='cuda:4'), covar=tensor([0.0061, 0.0057, 0.0104, 0.0218, 0.0067, 0.1964, 0.0086, 0.0128], device='cuda:4'), in_proj_covar=tensor([0.0096, 0.0082, 0.0129, 0.0124, 0.0097, 0.0151, 0.0112, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:44:03,757 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7672, 1.2992, 1.5638, 1.6194, 1.8481, 1.8292, 1.4297, 1.6966], device='cuda:4'), covar=tensor([0.0145, 0.0219, 0.0131, 0.0161, 0.0133, 0.0116, 0.0236, 0.0069], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0150, 0.0131, 0.0130, 0.0137, 0.0099, 0.0148, 0.0085], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 13:44:19,013 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=59225.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:44:35,831 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.70 vs. limit=2.0 2023-04-28 13:45:00,221 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.84 vs. limit=5.0 2023-04-28 13:45:00,570 INFO [train.py:904] (4/8) Epoch 6, batch 8500, loss[loss=0.1853, simple_loss=0.2798, pruned_loss=0.04538, over 16818.00 frames. ], tot_loss[loss=0.2173, simple_loss=0.3007, pruned_loss=0.06699, over 3038497.68 frames. ], batch size: 96, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:45:19,656 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59262.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:45:32,447 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-28 13:45:42,235 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-04-28 13:45:46,051 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.033e+02 2.693e+02 3.373e+02 4.229e+02 1.183e+03, threshold=6.745e+02, percent-clipped=1.0 2023-04-28 13:46:11,692 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3662, 2.9296, 2.6334, 2.2252, 2.1954, 2.1505, 2.9527, 2.9137], device='cuda:4'), covar=tensor([0.1893, 0.0647, 0.1164, 0.1639, 0.1947, 0.1567, 0.0417, 0.0796], device='cuda:4'), in_proj_covar=tensor([0.0275, 0.0237, 0.0260, 0.0244, 0.0261, 0.0198, 0.0237, 0.0245], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:46:25,537 INFO [train.py:904] (4/8) Epoch 6, batch 8550, loss[loss=0.2149, simple_loss=0.3093, pruned_loss=0.06023, over 16626.00 frames. ], tot_loss[loss=0.2142, simple_loss=0.2977, pruned_loss=0.06532, over 3025741.67 frames. ], batch size: 134, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:46:51,762 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.22 vs. limit=5.0 2023-04-28 13:47:23,317 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.47 vs. limit=5.0 2023-04-28 13:47:38,280 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=59337.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:48:07,969 INFO [train.py:904] (4/8) Epoch 6, batch 8600, loss[loss=0.1828, simple_loss=0.2736, pruned_loss=0.046, over 16527.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.2981, pruned_loss=0.06423, over 3030516.09 frames. ], batch size: 62, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:48:41,893 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5344, 2.6296, 2.1170, 2.2264, 3.0347, 2.7013, 3.4003, 3.2370], device='cuda:4'), covar=tensor([0.0025, 0.0193, 0.0235, 0.0236, 0.0115, 0.0164, 0.0077, 0.0100], device='cuda:4'), in_proj_covar=tensor([0.0085, 0.0162, 0.0163, 0.0162, 0.0158, 0.0163, 0.0143, 0.0142], device='cuda:4'), out_proj_covar=tensor([9.8749e-05, 1.8788e-04, 1.8426e-04, 1.8456e-04, 1.8501e-04, 1.8920e-04, 1.6083e-04, 1.6429e-04], device='cuda:4') 2023-04-28 13:48:47,012 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9662, 2.4768, 2.2904, 3.0643, 2.3624, 3.3388, 1.5724, 2.7898], device='cuda:4'), covar=tensor([0.1123, 0.0492, 0.0907, 0.0083, 0.0105, 0.0328, 0.1331, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0142, 0.0167, 0.0093, 0.0181, 0.0186, 0.0164, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 13:49:03,381 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.797e+02 2.892e+02 3.852e+02 5.028e+02 1.593e+03, threshold=7.704e+02, percent-clipped=11.0 2023-04-28 13:49:04,332 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7064, 3.7827, 3.0202, 2.1911, 2.8158, 2.2371, 3.9879, 3.7712], device='cuda:4'), covar=tensor([0.2159, 0.0524, 0.1112, 0.1700, 0.1785, 0.1515, 0.0319, 0.0558], device='cuda:4'), in_proj_covar=tensor([0.0273, 0.0235, 0.0258, 0.0243, 0.0255, 0.0195, 0.0235, 0.0243], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 13:49:16,114 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=59385.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:49:16,545 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-28 13:49:26,898 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-28 13:49:46,456 INFO [train.py:904] (4/8) Epoch 6, batch 8650, loss[loss=0.1907, simple_loss=0.275, pruned_loss=0.05321, over 11988.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2956, pruned_loss=0.06223, over 3030620.73 frames. ], batch size: 247, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:50:40,153 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-28 13:50:55,019 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=59431.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:51:31,175 INFO [train.py:904] (4/8) Epoch 6, batch 8700, loss[loss=0.2033, simple_loss=0.2823, pruned_loss=0.06211, over 12492.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2918, pruned_loss=0.0601, over 3044774.11 frames. ], batch size: 248, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:51:45,092 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-28 13:52:21,220 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.834e+02 2.911e+02 3.776e+02 4.464e+02 8.360e+02, threshold=7.553e+02, percent-clipped=2.0 2023-04-28 13:52:22,278 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=59479.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:52:24,354 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8559, 3.5024, 3.3007, 1.8702, 2.9097, 2.2691, 3.1513, 3.3518], device='cuda:4'), covar=tensor([0.0254, 0.0503, 0.0463, 0.1566, 0.0628, 0.0924, 0.0734, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0122, 0.0149, 0.0138, 0.0129, 0.0123, 0.0133, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 13:53:05,570 INFO [train.py:904] (4/8) Epoch 6, batch 8750, loss[loss=0.1837, simple_loss=0.2749, pruned_loss=0.04624, over 17160.00 frames. ], tot_loss[loss=0.2049, simple_loss=0.2915, pruned_loss=0.05911, over 3060132.85 frames. ], batch size: 49, lr: 1.10e-02, grad_scale: 4.0 2023-04-28 13:53:08,826 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=59502.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:53:54,489 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59520.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:54:55,917 INFO [train.py:904] (4/8) Epoch 6, batch 8800, loss[loss=0.2309, simple_loss=0.3021, pruned_loss=0.07982, over 12562.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2907, pruned_loss=0.05825, over 3074509.44 frames. ], batch size: 248, lr: 1.10e-02, grad_scale: 8.0 2023-04-28 13:55:17,764 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=59562.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:55:19,862 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=59563.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:55:52,197 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.676e+02 2.649e+02 3.204e+02 3.980e+02 7.645e+02, threshold=6.408e+02, percent-clipped=1.0 2023-04-28 13:56:37,974 INFO [train.py:904] (4/8) Epoch 6, batch 8850, loss[loss=0.2091, simple_loss=0.3049, pruned_loss=0.05662, over 16293.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2925, pruned_loss=0.05762, over 3050333.59 frames. ], batch size: 146, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 13:56:56,207 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=59610.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 13:58:21,088 INFO [train.py:904] (4/8) Epoch 6, batch 8900, loss[loss=0.1883, simple_loss=0.2827, pruned_loss=0.04699, over 16694.00 frames. ], tot_loss[loss=0.2032, simple_loss=0.293, pruned_loss=0.05675, over 3068521.66 frames. ], batch size: 89, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 13:59:22,389 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.981e+02 2.897e+02 3.325e+02 4.292e+02 6.901e+02, threshold=6.651e+02, percent-clipped=1.0 2023-04-28 14:00:23,144 INFO [train.py:904] (4/8) Epoch 6, batch 8950, loss[loss=0.1804, simple_loss=0.2726, pruned_loss=0.04408, over 16301.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2927, pruned_loss=0.05697, over 3082779.01 frames. ], batch size: 165, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:02:12,299 INFO [train.py:904] (4/8) Epoch 6, batch 9000, loss[loss=0.1898, simple_loss=0.2786, pruned_loss=0.05053, over 16090.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2894, pruned_loss=0.05566, over 3065917.10 frames. ], batch size: 165, lr: 1.09e-02, grad_scale: 4.0 2023-04-28 14:02:12,299 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 14:02:22,190 INFO [train.py:938] (4/8) Epoch 6, validation: loss=0.1682, simple_loss=0.2716, pruned_loss=0.03235, over 944034.00 frames. 2023-04-28 14:02:22,191 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 14:03:21,582 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.989e+02 2.787e+02 3.475e+02 4.351e+02 1.064e+03, threshold=6.950e+02, percent-clipped=4.0 2023-04-28 14:04:06,173 INFO [train.py:904] (4/8) Epoch 6, batch 9050, loss[loss=0.2009, simple_loss=0.2872, pruned_loss=0.05734, over 15267.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2914, pruned_loss=0.05674, over 3076402.50 frames. ], batch size: 191, lr: 1.09e-02, grad_scale: 4.0 2023-04-28 14:04:23,526 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.69 vs. limit=2.0 2023-04-28 14:04:44,672 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=59820.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:05:08,018 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9762, 5.2366, 4.9963, 4.9959, 4.6796, 4.6883, 4.7770, 5.2927], device='cuda:4'), covar=tensor([0.0629, 0.0692, 0.0939, 0.0533, 0.0604, 0.0691, 0.0712, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0503, 0.0423, 0.0328, 0.0309, 0.0337, 0.0414, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:05:45,916 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=59848.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:05:50,594 INFO [train.py:904] (4/8) Epoch 6, batch 9100, loss[loss=0.2155, simple_loss=0.306, pruned_loss=0.06249, over 16411.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2908, pruned_loss=0.0571, over 3070644.69 frames. ], batch size: 146, lr: 1.09e-02, grad_scale: 4.0 2023-04-28 14:06:04,242 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59858.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:06:24,658 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=59868.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:06:55,731 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.924e+02 2.998e+02 3.675e+02 4.654e+02 1.190e+03, threshold=7.350e+02, percent-clipped=4.0 2023-04-28 14:07:19,865 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=59889.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:07:47,001 INFO [train.py:904] (4/8) Epoch 6, batch 9150, loss[loss=0.2009, simple_loss=0.2859, pruned_loss=0.05796, over 16348.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2911, pruned_loss=0.05674, over 3068855.22 frames. ], batch size: 146, lr: 1.09e-02, grad_scale: 4.0 2023-04-28 14:08:05,304 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9524, 2.9374, 1.9769, 2.3624, 3.3040, 2.8985, 3.9108, 3.6087], device='cuda:4'), covar=tensor([0.0021, 0.0181, 0.0309, 0.0241, 0.0114, 0.0189, 0.0073, 0.0082], device='cuda:4'), in_proj_covar=tensor([0.0082, 0.0162, 0.0164, 0.0162, 0.0158, 0.0164, 0.0141, 0.0143], device='cuda:4'), out_proj_covar=tensor([9.4866e-05, 1.8833e-04, 1.8508e-04, 1.8395e-04, 1.8489e-04, 1.8981e-04, 1.5698e-04, 1.6441e-04], device='cuda:4') 2023-04-28 14:08:06,719 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=59909.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:08:27,605 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-28 14:09:30,424 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=59950.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:09:31,017 INFO [train.py:904] (4/8) Epoch 6, batch 9200, loss[loss=0.1825, simple_loss=0.2608, pruned_loss=0.05215, over 12484.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2863, pruned_loss=0.05553, over 3056511.98 frames. ], batch size: 248, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:10:15,909 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7388, 2.7694, 1.6591, 2.8824, 2.0064, 2.8002, 1.8850, 2.3480], device='cuda:4'), covar=tensor([0.0195, 0.0330, 0.1365, 0.0125, 0.0726, 0.0618, 0.1270, 0.0531], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0145, 0.0174, 0.0081, 0.0154, 0.0178, 0.0186, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 14:10:22,280 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.923e+02 2.993e+02 3.687e+02 4.643e+02 6.569e+02, threshold=7.374e+02, percent-clipped=0.0 2023-04-28 14:11:09,067 INFO [train.py:904] (4/8) Epoch 6, batch 9250, loss[loss=0.1921, simple_loss=0.2827, pruned_loss=0.05073, over 15261.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2859, pruned_loss=0.05559, over 3062795.58 frames. ], batch size: 191, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:12:13,453 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1126, 4.2048, 4.0143, 3.8480, 3.7006, 4.0765, 3.8405, 3.8960], device='cuda:4'), covar=tensor([0.0416, 0.0314, 0.0201, 0.0191, 0.0671, 0.0271, 0.0493, 0.0391], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0201, 0.0215, 0.0187, 0.0236, 0.0218, 0.0152, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:12:20,563 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4649, 4.1447, 4.0514, 2.1780, 3.2412, 2.5602, 3.6691, 3.9509], device='cuda:4'), covar=tensor([0.0259, 0.0483, 0.0370, 0.1531, 0.0643, 0.0863, 0.0691, 0.0802], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0121, 0.0151, 0.0140, 0.0132, 0.0125, 0.0137, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 14:12:33,900 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60040.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:12:57,815 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9669, 3.2537, 3.4539, 3.4147, 3.4039, 3.2486, 2.8937, 3.3151], device='cuda:4'), covar=tensor([0.0596, 0.0778, 0.0652, 0.0784, 0.0821, 0.0716, 0.1571, 0.0598], device='cuda:4'), in_proj_covar=tensor([0.0238, 0.0233, 0.0242, 0.0234, 0.0281, 0.0255, 0.0345, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-28 14:12:58,478 INFO [train.py:904] (4/8) Epoch 6, batch 9300, loss[loss=0.1913, simple_loss=0.2624, pruned_loss=0.06015, over 12400.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2835, pruned_loss=0.05502, over 3024009.99 frames. ], batch size: 248, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:13:01,515 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60052.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:13:33,172 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60065.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:14:01,684 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60078.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:14:05,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.815e+02 2.894e+02 3.605e+02 4.509e+02 1.051e+03, threshold=7.210e+02, percent-clipped=5.0 2023-04-28 14:14:35,761 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4000, 3.2826, 3.4002, 3.5184, 3.5644, 3.2270, 3.5217, 3.5763], device='cuda:4'), covar=tensor([0.0795, 0.0754, 0.0916, 0.0504, 0.0544, 0.2244, 0.0810, 0.0546], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0461, 0.0582, 0.0470, 0.0358, 0.0350, 0.0368, 0.0399], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:14:41,131 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-28 14:14:44,853 INFO [train.py:904] (4/8) Epoch 6, batch 9350, loss[loss=0.2062, simple_loss=0.2874, pruned_loss=0.06253, over 16830.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2833, pruned_loss=0.05495, over 3038751.47 frames. ], batch size: 124, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:14:46,417 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60101.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 14:15:12,247 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60113.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 14:15:18,201 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1717, 3.2774, 3.6319, 3.5818, 3.6071, 3.3595, 3.4253, 3.4492], device='cuda:4'), covar=tensor([0.0321, 0.0545, 0.0395, 0.0504, 0.0423, 0.0415, 0.0741, 0.0372], device='cuda:4'), in_proj_covar=tensor([0.0237, 0.0231, 0.0241, 0.0233, 0.0278, 0.0255, 0.0342, 0.0208], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-28 14:15:37,197 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60126.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:16:03,927 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60139.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:16:25,945 INFO [train.py:904] (4/8) Epoch 6, batch 9400, loss[loss=0.2204, simple_loss=0.3116, pruned_loss=0.06459, over 16320.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2834, pruned_loss=0.05479, over 3033425.02 frames. ], batch size: 146, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:16:40,744 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60158.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:16:40,917 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5626, 3.3002, 3.1409, 1.7687, 2.7173, 2.2219, 2.9940, 3.1217], device='cuda:4'), covar=tensor([0.0277, 0.0462, 0.0413, 0.1587, 0.0663, 0.0896, 0.0709, 0.0658], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0121, 0.0150, 0.0139, 0.0131, 0.0124, 0.0135, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 14:17:01,379 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4118, 3.0122, 2.7265, 2.2877, 2.1397, 2.1969, 2.9179, 2.9635], device='cuda:4'), covar=tensor([0.1842, 0.0575, 0.1011, 0.1498, 0.1754, 0.1468, 0.0351, 0.0717], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0237, 0.0260, 0.0243, 0.0240, 0.0198, 0.0236, 0.0240], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:17:25,111 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.674e+02 3.031e+02 3.474e+02 4.387e+02 1.018e+03, threshold=6.948e+02, percent-clipped=2.0 2023-04-28 14:18:05,562 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60198.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:18:09,612 INFO [train.py:904] (4/8) Epoch 6, batch 9450, loss[loss=0.2152, simple_loss=0.2979, pruned_loss=0.06619, over 16909.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2854, pruned_loss=0.05554, over 3008477.27 frames. ], batch size: 116, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:18:14,582 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60204.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:18:17,553 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5968, 3.5035, 3.4449, 2.9752, 3.4186, 1.9331, 3.2243, 2.9649], device='cuda:4'), covar=tensor([0.0075, 0.0069, 0.0101, 0.0161, 0.0067, 0.1743, 0.0088, 0.0147], device='cuda:4'), in_proj_covar=tensor([0.0093, 0.0079, 0.0123, 0.0115, 0.0095, 0.0149, 0.0109, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:18:19,280 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=60206.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:19:39,942 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60245.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:19:50,582 INFO [train.py:904] (4/8) Epoch 6, batch 9500, loss[loss=0.1857, simple_loss=0.2767, pruned_loss=0.04737, over 15317.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2843, pruned_loss=0.05457, over 3027915.17 frames. ], batch size: 191, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:20:07,406 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0749, 4.0323, 4.5616, 4.4439, 4.4509, 4.1168, 4.1432, 3.9760], device='cuda:4'), covar=tensor([0.0260, 0.0460, 0.0319, 0.0465, 0.0458, 0.0336, 0.0786, 0.0390], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0232, 0.0240, 0.0233, 0.0276, 0.0253, 0.0342, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-28 14:20:11,484 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60259.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:20:51,246 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.861e+02 2.773e+02 3.351e+02 4.137e+02 1.111e+03, threshold=6.703e+02, percent-clipped=2.0 2023-04-28 14:21:39,640 INFO [train.py:904] (4/8) Epoch 6, batch 9550, loss[loss=0.2277, simple_loss=0.3155, pruned_loss=0.06999, over 15264.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2839, pruned_loss=0.0541, over 3057948.93 frames. ], batch size: 190, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:23:21,500 INFO [train.py:904] (4/8) Epoch 6, batch 9600, loss[loss=0.2107, simple_loss=0.3062, pruned_loss=0.05759, over 16218.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.2857, pruned_loss=0.05542, over 3050821.62 frames. ], batch size: 165, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:24:17,215 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.609e+02 2.768e+02 3.406e+02 4.197e+02 1.161e+03, threshold=6.812e+02, percent-clipped=3.0 2023-04-28 14:24:57,307 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60396.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 14:25:09,700 INFO [train.py:904] (4/8) Epoch 6, batch 9650, loss[loss=0.1797, simple_loss=0.2729, pruned_loss=0.04323, over 16487.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2872, pruned_loss=0.05528, over 3050707.07 frames. ], batch size: 75, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:25:28,082 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60408.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 14:25:40,628 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5275, 3.5916, 3.3683, 3.1276, 3.2179, 3.4759, 3.2711, 3.2162], device='cuda:4'), covar=tensor([0.0482, 0.0438, 0.0211, 0.0181, 0.0610, 0.0364, 0.0917, 0.0433], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0198, 0.0213, 0.0185, 0.0235, 0.0216, 0.0150, 0.0241], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:25:57,639 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60421.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:26:22,096 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60434.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:26:58,006 INFO [train.py:904] (4/8) Epoch 6, batch 9700, loss[loss=0.1919, simple_loss=0.2867, pruned_loss=0.04852, over 16890.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2866, pruned_loss=0.05505, over 3062879.22 frames. ], batch size: 102, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:27:59,429 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.972e+02 2.947e+02 3.794e+02 4.727e+02 9.838e+02, threshold=7.588e+02, percent-clipped=5.0 2023-04-28 14:28:40,966 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 14:28:41,228 INFO [train.py:904] (4/8) Epoch 6, batch 9750, loss[loss=0.2019, simple_loss=0.2946, pruned_loss=0.0546, over 16815.00 frames. ], tot_loss[loss=0.1981, simple_loss=0.2858, pruned_loss=0.05519, over 3075942.56 frames. ], batch size: 124, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:28:47,810 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60504.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:28:52,046 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.48 vs. limit=5.0 2023-04-28 14:29:02,481 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-28 14:29:53,052 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 14:30:11,557 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60545.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:30:19,565 INFO [train.py:904] (4/8) Epoch 6, batch 9800, loss[loss=0.213, simple_loss=0.2952, pruned_loss=0.06534, over 12249.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.2853, pruned_loss=0.0542, over 3082789.50 frames. ], batch size: 247, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:30:22,404 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=60552.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:30:27,286 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60554.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:31:14,367 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.779e+02 2.802e+02 3.388e+02 4.163e+02 9.547e+02, threshold=6.775e+02, percent-clipped=1.0 2023-04-28 14:31:48,143 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=60593.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:32:03,897 INFO [train.py:904] (4/8) Epoch 6, batch 9850, loss[loss=0.2142, simple_loss=0.2998, pruned_loss=0.06434, over 15394.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.2867, pruned_loss=0.05395, over 3071772.37 frames. ], batch size: 191, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:33:32,138 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7151, 2.7555, 1.7392, 2.8657, 2.0444, 2.8502, 1.9123, 2.4114], device='cuda:4'), covar=tensor([0.0225, 0.0364, 0.1347, 0.0112, 0.0745, 0.0459, 0.1392, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0146, 0.0175, 0.0082, 0.0154, 0.0176, 0.0184, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 14:33:56,848 INFO [train.py:904] (4/8) Epoch 6, batch 9900, loss[loss=0.2055, simple_loss=0.2826, pruned_loss=0.06421, over 12688.00 frames. ], tot_loss[loss=0.1979, simple_loss=0.2873, pruned_loss=0.05427, over 3066472.74 frames. ], batch size: 248, lr: 1.09e-02, grad_scale: 8.0 2023-04-28 14:35:04,552 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.548e+02 2.757e+02 3.371e+02 4.013e+02 6.843e+02, threshold=6.741e+02, percent-clipped=1.0 2023-04-28 14:35:42,311 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60696.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:35:52,849 INFO [train.py:904] (4/8) Epoch 6, batch 9950, loss[loss=0.1804, simple_loss=0.2728, pruned_loss=0.04397, over 16499.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2903, pruned_loss=0.05481, over 3091387.83 frames. ], batch size: 68, lr: 1.08e-02, grad_scale: 8.0 2023-04-28 14:36:09,993 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60708.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:36:34,700 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60718.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:36:42,911 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60721.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:37:13,621 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60734.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:37:38,496 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=60744.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:37:54,651 INFO [train.py:904] (4/8) Epoch 6, batch 10000, loss[loss=0.2037, simple_loss=0.2876, pruned_loss=0.05991, over 16789.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2884, pruned_loss=0.05425, over 3102807.11 frames. ], batch size: 124, lr: 1.08e-02, grad_scale: 8.0 2023-04-28 14:38:06,642 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=60756.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:38:18,452 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0224, 3.1980, 1.6946, 3.3097, 2.1783, 3.2836, 1.9459, 2.5849], device='cuda:4'), covar=tensor([0.0174, 0.0294, 0.1568, 0.0114, 0.0870, 0.0547, 0.1386, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0146, 0.0177, 0.0084, 0.0155, 0.0177, 0.0186, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 14:38:29,851 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=60769.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:38:50,600 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60779.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:38:51,222 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.741e+02 2.599e+02 3.122e+02 4.143e+02 8.746e+02, threshold=6.243e+02, percent-clipped=4.0 2023-04-28 14:38:57,159 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=60782.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:39:23,743 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-28 14:39:33,417 INFO [train.py:904] (4/8) Epoch 6, batch 10050, loss[loss=0.2154, simple_loss=0.3072, pruned_loss=0.06179, over 16232.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.288, pruned_loss=0.05386, over 3090837.76 frames. ], batch size: 165, lr: 1.08e-02, grad_scale: 8.0 2023-04-28 14:41:05,270 INFO [train.py:904] (4/8) Epoch 6, batch 10100, loss[loss=0.2003, simple_loss=0.2774, pruned_loss=0.06162, over 12548.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.289, pruned_loss=0.05473, over 3072092.39 frames. ], batch size: 248, lr: 1.08e-02, grad_scale: 8.0 2023-04-28 14:41:10,981 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60854.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:41:21,681 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-28 14:42:01,260 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.991e+02 2.889e+02 3.366e+02 3.931e+02 7.986e+02, threshold=6.733e+02, percent-clipped=7.0 2023-04-28 14:42:21,322 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6925, 2.7457, 1.7727, 2.8337, 2.1036, 2.8409, 1.9891, 2.4474], device='cuda:4'), covar=tensor([0.0188, 0.0335, 0.1268, 0.0108, 0.0681, 0.0432, 0.1161, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0148, 0.0178, 0.0084, 0.0157, 0.0179, 0.0188, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 14:42:22,737 INFO [train.py:904] (4/8) Epoch 6, batch 10150, loss[loss=0.2018, simple_loss=0.2762, pruned_loss=0.06372, over 11794.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2875, pruned_loss=0.05479, over 3053678.37 frames. ], batch size: 248, lr: 1.08e-02, grad_scale: 8.0 2023-04-28 14:42:48,314 INFO [train.py:904] (4/8) Epoch 7, batch 0, loss[loss=0.2447, simple_loss=0.3335, pruned_loss=0.07798, over 16695.00 frames. ], tot_loss[loss=0.2447, simple_loss=0.3335, pruned_loss=0.07798, over 16695.00 frames. ], batch size: 57, lr: 1.02e-02, grad_scale: 8.0 2023-04-28 14:42:48,314 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 14:42:55,778 INFO [train.py:938] (4/8) Epoch 7, validation: loss=0.1665, simple_loss=0.2702, pruned_loss=0.03141, over 944034.00 frames. 2023-04-28 14:42:55,779 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 14:42:55,980 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=60902.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:43:13,837 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 14:43:38,146 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5314, 4.0812, 4.1856, 2.1008, 3.3984, 2.6274, 3.8601, 3.8829], device='cuda:4'), covar=tensor([0.0257, 0.0485, 0.0372, 0.1436, 0.0558, 0.0818, 0.0616, 0.0974], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0119, 0.0150, 0.0137, 0.0130, 0.0122, 0.0134, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 14:43:43,503 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9115, 1.7229, 1.4958, 1.4453, 1.8975, 1.6273, 1.7584, 1.9849], device='cuda:4'), covar=tensor([0.0032, 0.0169, 0.0201, 0.0204, 0.0110, 0.0147, 0.0098, 0.0097], device='cuda:4'), in_proj_covar=tensor([0.0083, 0.0165, 0.0164, 0.0162, 0.0159, 0.0165, 0.0142, 0.0145], device='cuda:4'), out_proj_covar=tensor([9.4271e-05, 1.9052e-04, 1.8500e-04, 1.8335e-04, 1.8436e-04, 1.8964e-04, 1.5666e-04, 1.6599e-04], device='cuda:4') 2023-04-28 14:44:05,495 INFO [train.py:904] (4/8) Epoch 7, batch 50, loss[loss=0.2354, simple_loss=0.2994, pruned_loss=0.0857, over 16706.00 frames. ], tot_loss[loss=0.2373, simple_loss=0.3094, pruned_loss=0.08256, over 747848.30 frames. ], batch size: 134, lr: 1.01e-02, grad_scale: 1.0 2023-04-28 14:44:29,616 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4880, 4.6153, 4.6630, 4.7370, 4.6057, 5.2292, 4.8234, 4.5443], device='cuda:4'), covar=tensor([0.1115, 0.1656, 0.1723, 0.1680, 0.2867, 0.1120, 0.1344, 0.2424], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0402, 0.0407, 0.0344, 0.0458, 0.0439, 0.0335, 0.0465], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 14:44:49,441 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.919e+02 3.058e+02 3.620e+02 4.703e+02 1.122e+03, threshold=7.241e+02, percent-clipped=6.0 2023-04-28 14:45:08,326 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9243, 4.2199, 2.0668, 4.6594, 2.7950, 4.6288, 2.3248, 3.2437], device='cuda:4'), covar=tensor([0.0197, 0.0318, 0.1754, 0.0086, 0.0901, 0.0377, 0.1502, 0.0669], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0153, 0.0182, 0.0088, 0.0159, 0.0187, 0.0191, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 14:45:15,337 INFO [train.py:904] (4/8) Epoch 7, batch 100, loss[loss=0.2409, simple_loss=0.3153, pruned_loss=0.08325, over 15532.00 frames. ], tot_loss[loss=0.2292, simple_loss=0.3017, pruned_loss=0.07836, over 1311830.28 frames. ], batch size: 191, lr: 1.01e-02, grad_scale: 1.0 2023-04-28 14:46:12,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1121, 1.6721, 2.3205, 2.9875, 2.7701, 3.4186, 1.6419, 3.3236], device='cuda:4'), covar=tensor([0.0084, 0.0276, 0.0165, 0.0109, 0.0111, 0.0071, 0.0295, 0.0065], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0150, 0.0133, 0.0132, 0.0137, 0.0097, 0.0145, 0.0084], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 14:46:24,664 INFO [train.py:904] (4/8) Epoch 7, batch 150, loss[loss=0.2133, simple_loss=0.2763, pruned_loss=0.07516, over 16705.00 frames. ], tot_loss[loss=0.2229, simple_loss=0.2971, pruned_loss=0.07437, over 1755380.93 frames. ], batch size: 83, lr: 1.01e-02, grad_scale: 1.0 2023-04-28 14:46:37,945 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=61061.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:46:46,093 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4838, 4.4871, 4.3673, 4.2384, 3.9632, 4.3817, 4.3408, 4.0543], device='cuda:4'), covar=tensor([0.0503, 0.0308, 0.0229, 0.0193, 0.0795, 0.0308, 0.0362, 0.0505], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0216, 0.0227, 0.0201, 0.0258, 0.0232, 0.0161, 0.0260], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:46:56,416 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=61074.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:47:07,163 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.794e+02 2.994e+02 3.585e+02 4.192e+02 7.211e+02, threshold=7.170e+02, percent-clipped=0.0 2023-04-28 14:47:34,196 INFO [train.py:904] (4/8) Epoch 7, batch 200, loss[loss=0.1876, simple_loss=0.2725, pruned_loss=0.05131, over 16803.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.2967, pruned_loss=0.07419, over 2097595.96 frames. ], batch size: 42, lr: 1.01e-02, grad_scale: 1.0 2023-04-28 14:48:00,114 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-28 14:48:01,626 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=61122.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:48:24,646 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 14:48:44,070 INFO [train.py:904] (4/8) Epoch 7, batch 250, loss[loss=0.2192, simple_loss=0.3056, pruned_loss=0.06642, over 16753.00 frames. ], tot_loss[loss=0.2199, simple_loss=0.294, pruned_loss=0.07292, over 2371397.81 frames. ], batch size: 57, lr: 1.01e-02, grad_scale: 1.0 2023-04-28 14:49:24,842 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.754e+02 3.071e+02 3.589e+02 4.451e+02 1.172e+03, threshold=7.179e+02, percent-clipped=5.0 2023-04-28 14:49:41,570 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.63 vs. limit=5.0 2023-04-28 14:49:52,014 INFO [train.py:904] (4/8) Epoch 7, batch 300, loss[loss=0.2232, simple_loss=0.282, pruned_loss=0.08219, over 16696.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.2897, pruned_loss=0.07003, over 2581362.56 frames. ], batch size: 134, lr: 1.01e-02, grad_scale: 1.0 2023-04-28 14:50:45,046 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9390, 3.8747, 3.8727, 3.3911, 3.8985, 1.7453, 3.7088, 3.5272], device='cuda:4'), covar=tensor([0.0085, 0.0075, 0.0127, 0.0234, 0.0071, 0.1956, 0.0103, 0.0147], device='cuda:4'), in_proj_covar=tensor([0.0100, 0.0086, 0.0135, 0.0124, 0.0101, 0.0156, 0.0118, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:50:59,330 INFO [train.py:904] (4/8) Epoch 7, batch 350, loss[loss=0.1939, simple_loss=0.2748, pruned_loss=0.05652, over 16765.00 frames. ], tot_loss[loss=0.2109, simple_loss=0.2861, pruned_loss=0.06787, over 2741747.80 frames. ], batch size: 62, lr: 1.01e-02, grad_scale: 1.0 2023-04-28 14:51:41,731 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.762e+02 2.825e+02 3.324e+02 4.016e+02 8.512e+02, threshold=6.649e+02, percent-clipped=2.0 2023-04-28 14:51:49,503 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1986, 1.8159, 2.4074, 2.9671, 2.7152, 3.4934, 2.2587, 3.2302], device='cuda:4'), covar=tensor([0.0093, 0.0247, 0.0181, 0.0149, 0.0133, 0.0088, 0.0234, 0.0080], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0150, 0.0134, 0.0134, 0.0137, 0.0100, 0.0146, 0.0085], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 14:52:08,571 INFO [train.py:904] (4/8) Epoch 7, batch 400, loss[loss=0.235, simple_loss=0.2908, pruned_loss=0.0896, over 16678.00 frames. ], tot_loss[loss=0.2101, simple_loss=0.2851, pruned_loss=0.06753, over 2865386.89 frames. ], batch size: 89, lr: 1.01e-02, grad_scale: 2.0 2023-04-28 14:52:45,119 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0025, 4.6709, 4.8900, 5.1572, 5.3911, 4.6638, 5.3390, 5.3333], device='cuda:4'), covar=tensor([0.1072, 0.0961, 0.1536, 0.0612, 0.0418, 0.0697, 0.0434, 0.0383], device='cuda:4'), in_proj_covar=tensor([0.0441, 0.0538, 0.0680, 0.0536, 0.0412, 0.0405, 0.0430, 0.0465], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 14:53:17,203 INFO [train.py:904] (4/8) Epoch 7, batch 450, loss[loss=0.2306, simple_loss=0.2956, pruned_loss=0.08275, over 16367.00 frames. ], tot_loss[loss=0.2067, simple_loss=0.2822, pruned_loss=0.06556, over 2970486.04 frames. ], batch size: 146, lr: 1.01e-02, grad_scale: 2.0 2023-04-28 14:53:48,675 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=61374.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:54:00,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.039e+02 2.768e+02 3.236e+02 4.177e+02 7.606e+02, threshold=6.472e+02, percent-clipped=3.0 2023-04-28 14:54:01,623 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3876, 3.8698, 3.7466, 1.9905, 3.1110, 2.4000, 3.7293, 3.7200], device='cuda:4'), covar=tensor([0.0235, 0.0556, 0.0482, 0.1586, 0.0679, 0.0934, 0.0513, 0.0815], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0126, 0.0152, 0.0139, 0.0131, 0.0123, 0.0135, 0.0137], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 14:54:27,747 INFO [train.py:904] (4/8) Epoch 7, batch 500, loss[loss=0.2117, simple_loss=0.2784, pruned_loss=0.07251, over 16605.00 frames. ], tot_loss[loss=0.2049, simple_loss=0.2811, pruned_loss=0.06432, over 3045473.14 frames. ], batch size: 134, lr: 1.01e-02, grad_scale: 2.0 2023-04-28 14:54:47,691 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=61417.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:54:54,990 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=61422.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 14:55:35,823 INFO [train.py:904] (4/8) Epoch 7, batch 550, loss[loss=0.1736, simple_loss=0.2581, pruned_loss=0.04452, over 16858.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.2803, pruned_loss=0.06291, over 3108188.88 frames. ], batch size: 42, lr: 1.01e-02, grad_scale: 2.0 2023-04-28 14:56:17,396 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.772e+02 2.888e+02 3.437e+02 4.263e+02 9.933e+02, threshold=6.875e+02, percent-clipped=6.0 2023-04-28 14:56:18,196 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-28 14:56:43,947 INFO [train.py:904] (4/8) Epoch 7, batch 600, loss[loss=0.1629, simple_loss=0.2458, pruned_loss=0.03993, over 16982.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2799, pruned_loss=0.06343, over 3145286.27 frames. ], batch size: 41, lr: 1.01e-02, grad_scale: 2.0 2023-04-28 14:57:04,897 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6393, 4.7527, 4.8900, 4.7927, 4.7092, 5.3568, 5.0052, 4.7057], device='cuda:4'), covar=tensor([0.1161, 0.1689, 0.1519, 0.1869, 0.2949, 0.1113, 0.1300, 0.2438], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0444, 0.0442, 0.0378, 0.0508, 0.0471, 0.0360, 0.0508], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 14:57:11,306 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5236, 4.4026, 4.2837, 1.6510, 4.5786, 4.6668, 3.2303, 3.4744], device='cuda:4'), covar=tensor([0.0969, 0.0090, 0.0156, 0.1419, 0.0086, 0.0067, 0.0340, 0.0424], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0094, 0.0083, 0.0140, 0.0069, 0.0087, 0.0117, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 14:57:53,637 INFO [train.py:904] (4/8) Epoch 7, batch 650, loss[loss=0.1889, simple_loss=0.2708, pruned_loss=0.05344, over 16670.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2786, pruned_loss=0.06295, over 3179858.17 frames. ], batch size: 62, lr: 1.01e-02, grad_scale: 2.0 2023-04-28 14:58:35,761 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.778e+02 2.890e+02 3.388e+02 4.314e+02 1.492e+03, threshold=6.776e+02, percent-clipped=7.0 2023-04-28 14:59:01,942 INFO [train.py:904] (4/8) Epoch 7, batch 700, loss[loss=0.2001, simple_loss=0.2796, pruned_loss=0.06031, over 15958.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.278, pruned_loss=0.06217, over 3215515.08 frames. ], batch size: 35, lr: 1.01e-02, grad_scale: 2.0 2023-04-28 14:59:21,954 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-28 14:59:24,159 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5851, 4.0335, 4.4298, 3.2573, 3.8822, 4.1949, 4.0293, 2.4364], device='cuda:4'), covar=tensor([0.0294, 0.0029, 0.0021, 0.0189, 0.0043, 0.0052, 0.0039, 0.0287], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0062, 0.0061, 0.0116, 0.0064, 0.0076, 0.0066, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:00:08,130 INFO [train.py:904] (4/8) Epoch 7, batch 750, loss[loss=0.1799, simple_loss=0.2782, pruned_loss=0.04083, over 17111.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2769, pruned_loss=0.06109, over 3248593.47 frames. ], batch size: 47, lr: 1.01e-02, grad_scale: 2.0 2023-04-28 15:00:30,587 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9545, 4.9043, 4.8064, 4.6146, 4.3679, 4.8480, 4.7725, 4.4562], device='cuda:4'), covar=tensor([0.0469, 0.0330, 0.0218, 0.0209, 0.0866, 0.0328, 0.0350, 0.0531], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0235, 0.0246, 0.0218, 0.0284, 0.0251, 0.0174, 0.0283], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:00:51,558 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.695e+02 2.818e+02 3.177e+02 3.782e+02 5.959e+02, threshold=6.355e+02, percent-clipped=0.0 2023-04-28 15:01:16,765 INFO [train.py:904] (4/8) Epoch 7, batch 800, loss[loss=0.2397, simple_loss=0.3003, pruned_loss=0.08958, over 12516.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2777, pruned_loss=0.06172, over 3259630.43 frames. ], batch size: 247, lr: 1.01e-02, grad_scale: 4.0 2023-04-28 15:01:37,409 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=61717.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:01:57,076 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=61732.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:02:13,234 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=61743.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:02:24,040 INFO [train.py:904] (4/8) Epoch 7, batch 850, loss[loss=0.1711, simple_loss=0.2452, pruned_loss=0.04851, over 16771.00 frames. ], tot_loss[loss=0.199, simple_loss=0.2763, pruned_loss=0.06086, over 3270105.23 frames. ], batch size: 39, lr: 1.01e-02, grad_scale: 4.0 2023-04-28 15:02:42,573 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=61765.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:02:44,071 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4562, 4.0385, 3.8319, 2.1068, 3.2325, 2.4244, 3.8548, 3.7404], device='cuda:4'), covar=tensor([0.0280, 0.0599, 0.0462, 0.1504, 0.0656, 0.0908, 0.0598, 0.0920], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0131, 0.0153, 0.0139, 0.0132, 0.0124, 0.0137, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 15:02:49,980 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9116, 2.1389, 2.2702, 4.6020, 1.9515, 2.9586, 2.2727, 2.3555], device='cuda:4'), covar=tensor([0.0593, 0.2744, 0.1624, 0.0283, 0.3368, 0.1624, 0.2431, 0.3048], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0346, 0.0291, 0.0322, 0.0389, 0.0375, 0.0314, 0.0415], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:03:06,806 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.892e+02 2.876e+02 3.503e+02 4.271e+02 9.516e+02, threshold=7.005e+02, percent-clipped=8.0 2023-04-28 15:03:19,363 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=61793.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:03:33,700 INFO [train.py:904] (4/8) Epoch 7, batch 900, loss[loss=0.2295, simple_loss=0.291, pruned_loss=0.08398, over 16714.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2752, pruned_loss=0.06006, over 3284733.84 frames. ], batch size: 134, lr: 1.01e-02, grad_scale: 4.0 2023-04-28 15:03:36,415 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=61804.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:04:08,026 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3000, 3.6941, 3.5134, 5.2993, 4.7314, 4.8954, 2.0130, 3.8445], device='cuda:4'), covar=tensor([0.1085, 0.0462, 0.0737, 0.0113, 0.0252, 0.0275, 0.1241, 0.0521], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0144, 0.0167, 0.0101, 0.0185, 0.0194, 0.0164, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 15:04:24,064 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-28 15:04:40,613 INFO [train.py:904] (4/8) Epoch 7, batch 950, loss[loss=0.2014, simple_loss=0.2689, pruned_loss=0.06698, over 12056.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2763, pruned_loss=0.0603, over 3289536.37 frames. ], batch size: 246, lr: 1.01e-02, grad_scale: 4.0 2023-04-28 15:05:20,754 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.798e+02 2.637e+02 3.238e+02 3.897e+02 7.916e+02, threshold=6.477e+02, percent-clipped=2.0 2023-04-28 15:05:46,847 INFO [train.py:904] (4/8) Epoch 7, batch 1000, loss[loss=0.1882, simple_loss=0.2644, pruned_loss=0.056, over 15556.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.2752, pruned_loss=0.05971, over 3301389.29 frames. ], batch size: 191, lr: 1.01e-02, grad_scale: 4.0 2023-04-28 15:06:05,339 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=61914.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:06:13,447 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6966, 1.7427, 2.1221, 2.6209, 2.6448, 2.5840, 1.6414, 2.7210], device='cuda:4'), covar=tensor([0.0087, 0.0243, 0.0170, 0.0130, 0.0117, 0.0125, 0.0247, 0.0071], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0153, 0.0137, 0.0137, 0.0141, 0.0100, 0.0148, 0.0089], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 15:06:26,593 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-28 15:06:44,630 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5024, 3.9278, 4.4051, 2.9362, 3.7965, 4.2377, 3.9762, 2.5085], device='cuda:4'), covar=tensor([0.0295, 0.0034, 0.0017, 0.0219, 0.0045, 0.0043, 0.0030, 0.0274], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0062, 0.0061, 0.0117, 0.0065, 0.0076, 0.0067, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:06:56,255 INFO [train.py:904] (4/8) Epoch 7, batch 1050, loss[loss=0.1874, simple_loss=0.2646, pruned_loss=0.05513, over 16867.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2749, pruned_loss=0.05927, over 3305968.23 frames. ], batch size: 42, lr: 1.01e-02, grad_scale: 4.0 2023-04-28 15:07:28,173 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=61975.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:07:37,640 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.626e+02 2.768e+02 3.346e+02 4.170e+02 9.065e+02, threshold=6.692e+02, percent-clipped=2.0 2023-04-28 15:08:07,086 INFO [train.py:904] (4/8) Epoch 7, batch 1100, loss[loss=0.1857, simple_loss=0.2573, pruned_loss=0.05707, over 16445.00 frames. ], tot_loss[loss=0.1956, simple_loss=0.2736, pruned_loss=0.05877, over 3313111.62 frames. ], batch size: 146, lr: 1.01e-02, grad_scale: 4.0 2023-04-28 15:08:53,775 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0926, 5.4689, 5.6326, 5.5115, 5.4953, 6.0796, 5.6541, 5.4196], device='cuda:4'), covar=tensor([0.0716, 0.1773, 0.1503, 0.1558, 0.2481, 0.0919, 0.1099, 0.2055], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0446, 0.0447, 0.0376, 0.0510, 0.0478, 0.0355, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:08:58,999 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3896, 2.3309, 1.8572, 1.9999, 2.7920, 2.6229, 2.8743, 2.8982], device='cuda:4'), covar=tensor([0.0102, 0.0219, 0.0281, 0.0270, 0.0102, 0.0167, 0.0136, 0.0127], device='cuda:4'), in_proj_covar=tensor([0.0103, 0.0174, 0.0173, 0.0173, 0.0170, 0.0176, 0.0165, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:09:15,540 INFO [train.py:904] (4/8) Epoch 7, batch 1150, loss[loss=0.2251, simple_loss=0.308, pruned_loss=0.07104, over 17052.00 frames. ], tot_loss[loss=0.1957, simple_loss=0.274, pruned_loss=0.05868, over 3315304.29 frames. ], batch size: 53, lr: 1.01e-02, grad_scale: 4.0 2023-04-28 15:09:57,382 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.924e+02 2.657e+02 3.127e+02 3.921e+02 1.096e+03, threshold=6.253e+02, percent-clipped=5.0 2023-04-28 15:10:04,071 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62088.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:10:18,762 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62099.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:10:22,593 INFO [train.py:904] (4/8) Epoch 7, batch 1200, loss[loss=0.217, simple_loss=0.3063, pruned_loss=0.06387, over 16680.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2739, pruned_loss=0.05837, over 3319209.69 frames. ], batch size: 62, lr: 1.01e-02, grad_scale: 8.0 2023-04-28 15:10:28,177 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.85 vs. limit=5.0 2023-04-28 15:10:45,097 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-28 15:10:51,405 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62122.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:10:52,543 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6816, 4.6369, 5.1163, 5.1331, 5.1182, 4.7407, 4.7563, 4.4843], device='cuda:4'), covar=tensor([0.0267, 0.0466, 0.0421, 0.0361, 0.0378, 0.0299, 0.0751, 0.0400], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0281, 0.0289, 0.0273, 0.0330, 0.0300, 0.0403, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 15:11:15,835 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-28 15:11:30,498 INFO [train.py:904] (4/8) Epoch 7, batch 1250, loss[loss=0.211, simple_loss=0.3051, pruned_loss=0.05849, over 17251.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2744, pruned_loss=0.05893, over 3313874.60 frames. ], batch size: 52, lr: 1.01e-02, grad_scale: 8.0 2023-04-28 15:11:36,166 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-28 15:11:59,850 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1456, 1.8166, 2.5237, 2.9619, 2.7703, 3.2874, 1.9054, 3.3197], device='cuda:4'), covar=tensor([0.0115, 0.0261, 0.0175, 0.0153, 0.0138, 0.0101, 0.0246, 0.0071], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0153, 0.0137, 0.0138, 0.0142, 0.0100, 0.0147, 0.0089], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 15:12:13,502 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.944e+02 2.700e+02 3.234e+02 4.045e+02 7.667e+02, threshold=6.469e+02, percent-clipped=6.0 2023-04-28 15:12:13,915 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62183.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:12:39,185 INFO [train.py:904] (4/8) Epoch 7, batch 1300, loss[loss=0.1981, simple_loss=0.2825, pruned_loss=0.05685, over 17081.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2741, pruned_loss=0.05906, over 3320867.84 frames. ], batch size: 53, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:12:53,632 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7683, 4.4687, 4.4747, 3.1045, 3.9852, 4.4370, 4.1753, 2.8253], device='cuda:4'), covar=tensor([0.0318, 0.0021, 0.0024, 0.0239, 0.0037, 0.0040, 0.0037, 0.0260], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0063, 0.0063, 0.0116, 0.0065, 0.0077, 0.0068, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:13:17,928 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4182, 2.3887, 2.0068, 2.0741, 2.8085, 2.5990, 3.3607, 3.1044], device='cuda:4'), covar=tensor([0.0044, 0.0213, 0.0279, 0.0278, 0.0145, 0.0216, 0.0122, 0.0119], device='cuda:4'), in_proj_covar=tensor([0.0103, 0.0173, 0.0172, 0.0174, 0.0170, 0.0176, 0.0166, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:13:49,052 INFO [train.py:904] (4/8) Epoch 7, batch 1350, loss[loss=0.2044, simple_loss=0.2702, pruned_loss=0.0693, over 16848.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.275, pruned_loss=0.05938, over 3311373.19 frames. ], batch size: 116, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:14:15,363 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62270.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:14:19,020 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62273.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:14:31,590 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.689e+02 2.487e+02 3.145e+02 3.874e+02 9.778e+02, threshold=6.290e+02, percent-clipped=3.0 2023-04-28 15:14:56,760 INFO [train.py:904] (4/8) Epoch 7, batch 1400, loss[loss=0.1563, simple_loss=0.2403, pruned_loss=0.03612, over 16856.00 frames. ], tot_loss[loss=0.1962, simple_loss=0.2743, pruned_loss=0.05908, over 3312301.71 frames. ], batch size: 42, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:15:42,020 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62334.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:16:04,667 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-28 15:16:06,306 INFO [train.py:904] (4/8) Epoch 7, batch 1450, loss[loss=0.2222, simple_loss=0.2804, pruned_loss=0.08201, over 16934.00 frames. ], tot_loss[loss=0.1957, simple_loss=0.2738, pruned_loss=0.05877, over 3318451.51 frames. ], batch size: 109, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:16:13,156 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-28 15:16:23,107 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 15:16:47,643 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.994e+02 2.891e+02 3.326e+02 4.338e+02 8.815e+02, threshold=6.653e+02, percent-clipped=8.0 2023-04-28 15:16:56,083 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62388.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:16:59,994 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62391.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:17:10,970 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62399.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:17:14,668 INFO [train.py:904] (4/8) Epoch 7, batch 1500, loss[loss=0.218, simple_loss=0.2844, pruned_loss=0.07582, over 16819.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2736, pruned_loss=0.05906, over 3316595.41 frames. ], batch size: 124, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:17:50,095 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8380, 4.8703, 5.4234, 5.4005, 5.3638, 5.0100, 4.9670, 4.7573], device='cuda:4'), covar=tensor([0.0254, 0.0355, 0.0382, 0.0373, 0.0389, 0.0289, 0.0677, 0.0335], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0291, 0.0301, 0.0280, 0.0341, 0.0311, 0.0413, 0.0254], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 15:18:01,111 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=62436.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:18:16,545 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=62447.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:18:23,648 INFO [train.py:904] (4/8) Epoch 7, batch 1550, loss[loss=0.1996, simple_loss=0.2922, pruned_loss=0.05352, over 17114.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2751, pruned_loss=0.06015, over 3320213.16 frames. ], batch size: 48, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:18:24,114 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62452.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:18:39,304 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7163, 3.8084, 2.8735, 2.3671, 2.7796, 2.2191, 3.7967, 3.6976], device='cuda:4'), covar=tensor([0.2085, 0.0605, 0.1202, 0.1728, 0.1802, 0.1603, 0.0440, 0.0806], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0255, 0.0270, 0.0253, 0.0280, 0.0207, 0.0248, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:18:59,836 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62478.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:19:06,019 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.614e+02 2.773e+02 3.245e+02 3.904e+02 9.637e+02, threshold=6.490e+02, percent-clipped=1.0 2023-04-28 15:19:32,091 INFO [train.py:904] (4/8) Epoch 7, batch 1600, loss[loss=0.1832, simple_loss=0.2653, pruned_loss=0.05051, over 17212.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2768, pruned_loss=0.0609, over 3324724.88 frames. ], batch size: 44, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:19:57,121 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-28 15:20:31,537 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62545.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:20:40,967 INFO [train.py:904] (4/8) Epoch 7, batch 1650, loss[loss=0.1945, simple_loss=0.2878, pruned_loss=0.05059, over 17044.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2792, pruned_loss=0.06269, over 3324543.13 frames. ], batch size: 55, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:21:04,856 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62570.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:21:22,701 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.048e+02 2.793e+02 3.410e+02 4.977e+02 9.439e+02, threshold=6.821e+02, percent-clipped=5.0 2023-04-28 15:21:25,424 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62585.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:21:49,315 INFO [train.py:904] (4/8) Epoch 7, batch 1700, loss[loss=0.2169, simple_loss=0.2891, pruned_loss=0.07238, over 16846.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2805, pruned_loss=0.0626, over 3322773.23 frames. ], batch size: 83, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:21:55,162 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62606.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:22:12,978 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=62618.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:22:27,849 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62629.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:22:44,380 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-28 15:22:51,252 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62646.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:22:58,544 INFO [train.py:904] (4/8) Epoch 7, batch 1750, loss[loss=0.1973, simple_loss=0.2768, pruned_loss=0.05893, over 15978.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2815, pruned_loss=0.06316, over 3318152.47 frames. ], batch size: 35, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:23:04,880 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7969, 4.2598, 3.1016, 2.4783, 3.0169, 2.4270, 4.6408, 4.0787], device='cuda:4'), covar=tensor([0.2419, 0.0726, 0.1411, 0.1862, 0.2387, 0.1609, 0.0355, 0.0761], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0256, 0.0272, 0.0256, 0.0282, 0.0209, 0.0250, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:23:04,978 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.79 vs. limit=5.0 2023-04-28 15:23:18,045 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62666.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:23:41,546 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.869e+02 2.853e+02 3.319e+02 4.250e+02 7.879e+02, threshold=6.638e+02, percent-clipped=1.0 2023-04-28 15:24:08,264 INFO [train.py:904] (4/8) Epoch 7, batch 1800, loss[loss=0.2246, simple_loss=0.2889, pruned_loss=0.08014, over 16850.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.283, pruned_loss=0.06362, over 3314893.11 frames. ], batch size: 109, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:24:42,422 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62727.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:25:10,915 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62747.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:25:17,218 INFO [train.py:904] (4/8) Epoch 7, batch 1850, loss[loss=0.1943, simple_loss=0.2944, pruned_loss=0.04713, over 17033.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2843, pruned_loss=0.06385, over 3315178.28 frames. ], batch size: 50, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:25:45,865 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62773.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:25:52,391 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62778.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:25:59,964 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.663e+02 2.724e+02 3.242e+02 3.816e+02 7.402e+02, threshold=6.484e+02, percent-clipped=2.0 2023-04-28 15:26:26,217 INFO [train.py:904] (4/8) Epoch 7, batch 1900, loss[loss=0.1982, simple_loss=0.2707, pruned_loss=0.06284, over 16865.00 frames. ], tot_loss[loss=0.204, simple_loss=0.283, pruned_loss=0.06256, over 3313165.67 frames. ], batch size: 109, lr: 1.00e-02, grad_scale: 8.0 2023-04-28 15:26:38,984 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0722, 5.0648, 4.9431, 4.7160, 4.5253, 4.9806, 4.9602, 4.6064], device='cuda:4'), covar=tensor([0.0442, 0.0290, 0.0197, 0.0206, 0.0828, 0.0289, 0.0261, 0.0496], device='cuda:4'), in_proj_covar=tensor([0.0215, 0.0251, 0.0255, 0.0229, 0.0291, 0.0258, 0.0181, 0.0293], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:27:00,818 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=62826.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:27:11,618 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62834.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:27:36,783 INFO [train.py:904] (4/8) Epoch 7, batch 1950, loss[loss=0.2229, simple_loss=0.2909, pruned_loss=0.07742, over 16861.00 frames. ], tot_loss[loss=0.2042, simple_loss=0.2838, pruned_loss=0.06227, over 3309043.48 frames. ], batch size: 109, lr: 9.99e-03, grad_scale: 8.0 2023-04-28 15:27:43,355 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-28 15:28:19,259 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.671e+02 2.715e+02 3.351e+02 4.043e+02 9.306e+02, threshold=6.703e+02, percent-clipped=2.0 2023-04-28 15:28:44,390 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62901.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:28:45,196 INFO [train.py:904] (4/8) Epoch 7, batch 2000, loss[loss=0.1746, simple_loss=0.2494, pruned_loss=0.04988, over 16802.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.2828, pruned_loss=0.06172, over 3312074.91 frames. ], batch size: 39, lr: 9.99e-03, grad_scale: 8.0 2023-04-28 15:29:24,397 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62929.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:29:39,594 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62941.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:29:55,689 INFO [train.py:904] (4/8) Epoch 7, batch 2050, loss[loss=0.2168, simple_loss=0.3014, pruned_loss=0.06613, over 17068.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2826, pruned_loss=0.06153, over 3308255.09 frames. ], batch size: 50, lr: 9.99e-03, grad_scale: 16.0 2023-04-28 15:30:30,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1328, 4.5248, 3.4237, 2.5879, 3.2357, 2.5785, 4.8488, 4.2279], device='cuda:4'), covar=tensor([0.1881, 0.0526, 0.1119, 0.1675, 0.2224, 0.1482, 0.0279, 0.0772], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0258, 0.0273, 0.0257, 0.0285, 0.0210, 0.0253, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:30:31,319 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=62977.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:30:39,208 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.104e+02 2.968e+02 3.465e+02 4.226e+02 9.943e+02, threshold=6.931e+02, percent-clipped=3.0 2023-04-28 15:31:05,139 INFO [train.py:904] (4/8) Epoch 7, batch 2100, loss[loss=0.2361, simple_loss=0.3089, pruned_loss=0.08171, over 16755.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.284, pruned_loss=0.06323, over 3303598.87 frames. ], batch size: 124, lr: 9.98e-03, grad_scale: 8.0 2023-04-28 15:31:33,249 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63022.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:32:09,455 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63047.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:32:16,868 INFO [train.py:904] (4/8) Epoch 7, batch 2150, loss[loss=0.1718, simple_loss=0.2492, pruned_loss=0.04716, over 16841.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.284, pruned_loss=0.06332, over 3308531.91 frames. ], batch size: 39, lr: 9.98e-03, grad_scale: 8.0 2023-04-28 15:33:00,561 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.945e+02 2.751e+02 3.323e+02 4.118e+02 9.925e+02, threshold=6.647e+02, percent-clipped=2.0 2023-04-28 15:33:17,098 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=63095.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:33:25,757 INFO [train.py:904] (4/8) Epoch 7, batch 2200, loss[loss=0.1846, simple_loss=0.2622, pruned_loss=0.05348, over 16994.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2847, pruned_loss=0.06397, over 3315983.80 frames. ], batch size: 41, lr: 9.98e-03, grad_scale: 8.0 2023-04-28 15:34:03,305 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63129.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:34:33,869 INFO [train.py:904] (4/8) Epoch 7, batch 2250, loss[loss=0.2137, simple_loss=0.3119, pruned_loss=0.05769, over 17131.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2855, pruned_loss=0.06368, over 3326566.94 frames. ], batch size: 49, lr: 9.97e-03, grad_scale: 8.0 2023-04-28 15:35:18,521 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.776e+02 2.937e+02 3.503e+02 4.061e+02 7.762e+02, threshold=7.006e+02, percent-clipped=5.0 2023-04-28 15:35:39,709 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8189, 4.4261, 4.5833, 2.0819, 4.9117, 4.8761, 3.3520, 3.7771], device='cuda:4'), covar=tensor([0.0708, 0.0138, 0.0164, 0.1125, 0.0040, 0.0067, 0.0326, 0.0293], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0100, 0.0086, 0.0145, 0.0074, 0.0092, 0.0124, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 15:35:42,254 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63201.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:35:43,088 INFO [train.py:904] (4/8) Epoch 7, batch 2300, loss[loss=0.1948, simple_loss=0.2744, pruned_loss=0.05759, over 17218.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2847, pruned_loss=0.0632, over 3326242.82 frames. ], batch size: 43, lr: 9.97e-03, grad_scale: 8.0 2023-04-28 15:36:10,152 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63221.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:36:36,754 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63241.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:36:48,219 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=63249.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:36:51,537 INFO [train.py:904] (4/8) Epoch 7, batch 2350, loss[loss=0.2013, simple_loss=0.2955, pruned_loss=0.05359, over 17142.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2852, pruned_loss=0.0639, over 3320522.44 frames. ], batch size: 48, lr: 9.96e-03, grad_scale: 4.0 2023-04-28 15:37:33,970 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63282.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:37:38,003 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.053e+02 2.699e+02 3.274e+02 4.399e+02 8.377e+02, threshold=6.549e+02, percent-clipped=2.0 2023-04-28 15:37:43,610 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=63289.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:37:53,927 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-28 15:38:00,171 INFO [train.py:904] (4/8) Epoch 7, batch 2400, loss[loss=0.2429, simple_loss=0.3052, pruned_loss=0.09032, over 16796.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2866, pruned_loss=0.06471, over 3313163.47 frames. ], batch size: 116, lr: 9.96e-03, grad_scale: 8.0 2023-04-28 15:38:27,646 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63322.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:38:54,505 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-28 15:39:10,358 INFO [train.py:904] (4/8) Epoch 7, batch 2450, loss[loss=0.2313, simple_loss=0.3052, pruned_loss=0.0787, over 16546.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.287, pruned_loss=0.06411, over 3302188.45 frames. ], batch size: 75, lr: 9.96e-03, grad_scale: 8.0 2023-04-28 15:39:34,203 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=63370.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:39:55,096 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.710e+02 2.805e+02 3.565e+02 4.384e+02 1.263e+03, threshold=7.129e+02, percent-clipped=5.0 2023-04-28 15:40:19,245 INFO [train.py:904] (4/8) Epoch 7, batch 2500, loss[loss=0.2692, simple_loss=0.3352, pruned_loss=0.1015, over 12255.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2863, pruned_loss=0.06329, over 3308763.75 frames. ], batch size: 247, lr: 9.95e-03, grad_scale: 4.0 2023-04-28 15:40:44,245 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63420.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:40:46,510 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63422.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:40:55,627 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63429.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:41:26,923 INFO [train.py:904] (4/8) Epoch 7, batch 2550, loss[loss=0.1844, simple_loss=0.2697, pruned_loss=0.04953, over 17121.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.2862, pruned_loss=0.06266, over 3308064.18 frames. ], batch size: 49, lr: 9.95e-03, grad_scale: 4.0 2023-04-28 15:42:01,822 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=63477.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:42:08,279 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63481.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:42:11,021 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63483.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:42:14,138 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.872e+02 3.238e+02 3.870e+02 4.521e+02 8.304e+02, threshold=7.740e+02, percent-clipped=2.0 2023-04-28 15:42:35,897 INFO [train.py:904] (4/8) Epoch 7, batch 2600, loss[loss=0.2089, simple_loss=0.3004, pruned_loss=0.05872, over 16707.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2857, pruned_loss=0.06269, over 3315845.25 frames. ], batch size: 57, lr: 9.94e-03, grad_scale: 4.0 2023-04-28 15:42:51,590 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8713, 4.8357, 5.3989, 5.4457, 5.4462, 5.0112, 4.9975, 4.7835], device='cuda:4'), covar=tensor([0.0240, 0.0358, 0.0313, 0.0312, 0.0320, 0.0262, 0.0787, 0.0333], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0287, 0.0289, 0.0273, 0.0329, 0.0305, 0.0410, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 15:43:42,047 INFO [train.py:904] (4/8) Epoch 7, batch 2650, loss[loss=0.196, simple_loss=0.2877, pruned_loss=0.05216, over 16536.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2862, pruned_loss=0.06233, over 3317605.64 frames. ], batch size: 68, lr: 9.94e-03, grad_scale: 4.0 2023-04-28 15:44:06,160 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63569.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:44:15,583 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63577.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:44:27,534 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.875e+02 2.764e+02 3.348e+02 3.954e+02 9.438e+02, threshold=6.696e+02, percent-clipped=1.0 2023-04-28 15:44:27,928 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63586.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:44:48,763 INFO [train.py:904] (4/8) Epoch 7, batch 2700, loss[loss=0.1976, simple_loss=0.2697, pruned_loss=0.06276, over 16849.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2862, pruned_loss=0.06232, over 3321749.96 frames. ], batch size: 102, lr: 9.94e-03, grad_scale: 4.0 2023-04-28 15:45:28,089 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63630.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:45:45,096 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9994, 4.8130, 4.7913, 4.5684, 4.3929, 4.8476, 4.7985, 4.5274], device='cuda:4'), covar=tensor([0.0441, 0.0365, 0.0243, 0.0243, 0.0938, 0.0319, 0.0303, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0258, 0.0264, 0.0237, 0.0299, 0.0266, 0.0187, 0.0302], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:45:47,248 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3924, 5.7347, 5.4776, 5.5421, 5.0401, 4.7912, 5.1170, 5.8237], device='cuda:4'), covar=tensor([0.0927, 0.0737, 0.0908, 0.0572, 0.0699, 0.0703, 0.0820, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0460, 0.0595, 0.0499, 0.0393, 0.0374, 0.0385, 0.0493, 0.0439], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:45:51,990 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63647.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:45:57,996 INFO [train.py:904] (4/8) Epoch 7, batch 2750, loss[loss=0.2042, simple_loss=0.2785, pruned_loss=0.06491, over 15465.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2857, pruned_loss=0.06144, over 3331124.76 frames. ], batch size: 191, lr: 9.93e-03, grad_scale: 4.0 2023-04-28 15:46:09,301 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6619, 4.6491, 4.8125, 4.7531, 4.7161, 5.3220, 4.8909, 4.5757], device='cuda:4'), covar=tensor([0.1108, 0.1710, 0.1693, 0.1706, 0.2490, 0.0972, 0.1253, 0.2604], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0463, 0.0459, 0.0388, 0.0526, 0.0492, 0.0370, 0.0528], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:46:45,295 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3193, 3.9389, 3.6439, 1.9848, 3.0595, 2.4379, 3.6690, 3.8533], device='cuda:4'), covar=tensor([0.0269, 0.0554, 0.0517, 0.1599, 0.0693, 0.0870, 0.0587, 0.0762], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0139, 0.0156, 0.0140, 0.0133, 0.0123, 0.0138, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 15:46:45,843 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.888e+02 2.696e+02 3.173e+02 3.925e+02 8.532e+02, threshold=6.347e+02, percent-clipped=1.0 2023-04-28 15:46:47,515 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2317, 2.0261, 1.7426, 1.9239, 2.4489, 2.2473, 2.4861, 2.5799], device='cuda:4'), covar=tensor([0.0079, 0.0200, 0.0215, 0.0229, 0.0096, 0.0171, 0.0119, 0.0119], device='cuda:4'), in_proj_covar=tensor([0.0110, 0.0177, 0.0173, 0.0176, 0.0174, 0.0179, 0.0175, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:47:08,527 INFO [train.py:904] (4/8) Epoch 7, batch 2800, loss[loss=0.1936, simple_loss=0.2774, pruned_loss=0.05493, over 17160.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2861, pruned_loss=0.06237, over 3328300.70 frames. ], batch size: 46, lr: 9.93e-03, grad_scale: 8.0 2023-04-28 15:47:57,706 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8985, 4.3034, 3.3408, 2.4190, 3.1366, 2.4268, 4.6021, 4.2427], device='cuda:4'), covar=tensor([0.2238, 0.0586, 0.1282, 0.1820, 0.1996, 0.1648, 0.0306, 0.0656], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0258, 0.0273, 0.0259, 0.0287, 0.0211, 0.0254, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:48:12,842 INFO [train.py:904] (4/8) Epoch 7, batch 2850, loss[loss=0.196, simple_loss=0.2682, pruned_loss=0.06192, over 16696.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2848, pruned_loss=0.06141, over 3338824.44 frames. ], batch size: 89, lr: 9.92e-03, grad_scale: 8.0 2023-04-28 15:48:40,443 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6945, 3.5976, 3.7442, 3.4948, 3.6233, 4.1208, 3.9064, 3.6489], device='cuda:4'), covar=tensor([0.2349, 0.2537, 0.1980, 0.2875, 0.3195, 0.1899, 0.1615, 0.3041], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0457, 0.0455, 0.0387, 0.0521, 0.0485, 0.0367, 0.0522], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 15:48:47,329 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63776.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:48:49,259 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3870, 2.3384, 1.8628, 2.2342, 2.8188, 2.5710, 2.9408, 2.9410], device='cuda:4'), covar=tensor([0.0086, 0.0194, 0.0236, 0.0221, 0.0088, 0.0149, 0.0110, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0108, 0.0175, 0.0171, 0.0173, 0.0171, 0.0176, 0.0173, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:48:50,124 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63778.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:48:55,913 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4417, 3.6327, 3.6543, 1.7585, 3.8738, 3.8281, 2.9811, 2.9635], device='cuda:4'), covar=tensor([0.0656, 0.0107, 0.0122, 0.1048, 0.0057, 0.0092, 0.0340, 0.0343], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0097, 0.0084, 0.0139, 0.0072, 0.0091, 0.0120, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 15:49:02,034 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.714e+02 3.034e+02 3.568e+02 4.214e+02 8.361e+02, threshold=7.135e+02, percent-clipped=4.0 2023-04-28 15:49:23,212 INFO [train.py:904] (4/8) Epoch 7, batch 2900, loss[loss=0.1938, simple_loss=0.2721, pruned_loss=0.05772, over 17215.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2848, pruned_loss=0.06287, over 3322279.11 frames. ], batch size: 45, lr: 9.92e-03, grad_scale: 8.0 2023-04-28 15:49:35,928 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63810.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:49:37,324 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-28 15:49:52,764 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-28 15:50:22,278 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63845.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:50:32,271 INFO [train.py:904] (4/8) Epoch 7, batch 2950, loss[loss=0.2342, simple_loss=0.2957, pruned_loss=0.08636, over 16923.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2847, pruned_loss=0.06324, over 3320161.46 frames. ], batch size: 116, lr: 9.92e-03, grad_scale: 8.0 2023-04-28 15:50:58,351 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63871.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:51:05,395 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63877.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:51:17,102 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.041e+02 3.004e+02 3.573e+02 4.238e+02 9.088e+02, threshold=7.147e+02, percent-clipped=1.0 2023-04-28 15:51:38,108 INFO [train.py:904] (4/8) Epoch 7, batch 3000, loss[loss=0.1984, simple_loss=0.2687, pruned_loss=0.06411, over 16798.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2833, pruned_loss=0.06266, over 3323283.44 frames. ], batch size: 102, lr: 9.91e-03, grad_scale: 8.0 2023-04-28 15:51:38,108 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 15:51:43,539 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1844, 4.9825, 5.1824, 5.3998, 5.5546, 5.0045, 5.5264, 5.5693], device='cuda:4'), covar=tensor([0.1074, 0.0856, 0.1179, 0.0418, 0.0317, 0.0460, 0.0281, 0.0286], device='cuda:4'), in_proj_covar=tensor([0.0481, 0.0589, 0.0753, 0.0594, 0.0453, 0.0455, 0.0466, 0.0514], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:51:44,607 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3882, 1.9590, 2.2355, 3.8417, 1.9864, 2.6021, 2.2405, 2.0153], device='cuda:4'), covar=tensor([0.0687, 0.2937, 0.1670, 0.0367, 0.3443, 0.1673, 0.2363, 0.3301], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0355, 0.0297, 0.0327, 0.0389, 0.0391, 0.0322, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:51:46,841 INFO [train.py:938] (4/8) Epoch 7, validation: loss=0.1489, simple_loss=0.2553, pruned_loss=0.02124, over 944034.00 frames. 2023-04-28 15:51:46,842 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 15:51:54,289 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63906.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:52:07,299 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63916.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:52:19,938 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=63925.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:52:19,951 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63925.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:52:42,088 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63942.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:52:56,533 INFO [train.py:904] (4/8) Epoch 7, batch 3050, loss[loss=0.2001, simple_loss=0.2702, pruned_loss=0.06498, over 16827.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2837, pruned_loss=0.06286, over 3317081.48 frames. ], batch size: 102, lr: 9.91e-03, grad_scale: 4.0 2023-04-28 15:53:30,262 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63977.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:53:43,705 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.653e+02 2.789e+02 3.428e+02 3.925e+02 6.614e+02, threshold=6.856e+02, percent-clipped=0.0 2023-04-28 15:54:06,475 INFO [train.py:904] (4/8) Epoch 7, batch 3100, loss[loss=0.1965, simple_loss=0.2739, pruned_loss=0.05951, over 16486.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2824, pruned_loss=0.06208, over 3322093.38 frames. ], batch size: 75, lr: 9.91e-03, grad_scale: 4.0 2023-04-28 15:54:36,538 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64023.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:55:16,375 INFO [train.py:904] (4/8) Epoch 7, batch 3150, loss[loss=0.2105, simple_loss=0.2842, pruned_loss=0.06845, over 16559.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2825, pruned_loss=0.06274, over 3322055.61 frames. ], batch size: 75, lr: 9.90e-03, grad_scale: 4.0 2023-04-28 15:55:49,709 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64076.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:55:52,147 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64078.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:56:00,069 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64084.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:56:03,681 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.615e+02 2.613e+02 3.331e+02 4.706e+02 9.304e+02, threshold=6.662e+02, percent-clipped=4.0 2023-04-28 15:56:24,282 INFO [train.py:904] (4/8) Epoch 7, batch 3200, loss[loss=0.2077, simple_loss=0.2983, pruned_loss=0.05856, over 16688.00 frames. ], tot_loss[loss=0.2032, simple_loss=0.2816, pruned_loss=0.06238, over 3320790.46 frames. ], batch size: 57, lr: 9.90e-03, grad_scale: 8.0 2023-04-28 15:56:54,783 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64124.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:56:57,816 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64126.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:57:23,723 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 15:57:28,672 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8743, 4.8436, 5.3981, 5.3460, 5.3285, 4.9391, 4.9094, 4.6578], device='cuda:4'), covar=tensor([0.0240, 0.0409, 0.0282, 0.0384, 0.0354, 0.0282, 0.0800, 0.0352], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0298, 0.0296, 0.0286, 0.0345, 0.0318, 0.0427, 0.0256], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 15:57:33,484 INFO [train.py:904] (4/8) Epoch 7, batch 3250, loss[loss=0.2746, simple_loss=0.3438, pruned_loss=0.1027, over 16358.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.2814, pruned_loss=0.06201, over 3323883.47 frames. ], batch size: 146, lr: 9.89e-03, grad_scale: 8.0 2023-04-28 15:57:53,162 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64166.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:58:21,284 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.755e+02 2.809e+02 3.257e+02 3.904e+02 7.986e+02, threshold=6.515e+02, percent-clipped=1.0 2023-04-28 15:58:33,444 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4221, 4.3887, 4.3176, 3.8174, 4.3441, 1.8253, 4.1025, 4.1569], device='cuda:4'), covar=tensor([0.0075, 0.0063, 0.0106, 0.0263, 0.0062, 0.1820, 0.0095, 0.0134], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0100, 0.0152, 0.0148, 0.0117, 0.0160, 0.0135, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 15:58:42,121 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64201.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:58:43,054 INFO [train.py:904] (4/8) Epoch 7, batch 3300, loss[loss=0.1921, simple_loss=0.2794, pruned_loss=0.05241, over 17059.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2823, pruned_loss=0.06214, over 3326544.31 frames. ], batch size: 55, lr: 9.89e-03, grad_scale: 8.0 2023-04-28 15:59:16,266 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64225.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:59:18,149 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.80 vs. limit=2.0 2023-04-28 15:59:25,446 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64232.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:59:40,399 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64242.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 15:59:51,969 INFO [train.py:904] (4/8) Epoch 7, batch 3350, loss[loss=0.1986, simple_loss=0.289, pruned_loss=0.05408, over 17079.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2831, pruned_loss=0.0619, over 3329176.09 frames. ], batch size: 55, lr: 9.89e-03, grad_scale: 8.0 2023-04-28 16:00:20,499 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64272.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:00:22,691 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64273.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:00:31,195 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7678, 4.2183, 3.3150, 2.4342, 3.0801, 2.4620, 4.6142, 4.0395], device='cuda:4'), covar=tensor([0.2308, 0.0584, 0.1138, 0.1725, 0.2010, 0.1459, 0.0309, 0.0697], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0256, 0.0271, 0.0257, 0.0287, 0.0211, 0.0253, 0.0280], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:00:42,046 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.854e+02 2.871e+02 3.440e+02 4.131e+02 8.469e+02, threshold=6.880e+02, percent-clipped=3.0 2023-04-28 16:00:45,938 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64290.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:00:49,504 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64293.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:01:02,162 INFO [train.py:904] (4/8) Epoch 7, batch 3400, loss[loss=0.2298, simple_loss=0.3062, pruned_loss=0.07674, over 15503.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2826, pruned_loss=0.06137, over 3328850.69 frames. ], batch size: 191, lr: 9.88e-03, grad_scale: 8.0 2023-04-28 16:02:00,681 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4967, 3.3253, 3.7705, 2.6040, 3.3824, 3.7225, 3.6256, 2.0785], device='cuda:4'), covar=tensor([0.0287, 0.0086, 0.0027, 0.0208, 0.0056, 0.0056, 0.0038, 0.0295], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0062, 0.0062, 0.0115, 0.0066, 0.0077, 0.0068, 0.0112], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:02:02,933 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64345.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:02:03,014 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8252, 1.7132, 1.4392, 1.5072, 1.9232, 1.6063, 1.7631, 1.9406], device='cuda:4'), covar=tensor([0.0074, 0.0172, 0.0217, 0.0219, 0.0105, 0.0170, 0.0121, 0.0109], device='cuda:4'), in_proj_covar=tensor([0.0111, 0.0177, 0.0174, 0.0175, 0.0174, 0.0178, 0.0177, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:02:10,853 INFO [train.py:904] (4/8) Epoch 7, batch 3450, loss[loss=0.15, simple_loss=0.2295, pruned_loss=0.0353, over 16787.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.281, pruned_loss=0.06059, over 3324360.70 frames. ], batch size: 39, lr: 9.88e-03, grad_scale: 8.0 2023-04-28 16:02:44,317 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4256, 4.2640, 3.8672, 2.0948, 2.9446, 2.6612, 3.7902, 3.8643], device='cuda:4'), covar=tensor([0.0327, 0.0556, 0.0542, 0.1624, 0.0869, 0.0907, 0.0730, 0.0956], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0140, 0.0155, 0.0139, 0.0133, 0.0124, 0.0138, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 16:02:45,464 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7019, 4.6259, 4.5586, 4.3605, 4.1521, 4.6613, 4.5025, 4.3466], device='cuda:4'), covar=tensor([0.0506, 0.0530, 0.0257, 0.0239, 0.1010, 0.0444, 0.0360, 0.0545], device='cuda:4'), in_proj_covar=tensor([0.0220, 0.0259, 0.0268, 0.0240, 0.0303, 0.0268, 0.0186, 0.0303], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:02:49,997 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64379.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:02:52,853 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1201, 5.6619, 5.7891, 5.6108, 5.6887, 6.2130, 5.7838, 5.5341], device='cuda:4'), covar=tensor([0.0724, 0.1633, 0.1412, 0.1681, 0.2348, 0.0843, 0.1062, 0.2060], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0453, 0.0453, 0.0383, 0.0516, 0.0486, 0.0366, 0.0518], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:03:01,784 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.744e+02 2.661e+02 3.216e+02 3.896e+02 8.084e+02, threshold=6.431e+02, percent-clipped=1.0 2023-04-28 16:03:07,092 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-28 16:03:21,642 INFO [train.py:904] (4/8) Epoch 7, batch 3500, loss[loss=0.241, simple_loss=0.3121, pruned_loss=0.08501, over 16308.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2794, pruned_loss=0.0602, over 3327310.71 frames. ], batch size: 165, lr: 9.87e-03, grad_scale: 8.0 2023-04-28 16:03:28,802 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64406.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:04:30,074 INFO [train.py:904] (4/8) Epoch 7, batch 3550, loss[loss=0.2054, simple_loss=0.2777, pruned_loss=0.06654, over 16895.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2791, pruned_loss=0.06016, over 3325255.34 frames. ], batch size: 116, lr: 9.87e-03, grad_scale: 8.0 2023-04-28 16:04:48,203 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64465.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:04:50,574 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64466.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:05:18,416 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.738e+02 2.678e+02 3.341e+02 4.465e+02 1.003e+03, threshold=6.682e+02, percent-clipped=6.0 2023-04-28 16:05:38,578 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64501.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:05:39,312 INFO [train.py:904] (4/8) Epoch 7, batch 3600, loss[loss=0.1716, simple_loss=0.2564, pruned_loss=0.04339, over 17192.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2796, pruned_loss=0.06071, over 3307079.98 frames. ], batch size: 46, lr: 9.87e-03, grad_scale: 8.0 2023-04-28 16:05:56,510 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64514.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:06:07,081 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9220, 2.6414, 2.6984, 1.8357, 2.5098, 2.6925, 2.6208, 1.8096], device='cuda:4'), covar=tensor([0.0277, 0.0050, 0.0038, 0.0240, 0.0058, 0.0064, 0.0047, 0.0250], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0061, 0.0061, 0.0115, 0.0065, 0.0076, 0.0068, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:06:14,276 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64526.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:06:22,257 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-28 16:06:45,516 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64549.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:06:48,844 INFO [train.py:904] (4/8) Epoch 7, batch 3650, loss[loss=0.2011, simple_loss=0.2581, pruned_loss=0.07208, over 16727.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2786, pruned_loss=0.06146, over 3301015.00 frames. ], batch size: 83, lr: 9.86e-03, grad_scale: 8.0 2023-04-28 16:07:19,300 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64572.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:07:40,346 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.014e+02 2.562e+02 3.124e+02 4.018e+02 1.198e+03, threshold=6.249e+02, percent-clipped=5.0 2023-04-28 16:07:42,459 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64588.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:07:44,516 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64589.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:08:01,998 INFO [train.py:904] (4/8) Epoch 7, batch 3700, loss[loss=0.1886, simple_loss=0.2588, pruned_loss=0.05922, over 16853.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2772, pruned_loss=0.06287, over 3286859.64 frames. ], batch size: 96, lr: 9.86e-03, grad_scale: 8.0 2023-04-28 16:08:11,413 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6552, 4.5264, 4.5250, 4.3669, 4.1923, 4.5720, 4.4318, 4.2901], device='cuda:4'), covar=tensor([0.0434, 0.0442, 0.0211, 0.0188, 0.0752, 0.0352, 0.0343, 0.0486], device='cuda:4'), in_proj_covar=tensor([0.0216, 0.0253, 0.0261, 0.0235, 0.0298, 0.0261, 0.0182, 0.0295], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:08:27,962 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64620.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:08:29,422 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5630, 3.8493, 4.1312, 1.8428, 4.3002, 4.3668, 3.2213, 3.2833], device='cuda:4'), covar=tensor([0.0828, 0.0147, 0.0139, 0.1318, 0.0077, 0.0091, 0.0329, 0.0428], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0098, 0.0085, 0.0139, 0.0073, 0.0093, 0.0120, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 16:08:43,524 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9829, 3.4825, 2.4883, 4.6301, 4.1420, 4.2458, 1.6239, 2.8582], device='cuda:4'), covar=tensor([0.1079, 0.0357, 0.0996, 0.0076, 0.0215, 0.0367, 0.1126, 0.0727], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0145, 0.0168, 0.0107, 0.0198, 0.0199, 0.0165, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 16:08:53,073 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6541, 3.7396, 2.8205, 2.1928, 2.5586, 2.0336, 3.6042, 3.4705], device='cuda:4'), covar=tensor([0.2185, 0.0566, 0.1260, 0.2035, 0.2192, 0.1822, 0.0491, 0.0887], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0261, 0.0277, 0.0262, 0.0293, 0.0215, 0.0256, 0.0286], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:08:54,682 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.64 vs. limit=5.0 2023-04-28 16:09:09,894 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64648.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:09:12,276 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64650.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:09:14,069 INFO [train.py:904] (4/8) Epoch 7, batch 3750, loss[loss=0.2312, simple_loss=0.2971, pruned_loss=0.08261, over 16443.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2771, pruned_loss=0.06401, over 3279017.58 frames. ], batch size: 146, lr: 9.86e-03, grad_scale: 8.0 2023-04-28 16:09:55,061 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64679.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:10:04,721 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.745e+02 2.843e+02 3.454e+02 4.284e+02 7.405e+02, threshold=6.908e+02, percent-clipped=2.0 2023-04-28 16:10:24,925 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64701.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:10:25,829 INFO [train.py:904] (4/8) Epoch 7, batch 3800, loss[loss=0.222, simple_loss=0.2906, pruned_loss=0.07671, over 16330.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2787, pruned_loss=0.06545, over 3282421.15 frames. ], batch size: 165, lr: 9.85e-03, grad_scale: 4.0 2023-04-28 16:10:27,928 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64703.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:10:38,642 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64709.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:11:05,637 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64727.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:11:23,127 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 16:11:31,481 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64745.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:11:39,994 INFO [train.py:904] (4/8) Epoch 7, batch 3850, loss[loss=0.2172, simple_loss=0.2901, pruned_loss=0.07214, over 16276.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2787, pruned_loss=0.06619, over 3271831.26 frames. ], batch size: 35, lr: 9.85e-03, grad_scale: 4.0 2023-04-28 16:11:41,086 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64752.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:11:58,590 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64764.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:12:06,811 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0899, 3.4895, 3.1461, 1.9555, 2.7834, 2.2000, 3.4256, 3.4160], device='cuda:4'), covar=tensor([0.0192, 0.0534, 0.0507, 0.1460, 0.0685, 0.0897, 0.0433, 0.0665], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0138, 0.0154, 0.0139, 0.0132, 0.0123, 0.0136, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 16:12:09,671 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8868, 1.9616, 2.2243, 2.7722, 2.7943, 2.7419, 1.5539, 2.9691], device='cuda:4'), covar=tensor([0.0084, 0.0247, 0.0169, 0.0133, 0.0121, 0.0126, 0.0291, 0.0057], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0155, 0.0139, 0.0142, 0.0146, 0.0105, 0.0148, 0.0095], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 16:12:33,098 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.798e+02 2.797e+02 3.034e+02 3.668e+02 1.110e+03, threshold=6.069e+02, percent-clipped=1.0 2023-04-28 16:12:52,610 INFO [train.py:904] (4/8) Epoch 7, batch 3900, loss[loss=0.2052, simple_loss=0.2701, pruned_loss=0.07017, over 17020.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2778, pruned_loss=0.06652, over 3284801.45 frames. ], batch size: 110, lr: 9.84e-03, grad_scale: 4.0 2023-04-28 16:13:00,366 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64806.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:13:05,072 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-28 16:13:08,261 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64813.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:13:20,517 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64821.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:14:02,898 INFO [train.py:904] (4/8) Epoch 7, batch 3950, loss[loss=0.1922, simple_loss=0.2624, pruned_loss=0.06105, over 16522.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2771, pruned_loss=0.06691, over 3286997.67 frames. ], batch size: 75, lr: 9.84e-03, grad_scale: 4.0 2023-04-28 16:14:05,031 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1052, 5.0395, 4.8194, 4.1532, 4.9999, 1.7925, 4.7178, 4.6988], device='cuda:4'), covar=tensor([0.0051, 0.0035, 0.0096, 0.0283, 0.0046, 0.1946, 0.0081, 0.0130], device='cuda:4'), in_proj_covar=tensor([0.0112, 0.0099, 0.0150, 0.0146, 0.0115, 0.0158, 0.0133, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:14:53,035 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.973e+02 2.882e+02 3.338e+02 4.272e+02 6.667e+02, threshold=6.675e+02, percent-clipped=4.0 2023-04-28 16:14:53,479 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64888.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:14:57,046 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7605, 4.7305, 5.2237, 5.1772, 5.2042, 4.7938, 4.7922, 4.3773], device='cuda:4'), covar=tensor([0.0257, 0.0410, 0.0298, 0.0399, 0.0418, 0.0239, 0.0777, 0.0421], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0293, 0.0291, 0.0282, 0.0337, 0.0306, 0.0413, 0.0250], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 16:15:12,235 INFO [train.py:904] (4/8) Epoch 7, batch 4000, loss[loss=0.2006, simple_loss=0.2699, pruned_loss=0.0657, over 17070.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2769, pruned_loss=0.06716, over 3291746.86 frames. ], batch size: 55, lr: 9.84e-03, grad_scale: 8.0 2023-04-28 16:15:53,649 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-28 16:16:02,045 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=64936.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:16:16,961 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64945.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:16:24,725 INFO [train.py:904] (4/8) Epoch 7, batch 4050, loss[loss=0.1999, simple_loss=0.2756, pruned_loss=0.06213, over 16696.00 frames. ], tot_loss[loss=0.2032, simple_loss=0.2759, pruned_loss=0.06531, over 3294613.77 frames. ], batch size: 134, lr: 9.83e-03, grad_scale: 8.0 2023-04-28 16:16:59,703 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64976.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:17:14,228 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0730, 5.0676, 5.5768, 5.4981, 5.4992, 5.1155, 5.0773, 4.6547], device='cuda:4'), covar=tensor([0.0208, 0.0330, 0.0211, 0.0318, 0.0331, 0.0207, 0.0659, 0.0338], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0288, 0.0287, 0.0278, 0.0335, 0.0302, 0.0407, 0.0244], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 16:17:16,149 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.366e+02 2.141e+02 2.464e+02 3.078e+02 9.943e+02, threshold=4.928e+02, percent-clipped=3.0 2023-04-28 16:17:35,693 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65001.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:17:36,414 INFO [train.py:904] (4/8) Epoch 7, batch 4100, loss[loss=0.2202, simple_loss=0.2998, pruned_loss=0.0703, over 16302.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2768, pruned_loss=0.06423, over 3262681.33 frames. ], batch size: 165, lr: 9.83e-03, grad_scale: 8.0 2023-04-28 16:17:39,308 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65004.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:18:29,669 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65037.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 16:18:31,552 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4948, 3.6512, 2.5898, 2.2215, 2.5819, 2.1187, 3.6617, 3.5417], device='cuda:4'), covar=tensor([0.2503, 0.0785, 0.1612, 0.1893, 0.2332, 0.1818, 0.0544, 0.0805], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0258, 0.0277, 0.0264, 0.0299, 0.0215, 0.0258, 0.0283], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:18:47,164 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65049.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:18:51,077 INFO [train.py:904] (4/8) Epoch 7, batch 4150, loss[loss=0.1975, simple_loss=0.2827, pruned_loss=0.0562, over 16665.00 frames. ], tot_loss[loss=0.2104, simple_loss=0.2851, pruned_loss=0.06782, over 3239410.76 frames. ], batch size: 62, lr: 9.83e-03, grad_scale: 8.0 2023-04-28 16:18:56,785 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=65055.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:19:01,896 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65059.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:19:44,054 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.893e+02 2.789e+02 3.341e+02 3.997e+02 8.596e+02, threshold=6.682e+02, percent-clipped=9.0 2023-04-28 16:20:03,712 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65101.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:20:04,500 INFO [train.py:904] (4/8) Epoch 7, batch 4200, loss[loss=0.266, simple_loss=0.335, pruned_loss=0.09851, over 11471.00 frames. ], tot_loss[loss=0.2171, simple_loss=0.2931, pruned_loss=0.07054, over 3198656.13 frames. ], batch size: 246, lr: 9.82e-03, grad_scale: 8.0 2023-04-28 16:20:13,269 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65108.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:20:24,683 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65116.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:20:32,673 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65121.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:21:03,991 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1162, 1.9143, 2.0879, 3.7354, 1.8562, 2.5185, 2.0871, 2.0503], device='cuda:4'), covar=tensor([0.0779, 0.3062, 0.1642, 0.0341, 0.3285, 0.1692, 0.2551, 0.2826], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0355, 0.0294, 0.0318, 0.0382, 0.0386, 0.0318, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:21:16,725 INFO [train.py:904] (4/8) Epoch 7, batch 4250, loss[loss=0.2159, simple_loss=0.289, pruned_loss=0.07142, over 11836.00 frames. ], tot_loss[loss=0.2189, simple_loss=0.2967, pruned_loss=0.07051, over 3182302.22 frames. ], batch size: 248, lr: 9.82e-03, grad_scale: 8.0 2023-04-28 16:21:42,348 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65169.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:22:08,030 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.682e+02 2.671e+02 3.307e+02 3.991e+02 9.211e+02, threshold=6.614e+02, percent-clipped=3.0 2023-04-28 16:22:29,059 INFO [train.py:904] (4/8) Epoch 7, batch 4300, loss[loss=0.246, simple_loss=0.3231, pruned_loss=0.08443, over 16355.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.2979, pruned_loss=0.06954, over 3177325.35 frames. ], batch size: 146, lr: 9.81e-03, grad_scale: 8.0 2023-04-28 16:23:07,909 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1395, 5.0979, 4.9483, 4.3564, 5.0078, 2.1048, 4.7137, 4.8844], device='cuda:4'), covar=tensor([0.0037, 0.0031, 0.0076, 0.0242, 0.0041, 0.1687, 0.0074, 0.0107], device='cuda:4'), in_proj_covar=tensor([0.0107, 0.0094, 0.0142, 0.0141, 0.0111, 0.0154, 0.0127, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:23:30,895 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-28 16:23:31,791 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65245.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:23:41,835 INFO [train.py:904] (4/8) Epoch 7, batch 4350, loss[loss=0.2255, simple_loss=0.3065, pruned_loss=0.07221, over 17216.00 frames. ], tot_loss[loss=0.2208, simple_loss=0.3008, pruned_loss=0.07035, over 3180570.55 frames. ], batch size: 52, lr: 9.81e-03, grad_scale: 8.0 2023-04-28 16:24:08,558 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6600, 4.6905, 5.1875, 5.1186, 5.1579, 4.6548, 4.7012, 4.3659], device='cuda:4'), covar=tensor([0.0200, 0.0339, 0.0243, 0.0343, 0.0347, 0.0248, 0.0755, 0.0405], device='cuda:4'), in_proj_covar=tensor([0.0277, 0.0277, 0.0277, 0.0269, 0.0324, 0.0294, 0.0394, 0.0236], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 16:24:30,778 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=65285.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:24:33,879 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.656e+02 2.885e+02 3.415e+02 4.232e+02 9.360e+02, threshold=6.829e+02, percent-clipped=2.0 2023-04-28 16:24:42,764 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65293.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:24:54,960 INFO [train.py:904] (4/8) Epoch 7, batch 4400, loss[loss=0.2095, simple_loss=0.2896, pruned_loss=0.06469, over 17256.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.3026, pruned_loss=0.07121, over 3195386.96 frames. ], batch size: 52, lr: 9.81e-03, grad_scale: 8.0 2023-04-28 16:24:57,870 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65304.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:25:39,701 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65332.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 16:25:41,065 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7776, 1.7263, 2.1115, 2.6210, 2.5287, 2.9546, 1.5725, 2.8470], device='cuda:4'), covar=tensor([0.0092, 0.0260, 0.0166, 0.0138, 0.0139, 0.0083, 0.0274, 0.0062], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0157, 0.0140, 0.0142, 0.0150, 0.0107, 0.0153, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 16:25:56,860 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-04-28 16:25:59,700 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65346.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:26:07,813 INFO [train.py:904] (4/8) Epoch 7, batch 4450, loss[loss=0.2639, simple_loss=0.3421, pruned_loss=0.09283, over 16170.00 frames. ], tot_loss[loss=0.224, simple_loss=0.3053, pruned_loss=0.0713, over 3211774.55 frames. ], batch size: 35, lr: 9.80e-03, grad_scale: 8.0 2023-04-28 16:26:08,114 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65352.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:26:17,488 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65359.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:26:17,647 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6719, 2.9379, 2.5763, 4.2172, 3.2037, 4.1563, 1.4073, 3.0654], device='cuda:4'), covar=tensor([0.1176, 0.0512, 0.0954, 0.0096, 0.0198, 0.0233, 0.1407, 0.0594], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0145, 0.0168, 0.0104, 0.0198, 0.0193, 0.0165, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 16:26:35,205 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0931, 5.1300, 4.9422, 4.8056, 4.5727, 4.9458, 4.8494, 4.6179], device='cuda:4'), covar=tensor([0.0334, 0.0172, 0.0146, 0.0134, 0.0717, 0.0214, 0.0223, 0.0442], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0229, 0.0237, 0.0211, 0.0268, 0.0235, 0.0166, 0.0267], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:26:58,766 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.442e+02 2.495e+02 2.995e+02 3.500e+02 6.394e+02, threshold=5.989e+02, percent-clipped=0.0 2023-04-28 16:27:16,278 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8390, 5.1521, 4.9272, 4.8990, 4.6880, 4.4363, 4.6103, 5.2160], device='cuda:4'), covar=tensor([0.0792, 0.0636, 0.0846, 0.0541, 0.0579, 0.0790, 0.0709, 0.0667], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0560, 0.0469, 0.0368, 0.0353, 0.0372, 0.0463, 0.0412], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:27:19,324 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65401.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:27:20,139 INFO [train.py:904] (4/8) Epoch 7, batch 4500, loss[loss=0.2138, simple_loss=0.3017, pruned_loss=0.06292, over 16597.00 frames. ], tot_loss[loss=0.2235, simple_loss=0.3051, pruned_loss=0.07095, over 3212902.63 frames. ], batch size: 62, lr: 9.80e-03, grad_scale: 8.0 2023-04-28 16:27:26,479 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65407.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:27:28,734 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65408.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:27:32,338 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65411.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:28:00,100 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-04-28 16:28:27,210 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65449.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:28:31,285 INFO [train.py:904] (4/8) Epoch 7, batch 4550, loss[loss=0.2535, simple_loss=0.3363, pruned_loss=0.0854, over 15469.00 frames. ], tot_loss[loss=0.2241, simple_loss=0.3056, pruned_loss=0.07127, over 3220722.75 frames. ], batch size: 191, lr: 9.80e-03, grad_scale: 8.0 2023-04-28 16:28:38,294 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65456.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:29:09,970 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9164, 4.2767, 3.2751, 2.4195, 3.1395, 2.3785, 4.5331, 4.1004], device='cuda:4'), covar=tensor([0.2178, 0.0525, 0.1298, 0.1695, 0.2024, 0.1531, 0.0335, 0.0565], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0254, 0.0273, 0.0262, 0.0291, 0.0212, 0.0254, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:29:21,410 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.690e+02 2.206e+02 2.596e+02 3.051e+02 6.081e+02, threshold=5.193e+02, percent-clipped=1.0 2023-04-28 16:29:41,302 INFO [train.py:904] (4/8) Epoch 7, batch 4600, loss[loss=0.2228, simple_loss=0.3014, pruned_loss=0.07207, over 16173.00 frames. ], tot_loss[loss=0.2239, simple_loss=0.306, pruned_loss=0.0709, over 3228204.56 frames. ], batch size: 35, lr: 9.79e-03, grad_scale: 8.0 2023-04-28 16:29:55,489 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9383, 5.2498, 5.0184, 5.0396, 4.7037, 4.4358, 4.7000, 5.3645], device='cuda:4'), covar=tensor([0.0812, 0.0680, 0.0851, 0.0501, 0.0633, 0.0771, 0.0759, 0.0702], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0540, 0.0456, 0.0354, 0.0341, 0.0361, 0.0448, 0.0396], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:30:24,203 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=65532.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 16:30:50,580 INFO [train.py:904] (4/8) Epoch 7, batch 4650, loss[loss=0.1997, simple_loss=0.2811, pruned_loss=0.05912, over 17048.00 frames. ], tot_loss[loss=0.2233, simple_loss=0.3048, pruned_loss=0.07086, over 3208733.13 frames. ], batch size: 55, lr: 9.79e-03, grad_scale: 8.0 2023-04-28 16:31:05,247 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7326, 3.9587, 4.2462, 2.0226, 4.5523, 4.4941, 3.2293, 3.3024], device='cuda:4'), covar=tensor([0.0723, 0.0146, 0.0119, 0.1104, 0.0033, 0.0044, 0.0328, 0.0401], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0096, 0.0081, 0.0137, 0.0069, 0.0086, 0.0118, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 16:31:41,970 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.329e+02 2.212e+02 2.707e+02 3.141e+02 8.468e+02, threshold=5.414e+02, percent-clipped=4.0 2023-04-28 16:31:49,839 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65593.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 16:32:03,152 INFO [train.py:904] (4/8) Epoch 7, batch 4700, loss[loss=0.2074, simple_loss=0.2766, pruned_loss=0.06916, over 11553.00 frames. ], tot_loss[loss=0.2209, simple_loss=0.3019, pruned_loss=0.06993, over 3200786.73 frames. ], batch size: 248, lr: 9.78e-03, grad_scale: 8.0 2023-04-28 16:32:45,696 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65632.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:32:55,394 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-28 16:32:56,219 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65641.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:33:00,909 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7248, 4.5988, 5.1665, 5.1328, 5.1251, 4.7469, 4.6570, 4.3902], device='cuda:4'), covar=tensor([0.0206, 0.0390, 0.0294, 0.0314, 0.0335, 0.0243, 0.0816, 0.0357], device='cuda:4'), in_proj_covar=tensor([0.0273, 0.0273, 0.0272, 0.0267, 0.0320, 0.0291, 0.0394, 0.0232], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 16:33:12,631 INFO [train.py:904] (4/8) Epoch 7, batch 4750, loss[loss=0.2658, simple_loss=0.3264, pruned_loss=0.1025, over 11953.00 frames. ], tot_loss[loss=0.217, simple_loss=0.2981, pruned_loss=0.06799, over 3202072.18 frames. ], batch size: 248, lr: 9.78e-03, grad_scale: 8.0 2023-04-28 16:33:53,638 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65680.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:34:03,333 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.519e+02 2.191e+02 2.638e+02 3.130e+02 6.213e+02, threshold=5.275e+02, percent-clipped=2.0 2023-04-28 16:34:22,789 INFO [train.py:904] (4/8) Epoch 7, batch 4800, loss[loss=0.1996, simple_loss=0.2918, pruned_loss=0.0537, over 15250.00 frames. ], tot_loss[loss=0.2141, simple_loss=0.2953, pruned_loss=0.06642, over 3181817.97 frames. ], batch size: 190, lr: 9.78e-03, grad_scale: 8.0 2023-04-28 16:34:36,308 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65711.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:34:48,158 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9988, 1.8463, 2.0915, 3.5912, 1.8745, 2.4104, 2.0377, 2.0139], device='cuda:4'), covar=tensor([0.0830, 0.2637, 0.1597, 0.0370, 0.3231, 0.1663, 0.2559, 0.2526], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0352, 0.0296, 0.0317, 0.0387, 0.0384, 0.0315, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:35:36,916 INFO [train.py:904] (4/8) Epoch 7, batch 4850, loss[loss=0.2221, simple_loss=0.3103, pruned_loss=0.06694, over 16359.00 frames. ], tot_loss[loss=0.2138, simple_loss=0.2963, pruned_loss=0.0657, over 3171194.39 frames. ], batch size: 146, lr: 9.77e-03, grad_scale: 8.0 2023-04-28 16:35:45,500 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65759.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:36:28,553 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.411e+02 2.781e+02 3.259e+02 6.496e+02, threshold=5.561e+02, percent-clipped=1.0 2023-04-28 16:36:48,885 INFO [train.py:904] (4/8) Epoch 7, batch 4900, loss[loss=0.2218, simple_loss=0.3043, pruned_loss=0.06966, over 16861.00 frames. ], tot_loss[loss=0.2127, simple_loss=0.2958, pruned_loss=0.06477, over 3173577.70 frames. ], batch size: 116, lr: 9.77e-03, grad_scale: 8.0 2023-04-28 16:37:28,019 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5503, 4.3368, 4.6032, 4.8188, 4.9468, 4.4621, 4.9767, 4.9067], device='cuda:4'), covar=tensor([0.1148, 0.0941, 0.1214, 0.0434, 0.0338, 0.0769, 0.0313, 0.0382], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0524, 0.0659, 0.0529, 0.0398, 0.0409, 0.0408, 0.0453], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:37:29,266 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5501, 3.7676, 1.8662, 4.0767, 2.5463, 3.9603, 2.0560, 2.7739], device='cuda:4'), covar=tensor([0.0169, 0.0229, 0.1597, 0.0046, 0.0765, 0.0300, 0.1434, 0.0675], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0156, 0.0178, 0.0093, 0.0161, 0.0194, 0.0187, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 16:37:43,996 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=65838.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 16:38:02,371 INFO [train.py:904] (4/8) Epoch 7, batch 4950, loss[loss=0.199, simple_loss=0.2798, pruned_loss=0.05915, over 16280.00 frames. ], tot_loss[loss=0.211, simple_loss=0.2947, pruned_loss=0.06367, over 3176552.94 frames. ], batch size: 35, lr: 9.77e-03, grad_scale: 8.0 2023-04-28 16:38:50,888 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6596, 4.4694, 4.7027, 4.9486, 5.0495, 4.5211, 5.0447, 4.9698], device='cuda:4'), covar=tensor([0.1222, 0.0904, 0.1231, 0.0454, 0.0349, 0.0665, 0.0356, 0.0414], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0522, 0.0655, 0.0531, 0.0397, 0.0408, 0.0407, 0.0453], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:38:53,490 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.703e+02 2.669e+02 3.212e+02 3.846e+02 7.226e+02, threshold=6.424e+02, percent-clipped=6.0 2023-04-28 16:38:53,917 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65888.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 16:39:08,369 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65899.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 16:39:12,680 INFO [train.py:904] (4/8) Epoch 7, batch 5000, loss[loss=0.2128, simple_loss=0.3054, pruned_loss=0.06014, over 16883.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.2968, pruned_loss=0.0639, over 3196893.63 frames. ], batch size: 116, lr: 9.76e-03, grad_scale: 8.0 2023-04-28 16:40:09,098 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65941.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:40:14,047 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5748, 4.6500, 4.9856, 4.9277, 4.9914, 4.6738, 4.4215, 4.4414], device='cuda:4'), covar=tensor([0.0340, 0.0508, 0.0405, 0.0578, 0.0556, 0.0320, 0.1167, 0.0368], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0265, 0.0265, 0.0262, 0.0316, 0.0286, 0.0383, 0.0228], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 16:40:19,576 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8828, 2.2971, 2.5395, 4.7165, 1.9818, 3.0556, 2.4548, 2.6067], device='cuda:4'), covar=tensor([0.0630, 0.2669, 0.1461, 0.0254, 0.3296, 0.1509, 0.2310, 0.2318], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0352, 0.0297, 0.0320, 0.0389, 0.0385, 0.0317, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:40:23,710 INFO [train.py:904] (4/8) Epoch 7, batch 5050, loss[loss=0.1878, simple_loss=0.2756, pruned_loss=0.05003, over 16459.00 frames. ], tot_loss[loss=0.2117, simple_loss=0.2967, pruned_loss=0.06338, over 3215986.89 frames. ], batch size: 68, lr: 9.76e-03, grad_scale: 8.0 2023-04-28 16:41:14,575 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.911e+02 2.609e+02 2.967e+02 3.441e+02 6.265e+02, threshold=5.933e+02, percent-clipped=0.0 2023-04-28 16:41:16,132 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=65989.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:41:35,929 INFO [train.py:904] (4/8) Epoch 7, batch 5100, loss[loss=0.2109, simple_loss=0.2866, pruned_loss=0.06763, over 16701.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.2946, pruned_loss=0.06252, over 3215730.49 frames. ], batch size: 62, lr: 9.75e-03, grad_scale: 8.0 2023-04-28 16:42:17,473 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-04-28 16:42:46,398 INFO [train.py:904] (4/8) Epoch 7, batch 5150, loss[loss=0.2212, simple_loss=0.3131, pruned_loss=0.0646, over 16334.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2946, pruned_loss=0.06205, over 3188640.17 frames. ], batch size: 146, lr: 9.75e-03, grad_scale: 8.0 2023-04-28 16:43:37,190 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.675e+02 2.591e+02 3.145e+02 3.855e+02 6.325e+02, threshold=6.290e+02, percent-clipped=3.0 2023-04-28 16:43:55,948 INFO [train.py:904] (4/8) Epoch 7, batch 5200, loss[loss=0.1936, simple_loss=0.2801, pruned_loss=0.05352, over 16952.00 frames. ], tot_loss[loss=0.2087, simple_loss=0.2938, pruned_loss=0.06182, over 3184631.84 frames. ], batch size: 96, lr: 9.75e-03, grad_scale: 8.0 2023-04-28 16:44:56,223 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66144.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:45:06,331 INFO [train.py:904] (4/8) Epoch 7, batch 5250, loss[loss=0.228, simple_loss=0.3079, pruned_loss=0.07409, over 15319.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2908, pruned_loss=0.06103, over 3203940.44 frames. ], batch size: 190, lr: 9.74e-03, grad_scale: 8.0 2023-04-28 16:45:15,552 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0809, 4.3650, 2.0746, 4.9476, 2.9911, 4.8026, 2.4183, 3.3120], device='cuda:4'), covar=tensor([0.0137, 0.0178, 0.1561, 0.0034, 0.0689, 0.0193, 0.1333, 0.0523], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0153, 0.0176, 0.0092, 0.0160, 0.0190, 0.0186, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 16:45:19,002 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66160.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 16:45:27,264 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5804, 4.5753, 5.0956, 5.0711, 5.0104, 4.6231, 4.6418, 4.5339], device='cuda:4'), covar=tensor([0.0239, 0.0375, 0.0233, 0.0295, 0.0356, 0.0244, 0.0758, 0.0316], device='cuda:4'), in_proj_covar=tensor([0.0273, 0.0270, 0.0273, 0.0267, 0.0323, 0.0293, 0.0393, 0.0232], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 16:45:52,850 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-28 16:45:59,716 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.672e+02 2.502e+02 2.872e+02 3.517e+02 7.118e+02, threshold=5.743e+02, percent-clipped=1.0 2023-04-28 16:46:00,035 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66188.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 16:46:06,950 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66194.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 16:46:11,147 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0939, 5.0397, 4.8740, 4.7100, 4.4787, 4.9313, 4.8723, 4.5617], device='cuda:4'), covar=tensor([0.0449, 0.0308, 0.0216, 0.0180, 0.0904, 0.0288, 0.0231, 0.0514], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0247, 0.0251, 0.0223, 0.0284, 0.0254, 0.0173, 0.0286], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:46:18,392 INFO [train.py:904] (4/8) Epoch 7, batch 5300, loss[loss=0.1808, simple_loss=0.2515, pruned_loss=0.05503, over 16771.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2865, pruned_loss=0.05978, over 3213476.53 frames. ], batch size: 39, lr: 9.74e-03, grad_scale: 8.0 2023-04-28 16:46:23,340 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66205.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 16:46:45,746 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66221.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 16:47:06,139 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=66236.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 16:47:26,597 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8519, 3.9670, 3.7185, 3.5889, 3.3819, 3.8535, 3.6184, 3.5419], device='cuda:4'), covar=tensor([0.0541, 0.0365, 0.0265, 0.0234, 0.0908, 0.0355, 0.0804, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0245, 0.0250, 0.0222, 0.0283, 0.0253, 0.0173, 0.0285], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:47:28,948 INFO [train.py:904] (4/8) Epoch 7, batch 5350, loss[loss=0.2073, simple_loss=0.3004, pruned_loss=0.05707, over 16792.00 frames. ], tot_loss[loss=0.2019, simple_loss=0.2853, pruned_loss=0.05928, over 3207178.60 frames. ], batch size: 83, lr: 9.74e-03, grad_scale: 8.0 2023-04-28 16:47:36,902 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66258.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:48:20,487 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.718e+02 2.423e+02 2.781e+02 3.267e+02 5.283e+02, threshold=5.561e+02, percent-clipped=0.0 2023-04-28 16:48:40,523 INFO [train.py:904] (4/8) Epoch 7, batch 5400, loss[loss=0.2154, simple_loss=0.2956, pruned_loss=0.06765, over 17169.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2883, pruned_loss=0.06041, over 3200066.27 frames. ], batch size: 46, lr: 9.73e-03, grad_scale: 8.0 2023-04-28 16:49:01,655 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2682, 5.1966, 5.1164, 4.4341, 5.2011, 1.9574, 4.9734, 5.2410], device='cuda:4'), covar=tensor([0.0054, 0.0050, 0.0073, 0.0379, 0.0051, 0.1821, 0.0070, 0.0108], device='cuda:4'), in_proj_covar=tensor([0.0106, 0.0093, 0.0142, 0.0141, 0.0109, 0.0155, 0.0125, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:49:05,934 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66319.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:49:55,997 INFO [train.py:904] (4/8) Epoch 7, batch 5450, loss[loss=0.2612, simple_loss=0.3288, pruned_loss=0.0968, over 11932.00 frames. ], tot_loss[loss=0.2092, simple_loss=0.2927, pruned_loss=0.06282, over 3189530.23 frames. ], batch size: 248, lr: 9.73e-03, grad_scale: 8.0 2023-04-28 16:50:00,562 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6622, 2.1442, 2.3618, 4.3216, 2.0908, 2.8840, 2.3867, 2.4571], device='cuda:4'), covar=tensor([0.0688, 0.2684, 0.1521, 0.0275, 0.3149, 0.1611, 0.2181, 0.2592], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0347, 0.0292, 0.0317, 0.0386, 0.0380, 0.0312, 0.0412], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:50:50,980 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.058e+02 3.043e+02 3.833e+02 5.336e+02 9.922e+02, threshold=7.666e+02, percent-clipped=19.0 2023-04-28 16:51:12,546 INFO [train.py:904] (4/8) Epoch 7, batch 5500, loss[loss=0.2491, simple_loss=0.325, pruned_loss=0.08665, over 16416.00 frames. ], tot_loss[loss=0.2211, simple_loss=0.3021, pruned_loss=0.07, over 3142837.53 frames. ], batch size: 146, lr: 9.73e-03, grad_scale: 8.0 2023-04-28 16:51:58,577 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7645, 4.5216, 4.6958, 4.9128, 5.1247, 4.5096, 5.0776, 5.0320], device='cuda:4'), covar=tensor([0.1186, 0.0874, 0.1422, 0.0584, 0.0492, 0.0850, 0.0481, 0.0454], device='cuda:4'), in_proj_covar=tensor([0.0445, 0.0543, 0.0684, 0.0550, 0.0415, 0.0418, 0.0422, 0.0471], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 16:52:29,861 INFO [train.py:904] (4/8) Epoch 7, batch 5550, loss[loss=0.2508, simple_loss=0.3213, pruned_loss=0.09012, over 16610.00 frames. ], tot_loss[loss=0.2324, simple_loss=0.3108, pruned_loss=0.07694, over 3117283.21 frames. ], batch size: 134, lr: 9.72e-03, grad_scale: 8.0 2023-04-28 16:53:27,052 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.385e+02 3.933e+02 4.760e+02 6.147e+02 1.111e+03, threshold=9.520e+02, percent-clipped=8.0 2023-04-28 16:53:33,415 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9119, 1.5609, 2.3825, 2.7792, 2.6437, 3.1508, 1.8609, 3.1938], device='cuda:4'), covar=tensor([0.0097, 0.0328, 0.0179, 0.0148, 0.0160, 0.0086, 0.0290, 0.0055], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0159, 0.0141, 0.0142, 0.0151, 0.0106, 0.0156, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 16:53:36,368 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66494.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 16:53:47,082 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66500.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 16:53:49,151 INFO [train.py:904] (4/8) Epoch 7, batch 5600, loss[loss=0.2198, simple_loss=0.2972, pruned_loss=0.07117, over 16622.00 frames. ], tot_loss[loss=0.2418, simple_loss=0.3177, pruned_loss=0.08301, over 3074818.42 frames. ], batch size: 57, lr: 9.72e-03, grad_scale: 8.0 2023-04-28 16:54:11,811 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66516.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 16:54:20,938 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66521.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 16:54:24,838 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0773, 3.2510, 3.5535, 3.5103, 3.5181, 3.2684, 3.3265, 3.3723], device='cuda:4'), covar=tensor([0.0343, 0.0595, 0.0344, 0.0424, 0.0416, 0.0458, 0.0801, 0.0452], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0267, 0.0269, 0.0263, 0.0319, 0.0291, 0.0388, 0.0231], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 16:54:54,202 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=66542.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 16:55:08,328 INFO [train.py:904] (4/8) Epoch 7, batch 5650, loss[loss=0.284, simple_loss=0.3404, pruned_loss=0.1138, over 11587.00 frames. ], tot_loss[loss=0.2498, simple_loss=0.3238, pruned_loss=0.08794, over 3054949.34 frames. ], batch size: 248, lr: 9.71e-03, grad_scale: 8.0 2023-04-28 16:55:26,380 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-28 16:55:53,660 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66582.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 16:56:03,089 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.602e+02 4.256e+02 5.191e+02 6.863e+02 2.150e+03, threshold=1.038e+03, percent-clipped=9.0 2023-04-28 16:56:23,903 INFO [train.py:904] (4/8) Epoch 7, batch 5700, loss[loss=0.2253, simple_loss=0.3024, pruned_loss=0.07407, over 17149.00 frames. ], tot_loss[loss=0.2507, simple_loss=0.3246, pruned_loss=0.08842, over 3071655.44 frames. ], batch size: 46, lr: 9.71e-03, grad_scale: 8.0 2023-04-28 16:56:42,487 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66614.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 16:56:53,300 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-28 16:57:20,221 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6394, 5.0034, 5.1849, 5.0036, 5.0598, 5.5828, 5.0790, 4.9313], device='cuda:4'), covar=tensor([0.0849, 0.1344, 0.1141, 0.1313, 0.1915, 0.0694, 0.1069, 0.2210], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0436, 0.0438, 0.0370, 0.0497, 0.0471, 0.0350, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:57:25,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7053, 3.6928, 3.8384, 3.6831, 3.7993, 4.1649, 3.9058, 3.7045], device='cuda:4'), covar=tensor([0.1937, 0.1816, 0.1536, 0.1893, 0.2264, 0.1406, 0.1352, 0.2570], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0435, 0.0438, 0.0370, 0.0496, 0.0471, 0.0350, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 16:57:32,645 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0658, 1.3371, 1.8356, 2.0064, 2.1772, 2.2439, 1.5335, 2.2077], device='cuda:4'), covar=tensor([0.0115, 0.0290, 0.0158, 0.0179, 0.0156, 0.0105, 0.0286, 0.0069], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0158, 0.0143, 0.0142, 0.0151, 0.0106, 0.0158, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 16:57:41,027 INFO [train.py:904] (4/8) Epoch 7, batch 5750, loss[loss=0.2497, simple_loss=0.3271, pruned_loss=0.08612, over 16242.00 frames. ], tot_loss[loss=0.2537, simple_loss=0.3269, pruned_loss=0.09024, over 3033847.98 frames. ], batch size: 165, lr: 9.71e-03, grad_scale: 8.0 2023-04-28 16:58:19,780 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-28 16:58:23,914 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0497, 1.6481, 2.3002, 2.8503, 2.7721, 3.1551, 1.9651, 3.1049], device='cuda:4'), covar=tensor([0.0078, 0.0277, 0.0203, 0.0132, 0.0127, 0.0084, 0.0273, 0.0074], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0157, 0.0143, 0.0141, 0.0150, 0.0106, 0.0157, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 16:58:39,453 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.231e+02 3.661e+02 4.395e+02 5.448e+02 7.543e+02, threshold=8.790e+02, percent-clipped=0.0 2023-04-28 16:59:02,126 INFO [train.py:904] (4/8) Epoch 7, batch 5800, loss[loss=0.2921, simple_loss=0.3413, pruned_loss=0.1214, over 11615.00 frames. ], tot_loss[loss=0.2531, simple_loss=0.3271, pruned_loss=0.08948, over 3031509.60 frames. ], batch size: 247, lr: 9.70e-03, grad_scale: 16.0 2023-04-28 16:59:48,300 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66731.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 16:59:48,401 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7016, 2.5143, 2.3610, 3.4867, 2.6585, 3.6679, 1.4172, 2.8444], device='cuda:4'), covar=tensor([0.1362, 0.0619, 0.1089, 0.0111, 0.0266, 0.0354, 0.1485, 0.0701], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0148, 0.0171, 0.0104, 0.0198, 0.0196, 0.0168, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 17:00:19,932 INFO [train.py:904] (4/8) Epoch 7, batch 5850, loss[loss=0.2128, simple_loss=0.3007, pruned_loss=0.0625, over 16680.00 frames. ], tot_loss[loss=0.2486, simple_loss=0.324, pruned_loss=0.08657, over 3041012.92 frames. ], batch size: 83, lr: 9.70e-03, grad_scale: 8.0 2023-04-28 17:00:41,616 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.78 vs. limit=5.0 2023-04-28 17:00:44,493 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3314, 3.8088, 3.8940, 2.6108, 3.4623, 3.8612, 3.6596, 2.0948], device='cuda:4'), covar=tensor([0.0335, 0.0023, 0.0026, 0.0256, 0.0050, 0.0072, 0.0036, 0.0299], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0059, 0.0060, 0.0120, 0.0068, 0.0078, 0.0068, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 17:00:51,594 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-28 17:01:20,003 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.089e+02 3.451e+02 4.238e+02 5.428e+02 1.064e+03, threshold=8.477e+02, percent-clipped=4.0 2023-04-28 17:01:25,708 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66792.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 17:01:38,961 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66800.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 17:01:41,100 INFO [train.py:904] (4/8) Epoch 7, batch 5900, loss[loss=0.2293, simple_loss=0.3169, pruned_loss=0.07087, over 16682.00 frames. ], tot_loss[loss=0.247, simple_loss=0.3231, pruned_loss=0.08541, over 3065737.06 frames. ], batch size: 89, lr: 9.70e-03, grad_scale: 8.0 2023-04-28 17:02:07,235 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66816.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 17:02:56,754 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=66848.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:03:02,870 INFO [train.py:904] (4/8) Epoch 7, batch 5950, loss[loss=0.2607, simple_loss=0.3334, pruned_loss=0.09401, over 11643.00 frames. ], tot_loss[loss=0.2468, simple_loss=0.3242, pruned_loss=0.08468, over 3052505.83 frames. ], batch size: 247, lr: 9.69e-03, grad_scale: 8.0 2023-04-28 17:03:19,236 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2537, 4.2072, 4.1287, 3.2247, 4.1714, 1.5563, 3.8872, 3.8494], device='cuda:4'), covar=tensor([0.0091, 0.0078, 0.0123, 0.0399, 0.0077, 0.2340, 0.0123, 0.0209], device='cuda:4'), in_proj_covar=tensor([0.0104, 0.0092, 0.0139, 0.0137, 0.0106, 0.0154, 0.0122, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:03:23,046 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=66864.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 17:03:41,724 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66877.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 17:04:00,333 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.040e+02 3.507e+02 4.238e+02 5.487e+02 1.167e+03, threshold=8.477e+02, percent-clipped=3.0 2023-04-28 17:04:22,260 INFO [train.py:904] (4/8) Epoch 7, batch 6000, loss[loss=0.2292, simple_loss=0.3089, pruned_loss=0.07476, over 16179.00 frames. ], tot_loss[loss=0.2458, simple_loss=0.3233, pruned_loss=0.08411, over 3065055.95 frames. ], batch size: 165, lr: 9.69e-03, grad_scale: 8.0 2023-04-28 17:04:22,260 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 17:04:32,878 INFO [train.py:938] (4/8) Epoch 7, validation: loss=0.1758, simple_loss=0.2891, pruned_loss=0.03127, over 944034.00 frames. 2023-04-28 17:04:32,879 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 17:04:45,680 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9450, 4.0807, 2.1320, 4.5996, 2.8328, 4.5937, 2.2299, 2.9284], device='cuda:4'), covar=tensor([0.0170, 0.0266, 0.1635, 0.0053, 0.0711, 0.0253, 0.1453, 0.0649], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0157, 0.0180, 0.0096, 0.0163, 0.0194, 0.0188, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 17:04:51,834 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66914.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:05:28,628 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7225, 4.7077, 4.5696, 3.8570, 4.5606, 1.7561, 4.4262, 4.5054], device='cuda:4'), covar=tensor([0.0073, 0.0064, 0.0105, 0.0358, 0.0069, 0.2045, 0.0102, 0.0151], device='cuda:4'), in_proj_covar=tensor([0.0104, 0.0092, 0.0139, 0.0137, 0.0105, 0.0154, 0.0122, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:05:34,293 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-28 17:05:53,520 INFO [train.py:904] (4/8) Epoch 7, batch 6050, loss[loss=0.2212, simple_loss=0.3105, pruned_loss=0.06597, over 16588.00 frames. ], tot_loss[loss=0.2434, simple_loss=0.3214, pruned_loss=0.08269, over 3074946.91 frames. ], batch size: 68, lr: 9.69e-03, grad_scale: 8.0 2023-04-28 17:06:02,592 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66957.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:06:09,560 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=66962.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:06:50,891 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.964e+02 3.435e+02 4.153e+02 5.087e+02 8.285e+02, threshold=8.306e+02, percent-clipped=0.0 2023-04-28 17:07:12,266 INFO [train.py:904] (4/8) Epoch 7, batch 6100, loss[loss=0.2723, simple_loss=0.332, pruned_loss=0.1063, over 11580.00 frames. ], tot_loss[loss=0.2415, simple_loss=0.3201, pruned_loss=0.08143, over 3066469.66 frames. ], batch size: 246, lr: 9.68e-03, grad_scale: 8.0 2023-04-28 17:07:39,944 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67018.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 17:07:54,109 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6830, 3.7304, 3.7964, 3.7079, 3.7150, 4.1573, 3.9118, 3.6015], device='cuda:4'), covar=tensor([0.1872, 0.1673, 0.1774, 0.2185, 0.2877, 0.1568, 0.1265, 0.2757], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0436, 0.0443, 0.0375, 0.0509, 0.0479, 0.0354, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 17:08:32,436 INFO [train.py:904] (4/8) Epoch 7, batch 6150, loss[loss=0.2091, simple_loss=0.2904, pruned_loss=0.06389, over 16816.00 frames. ], tot_loss[loss=0.2391, simple_loss=0.3174, pruned_loss=0.08044, over 3070109.10 frames. ], batch size: 83, lr: 9.68e-03, grad_scale: 8.0 2023-04-28 17:09:23,186 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67083.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:09:29,182 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=67087.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 17:09:31,359 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.004e+02 3.456e+02 4.175e+02 5.279e+02 9.590e+02, threshold=8.351e+02, percent-clipped=3.0 2023-04-28 17:09:51,778 INFO [train.py:904] (4/8) Epoch 7, batch 6200, loss[loss=0.2862, simple_loss=0.332, pruned_loss=0.1203, over 11586.00 frames. ], tot_loss[loss=0.2379, simple_loss=0.3157, pruned_loss=0.08004, over 3061943.36 frames. ], batch size: 248, lr: 9.67e-03, grad_scale: 4.0 2023-04-28 17:09:53,407 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-28 17:11:02,475 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67144.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:11:14,123 INFO [train.py:904] (4/8) Epoch 7, batch 6250, loss[loss=0.2464, simple_loss=0.3238, pruned_loss=0.08445, over 16808.00 frames. ], tot_loss[loss=0.2369, simple_loss=0.3146, pruned_loss=0.07957, over 3074959.97 frames. ], batch size: 124, lr: 9.67e-03, grad_scale: 4.0 2023-04-28 17:11:52,511 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=67177.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 17:12:11,781 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.341e+02 3.226e+02 4.153e+02 4.983e+02 8.969e+02, threshold=8.306e+02, percent-clipped=3.0 2023-04-28 17:12:27,128 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1351, 3.2772, 3.5569, 3.5320, 3.5261, 3.2849, 3.3159, 3.3865], device='cuda:4'), covar=tensor([0.0393, 0.0536, 0.0417, 0.0453, 0.0491, 0.0429, 0.0867, 0.0471], device='cuda:4'), in_proj_covar=tensor([0.0277, 0.0277, 0.0281, 0.0272, 0.0329, 0.0300, 0.0402, 0.0240], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 17:12:30,902 INFO [train.py:904] (4/8) Epoch 7, batch 6300, loss[loss=0.2306, simple_loss=0.309, pruned_loss=0.07609, over 16727.00 frames. ], tot_loss[loss=0.237, simple_loss=0.3153, pruned_loss=0.07939, over 3071980.81 frames. ], batch size: 134, lr: 9.67e-03, grad_scale: 4.0 2023-04-28 17:13:08,746 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=67225.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 17:13:49,867 INFO [train.py:904] (4/8) Epoch 7, batch 6350, loss[loss=0.2392, simple_loss=0.3181, pruned_loss=0.08014, over 16417.00 frames. ], tot_loss[loss=0.2401, simple_loss=0.3172, pruned_loss=0.08146, over 3063583.56 frames. ], batch size: 146, lr: 9.66e-03, grad_scale: 4.0 2023-04-28 17:14:31,928 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8100, 3.8853, 1.6497, 4.2333, 2.5509, 4.1546, 2.0170, 2.8211], device='cuda:4'), covar=tensor([0.0142, 0.0263, 0.1850, 0.0065, 0.0761, 0.0334, 0.1565, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0156, 0.0180, 0.0095, 0.0163, 0.0194, 0.0188, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 17:14:47,872 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.015e+02 3.646e+02 4.526e+02 5.415e+02 1.244e+03, threshold=9.052e+02, percent-clipped=4.0 2023-04-28 17:15:05,415 INFO [train.py:904] (4/8) Epoch 7, batch 6400, loss[loss=0.306, simple_loss=0.3591, pruned_loss=0.1265, over 11182.00 frames. ], tot_loss[loss=0.2418, simple_loss=0.3176, pruned_loss=0.08299, over 3047916.56 frames. ], batch size: 248, lr: 9.66e-03, grad_scale: 8.0 2023-04-28 17:15:22,339 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=67313.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 17:16:20,297 INFO [train.py:904] (4/8) Epoch 7, batch 6450, loss[loss=0.2119, simple_loss=0.3016, pruned_loss=0.06107, over 16745.00 frames. ], tot_loss[loss=0.2408, simple_loss=0.3171, pruned_loss=0.08222, over 3028193.43 frames. ], batch size: 89, lr: 9.66e-03, grad_scale: 4.0 2023-04-28 17:17:16,539 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=67387.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 17:17:22,054 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.155e+02 3.316e+02 4.009e+02 5.317e+02 1.482e+03, threshold=8.018e+02, percent-clipped=5.0 2023-04-28 17:17:38,434 INFO [train.py:904] (4/8) Epoch 7, batch 6500, loss[loss=0.2166, simple_loss=0.2938, pruned_loss=0.0697, over 16588.00 frames. ], tot_loss[loss=0.2378, simple_loss=0.3142, pruned_loss=0.0807, over 3050190.21 frames. ], batch size: 62, lr: 9.65e-03, grad_scale: 4.0 2023-04-28 17:17:45,393 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.66 vs. limit=2.0 2023-04-28 17:18:29,446 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=67435.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 17:18:36,061 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=67439.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:18:47,979 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-28 17:18:57,738 INFO [train.py:904] (4/8) Epoch 7, batch 6550, loss[loss=0.2289, simple_loss=0.3229, pruned_loss=0.06747, over 16278.00 frames. ], tot_loss[loss=0.2406, simple_loss=0.3175, pruned_loss=0.08188, over 3050383.46 frames. ], batch size: 165, lr: 9.65e-03, grad_scale: 4.0 2023-04-28 17:19:56,001 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.411e+02 3.629e+02 4.533e+02 5.660e+02 1.346e+03, threshold=9.066e+02, percent-clipped=11.0 2023-04-28 17:20:13,102 INFO [train.py:904] (4/8) Epoch 7, batch 6600, loss[loss=0.2471, simple_loss=0.3203, pruned_loss=0.08694, over 17027.00 frames. ], tot_loss[loss=0.2428, simple_loss=0.3201, pruned_loss=0.08273, over 3056733.28 frames. ], batch size: 55, lr: 9.65e-03, grad_scale: 4.0 2023-04-28 17:20:30,835 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67513.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:20:46,995 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6975, 3.6958, 3.7644, 3.7381, 3.8012, 4.2090, 3.9291, 3.6655], device='cuda:4'), covar=tensor([0.1837, 0.1875, 0.1795, 0.1831, 0.2487, 0.1355, 0.1330, 0.2126], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0433, 0.0446, 0.0369, 0.0501, 0.0477, 0.0357, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 17:21:09,001 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3501, 3.2763, 3.3431, 3.4334, 3.4413, 3.2063, 3.4514, 3.5110], device='cuda:4'), covar=tensor([0.0907, 0.0770, 0.1132, 0.0548, 0.0711, 0.2094, 0.0919, 0.0634], device='cuda:4'), in_proj_covar=tensor([0.0443, 0.0545, 0.0680, 0.0545, 0.0418, 0.0412, 0.0425, 0.0474], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:21:29,932 INFO [train.py:904] (4/8) Epoch 7, batch 6650, loss[loss=0.2693, simple_loss=0.3254, pruned_loss=0.1066, over 11562.00 frames. ], tot_loss[loss=0.2441, simple_loss=0.3208, pruned_loss=0.08367, over 3044624.16 frames. ], batch size: 246, lr: 9.64e-03, grad_scale: 4.0 2023-04-28 17:22:02,868 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67574.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 17:22:27,317 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.256e+02 3.378e+02 3.968e+02 4.972e+02 1.148e+03, threshold=7.935e+02, percent-clipped=1.0 2023-04-28 17:22:43,252 INFO [train.py:904] (4/8) Epoch 7, batch 6700, loss[loss=0.2316, simple_loss=0.3136, pruned_loss=0.07483, over 16805.00 frames. ], tot_loss[loss=0.2439, simple_loss=0.3199, pruned_loss=0.08397, over 3025472.26 frames. ], batch size: 83, lr: 9.64e-03, grad_scale: 4.0 2023-04-28 17:23:00,342 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=67613.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:23:04,295 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 17:23:56,588 INFO [train.py:904] (4/8) Epoch 7, batch 6750, loss[loss=0.2134, simple_loss=0.2966, pruned_loss=0.06509, over 16774.00 frames. ], tot_loss[loss=0.2431, simple_loss=0.319, pruned_loss=0.08358, over 3040134.18 frames. ], batch size: 124, lr: 9.64e-03, grad_scale: 4.0 2023-04-28 17:24:11,054 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=67661.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:24:37,969 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67680.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:24:54,245 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.424e+02 3.570e+02 4.418e+02 5.557e+02 1.177e+03, threshold=8.835e+02, percent-clipped=2.0 2023-04-28 17:24:55,988 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67692.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:25:10,438 INFO [train.py:904] (4/8) Epoch 7, batch 6800, loss[loss=0.2769, simple_loss=0.3364, pruned_loss=0.1087, over 11663.00 frames. ], tot_loss[loss=0.2416, simple_loss=0.318, pruned_loss=0.08263, over 3060356.54 frames. ], batch size: 246, lr: 9.63e-03, grad_scale: 8.0 2023-04-28 17:25:17,902 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.63 vs. limit=2.0 2023-04-28 17:26:06,823 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=67739.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:26:10,072 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67741.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:26:23,554 INFO [train.py:904] (4/8) Epoch 7, batch 6850, loss[loss=0.2309, simple_loss=0.3311, pruned_loss=0.06532, over 16429.00 frames. ], tot_loss[loss=0.2431, simple_loss=0.3195, pruned_loss=0.08339, over 3056191.32 frames. ], batch size: 68, lr: 9.63e-03, grad_scale: 8.0 2023-04-28 17:26:25,728 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67753.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 17:27:14,419 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=67787.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:27:20,977 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.331e+02 3.323e+02 3.993e+02 4.819e+02 7.017e+02, threshold=7.986e+02, percent-clipped=0.0 2023-04-28 17:27:35,614 INFO [train.py:904] (4/8) Epoch 7, batch 6900, loss[loss=0.2719, simple_loss=0.3443, pruned_loss=0.0997, over 16876.00 frames. ], tot_loss[loss=0.2432, simple_loss=0.3221, pruned_loss=0.08217, over 3099145.89 frames. ], batch size: 109, lr: 9.63e-03, grad_scale: 2.0 2023-04-28 17:28:46,543 INFO [train.py:904] (4/8) Epoch 7, batch 6950, loss[loss=0.2558, simple_loss=0.3258, pruned_loss=0.09291, over 15169.00 frames. ], tot_loss[loss=0.2457, simple_loss=0.3235, pruned_loss=0.08393, over 3085221.93 frames. ], batch size: 190, lr: 9.62e-03, grad_scale: 2.0 2023-04-28 17:28:52,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0819, 1.7389, 2.4188, 2.8926, 2.8389, 3.3364, 1.8595, 3.2073], device='cuda:4'), covar=tensor([0.0094, 0.0295, 0.0162, 0.0128, 0.0137, 0.0079, 0.0297, 0.0074], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0152, 0.0134, 0.0136, 0.0145, 0.0101, 0.0153, 0.0093], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 17:29:13,145 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=67869.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 17:29:46,387 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.351e+02 3.618e+02 4.486e+02 5.724e+02 1.200e+03, threshold=8.972e+02, percent-clipped=9.0 2023-04-28 17:29:59,772 INFO [train.py:904] (4/8) Epoch 7, batch 7000, loss[loss=0.239, simple_loss=0.3231, pruned_loss=0.07745, over 16858.00 frames. ], tot_loss[loss=0.2435, simple_loss=0.3223, pruned_loss=0.08236, over 3095885.34 frames. ], batch size: 116, lr: 9.62e-03, grad_scale: 2.0 2023-04-28 17:31:13,271 INFO [train.py:904] (4/8) Epoch 7, batch 7050, loss[loss=0.2352, simple_loss=0.3194, pruned_loss=0.07552, over 17197.00 frames. ], tot_loss[loss=0.2434, simple_loss=0.3226, pruned_loss=0.08209, over 3090621.60 frames. ], batch size: 46, lr: 9.61e-03, grad_scale: 2.0 2023-04-28 17:31:33,264 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.85 vs. limit=2.0 2023-04-28 17:31:57,372 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67981.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:32:14,207 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.912e+02 3.870e+02 4.768e+02 5.699e+02 1.240e+03, threshold=9.537e+02, percent-clipped=2.0 2023-04-28 17:32:20,216 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2347, 4.2694, 4.1137, 3.9905, 3.7822, 4.0958, 3.9482, 3.9018], device='cuda:4'), covar=tensor([0.0455, 0.0326, 0.0204, 0.0171, 0.0736, 0.0305, 0.0488, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0233, 0.0235, 0.0208, 0.0265, 0.0239, 0.0167, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:32:29,898 INFO [train.py:904] (4/8) Epoch 7, batch 7100, loss[loss=0.2383, simple_loss=0.3212, pruned_loss=0.07766, over 16394.00 frames. ], tot_loss[loss=0.2438, simple_loss=0.3218, pruned_loss=0.08285, over 3064784.95 frames. ], batch size: 146, lr: 9.61e-03, grad_scale: 2.0 2023-04-28 17:33:21,414 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68036.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:33:30,305 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68042.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:33:39,218 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68048.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 17:33:44,541 INFO [train.py:904] (4/8) Epoch 7, batch 7150, loss[loss=0.2176, simple_loss=0.3064, pruned_loss=0.06435, over 16885.00 frames. ], tot_loss[loss=0.2426, simple_loss=0.3199, pruned_loss=0.0826, over 3070357.23 frames. ], batch size: 96, lr: 9.61e-03, grad_scale: 2.0 2023-04-28 17:33:55,961 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9320, 3.9538, 3.8680, 3.0642, 3.8623, 1.7372, 3.6648, 3.4401], device='cuda:4'), covar=tensor([0.0098, 0.0078, 0.0139, 0.0388, 0.0088, 0.2225, 0.0120, 0.0223], device='cuda:4'), in_proj_covar=tensor([0.0103, 0.0090, 0.0137, 0.0134, 0.0105, 0.0154, 0.0122, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:34:37,153 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=68087.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 17:34:44,522 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.291e+02 3.518e+02 4.375e+02 5.869e+02 1.278e+03, threshold=8.749e+02, percent-clipped=4.0 2023-04-28 17:34:58,212 INFO [train.py:904] (4/8) Epoch 7, batch 7200, loss[loss=0.1845, simple_loss=0.2781, pruned_loss=0.04544, over 16739.00 frames. ], tot_loss[loss=0.2399, simple_loss=0.3179, pruned_loss=0.081, over 3056395.95 frames. ], batch size: 83, lr: 9.60e-03, grad_scale: 4.0 2023-04-28 17:35:00,322 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7328, 5.1231, 5.2533, 5.2181, 5.1932, 5.7228, 5.2079, 4.9945], device='cuda:4'), covar=tensor([0.0898, 0.1414, 0.1405, 0.1483, 0.2340, 0.0855, 0.1123, 0.2000], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0431, 0.0442, 0.0369, 0.0495, 0.0470, 0.0356, 0.0507], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 17:35:37,069 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2978, 4.3165, 4.8521, 4.8465, 4.7486, 4.4226, 4.4551, 4.2131], device='cuda:4'), covar=tensor([0.0268, 0.0373, 0.0271, 0.0307, 0.0474, 0.0296, 0.0858, 0.0401], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0274, 0.0274, 0.0269, 0.0321, 0.0294, 0.0391, 0.0238], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 17:36:11,120 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68148.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 17:36:16,027 INFO [train.py:904] (4/8) Epoch 7, batch 7250, loss[loss=0.2602, simple_loss=0.3201, pruned_loss=0.1001, over 11911.00 frames. ], tot_loss[loss=0.2363, simple_loss=0.3147, pruned_loss=0.079, over 3076256.42 frames. ], batch size: 248, lr: 9.60e-03, grad_scale: 4.0 2023-04-28 17:36:41,109 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68169.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:37:16,279 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.105e+02 3.214e+02 3.991e+02 5.016e+02 1.051e+03, threshold=7.982e+02, percent-clipped=2.0 2023-04-28 17:37:29,946 INFO [train.py:904] (4/8) Epoch 7, batch 7300, loss[loss=0.2805, simple_loss=0.3305, pruned_loss=0.1152, over 11532.00 frames. ], tot_loss[loss=0.2372, simple_loss=0.315, pruned_loss=0.07972, over 3060620.12 frames. ], batch size: 248, lr: 9.60e-03, grad_scale: 4.0 2023-04-28 17:37:36,775 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0674, 3.2208, 3.3335, 1.5488, 3.6004, 3.6867, 2.7264, 2.5812], device='cuda:4'), covar=tensor([0.0887, 0.0198, 0.0213, 0.1256, 0.0062, 0.0088, 0.0432, 0.0495], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0094, 0.0082, 0.0137, 0.0069, 0.0085, 0.0118, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 17:37:52,950 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=68217.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:38:00,356 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6231, 3.4976, 3.7189, 3.6490, 3.7332, 4.1231, 3.8661, 3.5889], device='cuda:4'), covar=tensor([0.1981, 0.2139, 0.1478, 0.2170, 0.2411, 0.1352, 0.1269, 0.2366], device='cuda:4'), in_proj_covar=tensor([0.0304, 0.0425, 0.0435, 0.0365, 0.0490, 0.0463, 0.0350, 0.0499], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 17:38:44,625 INFO [train.py:904] (4/8) Epoch 7, batch 7350, loss[loss=0.2236, simple_loss=0.3068, pruned_loss=0.07016, over 16456.00 frames. ], tot_loss[loss=0.238, simple_loss=0.3153, pruned_loss=0.08033, over 3051788.45 frames. ], batch size: 68, lr: 9.59e-03, grad_scale: 4.0 2023-04-28 17:39:06,014 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 17:39:48,567 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.908e+02 3.085e+02 3.772e+02 4.630e+02 1.239e+03, threshold=7.544e+02, percent-clipped=2.0 2023-04-28 17:40:02,136 INFO [train.py:904] (4/8) Epoch 7, batch 7400, loss[loss=0.2567, simple_loss=0.3366, pruned_loss=0.08835, over 16920.00 frames. ], tot_loss[loss=0.2393, simple_loss=0.3163, pruned_loss=0.08114, over 3033722.79 frames. ], batch size: 116, lr: 9.59e-03, grad_scale: 4.0 2023-04-28 17:40:08,578 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7387, 4.6627, 4.5714, 4.4022, 4.1741, 4.5709, 4.6386, 4.2920], device='cuda:4'), covar=tensor([0.0515, 0.0433, 0.0226, 0.0210, 0.0830, 0.0357, 0.0263, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0233, 0.0234, 0.0208, 0.0263, 0.0237, 0.0165, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:40:53,822 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68336.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:40:54,927 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68337.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:41:14,161 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68348.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:41:19,088 INFO [train.py:904] (4/8) Epoch 7, batch 7450, loss[loss=0.234, simple_loss=0.3155, pruned_loss=0.07624, over 16705.00 frames. ], tot_loss[loss=0.2398, simple_loss=0.3171, pruned_loss=0.08128, over 3054746.24 frames. ], batch size: 134, lr: 9.59e-03, grad_scale: 4.0 2023-04-28 17:42:11,320 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=68384.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:42:26,328 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.512e+02 3.665e+02 4.367e+02 5.646e+02 8.920e+02, threshold=8.734e+02, percent-clipped=5.0 2023-04-28 17:42:30,798 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=68396.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:42:34,229 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 17:42:35,731 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=68399.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:42:39,839 INFO [train.py:904] (4/8) Epoch 7, batch 7500, loss[loss=0.2208, simple_loss=0.3078, pruned_loss=0.06693, over 16589.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.3168, pruned_loss=0.08052, over 3044453.36 frames. ], batch size: 62, lr: 9.58e-03, grad_scale: 4.0 2023-04-28 17:43:02,324 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5132, 4.2522, 4.5243, 4.7276, 4.8711, 4.3433, 4.8303, 4.8527], device='cuda:4'), covar=tensor([0.1133, 0.0998, 0.1363, 0.0534, 0.0434, 0.0795, 0.0463, 0.0448], device='cuda:4'), in_proj_covar=tensor([0.0436, 0.0539, 0.0672, 0.0541, 0.0412, 0.0413, 0.0432, 0.0468], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:43:44,382 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68443.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 17:43:57,919 INFO [train.py:904] (4/8) Epoch 7, batch 7550, loss[loss=0.2089, simple_loss=0.2939, pruned_loss=0.06194, over 16230.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.316, pruned_loss=0.08094, over 3046640.05 frames. ], batch size: 35, lr: 9.58e-03, grad_scale: 4.0 2023-04-28 17:44:10,843 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68460.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:44:29,132 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9594, 5.2703, 5.0446, 5.0140, 4.6945, 4.5662, 4.7816, 5.3905], device='cuda:4'), covar=tensor([0.0832, 0.0678, 0.0934, 0.0537, 0.0636, 0.0752, 0.0744, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0444, 0.0560, 0.0478, 0.0365, 0.0349, 0.0379, 0.0459, 0.0414], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:44:59,735 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.309e+02 3.638e+02 4.521e+02 5.512e+02 1.190e+03, threshold=9.043e+02, percent-clipped=2.0 2023-04-28 17:45:12,771 INFO [train.py:904] (4/8) Epoch 7, batch 7600, loss[loss=0.2165, simple_loss=0.2976, pruned_loss=0.06774, over 16974.00 frames. ], tot_loss[loss=0.2378, simple_loss=0.3146, pruned_loss=0.08047, over 3055535.33 frames. ], batch size: 55, lr: 9.58e-03, grad_scale: 8.0 2023-04-28 17:46:01,166 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9402, 1.6111, 2.3817, 2.8718, 2.7008, 3.2458, 1.8882, 3.1852], device='cuda:4'), covar=tensor([0.0095, 0.0305, 0.0158, 0.0137, 0.0134, 0.0073, 0.0295, 0.0056], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0152, 0.0133, 0.0134, 0.0143, 0.0100, 0.0152, 0.0090], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 17:46:13,500 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6917, 3.6151, 3.7059, 3.7348, 3.7549, 4.1754, 3.9499, 3.6461], device='cuda:4'), covar=tensor([0.2320, 0.2333, 0.2128, 0.2391, 0.2792, 0.1630, 0.1495, 0.2862], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0434, 0.0451, 0.0376, 0.0501, 0.0475, 0.0358, 0.0514], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 17:46:28,266 INFO [train.py:904] (4/8) Epoch 7, batch 7650, loss[loss=0.2275, simple_loss=0.306, pruned_loss=0.07451, over 16254.00 frames. ], tot_loss[loss=0.2374, simple_loss=0.3146, pruned_loss=0.08014, over 3082380.06 frames. ], batch size: 165, lr: 9.57e-03, grad_scale: 8.0 2023-04-28 17:47:08,487 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=68578.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:47:12,113 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2804, 3.7384, 3.5817, 1.9873, 2.9614, 2.4218, 3.5132, 3.7801], device='cuda:4'), covar=tensor([0.0259, 0.0646, 0.0500, 0.1714, 0.0777, 0.0872, 0.0649, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0132, 0.0153, 0.0140, 0.0134, 0.0124, 0.0135, 0.0140], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 17:47:29,673 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.494e+02 3.559e+02 4.201e+02 5.687e+02 1.184e+03, threshold=8.403e+02, percent-clipped=4.0 2023-04-28 17:47:42,700 INFO [train.py:904] (4/8) Epoch 7, batch 7700, loss[loss=0.2401, simple_loss=0.3214, pruned_loss=0.07941, over 15413.00 frames. ], tot_loss[loss=0.24, simple_loss=0.3161, pruned_loss=0.08195, over 3070349.15 frames. ], batch size: 191, lr: 9.57e-03, grad_scale: 8.0 2023-04-28 17:48:03,851 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4694, 3.8441, 3.6882, 2.0189, 3.0587, 2.5569, 3.7000, 3.8777], device='cuda:4'), covar=tensor([0.0197, 0.0526, 0.0507, 0.1606, 0.0711, 0.0821, 0.0561, 0.0759], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0132, 0.0153, 0.0140, 0.0134, 0.0124, 0.0135, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 17:48:35,361 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68637.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:48:38,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68639.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:48:57,122 INFO [train.py:904] (4/8) Epoch 7, batch 7750, loss[loss=0.2713, simple_loss=0.3266, pruned_loss=0.108, over 11487.00 frames. ], tot_loss[loss=0.2408, simple_loss=0.317, pruned_loss=0.08228, over 3062893.06 frames. ], batch size: 246, lr: 9.57e-03, grad_scale: 8.0 2023-04-28 17:49:44,494 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=68685.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:49:44,666 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=68685.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:49:46,548 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0738, 2.9310, 2.6782, 2.0127, 2.5916, 2.1408, 2.7629, 2.9788], device='cuda:4'), covar=tensor([0.0301, 0.0495, 0.0531, 0.1424, 0.0687, 0.0878, 0.0525, 0.0574], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0132, 0.0153, 0.0141, 0.0134, 0.0124, 0.0136, 0.0140], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 17:49:57,427 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.246e+02 3.570e+02 4.272e+02 5.150e+02 9.085e+02, threshold=8.545e+02, percent-clipped=2.0 2023-04-28 17:50:11,061 INFO [train.py:904] (4/8) Epoch 7, batch 7800, loss[loss=0.2327, simple_loss=0.3176, pruned_loss=0.07388, over 17199.00 frames. ], tot_loss[loss=0.2423, simple_loss=0.3182, pruned_loss=0.08326, over 3047911.02 frames. ], batch size: 44, lr: 9.56e-03, grad_scale: 8.0 2023-04-28 17:50:48,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9710, 3.5011, 3.3893, 1.8740, 2.9420, 2.4177, 3.4055, 3.5443], device='cuda:4'), covar=tensor([0.0222, 0.0442, 0.0512, 0.1639, 0.0673, 0.0808, 0.0525, 0.0668], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0131, 0.0151, 0.0139, 0.0133, 0.0123, 0.0134, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 17:51:13,475 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68743.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 17:51:17,571 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68746.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:51:25,161 INFO [train.py:904] (4/8) Epoch 7, batch 7850, loss[loss=0.2223, simple_loss=0.3088, pruned_loss=0.06791, over 16923.00 frames. ], tot_loss[loss=0.242, simple_loss=0.3189, pruned_loss=0.08254, over 3068026.68 frames. ], batch size: 116, lr: 9.56e-03, grad_scale: 8.0 2023-04-28 17:51:30,909 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68755.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:52:24,300 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=68791.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 17:52:26,860 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.225e+02 3.517e+02 4.266e+02 5.466e+02 1.247e+03, threshold=8.532e+02, percent-clipped=3.0 2023-04-28 17:52:40,546 INFO [train.py:904] (4/8) Epoch 7, batch 7900, loss[loss=0.2122, simple_loss=0.2958, pruned_loss=0.06427, over 16832.00 frames. ], tot_loss[loss=0.2406, simple_loss=0.3177, pruned_loss=0.08171, over 3071000.95 frames. ], batch size: 116, lr: 9.56e-03, grad_scale: 8.0 2023-04-28 17:52:42,675 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9959, 4.1534, 3.8706, 3.8178, 3.3701, 3.9716, 3.7970, 3.6146], device='cuda:4'), covar=tensor([0.0640, 0.0378, 0.0282, 0.0234, 0.1049, 0.0423, 0.0596, 0.0609], device='cuda:4'), in_proj_covar=tensor([0.0206, 0.0239, 0.0238, 0.0212, 0.0267, 0.0240, 0.0169, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:52:48,388 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6895, 1.2522, 2.1125, 2.6147, 2.5062, 3.0530, 1.6187, 2.9070], device='cuda:4'), covar=tensor([0.0098, 0.0305, 0.0170, 0.0138, 0.0139, 0.0075, 0.0315, 0.0066], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0154, 0.0135, 0.0135, 0.0146, 0.0101, 0.0154, 0.0092], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 17:53:58,005 INFO [train.py:904] (4/8) Epoch 7, batch 7950, loss[loss=0.2146, simple_loss=0.2931, pruned_loss=0.06805, over 16265.00 frames. ], tot_loss[loss=0.2411, simple_loss=0.3177, pruned_loss=0.08226, over 3060789.78 frames. ], batch size: 165, lr: 9.55e-03, grad_scale: 2.0 2023-04-28 17:54:35,698 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 17:54:59,226 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6293, 2.6723, 1.6501, 2.7180, 2.1651, 2.7365, 1.9855, 2.4065], device='cuda:4'), covar=tensor([0.0231, 0.0371, 0.1280, 0.0144, 0.0557, 0.0547, 0.1116, 0.0543], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0157, 0.0182, 0.0097, 0.0166, 0.0196, 0.0190, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 17:55:03,363 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.354e+02 3.365e+02 3.868e+02 4.958e+02 9.415e+02, threshold=7.735e+02, percent-clipped=1.0 2023-04-28 17:55:12,758 INFO [train.py:904] (4/8) Epoch 7, batch 8000, loss[loss=0.2959, simple_loss=0.3546, pruned_loss=0.1187, over 11687.00 frames. ], tot_loss[loss=0.2421, simple_loss=0.3185, pruned_loss=0.08283, over 3058809.19 frames. ], batch size: 250, lr: 9.55e-03, grad_scale: 4.0 2023-04-28 17:56:00,146 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68934.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:56:07,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8973, 2.2521, 2.2658, 4.5712, 2.1115, 2.8864, 2.3898, 2.5475], device='cuda:4'), covar=tensor([0.0710, 0.2958, 0.1778, 0.0302, 0.3492, 0.1729, 0.2429, 0.2866], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0357, 0.0296, 0.0322, 0.0394, 0.0385, 0.0316, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 17:56:25,017 INFO [train.py:904] (4/8) Epoch 7, batch 8050, loss[loss=0.2762, simple_loss=0.3537, pruned_loss=0.09937, over 15249.00 frames. ], tot_loss[loss=0.2423, simple_loss=0.3186, pruned_loss=0.08296, over 3048581.24 frames. ], batch size: 190, lr: 9.54e-03, grad_scale: 4.0 2023-04-28 17:56:53,432 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0514, 1.3535, 1.6994, 1.9986, 2.0867, 2.2866, 1.5559, 2.2187], device='cuda:4'), covar=tensor([0.0123, 0.0272, 0.0159, 0.0175, 0.0152, 0.0086, 0.0267, 0.0064], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0153, 0.0134, 0.0135, 0.0146, 0.0101, 0.0153, 0.0092], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 17:57:30,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.364e+02 3.694e+02 4.392e+02 5.257e+02 9.021e+02, threshold=8.783e+02, percent-clipped=4.0 2023-04-28 17:57:41,497 INFO [train.py:904] (4/8) Epoch 7, batch 8100, loss[loss=0.2172, simple_loss=0.2986, pruned_loss=0.06795, over 16733.00 frames. ], tot_loss[loss=0.2407, simple_loss=0.3175, pruned_loss=0.08194, over 3062924.99 frames. ], batch size: 124, lr: 9.54e-03, grad_scale: 4.0 2023-04-28 17:58:40,437 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69041.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:58:43,478 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69043.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:58:55,987 INFO [train.py:904] (4/8) Epoch 7, batch 8150, loss[loss=0.2118, simple_loss=0.295, pruned_loss=0.06433, over 16773.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.3156, pruned_loss=0.08108, over 3065105.22 frames. ], batch size: 89, lr: 9.54e-03, grad_scale: 4.0 2023-04-28 17:58:57,648 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8832, 4.1389, 2.2754, 4.8946, 3.0207, 4.7748, 2.5272, 3.3101], device='cuda:4'), covar=tensor([0.0205, 0.0316, 0.1505, 0.0049, 0.0683, 0.0360, 0.1379, 0.0572], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0156, 0.0180, 0.0096, 0.0163, 0.0193, 0.0189, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 17:59:00,697 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69055.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 17:59:59,699 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.706e+02 3.772e+02 4.723e+02 6.532e+02 2.181e+03, threshold=9.446e+02, percent-clipped=10.0 2023-04-28 18:00:10,161 INFO [train.py:904] (4/8) Epoch 7, batch 8200, loss[loss=0.2395, simple_loss=0.3217, pruned_loss=0.07867, over 15134.00 frames. ], tot_loss[loss=0.2361, simple_loss=0.3125, pruned_loss=0.0798, over 3082112.30 frames. ], batch size: 190, lr: 9.53e-03, grad_scale: 4.0 2023-04-28 18:00:12,049 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69103.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:00:13,535 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69104.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:01:30,736 INFO [train.py:904] (4/8) Epoch 7, batch 8250, loss[loss=0.1916, simple_loss=0.2907, pruned_loss=0.04621, over 16803.00 frames. ], tot_loss[loss=0.2334, simple_loss=0.3113, pruned_loss=0.07776, over 3058848.02 frames. ], batch size: 102, lr: 9.53e-03, grad_scale: 4.0 2023-04-28 18:01:47,258 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69162.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:01:58,259 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69168.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:02:40,829 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.107e+02 2.942e+02 3.716e+02 4.471e+02 8.485e+02, threshold=7.433e+02, percent-clipped=0.0 2023-04-28 18:02:49,964 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5562, 3.5721, 2.9735, 2.1410, 2.2969, 2.1333, 3.6582, 3.3631], device='cuda:4'), covar=tensor([0.2312, 0.0596, 0.1188, 0.1930, 0.1982, 0.1675, 0.0334, 0.0784], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0246, 0.0268, 0.0254, 0.0273, 0.0205, 0.0245, 0.0263], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:02:52,449 INFO [train.py:904] (4/8) Epoch 7, batch 8300, loss[loss=0.181, simple_loss=0.263, pruned_loss=0.04953, over 11749.00 frames. ], tot_loss[loss=0.2272, simple_loss=0.3074, pruned_loss=0.07349, over 3063205.84 frames. ], batch size: 247, lr: 9.53e-03, grad_scale: 4.0 2023-04-28 18:03:26,096 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69223.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:03:29,166 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69225.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:03:34,958 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69229.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:03:42,698 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69234.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:04:00,658 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69245.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 18:04:11,894 INFO [train.py:904] (4/8) Epoch 7, batch 8350, loss[loss=0.2223, simple_loss=0.313, pruned_loss=0.06578, over 16680.00 frames. ], tot_loss[loss=0.2238, simple_loss=0.3058, pruned_loss=0.07088, over 3055955.84 frames. ], batch size: 134, lr: 9.52e-03, grad_scale: 4.0 2023-04-28 18:04:46,391 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69273.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:05:00,052 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69282.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:05:06,564 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69286.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:05:20,697 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.545e+02 2.640e+02 3.244e+02 4.107e+02 9.415e+02, threshold=6.488e+02, percent-clipped=2.0 2023-04-28 18:05:32,118 INFO [train.py:904] (4/8) Epoch 7, batch 8400, loss[loss=0.2052, simple_loss=0.2756, pruned_loss=0.06743, over 11802.00 frames. ], tot_loss[loss=0.2198, simple_loss=0.3029, pruned_loss=0.06835, over 3048847.12 frames. ], batch size: 247, lr: 9.52e-03, grad_scale: 8.0 2023-04-28 18:05:38,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2508, 3.1632, 3.1247, 1.7435, 3.3470, 3.3774, 2.6934, 2.7891], device='cuda:4'), covar=tensor([0.0636, 0.0163, 0.0186, 0.1115, 0.0061, 0.0097, 0.0382, 0.0349], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0091, 0.0079, 0.0136, 0.0066, 0.0082, 0.0115, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 18:05:39,375 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69306.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 18:06:02,579 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69320.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:06:23,791 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69334.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 18:06:26,640 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6177, 4.3821, 4.6909, 4.8499, 4.9880, 4.4775, 4.9936, 4.9812], device='cuda:4'), covar=tensor([0.1401, 0.0947, 0.1279, 0.0550, 0.0406, 0.0693, 0.0385, 0.0388], device='cuda:4'), in_proj_covar=tensor([0.0432, 0.0527, 0.0656, 0.0532, 0.0403, 0.0406, 0.0423, 0.0460], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:06:35,268 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69341.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:06:52,097 INFO [train.py:904] (4/8) Epoch 7, batch 8450, loss[loss=0.2183, simple_loss=0.292, pruned_loss=0.0723, over 12356.00 frames. ], tot_loss[loss=0.2169, simple_loss=0.3009, pruned_loss=0.0665, over 3044513.77 frames. ], batch size: 246, lr: 9.52e-03, grad_scale: 8.0 2023-04-28 18:06:59,352 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5453, 3.5914, 3.3671, 3.3073, 3.2067, 3.4628, 3.2764, 3.3537], device='cuda:4'), covar=tensor([0.0451, 0.0402, 0.0219, 0.0194, 0.0524, 0.0274, 0.0905, 0.0434], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0236, 0.0235, 0.0208, 0.0260, 0.0235, 0.0166, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:07:18,746 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69368.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:07:38,193 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69381.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:07:40,506 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.55 vs. limit=5.0 2023-04-28 18:07:50,678 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69389.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:07:58,591 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69394.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:07:59,250 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.689e+02 2.708e+02 3.167e+02 3.913e+02 6.077e+02, threshold=6.334e+02, percent-clipped=0.0 2023-04-28 18:08:06,556 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69399.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:08:10,405 INFO [train.py:904] (4/8) Epoch 7, batch 8500, loss[loss=0.1713, simple_loss=0.2471, pruned_loss=0.0477, over 11843.00 frames. ], tot_loss[loss=0.2115, simple_loss=0.2959, pruned_loss=0.06355, over 3029436.67 frames. ], batch size: 246, lr: 9.51e-03, grad_scale: 8.0 2023-04-28 18:08:25,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5004, 3.4976, 3.4269, 2.9165, 3.4094, 2.0733, 3.1061, 2.8436], device='cuda:4'), covar=tensor([0.0098, 0.0086, 0.0119, 0.0193, 0.0072, 0.1733, 0.0094, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0102, 0.0090, 0.0139, 0.0130, 0.0105, 0.0157, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:08:55,036 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69429.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:09:33,177 INFO [train.py:904] (4/8) Epoch 7, batch 8550, loss[loss=0.1828, simple_loss=0.2808, pruned_loss=0.0424, over 16891.00 frames. ], tot_loss[loss=0.2092, simple_loss=0.2936, pruned_loss=0.06247, over 3013202.21 frames. ], batch size: 96, lr: 9.51e-03, grad_scale: 8.0 2023-04-28 18:09:41,799 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69455.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:10:56,160 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1735, 4.2149, 4.0601, 3.9390, 3.7283, 4.1147, 3.8370, 3.8198], device='cuda:4'), covar=tensor([0.0450, 0.0336, 0.0217, 0.0186, 0.0680, 0.0306, 0.0529, 0.0501], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0232, 0.0230, 0.0205, 0.0255, 0.0232, 0.0163, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:10:58,765 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.959e+02 3.006e+02 3.588e+02 4.597e+02 9.217e+02, threshold=7.177e+02, percent-clipped=6.0 2023-04-28 18:11:12,035 INFO [train.py:904] (4/8) Epoch 7, batch 8600, loss[loss=0.2107, simple_loss=0.2967, pruned_loss=0.06235, over 16736.00 frames. ], tot_loss[loss=0.2084, simple_loss=0.2942, pruned_loss=0.06129, over 3036721.92 frames. ], batch size: 134, lr: 9.51e-03, grad_scale: 8.0 2023-04-28 18:11:46,333 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69518.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:11:56,250 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69524.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:12:47,749 INFO [train.py:904] (4/8) Epoch 7, batch 8650, loss[loss=0.1849, simple_loss=0.2842, pruned_loss=0.04281, over 16681.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2916, pruned_loss=0.05953, over 3012169.71 frames. ], batch size: 89, lr: 9.50e-03, grad_scale: 8.0 2023-04-28 18:13:25,856 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-28 18:13:54,359 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69581.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:14:19,150 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.635e+02 2.551e+02 2.997e+02 3.527e+02 6.166e+02, threshold=5.994e+02, percent-clipped=0.0 2023-04-28 18:14:31,858 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69601.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 18:14:32,612 INFO [train.py:904] (4/8) Epoch 7, batch 8700, loss[loss=0.2005, simple_loss=0.2877, pruned_loss=0.05664, over 16809.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2884, pruned_loss=0.05778, over 3033884.04 frames. ], batch size: 124, lr: 9.50e-03, grad_scale: 8.0 2023-04-28 18:15:17,091 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7290, 3.8482, 2.0789, 4.3205, 2.6097, 4.2267, 2.2245, 3.1199], device='cuda:4'), covar=tensor([0.0185, 0.0229, 0.1463, 0.0070, 0.0870, 0.0339, 0.1521, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0147, 0.0174, 0.0090, 0.0158, 0.0181, 0.0184, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 18:15:22,787 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69629.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 18:15:46,441 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.65 vs. limit=5.0 2023-04-28 18:16:08,631 INFO [train.py:904] (4/8) Epoch 7, batch 8750, loss[loss=0.1941, simple_loss=0.2976, pruned_loss=0.04533, over 16866.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2879, pruned_loss=0.05681, over 3039662.47 frames. ], batch size: 102, lr: 9.50e-03, grad_scale: 8.0 2023-04-28 18:17:04,338 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69674.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 18:17:07,493 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69676.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:17:14,879 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9973, 3.9909, 3.8915, 3.4544, 3.9641, 1.8265, 3.7485, 3.6772], device='cuda:4'), covar=tensor([0.0065, 0.0056, 0.0112, 0.0201, 0.0062, 0.1881, 0.0088, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0101, 0.0089, 0.0137, 0.0126, 0.0103, 0.0156, 0.0121, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:17:45,988 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.541e+02 2.592e+02 3.192e+02 4.101e+02 7.533e+02, threshold=6.383e+02, percent-clipped=6.0 2023-04-28 18:17:55,726 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69699.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:18:01,164 INFO [train.py:904] (4/8) Epoch 7, batch 8800, loss[loss=0.2177, simple_loss=0.2953, pruned_loss=0.06998, over 12634.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2863, pruned_loss=0.05574, over 3045031.69 frames. ], batch size: 250, lr: 9.49e-03, grad_scale: 8.0 2023-04-28 18:18:47,924 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69724.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:19:11,440 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2867, 4.2785, 4.0974, 3.9668, 3.8041, 4.1913, 3.9994, 3.9994], device='cuda:4'), covar=tensor([0.0412, 0.0227, 0.0215, 0.0186, 0.0722, 0.0288, 0.0462, 0.0419], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0227, 0.0228, 0.0202, 0.0250, 0.0228, 0.0160, 0.0263], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:19:11,508 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69735.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 18:19:36,494 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69747.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:19:36,982 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-28 18:19:43,877 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69750.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:19:46,919 INFO [train.py:904] (4/8) Epoch 7, batch 8850, loss[loss=0.204, simple_loss=0.2955, pruned_loss=0.05624, over 15369.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2883, pruned_loss=0.05491, over 3040724.78 frames. ], batch size: 191, lr: 9.49e-03, grad_scale: 4.0 2023-04-28 18:21:21,320 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.819e+02 2.840e+02 3.473e+02 4.436e+02 7.105e+02, threshold=6.946e+02, percent-clipped=4.0 2023-04-28 18:21:34,927 INFO [train.py:904] (4/8) Epoch 7, batch 8900, loss[loss=0.1986, simple_loss=0.2929, pruned_loss=0.05219, over 16655.00 frames. ], tot_loss[loss=0.199, simple_loss=0.2892, pruned_loss=0.05442, over 3053659.96 frames. ], batch size: 134, lr: 9.49e-03, grad_scale: 4.0 2023-04-28 18:22:06,688 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69818.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:22:21,950 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69824.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:22:50,108 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4571, 1.5209, 1.8574, 2.4094, 2.3488, 2.3953, 1.7204, 2.5108], device='cuda:4'), covar=tensor([0.0079, 0.0282, 0.0161, 0.0141, 0.0152, 0.0133, 0.0287, 0.0076], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0150, 0.0133, 0.0133, 0.0143, 0.0097, 0.0149, 0.0089], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 18:23:03,155 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1568, 1.9175, 2.1036, 3.6311, 1.8796, 2.4592, 2.0816, 2.0106], device='cuda:4'), covar=tensor([0.0739, 0.2807, 0.1699, 0.0321, 0.3550, 0.1591, 0.2608, 0.2613], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0340, 0.0290, 0.0303, 0.0383, 0.0364, 0.0307, 0.0398], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:23:38,919 INFO [train.py:904] (4/8) Epoch 7, batch 8950, loss[loss=0.1709, simple_loss=0.2701, pruned_loss=0.03584, over 16683.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2883, pruned_loss=0.05436, over 3047399.66 frames. ], batch size: 83, lr: 9.48e-03, grad_scale: 4.0 2023-04-28 18:23:42,141 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3509, 3.4310, 3.4185, 1.7566, 3.6771, 3.7212, 2.9295, 2.9362], device='cuda:4'), covar=tensor([0.0681, 0.0137, 0.0136, 0.1139, 0.0041, 0.0067, 0.0314, 0.0357], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0091, 0.0077, 0.0135, 0.0064, 0.0080, 0.0113, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 18:24:09,508 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69866.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:24:22,084 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69872.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:24:38,037 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.90 vs. limit=5.0 2023-04-28 18:24:42,901 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69881.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:24:47,791 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7061, 3.7008, 3.8714, 3.7393, 3.8373, 4.2304, 3.9113, 3.6566], device='cuda:4'), covar=tensor([0.1894, 0.1993, 0.1519, 0.2413, 0.2658, 0.1298, 0.1232, 0.2741], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0401, 0.0416, 0.0355, 0.0461, 0.0447, 0.0334, 0.0467], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:25:14,731 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.602e+02 2.573e+02 2.911e+02 3.451e+02 7.486e+02, threshold=5.822e+02, percent-clipped=1.0 2023-04-28 18:25:27,356 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69901.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 18:25:27,969 INFO [train.py:904] (4/8) Epoch 7, batch 9000, loss[loss=0.1969, simple_loss=0.2933, pruned_loss=0.05021, over 16343.00 frames. ], tot_loss[loss=0.1952, simple_loss=0.2847, pruned_loss=0.05278, over 3039126.05 frames. ], batch size: 146, lr: 9.48e-03, grad_scale: 4.0 2023-04-28 18:25:27,970 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 18:25:37,198 INFO [train.py:938] (4/8) Epoch 7, validation: loss=0.1647, simple_loss=0.2682, pruned_loss=0.03062, over 944034.00 frames. 2023-04-28 18:25:37,199 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 18:25:40,488 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3261, 3.6479, 3.8378, 2.5993, 3.3480, 3.6964, 3.6123, 1.9194], device='cuda:4'), covar=tensor([0.0334, 0.0018, 0.0022, 0.0236, 0.0055, 0.0041, 0.0032, 0.0346], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0055, 0.0058, 0.0116, 0.0064, 0.0073, 0.0066, 0.0112], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 18:26:33,161 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69929.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:26:33,217 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69929.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 18:27:14,335 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69949.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 18:27:19,719 INFO [train.py:904] (4/8) Epoch 7, batch 9050, loss[loss=0.1777, simple_loss=0.2665, pruned_loss=0.04445, over 15476.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2857, pruned_loss=0.05305, over 3066981.08 frames. ], batch size: 191, lr: 9.48e-03, grad_scale: 4.0 2023-04-28 18:28:08,902 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69976.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:28:09,987 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=69977.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:28:50,080 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.782e+02 2.800e+02 3.566e+02 4.608e+02 8.238e+02, threshold=7.132e+02, percent-clipped=9.0 2023-04-28 18:29:06,534 INFO [train.py:904] (4/8) Epoch 7, batch 9100, loss[loss=0.2099, simple_loss=0.3028, pruned_loss=0.0585, over 15392.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2853, pruned_loss=0.05378, over 3062858.07 frames. ], batch size: 191, lr: 9.47e-03, grad_scale: 4.0 2023-04-28 18:29:58,209 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=70024.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:29:58,322 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70024.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:30:13,168 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=70030.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 18:31:01,020 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70050.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:31:03,189 INFO [train.py:904] (4/8) Epoch 7, batch 9150, loss[loss=0.1957, simple_loss=0.2892, pruned_loss=0.05109, over 16364.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2855, pruned_loss=0.05315, over 3069779.19 frames. ], batch size: 146, lr: 9.47e-03, grad_scale: 4.0 2023-04-28 18:31:11,790 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8613, 4.1675, 3.9254, 4.0161, 3.6554, 3.8229, 3.8222, 4.1669], device='cuda:4'), covar=tensor([0.1036, 0.0916, 0.1026, 0.0570, 0.0774, 0.1245, 0.0852, 0.0924], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0539, 0.0455, 0.0362, 0.0340, 0.0364, 0.0447, 0.0397], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:31:47,574 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=70072.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:32:25,731 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6617, 3.3602, 3.1715, 1.8431, 2.7797, 2.1608, 3.0656, 3.1997], device='cuda:4'), covar=tensor([0.0316, 0.0515, 0.0455, 0.1654, 0.0709, 0.0924, 0.0749, 0.0857], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0122, 0.0150, 0.0137, 0.0131, 0.0123, 0.0132, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 18:32:35,334 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.681e+02 3.453e+02 4.314e+02 7.152e+02, threshold=6.905e+02, percent-clipped=1.0 2023-04-28 18:32:39,002 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=70098.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:32:45,140 INFO [train.py:904] (4/8) Epoch 7, batch 9200, loss[loss=0.1931, simple_loss=0.2834, pruned_loss=0.05143, over 15401.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.281, pruned_loss=0.05197, over 3066541.72 frames. ], batch size: 191, lr: 9.47e-03, grad_scale: 8.0 2023-04-28 18:33:01,822 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=70110.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:34:22,191 INFO [train.py:904] (4/8) Epoch 7, batch 9250, loss[loss=0.1801, simple_loss=0.2591, pruned_loss=0.05052, over 11700.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2809, pruned_loss=0.05218, over 3052237.50 frames. ], batch size: 247, lr: 9.46e-03, grad_scale: 8.0 2023-04-28 18:35:02,128 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=70171.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:35:39,203 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1326, 4.0433, 4.0275, 3.0577, 4.0258, 1.5179, 3.8147, 3.7387], device='cuda:4'), covar=tensor([0.0136, 0.0111, 0.0156, 0.0503, 0.0121, 0.2775, 0.0148, 0.0297], device='cuda:4'), in_proj_covar=tensor([0.0102, 0.0089, 0.0136, 0.0126, 0.0103, 0.0156, 0.0121, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:36:01,188 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.738e+02 2.717e+02 3.344e+02 3.929e+02 8.889e+02, threshold=6.688e+02, percent-clipped=4.0 2023-04-28 18:36:13,560 INFO [train.py:904] (4/8) Epoch 7, batch 9300, loss[loss=0.1745, simple_loss=0.2728, pruned_loss=0.03808, over 16484.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2787, pruned_loss=0.05118, over 3029968.65 frames. ], batch size: 68, lr: 9.46e-03, grad_scale: 8.0 2023-04-28 18:36:55,974 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8245, 3.5680, 3.4458, 1.8678, 2.8708, 2.2950, 3.2428, 3.5760], device='cuda:4'), covar=tensor([0.0341, 0.0553, 0.0440, 0.1684, 0.0706, 0.0904, 0.0728, 0.0693], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0124, 0.0151, 0.0138, 0.0132, 0.0123, 0.0133, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 18:37:04,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6511, 1.5274, 2.0339, 2.5936, 2.3732, 2.5139, 1.6671, 2.6474], device='cuda:4'), covar=tensor([0.0080, 0.0289, 0.0170, 0.0157, 0.0153, 0.0134, 0.0274, 0.0110], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0149, 0.0131, 0.0133, 0.0141, 0.0096, 0.0149, 0.0088], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 18:37:22,917 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5968, 4.7292, 4.7669, 4.8097, 4.8015, 5.3593, 4.9064, 4.5875], device='cuda:4'), covar=tensor([0.0933, 0.1609, 0.1560, 0.1738, 0.2695, 0.0925, 0.1315, 0.2519], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0393, 0.0412, 0.0347, 0.0454, 0.0437, 0.0333, 0.0462], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:37:59,027 INFO [train.py:904] (4/8) Epoch 7, batch 9350, loss[loss=0.2061, simple_loss=0.2904, pruned_loss=0.06091, over 17006.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2789, pruned_loss=0.05086, over 3064371.93 frames. ], batch size: 109, lr: 9.46e-03, grad_scale: 8.0 2023-04-28 18:38:18,309 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=70260.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:39:31,294 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.755e+02 2.689e+02 3.082e+02 3.847e+02 8.568e+02, threshold=6.165e+02, percent-clipped=3.0 2023-04-28 18:39:40,120 INFO [train.py:904] (4/8) Epoch 7, batch 9400, loss[loss=0.197, simple_loss=0.2948, pruned_loss=0.04961, over 16167.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2795, pruned_loss=0.05077, over 3075467.71 frames. ], batch size: 165, lr: 9.45e-03, grad_scale: 4.0 2023-04-28 18:39:43,903 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9671, 4.2192, 4.0415, 4.0052, 3.7514, 3.8030, 3.8851, 4.2014], device='cuda:4'), covar=tensor([0.0849, 0.0826, 0.0875, 0.0629, 0.0701, 0.1469, 0.0737, 0.0946], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0537, 0.0453, 0.0360, 0.0342, 0.0364, 0.0443, 0.0398], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:39:58,776 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.73 vs. limit=5.0 2023-04-28 18:40:18,708 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=70321.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:40:36,579 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70330.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 18:41:20,562 INFO [train.py:904] (4/8) Epoch 7, batch 9450, loss[loss=0.185, simple_loss=0.2742, pruned_loss=0.04788, over 16891.00 frames. ], tot_loss[loss=0.1914, simple_loss=0.2807, pruned_loss=0.05106, over 3066076.07 frames. ], batch size: 109, lr: 9.45e-03, grad_scale: 4.0 2023-04-28 18:41:33,829 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0453, 2.6474, 2.6084, 1.7640, 2.8387, 2.8676, 2.4566, 2.4832], device='cuda:4'), covar=tensor([0.0668, 0.0136, 0.0152, 0.1000, 0.0081, 0.0124, 0.0386, 0.0368], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0092, 0.0079, 0.0136, 0.0064, 0.0081, 0.0115, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 18:42:04,297 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0185, 3.2281, 2.9906, 4.7605, 3.8044, 4.5534, 1.4644, 3.5854], device='cuda:4'), covar=tensor([0.1277, 0.0522, 0.0862, 0.0064, 0.0144, 0.0263, 0.1455, 0.0514], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0147, 0.0169, 0.0103, 0.0175, 0.0196, 0.0170, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 18:42:14,168 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=70378.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 18:42:50,913 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.795e+02 2.717e+02 3.594e+02 4.518e+02 8.204e+02, threshold=7.189e+02, percent-clipped=4.0 2023-04-28 18:43:02,095 INFO [train.py:904] (4/8) Epoch 7, batch 9500, loss[loss=0.2164, simple_loss=0.2996, pruned_loss=0.06658, over 16859.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2802, pruned_loss=0.05064, over 3080254.55 frames. ], batch size: 116, lr: 9.45e-03, grad_scale: 4.0 2023-04-28 18:43:49,081 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4847, 3.6187, 1.6103, 3.8053, 2.5292, 3.7203, 1.7834, 2.7712], device='cuda:4'), covar=tensor([0.0154, 0.0217, 0.1814, 0.0095, 0.0745, 0.0392, 0.1839, 0.0610], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0147, 0.0173, 0.0089, 0.0156, 0.0178, 0.0183, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 18:43:50,738 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.95 vs. limit=2.0 2023-04-28 18:44:08,044 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1360, 3.0961, 3.1409, 1.5831, 3.3859, 3.3643, 2.7031, 2.6680], device='cuda:4'), covar=tensor([0.0825, 0.0178, 0.0139, 0.1335, 0.0058, 0.0102, 0.0408, 0.0412], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0093, 0.0080, 0.0138, 0.0065, 0.0083, 0.0116, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 18:44:45,640 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9247, 2.7115, 2.5588, 1.9440, 2.5153, 2.7249, 2.6033, 1.8398], device='cuda:4'), covar=tensor([0.0283, 0.0028, 0.0037, 0.0233, 0.0063, 0.0055, 0.0042, 0.0310], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0056, 0.0059, 0.0117, 0.0065, 0.0073, 0.0065, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 18:44:47,822 INFO [train.py:904] (4/8) Epoch 7, batch 9550, loss[loss=0.1945, simple_loss=0.2863, pruned_loss=0.05137, over 16844.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2798, pruned_loss=0.0508, over 3072941.84 frames. ], batch size: 90, lr: 9.44e-03, grad_scale: 4.0 2023-04-28 18:44:57,973 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-28 18:45:12,635 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-28 18:45:18,569 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=70466.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:46:18,719 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.501e+02 2.839e+02 3.500e+02 4.170e+02 7.519e+02, threshold=6.999e+02, percent-clipped=3.0 2023-04-28 18:46:23,053 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-28 18:46:26,841 INFO [train.py:904] (4/8) Epoch 7, batch 9600, loss[loss=0.1913, simple_loss=0.2799, pruned_loss=0.05133, over 16451.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2814, pruned_loss=0.05179, over 3057993.60 frames. ], batch size: 68, lr: 9.44e-03, grad_scale: 8.0 2023-04-28 18:48:15,889 INFO [train.py:904] (4/8) Epoch 7, batch 9650, loss[loss=0.1955, simple_loss=0.2766, pruned_loss=0.05721, over 12366.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2833, pruned_loss=0.05238, over 3051960.76 frames. ], batch size: 246, lr: 9.44e-03, grad_scale: 8.0 2023-04-28 18:49:13,803 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 18:49:55,419 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.918e+02 2.673e+02 3.296e+02 3.931e+02 1.187e+03, threshold=6.593e+02, percent-clipped=2.0 2023-04-28 18:50:06,082 INFO [train.py:904] (4/8) Epoch 7, batch 9700, loss[loss=0.1821, simple_loss=0.2647, pruned_loss=0.04972, over 12210.00 frames. ], tot_loss[loss=0.193, simple_loss=0.282, pruned_loss=0.05196, over 3053934.85 frames. ], batch size: 249, lr: 9.43e-03, grad_scale: 8.0 2023-04-28 18:50:23,546 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6466, 2.7685, 1.7595, 2.8165, 2.1185, 2.7874, 1.9354, 2.4707], device='cuda:4'), covar=tensor([0.0188, 0.0314, 0.1231, 0.0121, 0.0663, 0.0391, 0.1200, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0148, 0.0176, 0.0091, 0.0158, 0.0179, 0.0185, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 18:50:23,882 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-28 18:50:31,155 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4985, 3.4204, 3.4797, 2.6629, 3.4394, 1.8088, 3.2168, 2.9980], device='cuda:4'), covar=tensor([0.0171, 0.0136, 0.0161, 0.0397, 0.0111, 0.2558, 0.0159, 0.0304], device='cuda:4'), in_proj_covar=tensor([0.0101, 0.0089, 0.0135, 0.0124, 0.0102, 0.0156, 0.0120, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:50:33,466 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=70616.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:51:47,796 INFO [train.py:904] (4/8) Epoch 7, batch 9750, loss[loss=0.2045, simple_loss=0.2771, pruned_loss=0.06595, over 12111.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2811, pruned_loss=0.05205, over 3059497.41 frames. ], batch size: 248, lr: 9.43e-03, grad_scale: 8.0 2023-04-28 18:53:18,480 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.561e+02 2.655e+02 3.090e+02 3.865e+02 1.050e+03, threshold=6.180e+02, percent-clipped=3.0 2023-04-28 18:53:27,255 INFO [train.py:904] (4/8) Epoch 7, batch 9800, loss[loss=0.1799, simple_loss=0.2828, pruned_loss=0.03851, over 16775.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2812, pruned_loss=0.05065, over 3077719.86 frames. ], batch size: 83, lr: 9.43e-03, grad_scale: 8.0 2023-04-28 18:55:12,749 INFO [train.py:904] (4/8) Epoch 7, batch 9850, loss[loss=0.1802, simple_loss=0.2732, pruned_loss=0.04361, over 16938.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2822, pruned_loss=0.05048, over 3083765.41 frames. ], batch size: 109, lr: 9.42e-03, grad_scale: 8.0 2023-04-28 18:55:43,256 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70766.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:56:54,661 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.862e+02 2.569e+02 3.072e+02 3.785e+02 9.202e+02, threshold=6.144e+02, percent-clipped=3.0 2023-04-28 18:57:04,338 INFO [train.py:904] (4/8) Epoch 7, batch 9900, loss[loss=0.2222, simple_loss=0.3101, pruned_loss=0.06718, over 16697.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2829, pruned_loss=0.05088, over 3075163.56 frames. ], batch size: 134, lr: 9.42e-03, grad_scale: 8.0 2023-04-28 18:57:33,474 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=70814.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 18:58:36,733 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0513, 5.4231, 5.1835, 5.1746, 4.8193, 4.7605, 4.8658, 5.4528], device='cuda:4'), covar=tensor([0.0806, 0.0646, 0.0719, 0.0518, 0.0654, 0.0665, 0.0674, 0.0843], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0525, 0.0439, 0.0357, 0.0337, 0.0355, 0.0435, 0.0392], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 18:59:01,222 INFO [train.py:904] (4/8) Epoch 7, batch 9950, loss[loss=0.1926, simple_loss=0.2881, pruned_loss=0.04855, over 15417.00 frames. ], tot_loss[loss=0.1938, simple_loss=0.2851, pruned_loss=0.05127, over 3071613.61 frames. ], batch size: 191, lr: 9.42e-03, grad_scale: 8.0 2023-04-28 19:00:47,807 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.452e+02 2.726e+02 3.249e+02 3.973e+02 1.098e+03, threshold=6.498e+02, percent-clipped=6.0 2023-04-28 19:01:00,566 INFO [train.py:904] (4/8) Epoch 7, batch 10000, loss[loss=0.1774, simple_loss=0.2752, pruned_loss=0.03979, over 16985.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2836, pruned_loss=0.05083, over 3080323.19 frames. ], batch size: 109, lr: 9.41e-03, grad_scale: 8.0 2023-04-28 19:01:29,775 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70916.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:01:30,088 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-28 19:02:40,827 INFO [train.py:904] (4/8) Epoch 7, batch 10050, loss[loss=0.2032, simple_loss=0.2905, pruned_loss=0.05793, over 17003.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2833, pruned_loss=0.05053, over 3069521.82 frames. ], batch size: 109, lr: 9.41e-03, grad_scale: 8.0 2023-04-28 19:03:04,142 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=70964.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:03:56,283 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=70992.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:04:04,308 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.660e+02 2.613e+02 3.026e+02 3.875e+02 7.173e+02, threshold=6.053e+02, percent-clipped=4.0 2023-04-28 19:04:12,142 INFO [train.py:904] (4/8) Epoch 7, batch 10100, loss[loss=0.1904, simple_loss=0.27, pruned_loss=0.05541, over 12176.00 frames. ], tot_loss[loss=0.1928, simple_loss=0.2837, pruned_loss=0.05093, over 3069388.77 frames. ], batch size: 247, lr: 9.41e-03, grad_scale: 8.0 2023-04-28 19:05:07,018 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=71030.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:05:54,955 INFO [train.py:904] (4/8) Epoch 8, batch 0, loss[loss=0.3338, simple_loss=0.3824, pruned_loss=0.1426, over 15362.00 frames. ], tot_loss[loss=0.3338, simple_loss=0.3824, pruned_loss=0.1426, over 15362.00 frames. ], batch size: 190, lr: 8.86e-03, grad_scale: 8.0 2023-04-28 19:05:54,955 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 19:06:02,577 INFO [train.py:938] (4/8) Epoch 8, validation: loss=0.1627, simple_loss=0.2663, pruned_loss=0.02958, over 944034.00 frames. 2023-04-28 19:06:02,578 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 19:06:04,011 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=71053.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:06:28,308 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-28 19:06:55,604 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=71091.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:07:07,690 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.561e+02 3.480e+02 4.237e+02 5.161e+02 1.332e+03, threshold=8.475e+02, percent-clipped=15.0 2023-04-28 19:07:09,983 INFO [train.py:904] (4/8) Epoch 8, batch 50, loss[loss=0.2147, simple_loss=0.2769, pruned_loss=0.07629, over 16839.00 frames. ], tot_loss[loss=0.2303, simple_loss=0.3009, pruned_loss=0.07986, over 758946.13 frames. ], batch size: 96, lr: 8.86e-03, grad_scale: 1.0 2023-04-28 19:07:28,516 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8072, 3.9236, 2.9861, 2.3489, 2.5740, 2.2482, 3.7922, 3.6454], device='cuda:4'), covar=tensor([0.1949, 0.0459, 0.1229, 0.1804, 0.1916, 0.1485, 0.0394, 0.0815], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0239, 0.0268, 0.0251, 0.0243, 0.0204, 0.0242, 0.0257], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:08:10,224 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8944, 4.0882, 2.0904, 4.6161, 2.8745, 4.5548, 2.0414, 3.2580], device='cuda:4'), covar=tensor([0.0211, 0.0340, 0.1852, 0.0147, 0.0863, 0.0342, 0.1874, 0.0567], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0151, 0.0177, 0.0094, 0.0159, 0.0182, 0.0187, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 19:08:17,894 INFO [train.py:904] (4/8) Epoch 8, batch 100, loss[loss=0.227, simple_loss=0.2965, pruned_loss=0.0788, over 16551.00 frames. ], tot_loss[loss=0.2222, simple_loss=0.2964, pruned_loss=0.07401, over 1322228.70 frames. ], batch size: 146, lr: 8.85e-03, grad_scale: 1.0 2023-04-28 19:08:42,464 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7535, 3.7738, 3.8020, 3.7787, 3.8013, 4.2372, 3.9423, 3.6721], device='cuda:4'), covar=tensor([0.1773, 0.1831, 0.1830, 0.2257, 0.2862, 0.1578, 0.1309, 0.2650], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0427, 0.0445, 0.0374, 0.0491, 0.0467, 0.0355, 0.0496], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:09:23,827 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.200e+02 2.858e+02 3.485e+02 4.000e+02 7.217e+02, threshold=6.971e+02, percent-clipped=0.0 2023-04-28 19:09:26,722 INFO [train.py:904] (4/8) Epoch 8, batch 150, loss[loss=0.2339, simple_loss=0.3098, pruned_loss=0.07902, over 15444.00 frames. ], tot_loss[loss=0.2172, simple_loss=0.294, pruned_loss=0.07016, over 1762666.56 frames. ], batch size: 190, lr: 8.85e-03, grad_scale: 1.0 2023-04-28 19:09:41,140 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4374, 5.8682, 5.6005, 5.6582, 5.1534, 5.0367, 5.3571, 5.9829], device='cuda:4'), covar=tensor([0.1044, 0.0847, 0.1035, 0.0647, 0.0886, 0.0710, 0.0798, 0.0904], device='cuda:4'), in_proj_covar=tensor([0.0446, 0.0564, 0.0472, 0.0382, 0.0364, 0.0379, 0.0472, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:10:33,447 INFO [train.py:904] (4/8) Epoch 8, batch 200, loss[loss=0.2396, simple_loss=0.2972, pruned_loss=0.09098, over 16922.00 frames. ], tot_loss[loss=0.2147, simple_loss=0.2919, pruned_loss=0.06878, over 2117263.59 frames. ], batch size: 96, lr: 8.85e-03, grad_scale: 1.0 2023-04-28 19:11:38,547 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2569, 5.2425, 4.9121, 4.3706, 4.9538, 1.7868, 4.8028, 5.0519], device='cuda:4'), covar=tensor([0.0049, 0.0041, 0.0124, 0.0284, 0.0067, 0.2026, 0.0077, 0.0117], device='cuda:4'), in_proj_covar=tensor([0.0107, 0.0094, 0.0144, 0.0132, 0.0109, 0.0163, 0.0127, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:11:40,016 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.535e+02 2.597e+02 3.114e+02 3.652e+02 1.080e+03, threshold=6.228e+02, percent-clipped=1.0 2023-04-28 19:11:42,945 INFO [train.py:904] (4/8) Epoch 8, batch 250, loss[loss=0.2431, simple_loss=0.3133, pruned_loss=0.08647, over 16544.00 frames. ], tot_loss[loss=0.2119, simple_loss=0.2889, pruned_loss=0.06746, over 2388964.17 frames. ], batch size: 68, lr: 8.84e-03, grad_scale: 1.0 2023-04-28 19:11:43,400 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0679, 4.6863, 4.9754, 5.2551, 5.4906, 4.8343, 5.4129, 5.4137], device='cuda:4'), covar=tensor([0.1134, 0.1016, 0.1576, 0.0672, 0.0418, 0.0586, 0.0406, 0.0494], device='cuda:4'), in_proj_covar=tensor([0.0452, 0.0568, 0.0693, 0.0565, 0.0427, 0.0429, 0.0441, 0.0491], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:11:56,904 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5007, 3.4632, 2.6310, 2.1322, 2.2619, 2.1343, 3.3430, 3.2528], device='cuda:4'), covar=tensor([0.2334, 0.0609, 0.1321, 0.2025, 0.2046, 0.1653, 0.0477, 0.0989], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0245, 0.0275, 0.0257, 0.0258, 0.0209, 0.0250, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:12:02,806 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9477, 3.9076, 4.4129, 3.2147, 4.0125, 4.2999, 4.1091, 2.6608], device='cuda:4'), covar=tensor([0.0310, 0.0034, 0.0021, 0.0206, 0.0046, 0.0040, 0.0033, 0.0288], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0063, 0.0062, 0.0119, 0.0066, 0.0075, 0.0067, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:12:23,545 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.91 vs. limit=5.0 2023-04-28 19:12:46,244 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1163, 3.2609, 1.7312, 3.2542, 2.3791, 3.3325, 1.8543, 2.6272], device='cuda:4'), covar=tensor([0.0221, 0.0389, 0.1579, 0.0314, 0.0774, 0.0671, 0.1376, 0.0611], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0155, 0.0181, 0.0098, 0.0163, 0.0190, 0.0191, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 19:12:47,181 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=71348.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:12:52,125 INFO [train.py:904] (4/8) Epoch 8, batch 300, loss[loss=0.1708, simple_loss=0.2496, pruned_loss=0.04595, over 16776.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2849, pruned_loss=0.06529, over 2589612.16 frames. ], batch size: 39, lr: 8.84e-03, grad_scale: 1.0 2023-04-28 19:13:38,970 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=71386.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:13:58,499 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.899e+02 2.511e+02 3.030e+02 3.934e+02 1.129e+03, threshold=6.060e+02, percent-clipped=6.0 2023-04-28 19:14:01,156 INFO [train.py:904] (4/8) Epoch 8, batch 350, loss[loss=0.1741, simple_loss=0.2606, pruned_loss=0.04377, over 17161.00 frames. ], tot_loss[loss=0.205, simple_loss=0.2828, pruned_loss=0.06359, over 2756545.36 frames. ], batch size: 46, lr: 8.84e-03, grad_scale: 1.0 2023-04-28 19:14:44,216 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3177, 5.2365, 5.1266, 4.7230, 4.6275, 5.1225, 5.1486, 4.7601], device='cuda:4'), covar=tensor([0.0482, 0.0298, 0.0227, 0.0239, 0.0945, 0.0299, 0.0186, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0214, 0.0253, 0.0252, 0.0223, 0.0282, 0.0255, 0.0171, 0.0294], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:15:00,296 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.84 vs. limit=5.0 2023-04-28 19:15:10,453 INFO [train.py:904] (4/8) Epoch 8, batch 400, loss[loss=0.2013, simple_loss=0.2926, pruned_loss=0.05496, over 17251.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2803, pruned_loss=0.0631, over 2867803.31 frames. ], batch size: 52, lr: 8.83e-03, grad_scale: 2.0 2023-04-28 19:15:59,999 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8435, 3.1239, 2.7524, 4.4965, 3.9025, 4.3985, 1.6802, 3.0799], device='cuda:4'), covar=tensor([0.1253, 0.0507, 0.0946, 0.0107, 0.0238, 0.0304, 0.1267, 0.0701], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0149, 0.0171, 0.0109, 0.0184, 0.0200, 0.0170, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 19:16:08,587 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-28 19:16:17,822 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.026e+02 2.889e+02 3.429e+02 4.074e+02 1.363e+03, threshold=6.859e+02, percent-clipped=4.0 2023-04-28 19:16:20,152 INFO [train.py:904] (4/8) Epoch 8, batch 450, loss[loss=0.1932, simple_loss=0.2595, pruned_loss=0.06342, over 16827.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2787, pruned_loss=0.06219, over 2968619.21 frames. ], batch size: 83, lr: 8.83e-03, grad_scale: 2.0 2023-04-28 19:16:50,235 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7682, 5.0109, 5.1885, 5.0696, 4.9642, 5.6084, 5.1821, 4.8772], device='cuda:4'), covar=tensor([0.1046, 0.1869, 0.1706, 0.1914, 0.2868, 0.1061, 0.1313, 0.2537], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0449, 0.0463, 0.0392, 0.0517, 0.0490, 0.0372, 0.0526], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:17:07,427 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7900, 4.1001, 4.2893, 1.9539, 4.5711, 4.5926, 3.1649, 3.5155], device='cuda:4'), covar=tensor([0.0696, 0.0134, 0.0163, 0.1064, 0.0048, 0.0106, 0.0340, 0.0345], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0092, 0.0080, 0.0137, 0.0066, 0.0087, 0.0116, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 19:17:28,362 INFO [train.py:904] (4/8) Epoch 8, batch 500, loss[loss=0.1927, simple_loss=0.2854, pruned_loss=0.04998, over 17079.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.2759, pruned_loss=0.0603, over 3050322.58 frames. ], batch size: 53, lr: 8.83e-03, grad_scale: 2.0 2023-04-28 19:18:33,694 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.591e+02 2.499e+02 3.026e+02 3.865e+02 9.130e+02, threshold=6.053e+02, percent-clipped=1.0 2023-04-28 19:18:37,346 INFO [train.py:904] (4/8) Epoch 8, batch 550, loss[loss=0.2268, simple_loss=0.2951, pruned_loss=0.07921, over 16717.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2747, pruned_loss=0.05967, over 3116041.37 frames. ], batch size: 134, lr: 8.83e-03, grad_scale: 2.0 2023-04-28 19:18:56,814 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=71616.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:19:11,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5332, 5.8676, 5.6275, 5.7473, 5.2173, 5.0796, 5.3042, 6.0157], device='cuda:4'), covar=tensor([0.0997, 0.0857, 0.1060, 0.0538, 0.0796, 0.0680, 0.0840, 0.0753], device='cuda:4'), in_proj_covar=tensor([0.0473, 0.0604, 0.0503, 0.0406, 0.0385, 0.0400, 0.0502, 0.0446], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:19:41,707 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=71648.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:19:45,774 INFO [train.py:904] (4/8) Epoch 8, batch 600, loss[loss=0.2152, simple_loss=0.28, pruned_loss=0.07518, over 16863.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2744, pruned_loss=0.06035, over 3161749.00 frames. ], batch size: 96, lr: 8.82e-03, grad_scale: 2.0 2023-04-28 19:20:01,316 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-28 19:20:05,643 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3302, 1.9746, 2.2363, 3.9331, 2.0038, 2.4820, 2.1273, 2.1977], device='cuda:4'), covar=tensor([0.0912, 0.3041, 0.1752, 0.0410, 0.3225, 0.1892, 0.2787, 0.2620], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0358, 0.0303, 0.0323, 0.0395, 0.0394, 0.0323, 0.0426], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:20:15,674 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3293, 5.2214, 5.0750, 4.6526, 4.5669, 5.0694, 5.1044, 4.6966], device='cuda:4'), covar=tensor([0.0460, 0.0349, 0.0238, 0.0269, 0.1115, 0.0387, 0.0191, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0219, 0.0258, 0.0256, 0.0230, 0.0287, 0.0261, 0.0176, 0.0299], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:20:19,070 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=71677.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:20:34,029 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=71686.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:20:36,051 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 19:20:45,947 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=71696.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:20:50,939 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.658e+02 2.476e+02 3.033e+02 3.542e+02 7.789e+02, threshold=6.066e+02, percent-clipped=2.0 2023-04-28 19:20:53,317 INFO [train.py:904] (4/8) Epoch 8, batch 650, loss[loss=0.1803, simple_loss=0.2641, pruned_loss=0.04825, over 17209.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2725, pruned_loss=0.05956, over 3192326.63 frames. ], batch size: 44, lr: 8.82e-03, grad_scale: 2.0 2023-04-28 19:21:37,502 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=71734.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:22:01,894 INFO [train.py:904] (4/8) Epoch 8, batch 700, loss[loss=0.2186, simple_loss=0.2887, pruned_loss=0.07428, over 16731.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2726, pruned_loss=0.05873, over 3232499.87 frames. ], batch size: 134, lr: 8.82e-03, grad_scale: 2.0 2023-04-28 19:23:05,954 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.350e+02 2.585e+02 3.182e+02 3.748e+02 9.363e+02, threshold=6.364e+02, percent-clipped=4.0 2023-04-28 19:23:08,640 INFO [train.py:904] (4/8) Epoch 8, batch 750, loss[loss=0.1969, simple_loss=0.2815, pruned_loss=0.05613, over 17045.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2733, pruned_loss=0.05851, over 3249143.46 frames. ], batch size: 55, lr: 8.81e-03, grad_scale: 2.0 2023-04-28 19:23:15,307 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0197, 4.7452, 4.8682, 5.2603, 5.4117, 4.7424, 5.3240, 5.3070], device='cuda:4'), covar=tensor([0.1241, 0.1106, 0.2207, 0.0696, 0.0569, 0.0667, 0.0631, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0483, 0.0596, 0.0738, 0.0598, 0.0454, 0.0455, 0.0471, 0.0522], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:23:43,113 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-28 19:24:17,945 INFO [train.py:904] (4/8) Epoch 8, batch 800, loss[loss=0.1832, simple_loss=0.2719, pruned_loss=0.04724, over 16779.00 frames. ], tot_loss[loss=0.1957, simple_loss=0.2738, pruned_loss=0.05879, over 3259907.94 frames. ], batch size: 57, lr: 8.81e-03, grad_scale: 4.0 2023-04-28 19:24:32,595 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 19:24:45,213 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 19:25:05,935 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5229, 2.5287, 2.0025, 2.2581, 2.8929, 2.7001, 3.4585, 3.1390], device='cuda:4'), covar=tensor([0.0055, 0.0244, 0.0320, 0.0303, 0.0150, 0.0236, 0.0131, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0111, 0.0182, 0.0178, 0.0178, 0.0175, 0.0185, 0.0178, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:25:23,720 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.711e+02 2.663e+02 3.116e+02 3.838e+02 7.032e+02, threshold=6.231e+02, percent-clipped=1.0 2023-04-28 19:25:24,257 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4237, 3.6318, 1.9026, 3.7063, 2.5676, 3.6840, 1.9855, 2.8323], device='cuda:4'), covar=tensor([0.0200, 0.0318, 0.1378, 0.0159, 0.0722, 0.0506, 0.1349, 0.0545], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0156, 0.0177, 0.0099, 0.0162, 0.0194, 0.0190, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 19:25:25,971 INFO [train.py:904] (4/8) Epoch 8, batch 850, loss[loss=0.1961, simple_loss=0.2838, pruned_loss=0.05421, over 17013.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2728, pruned_loss=0.05801, over 3265875.90 frames. ], batch size: 50, lr: 8.81e-03, grad_scale: 4.0 2023-04-28 19:25:49,761 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=71919.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:26:32,698 INFO [train.py:904] (4/8) Epoch 8, batch 900, loss[loss=0.2066, simple_loss=0.279, pruned_loss=0.06714, over 16433.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.272, pruned_loss=0.05737, over 3284123.25 frames. ], batch size: 146, lr: 8.80e-03, grad_scale: 4.0 2023-04-28 19:26:50,507 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9406, 5.3420, 5.4302, 5.2164, 5.2554, 5.8619, 5.4348, 5.1553], device='cuda:4'), covar=tensor([0.0835, 0.1482, 0.1540, 0.1885, 0.2613, 0.0938, 0.1157, 0.2082], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0453, 0.0470, 0.0396, 0.0522, 0.0497, 0.0373, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:27:00,518 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=71972.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:27:00,619 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1586, 3.9660, 4.1506, 4.3797, 4.4518, 4.0156, 4.1620, 4.4406], device='cuda:4'), covar=tensor([0.1131, 0.0880, 0.1222, 0.0502, 0.0514, 0.1218, 0.1650, 0.0491], device='cuda:4'), in_proj_covar=tensor([0.0483, 0.0598, 0.0740, 0.0600, 0.0458, 0.0456, 0.0470, 0.0517], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:27:13,345 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=71980.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:27:38,431 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.818e+02 2.573e+02 3.042e+02 3.857e+02 8.032e+02, threshold=6.083e+02, percent-clipped=5.0 2023-04-28 19:27:44,214 INFO [train.py:904] (4/8) Epoch 8, batch 950, loss[loss=0.2123, simple_loss=0.2949, pruned_loss=0.06489, over 16501.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2719, pruned_loss=0.05707, over 3294539.59 frames. ], batch size: 68, lr: 8.80e-03, grad_scale: 4.0 2023-04-28 19:27:53,697 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7822, 2.9633, 2.7631, 4.9821, 4.0485, 4.4790, 1.6744, 3.1980], device='cuda:4'), covar=tensor([0.1418, 0.0718, 0.1193, 0.0134, 0.0267, 0.0383, 0.1526, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0152, 0.0174, 0.0114, 0.0192, 0.0205, 0.0171, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 19:28:52,161 INFO [train.py:904] (4/8) Epoch 8, batch 1000, loss[loss=0.2005, simple_loss=0.2656, pruned_loss=0.06775, over 16864.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2706, pruned_loss=0.05668, over 3307584.04 frames. ], batch size: 116, lr: 8.80e-03, grad_scale: 4.0 2023-04-28 19:29:58,311 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.793e+02 2.648e+02 3.149e+02 3.724e+02 6.885e+02, threshold=6.299e+02, percent-clipped=3.0 2023-04-28 19:30:01,626 INFO [train.py:904] (4/8) Epoch 8, batch 1050, loss[loss=0.2005, simple_loss=0.2721, pruned_loss=0.06445, over 12186.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2711, pruned_loss=0.05707, over 3309741.91 frames. ], batch size: 246, lr: 8.79e-03, grad_scale: 4.0 2023-04-28 19:30:40,254 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=72129.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:30:40,276 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8617, 1.6074, 2.2266, 2.7664, 2.6359, 3.2027, 2.0697, 3.2179], device='cuda:4'), covar=tensor([0.0152, 0.0358, 0.0210, 0.0180, 0.0172, 0.0113, 0.0279, 0.0091], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0159, 0.0143, 0.0143, 0.0150, 0.0106, 0.0157, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 19:30:50,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2973, 3.2922, 3.7423, 2.4042, 3.5050, 3.7438, 3.5872, 2.1768], device='cuda:4'), covar=tensor([0.0347, 0.0141, 0.0039, 0.0267, 0.0057, 0.0068, 0.0059, 0.0325], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0064, 0.0062, 0.0118, 0.0068, 0.0077, 0.0070, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:31:10,579 INFO [train.py:904] (4/8) Epoch 8, batch 1100, loss[loss=0.1956, simple_loss=0.2875, pruned_loss=0.05185, over 17082.00 frames. ], tot_loss[loss=0.1919, simple_loss=0.2703, pruned_loss=0.05671, over 3318445.48 frames. ], batch size: 55, lr: 8.79e-03, grad_scale: 4.0 2023-04-28 19:31:59,409 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5612, 5.4818, 5.3567, 5.0491, 4.9479, 5.3646, 5.3761, 5.0645], device='cuda:4'), covar=tensor([0.0385, 0.0321, 0.0190, 0.0203, 0.0880, 0.0296, 0.0186, 0.0528], device='cuda:4'), in_proj_covar=tensor([0.0228, 0.0267, 0.0266, 0.0240, 0.0299, 0.0271, 0.0182, 0.0308], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:32:03,604 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=72190.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:32:15,691 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.524e+02 2.462e+02 2.908e+02 3.717e+02 7.729e+02, threshold=5.817e+02, percent-clipped=1.0 2023-04-28 19:32:18,230 INFO [train.py:904] (4/8) Epoch 8, batch 1150, loss[loss=0.1943, simple_loss=0.2601, pruned_loss=0.06422, over 16897.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.2701, pruned_loss=0.05648, over 3305417.94 frames. ], batch size: 90, lr: 8.79e-03, grad_scale: 4.0 2023-04-28 19:33:22,634 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1235, 3.3050, 3.3013, 1.6177, 3.5089, 3.5527, 2.7588, 2.6554], device='cuda:4'), covar=tensor([0.0817, 0.0150, 0.0186, 0.1141, 0.0079, 0.0134, 0.0425, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0095, 0.0084, 0.0140, 0.0070, 0.0091, 0.0120, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 19:33:26,638 INFO [train.py:904] (4/8) Epoch 8, batch 1200, loss[loss=0.1933, simple_loss=0.2682, pruned_loss=0.05915, over 12387.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2697, pruned_loss=0.05648, over 3296191.72 frames. ], batch size: 246, lr: 8.79e-03, grad_scale: 8.0 2023-04-28 19:33:53,177 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=72272.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:33:56,993 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=72275.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:34:07,233 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9642, 5.6480, 5.7808, 5.5978, 5.6416, 6.1625, 5.7996, 5.4958], device='cuda:4'), covar=tensor([0.0787, 0.1540, 0.1738, 0.1756, 0.2416, 0.0918, 0.1196, 0.2343], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0455, 0.0475, 0.0404, 0.0527, 0.0505, 0.0378, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:34:30,639 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.919e+02 2.822e+02 3.117e+02 3.864e+02 9.153e+02, threshold=6.234e+02, percent-clipped=4.0 2023-04-28 19:34:32,371 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5732, 3.3437, 2.6780, 2.2024, 2.3186, 2.1680, 3.3833, 3.2678], device='cuda:4'), covar=tensor([0.2158, 0.0749, 0.1416, 0.2085, 0.2014, 0.1696, 0.0508, 0.0964], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0254, 0.0276, 0.0260, 0.0274, 0.0213, 0.0255, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:34:32,984 INFO [train.py:904] (4/8) Epoch 8, batch 1250, loss[loss=0.2118, simple_loss=0.2813, pruned_loss=0.07118, over 16315.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.27, pruned_loss=0.05683, over 3310680.13 frames. ], batch size: 165, lr: 8.78e-03, grad_scale: 8.0 2023-04-28 19:34:58,450 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=72320.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:35:06,144 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=72326.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:35:41,551 INFO [train.py:904] (4/8) Epoch 8, batch 1300, loss[loss=0.1728, simple_loss=0.2655, pruned_loss=0.04009, over 17222.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2695, pruned_loss=0.05628, over 3309301.68 frames. ], batch size: 44, lr: 8.78e-03, grad_scale: 8.0 2023-04-28 19:36:30,378 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=72387.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:36:49,363 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.738e+02 2.612e+02 3.072e+02 3.829e+02 7.494e+02, threshold=6.144e+02, percent-clipped=3.0 2023-04-28 19:36:52,118 INFO [train.py:904] (4/8) Epoch 8, batch 1350, loss[loss=0.1825, simple_loss=0.27, pruned_loss=0.0475, over 16501.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2696, pruned_loss=0.05592, over 3292552.07 frames. ], batch size: 68, lr: 8.78e-03, grad_scale: 8.0 2023-04-28 19:37:13,374 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6978, 2.5425, 2.1626, 2.4748, 3.0156, 2.7750, 3.6762, 3.2310], device='cuda:4'), covar=tensor([0.0048, 0.0250, 0.0309, 0.0259, 0.0153, 0.0219, 0.0108, 0.0132], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0187, 0.0182, 0.0182, 0.0181, 0.0188, 0.0186, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:37:19,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8768, 3.9288, 3.1286, 2.4111, 2.7840, 2.3860, 4.1650, 3.7704], device='cuda:4'), covar=tensor([0.2106, 0.0566, 0.1215, 0.1784, 0.1986, 0.1607, 0.0370, 0.0846], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0258, 0.0280, 0.0264, 0.0276, 0.0215, 0.0258, 0.0285], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:37:24,425 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3661, 3.3493, 3.7937, 2.7635, 3.4265, 3.7430, 3.5804, 2.1146], device='cuda:4'), covar=tensor([0.0376, 0.0144, 0.0031, 0.0227, 0.0069, 0.0068, 0.0066, 0.0346], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0065, 0.0062, 0.0117, 0.0068, 0.0076, 0.0070, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:38:01,923 INFO [train.py:904] (4/8) Epoch 8, batch 1400, loss[loss=0.2094, simple_loss=0.2896, pruned_loss=0.06464, over 16756.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2692, pruned_loss=0.05561, over 3291650.83 frames. ], batch size: 62, lr: 8.77e-03, grad_scale: 8.0 2023-04-28 19:38:11,391 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9722, 1.6967, 2.2942, 2.8464, 2.6754, 3.3151, 1.9725, 3.3391], device='cuda:4'), covar=tensor([0.0127, 0.0317, 0.0216, 0.0191, 0.0178, 0.0101, 0.0320, 0.0082], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0157, 0.0143, 0.0143, 0.0150, 0.0104, 0.0157, 0.0097], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 19:38:47,858 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=72485.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:39:07,467 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.787e+02 2.583e+02 2.954e+02 3.513e+02 6.005e+02, threshold=5.909e+02, percent-clipped=0.0 2023-04-28 19:39:11,104 INFO [train.py:904] (4/8) Epoch 8, batch 1450, loss[loss=0.1774, simple_loss=0.2607, pruned_loss=0.04707, over 16839.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2688, pruned_loss=0.0555, over 3292717.10 frames. ], batch size: 42, lr: 8.77e-03, grad_scale: 8.0 2023-04-28 19:39:28,256 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8983, 3.9542, 3.2389, 2.4183, 2.8467, 2.4933, 4.2167, 3.8344], device='cuda:4'), covar=tensor([0.2086, 0.0672, 0.1220, 0.1820, 0.2021, 0.1498, 0.0383, 0.0926], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0259, 0.0280, 0.0263, 0.0278, 0.0215, 0.0258, 0.0285], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:39:48,328 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1477, 4.0590, 4.6049, 4.6373, 4.6384, 4.2703, 4.2799, 4.1238], device='cuda:4'), covar=tensor([0.0309, 0.0535, 0.0328, 0.0351, 0.0345, 0.0337, 0.0789, 0.0499], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0311, 0.0305, 0.0292, 0.0347, 0.0325, 0.0424, 0.0264], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 19:40:17,878 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6819, 4.7449, 5.2710, 5.2766, 5.2525, 4.9123, 4.7762, 4.5506], device='cuda:4'), covar=tensor([0.0308, 0.0373, 0.0356, 0.0411, 0.0418, 0.0283, 0.0877, 0.0417], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0314, 0.0309, 0.0296, 0.0352, 0.0330, 0.0430, 0.0268], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 19:40:20,581 INFO [train.py:904] (4/8) Epoch 8, batch 1500, loss[loss=0.2116, simple_loss=0.2799, pruned_loss=0.0716, over 16536.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2693, pruned_loss=0.0558, over 3301601.38 frames. ], batch size: 68, lr: 8.77e-03, grad_scale: 8.0 2023-04-28 19:40:52,090 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=72575.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:40:53,735 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 19:41:25,445 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.638e+02 2.563e+02 3.021e+02 3.528e+02 6.404e+02, threshold=6.041e+02, percent-clipped=2.0 2023-04-28 19:41:28,321 INFO [train.py:904] (4/8) Epoch 8, batch 1550, loss[loss=0.2149, simple_loss=0.2822, pruned_loss=0.07376, over 16466.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.271, pruned_loss=0.05791, over 3314229.32 frames. ], batch size: 146, lr: 8.76e-03, grad_scale: 8.0 2023-04-28 19:41:58,081 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=72623.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:42:39,323 INFO [train.py:904] (4/8) Epoch 8, batch 1600, loss[loss=0.2154, simple_loss=0.2958, pruned_loss=0.06753, over 16523.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.2728, pruned_loss=0.05828, over 3307261.71 frames. ], batch size: 68, lr: 8.76e-03, grad_scale: 8.0 2023-04-28 19:43:20,418 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=72682.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:43:44,412 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.962e+02 2.814e+02 3.483e+02 4.056e+02 7.586e+02, threshold=6.967e+02, percent-clipped=5.0 2023-04-28 19:43:47,307 INFO [train.py:904] (4/8) Epoch 8, batch 1650, loss[loss=0.2275, simple_loss=0.2923, pruned_loss=0.08133, over 16807.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2741, pruned_loss=0.05881, over 3314050.90 frames. ], batch size: 124, lr: 8.76e-03, grad_scale: 8.0 2023-04-28 19:44:58,248 INFO [train.py:904] (4/8) Epoch 8, batch 1700, loss[loss=0.2139, simple_loss=0.2945, pruned_loss=0.0666, over 16656.00 frames. ], tot_loss[loss=0.1962, simple_loss=0.2751, pruned_loss=0.05861, over 3319761.49 frames. ], batch size: 134, lr: 8.76e-03, grad_scale: 8.0 2023-04-28 19:45:25,378 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9386, 3.2668, 2.7882, 5.2139, 4.5381, 4.8983, 1.6406, 3.4124], device='cuda:4'), covar=tensor([0.1304, 0.0608, 0.1100, 0.0118, 0.0263, 0.0306, 0.1448, 0.0655], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0154, 0.0174, 0.0117, 0.0196, 0.0206, 0.0172, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 19:45:44,137 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=72785.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:45:49,874 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 19:46:05,113 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.772e+02 2.648e+02 3.116e+02 3.582e+02 6.867e+02, threshold=6.231e+02, percent-clipped=0.0 2023-04-28 19:46:07,546 INFO [train.py:904] (4/8) Epoch 8, batch 1750, loss[loss=0.1697, simple_loss=0.2519, pruned_loss=0.0438, over 17199.00 frames. ], tot_loss[loss=0.196, simple_loss=0.2758, pruned_loss=0.05816, over 3321503.80 frames. ], batch size: 44, lr: 8.75e-03, grad_scale: 8.0 2023-04-28 19:46:07,839 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2094, 5.5618, 5.3127, 5.4051, 4.9934, 4.8738, 5.0331, 5.7251], device='cuda:4'), covar=tensor([0.1026, 0.0827, 0.1097, 0.0617, 0.0744, 0.0688, 0.0892, 0.0818], device='cuda:4'), in_proj_covar=tensor([0.0481, 0.0617, 0.0510, 0.0413, 0.0386, 0.0401, 0.0513, 0.0454], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:46:50,155 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=72833.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:47:11,464 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7786, 3.0083, 2.7060, 4.5865, 3.8789, 4.4040, 1.6391, 3.0835], device='cuda:4'), covar=tensor([0.1305, 0.0558, 0.0999, 0.0127, 0.0300, 0.0332, 0.1311, 0.0730], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0152, 0.0173, 0.0116, 0.0195, 0.0204, 0.0170, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 19:47:16,118 INFO [train.py:904] (4/8) Epoch 8, batch 1800, loss[loss=0.2114, simple_loss=0.3049, pruned_loss=0.05896, over 17092.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.2773, pruned_loss=0.05863, over 3326366.66 frames. ], batch size: 53, lr: 8.75e-03, grad_scale: 4.0 2023-04-28 19:47:16,511 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5230, 4.5191, 4.4700, 3.9222, 4.4595, 1.6913, 4.2364, 4.2174], device='cuda:4'), covar=tensor([0.0071, 0.0056, 0.0106, 0.0287, 0.0066, 0.2101, 0.0109, 0.0148], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0103, 0.0155, 0.0149, 0.0119, 0.0166, 0.0140, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:47:28,008 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5467, 3.6517, 2.0661, 3.8301, 2.6664, 3.7605, 2.1213, 2.8842], device='cuda:4'), covar=tensor([0.0149, 0.0381, 0.1142, 0.0121, 0.0621, 0.0488, 0.1076, 0.0479], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0163, 0.0180, 0.0107, 0.0164, 0.0202, 0.0190, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 19:48:15,239 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=72894.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:48:24,988 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.805e+02 2.735e+02 3.300e+02 4.209e+02 8.740e+02, threshold=6.600e+02, percent-clipped=4.0 2023-04-28 19:48:26,868 INFO [train.py:904] (4/8) Epoch 8, batch 1850, loss[loss=0.2136, simple_loss=0.2817, pruned_loss=0.07269, over 16399.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2783, pruned_loss=0.05905, over 3323309.16 frames. ], batch size: 146, lr: 8.75e-03, grad_scale: 4.0 2023-04-28 19:48:47,806 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-28 19:49:00,726 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-04-28 19:49:22,788 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-28 19:49:35,219 INFO [train.py:904] (4/8) Epoch 8, batch 1900, loss[loss=0.1686, simple_loss=0.2443, pruned_loss=0.0465, over 16851.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2777, pruned_loss=0.05888, over 3320206.79 frames. ], batch size: 42, lr: 8.74e-03, grad_scale: 4.0 2023-04-28 19:49:39,697 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=72955.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:49:52,011 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=72964.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 19:49:52,067 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2627, 2.3241, 1.7844, 2.0478, 2.7544, 2.4939, 3.3170, 2.9725], device='cuda:4'), covar=tensor([0.0063, 0.0286, 0.0386, 0.0315, 0.0180, 0.0257, 0.0152, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0185, 0.0183, 0.0182, 0.0184, 0.0189, 0.0187, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:50:05,647 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4715, 5.8969, 5.6741, 5.7607, 5.2845, 5.0280, 5.3599, 6.0621], device='cuda:4'), covar=tensor([0.1098, 0.0855, 0.0957, 0.0602, 0.0716, 0.0727, 0.0910, 0.0849], device='cuda:4'), in_proj_covar=tensor([0.0481, 0.0612, 0.0509, 0.0408, 0.0386, 0.0402, 0.0509, 0.0454], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:50:11,430 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-28 19:50:16,632 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=72982.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:50:18,549 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-28 19:50:41,094 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.639e+02 2.549e+02 2.878e+02 3.342e+02 6.835e+02, threshold=5.756e+02, percent-clipped=2.0 2023-04-28 19:50:43,027 INFO [train.py:904] (4/8) Epoch 8, batch 1950, loss[loss=0.2324, simple_loss=0.2995, pruned_loss=0.08266, over 16720.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2782, pruned_loss=0.05869, over 3324055.74 frames. ], batch size: 124, lr: 8.74e-03, grad_scale: 4.0 2023-04-28 19:51:14,274 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=73025.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 19:51:21,211 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=73030.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:51:24,820 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6739, 4.0298, 4.2942, 3.0563, 3.7530, 4.2517, 3.8928, 2.7944], device='cuda:4'), covar=tensor([0.0324, 0.0043, 0.0022, 0.0226, 0.0067, 0.0044, 0.0048, 0.0260], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0066, 0.0063, 0.0117, 0.0068, 0.0078, 0.0070, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:51:36,875 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4096, 1.7877, 2.6263, 3.2716, 3.0261, 3.5922, 2.0176, 3.4005], device='cuda:4'), covar=tensor([0.0078, 0.0310, 0.0167, 0.0120, 0.0146, 0.0082, 0.0326, 0.0075], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0159, 0.0145, 0.0146, 0.0152, 0.0107, 0.0159, 0.0099], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 19:51:46,654 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-28 19:51:50,195 INFO [train.py:904] (4/8) Epoch 8, batch 2000, loss[loss=0.2241, simple_loss=0.2983, pruned_loss=0.07499, over 16735.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2777, pruned_loss=0.05774, over 3327314.00 frames. ], batch size: 124, lr: 8.74e-03, grad_scale: 8.0 2023-04-28 19:52:27,207 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0146, 5.4414, 5.5606, 5.3784, 5.3878, 5.9806, 5.5807, 5.2987], device='cuda:4'), covar=tensor([0.0756, 0.1516, 0.1486, 0.1812, 0.2218, 0.0764, 0.1007, 0.1866], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0460, 0.0478, 0.0407, 0.0529, 0.0507, 0.0380, 0.0539], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:52:58,780 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.586e+02 2.718e+02 3.255e+02 3.787e+02 1.366e+03, threshold=6.510e+02, percent-clipped=4.0 2023-04-28 19:53:00,013 INFO [train.py:904] (4/8) Epoch 8, batch 2050, loss[loss=0.2025, simple_loss=0.2737, pruned_loss=0.06561, over 16746.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2775, pruned_loss=0.05806, over 3329729.05 frames. ], batch size: 134, lr: 8.73e-03, grad_scale: 8.0 2023-04-28 19:53:34,970 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4125, 3.2758, 3.4251, 2.8525, 3.3542, 2.0701, 2.9869, 2.6747], device='cuda:4'), covar=tensor([0.0105, 0.0088, 0.0137, 0.0188, 0.0074, 0.1721, 0.0116, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0105, 0.0157, 0.0151, 0.0121, 0.0169, 0.0142, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:53:45,911 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4091, 4.3143, 4.7757, 4.7847, 4.8136, 4.4500, 4.4959, 4.2938], device='cuda:4'), covar=tensor([0.0299, 0.0404, 0.0389, 0.0415, 0.0391, 0.0318, 0.0793, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0304, 0.0310, 0.0311, 0.0296, 0.0350, 0.0330, 0.0432, 0.0264], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 19:53:54,663 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6667, 3.8293, 4.0331, 3.0060, 3.6236, 4.1093, 3.8105, 2.5662], device='cuda:4'), covar=tensor([0.0331, 0.0121, 0.0030, 0.0217, 0.0065, 0.0053, 0.0046, 0.0279], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0067, 0.0064, 0.0119, 0.0069, 0.0080, 0.0072, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 19:54:09,330 INFO [train.py:904] (4/8) Epoch 8, batch 2100, loss[loss=0.2123, simple_loss=0.2984, pruned_loss=0.06311, over 16712.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2794, pruned_loss=0.05995, over 3320677.01 frames. ], batch size: 57, lr: 8.73e-03, grad_scale: 8.0 2023-04-28 19:55:14,761 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.921e+02 2.680e+02 3.272e+02 3.798e+02 7.634e+02, threshold=6.545e+02, percent-clipped=2.0 2023-04-28 19:55:16,487 INFO [train.py:904] (4/8) Epoch 8, batch 2150, loss[loss=0.2333, simple_loss=0.3009, pruned_loss=0.08287, over 16667.00 frames. ], tot_loss[loss=0.2007, simple_loss=0.28, pruned_loss=0.06066, over 3324129.37 frames. ], batch size: 134, lr: 8.73e-03, grad_scale: 8.0 2023-04-28 19:55:31,773 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1284, 4.5040, 3.3028, 2.5126, 3.1367, 2.5987, 4.7610, 4.1194], device='cuda:4'), covar=tensor([0.2054, 0.0538, 0.1282, 0.1976, 0.2193, 0.1527, 0.0297, 0.0764], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0255, 0.0275, 0.0262, 0.0279, 0.0213, 0.0254, 0.0286], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:56:08,706 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-28 19:56:21,729 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=73250.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:56:23,750 INFO [train.py:904] (4/8) Epoch 8, batch 2200, loss[loss=0.1861, simple_loss=0.2626, pruned_loss=0.05477, over 16812.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.2814, pruned_loss=0.0614, over 3319015.50 frames. ], batch size: 102, lr: 8.73e-03, grad_scale: 8.0 2023-04-28 19:56:46,945 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-28 19:57:12,040 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3424, 3.4651, 1.7988, 3.5857, 2.4793, 3.5744, 2.0017, 2.7376], device='cuda:4'), covar=tensor([0.0181, 0.0304, 0.1359, 0.0163, 0.0711, 0.0420, 0.1198, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0164, 0.0182, 0.0108, 0.0165, 0.0202, 0.0191, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 19:57:20,304 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=73293.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:57:29,787 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.707e+02 2.689e+02 3.130e+02 3.780e+02 7.244e+02, threshold=6.259e+02, percent-clipped=1.0 2023-04-28 19:57:31,974 INFO [train.py:904] (4/8) Epoch 8, batch 2250, loss[loss=0.1784, simple_loss=0.2563, pruned_loss=0.05027, over 16822.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.2808, pruned_loss=0.06083, over 3323565.67 frames. ], batch size: 39, lr: 8.72e-03, grad_scale: 8.0 2023-04-28 19:57:56,967 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=73320.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 19:58:40,764 INFO [train.py:904] (4/8) Epoch 8, batch 2300, loss[loss=0.1634, simple_loss=0.2543, pruned_loss=0.03624, over 16765.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.282, pruned_loss=0.06149, over 3309993.76 frames. ], batch size: 39, lr: 8.72e-03, grad_scale: 8.0 2023-04-28 19:58:41,765 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-28 19:58:44,245 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=73354.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 19:58:56,750 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0985, 1.7296, 2.3059, 2.7909, 2.7417, 3.3090, 2.2194, 3.2521], device='cuda:4'), covar=tensor([0.0103, 0.0279, 0.0187, 0.0151, 0.0147, 0.0084, 0.0226, 0.0077], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0160, 0.0146, 0.0149, 0.0154, 0.0108, 0.0160, 0.0100], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 19:59:28,553 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4858, 2.3725, 2.1360, 2.2862, 2.8029, 2.4845, 3.3763, 2.9802], device='cuda:4'), covar=tensor([0.0060, 0.0271, 0.0301, 0.0289, 0.0175, 0.0261, 0.0131, 0.0172], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0185, 0.0181, 0.0181, 0.0183, 0.0185, 0.0185, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 19:59:48,860 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.976e+02 2.910e+02 3.420e+02 4.127e+02 1.105e+03, threshold=6.841e+02, percent-clipped=4.0 2023-04-28 19:59:49,986 INFO [train.py:904] (4/8) Epoch 8, batch 2350, loss[loss=0.1999, simple_loss=0.2895, pruned_loss=0.05518, over 17120.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2818, pruned_loss=0.06103, over 3315201.76 frames. ], batch size: 49, lr: 8.72e-03, grad_scale: 8.0 2023-04-28 20:00:58,115 INFO [train.py:904] (4/8) Epoch 8, batch 2400, loss[loss=0.19, simple_loss=0.2781, pruned_loss=0.05093, over 16809.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2826, pruned_loss=0.06116, over 3317222.04 frames. ], batch size: 42, lr: 8.71e-03, grad_scale: 8.0 2023-04-28 20:02:04,599 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.977e+02 2.637e+02 3.145e+02 3.743e+02 9.455e+02, threshold=6.290e+02, percent-clipped=3.0 2023-04-28 20:02:05,796 INFO [train.py:904] (4/8) Epoch 8, batch 2450, loss[loss=0.2256, simple_loss=0.2969, pruned_loss=0.0771, over 16907.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2833, pruned_loss=0.0607, over 3314885.16 frames. ], batch size: 116, lr: 8.71e-03, grad_scale: 8.0 2023-04-28 20:02:17,856 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7470, 4.0503, 2.1009, 4.3580, 2.9117, 4.3845, 2.1671, 3.0632], device='cuda:4'), covar=tensor([0.0179, 0.0253, 0.1443, 0.0118, 0.0718, 0.0319, 0.1371, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0166, 0.0183, 0.0109, 0.0166, 0.0203, 0.0192, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 20:03:12,601 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=73550.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:03:14,450 INFO [train.py:904] (4/8) Epoch 8, batch 2500, loss[loss=0.2001, simple_loss=0.2627, pruned_loss=0.06876, over 16749.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2826, pruned_loss=0.06026, over 3318014.23 frames. ], batch size: 83, lr: 8.71e-03, grad_scale: 8.0 2023-04-28 20:04:07,463 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=73590.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:04:18,852 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=73598.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:04:22,383 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.774e+02 3.514e+02 4.177e+02 6.912e+02, threshold=7.028e+02, percent-clipped=4.0 2023-04-28 20:04:23,440 INFO [train.py:904] (4/8) Epoch 8, batch 2550, loss[loss=0.2292, simple_loss=0.3184, pruned_loss=0.07, over 16664.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2826, pruned_loss=0.06038, over 3310919.36 frames. ], batch size: 62, lr: 8.70e-03, grad_scale: 8.0 2023-04-28 20:04:38,287 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4494, 4.4204, 4.5780, 4.4510, 4.3865, 5.0408, 4.6585, 4.2860], device='cuda:4'), covar=tensor([0.1484, 0.1775, 0.1471, 0.1984, 0.2835, 0.1057, 0.1332, 0.2501], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0461, 0.0479, 0.0407, 0.0536, 0.0513, 0.0384, 0.0542], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:04:49,504 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=73620.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 20:05:29,133 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=73649.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:05:32,299 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=73651.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:05:32,975 INFO [train.py:904] (4/8) Epoch 8, batch 2600, loss[loss=0.1559, simple_loss=0.2474, pruned_loss=0.03216, over 16908.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2813, pruned_loss=0.05946, over 3321557.70 frames. ], batch size: 42, lr: 8.70e-03, grad_scale: 8.0 2023-04-28 20:05:55,317 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=73668.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 20:06:22,519 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4603, 3.6874, 3.7415, 1.7588, 3.9538, 3.9333, 3.1017, 2.7617], device='cuda:4'), covar=tensor([0.0691, 0.0102, 0.0123, 0.1104, 0.0060, 0.0100, 0.0309, 0.0448], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0095, 0.0084, 0.0138, 0.0070, 0.0093, 0.0120, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 20:06:43,000 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.895e+02 2.639e+02 3.290e+02 4.275e+02 7.096e+02, threshold=6.579e+02, percent-clipped=1.0 2023-04-28 20:06:43,015 INFO [train.py:904] (4/8) Epoch 8, batch 2650, loss[loss=0.1936, simple_loss=0.2823, pruned_loss=0.05247, over 16488.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2823, pruned_loss=0.05975, over 3326893.07 frames. ], batch size: 75, lr: 8.70e-03, grad_scale: 4.0 2023-04-28 20:07:23,279 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=73732.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:07:50,960 INFO [train.py:904] (4/8) Epoch 8, batch 2700, loss[loss=0.2035, simple_loss=0.2803, pruned_loss=0.06335, over 16278.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2822, pruned_loss=0.05903, over 3330523.70 frames. ], batch size: 165, lr: 8.70e-03, grad_scale: 4.0 2023-04-28 20:08:19,292 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6885, 4.9950, 4.7073, 4.7844, 4.5295, 4.4127, 4.4867, 5.0435], device='cuda:4'), covar=tensor([0.0930, 0.0822, 0.0954, 0.0588, 0.0667, 0.1010, 0.0811, 0.0804], device='cuda:4'), in_proj_covar=tensor([0.0495, 0.0633, 0.0523, 0.0422, 0.0395, 0.0409, 0.0522, 0.0469], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:08:46,256 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=73793.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:08:57,113 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.671e+02 2.547e+02 3.040e+02 3.898e+02 5.528e+02, threshold=6.080e+02, percent-clipped=0.0 2023-04-28 20:08:57,128 INFO [train.py:904] (4/8) Epoch 8, batch 2750, loss[loss=0.1954, simple_loss=0.2819, pruned_loss=0.05448, over 16765.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2827, pruned_loss=0.05921, over 3336416.26 frames. ], batch size: 89, lr: 8.69e-03, grad_scale: 4.0 2023-04-28 20:10:05,127 INFO [train.py:904] (4/8) Epoch 8, batch 2800, loss[loss=0.2071, simple_loss=0.2885, pruned_loss=0.06291, over 16504.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2829, pruned_loss=0.05938, over 3322216.02 frames. ], batch size: 68, lr: 8.69e-03, grad_scale: 8.0 2023-04-28 20:10:14,473 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0771, 4.7903, 5.0326, 5.2668, 5.4764, 4.7753, 5.4483, 5.3715], device='cuda:4'), covar=tensor([0.1341, 0.0919, 0.1389, 0.0542, 0.0403, 0.0692, 0.0371, 0.0449], device='cuda:4'), in_proj_covar=tensor([0.0505, 0.0622, 0.0774, 0.0624, 0.0474, 0.0481, 0.0490, 0.0542], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:10:32,793 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 20:11:14,609 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.850e+02 2.765e+02 3.378e+02 4.166e+02 1.030e+03, threshold=6.755e+02, percent-clipped=2.0 2023-04-28 20:11:14,626 INFO [train.py:904] (4/8) Epoch 8, batch 2850, loss[loss=0.1603, simple_loss=0.2399, pruned_loss=0.0403, over 16785.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2817, pruned_loss=0.05875, over 3331603.87 frames. ], batch size: 39, lr: 8.69e-03, grad_scale: 8.0 2023-04-28 20:12:16,873 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=73946.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:12:20,702 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=73949.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:12:23,979 INFO [train.py:904] (4/8) Epoch 8, batch 2900, loss[loss=0.1962, simple_loss=0.2666, pruned_loss=0.06289, over 16724.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2807, pruned_loss=0.05932, over 3329250.94 frames. ], batch size: 89, lr: 8.68e-03, grad_scale: 8.0 2023-04-28 20:12:35,680 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8047, 4.8259, 5.0212, 4.9057, 4.8351, 5.4852, 5.1456, 4.8145], device='cuda:4'), covar=tensor([0.1147, 0.1912, 0.1616, 0.1842, 0.2923, 0.1008, 0.1293, 0.2589], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0469, 0.0489, 0.0416, 0.0546, 0.0517, 0.0394, 0.0555], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:12:55,903 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-28 20:13:28,886 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=73997.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:13:38,233 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.885e+02 2.787e+02 3.423e+02 3.955e+02 9.086e+02, threshold=6.847e+02, percent-clipped=3.0 2023-04-28 20:13:38,248 INFO [train.py:904] (4/8) Epoch 8, batch 2950, loss[loss=0.1737, simple_loss=0.2499, pruned_loss=0.04877, over 16140.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2798, pruned_loss=0.05982, over 3322108.55 frames. ], batch size: 35, lr: 8.68e-03, grad_scale: 8.0 2023-04-28 20:13:45,535 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8259, 5.1421, 4.8618, 4.8923, 4.5790, 4.6138, 4.6773, 5.2218], device='cuda:4'), covar=tensor([0.0935, 0.0743, 0.0942, 0.0603, 0.0790, 0.0900, 0.0827, 0.0799], device='cuda:4'), in_proj_covar=tensor([0.0492, 0.0629, 0.0518, 0.0421, 0.0394, 0.0405, 0.0518, 0.0465], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:14:46,251 INFO [train.py:904] (4/8) Epoch 8, batch 3000, loss[loss=0.1765, simple_loss=0.2612, pruned_loss=0.04592, over 17006.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2801, pruned_loss=0.06037, over 3329050.25 frames. ], batch size: 41, lr: 8.68e-03, grad_scale: 8.0 2023-04-28 20:14:46,251 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 20:14:55,853 INFO [train.py:938] (4/8) Epoch 8, validation: loss=0.1462, simple_loss=0.2525, pruned_loss=0.01995, over 944034.00 frames. 2023-04-28 20:14:55,854 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 20:15:13,370 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74064.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:15:45,910 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=74088.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:16:02,275 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 20:16:06,763 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.517e+02 2.715e+02 3.274e+02 3.889e+02 8.736e+02, threshold=6.548e+02, percent-clipped=1.0 2023-04-28 20:16:06,779 INFO [train.py:904] (4/8) Epoch 8, batch 3050, loss[loss=0.2268, simple_loss=0.2916, pruned_loss=0.08101, over 16234.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2797, pruned_loss=0.06031, over 3334620.55 frames. ], batch size: 165, lr: 8.68e-03, grad_scale: 8.0 2023-04-28 20:16:24,821 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.64 vs. limit=5.0 2023-04-28 20:16:32,701 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6362, 3.2915, 2.7057, 4.9018, 4.1113, 4.7019, 1.4774, 3.2612], device='cuda:4'), covar=tensor([0.1444, 0.0595, 0.1150, 0.0199, 0.0349, 0.0344, 0.1577, 0.0731], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0153, 0.0174, 0.0121, 0.0200, 0.0207, 0.0172, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 20:16:38,569 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=74125.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:16:52,853 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4401, 4.0216, 3.9873, 2.0144, 3.1284, 2.4277, 3.7676, 3.6773], device='cuda:4'), covar=tensor([0.0246, 0.0546, 0.0425, 0.1570, 0.0689, 0.0885, 0.0624, 0.1013], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0140, 0.0155, 0.0139, 0.0135, 0.0123, 0.0136, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 20:16:53,828 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2929, 3.3130, 3.4678, 2.4927, 3.1128, 3.5097, 3.2670, 2.0193], device='cuda:4'), covar=tensor([0.0299, 0.0066, 0.0028, 0.0214, 0.0069, 0.0053, 0.0056, 0.0289], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0065, 0.0063, 0.0116, 0.0070, 0.0079, 0.0072, 0.0112], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:17:06,436 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0578, 5.3736, 5.0942, 5.1425, 4.7594, 4.7738, 4.8758, 5.4820], device='cuda:4'), covar=tensor([0.0912, 0.0808, 0.0899, 0.0595, 0.0732, 0.0753, 0.0960, 0.0767], device='cuda:4'), in_proj_covar=tensor([0.0494, 0.0632, 0.0518, 0.0422, 0.0395, 0.0407, 0.0526, 0.0467], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:17:13,363 INFO [train.py:904] (4/8) Epoch 8, batch 3100, loss[loss=0.2159, simple_loss=0.3006, pruned_loss=0.0656, over 16699.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2789, pruned_loss=0.05972, over 3331279.08 frames. ], batch size: 57, lr: 8.67e-03, grad_scale: 8.0 2023-04-28 20:18:08,629 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2302, 4.2090, 4.4445, 1.8615, 4.7075, 4.6239, 3.2487, 3.4319], device='cuda:4'), covar=tensor([0.0491, 0.0105, 0.0126, 0.1102, 0.0040, 0.0091, 0.0319, 0.0361], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0094, 0.0083, 0.0135, 0.0069, 0.0093, 0.0117, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 20:18:21,290 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.831e+02 2.546e+02 2.917e+02 3.424e+02 8.256e+02, threshold=5.834e+02, percent-clipped=1.0 2023-04-28 20:18:21,306 INFO [train.py:904] (4/8) Epoch 8, batch 3150, loss[loss=0.1831, simple_loss=0.2552, pruned_loss=0.05545, over 15927.00 frames. ], tot_loss[loss=0.198, simple_loss=0.278, pruned_loss=0.05899, over 3329598.94 frames. ], batch size: 35, lr: 8.67e-03, grad_scale: 8.0 2023-04-28 20:18:23,660 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0938, 4.8366, 5.0860, 5.3432, 5.5163, 4.8829, 5.4633, 5.4325], device='cuda:4'), covar=tensor([0.1400, 0.0971, 0.1558, 0.0588, 0.0420, 0.0684, 0.0439, 0.0433], device='cuda:4'), in_proj_covar=tensor([0.0501, 0.0614, 0.0768, 0.0625, 0.0470, 0.0471, 0.0488, 0.0535], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:19:15,050 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74240.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:19:23,403 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=74246.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:19:23,965 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.60 vs. limit=2.0 2023-04-28 20:19:32,368 INFO [train.py:904] (4/8) Epoch 8, batch 3200, loss[loss=0.1844, simple_loss=0.2679, pruned_loss=0.0505, over 17230.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2748, pruned_loss=0.05697, over 3333530.10 frames. ], batch size: 45, lr: 8.67e-03, grad_scale: 8.0 2023-04-28 20:19:54,235 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1583, 4.1375, 4.0909, 3.6200, 4.1520, 1.8224, 3.9661, 3.8565], device='cuda:4'), covar=tensor([0.0078, 0.0071, 0.0111, 0.0261, 0.0063, 0.1973, 0.0096, 0.0143], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0107, 0.0159, 0.0154, 0.0123, 0.0167, 0.0145, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:20:09,798 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1191, 5.5735, 5.7152, 5.5430, 5.4563, 6.0767, 5.7194, 5.3761], device='cuda:4'), covar=tensor([0.0771, 0.1624, 0.1856, 0.1930, 0.2918, 0.0998, 0.1269, 0.2444], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0472, 0.0486, 0.0412, 0.0544, 0.0515, 0.0390, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:20:17,983 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7002, 4.8043, 5.2400, 5.2526, 5.2303, 4.9018, 4.8523, 4.7332], device='cuda:4'), covar=tensor([0.0282, 0.0406, 0.0389, 0.0445, 0.0386, 0.0329, 0.0829, 0.0327], device='cuda:4'), in_proj_covar=tensor([0.0300, 0.0306, 0.0305, 0.0295, 0.0345, 0.0324, 0.0430, 0.0261], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 20:20:30,670 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=74294.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:20:40,631 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=74301.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:20:41,303 INFO [train.py:904] (4/8) Epoch 8, batch 3250, loss[loss=0.1913, simple_loss=0.2805, pruned_loss=0.05109, over 17136.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2752, pruned_loss=0.05738, over 3327284.91 frames. ], batch size: 48, lr: 8.66e-03, grad_scale: 4.0 2023-04-28 20:20:42,359 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.676e+02 2.591e+02 3.147e+02 3.802e+02 9.622e+02, threshold=6.293e+02, percent-clipped=5.0 2023-04-28 20:21:31,959 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9153, 2.0622, 2.2674, 3.2185, 2.0717, 2.3532, 2.2225, 2.1665], device='cuda:4'), covar=tensor([0.0776, 0.2257, 0.1323, 0.0400, 0.2763, 0.1464, 0.2223, 0.2254], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0367, 0.0308, 0.0328, 0.0398, 0.0414, 0.0330, 0.0439], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:21:49,532 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-28 20:21:52,434 INFO [train.py:904] (4/8) Epoch 8, batch 3300, loss[loss=0.2509, simple_loss=0.3172, pruned_loss=0.09227, over 16258.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2772, pruned_loss=0.05835, over 3315753.52 frames. ], batch size: 165, lr: 8.66e-03, grad_scale: 4.0 2023-04-28 20:22:41,317 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=74388.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:23:01,500 INFO [train.py:904] (4/8) Epoch 8, batch 3350, loss[loss=0.1872, simple_loss=0.2639, pruned_loss=0.05524, over 16726.00 frames. ], tot_loss[loss=0.1981, simple_loss=0.278, pruned_loss=0.05913, over 3306045.75 frames. ], batch size: 89, lr: 8.66e-03, grad_scale: 4.0 2023-04-28 20:23:02,760 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.835e+02 2.621e+02 3.247e+02 4.157e+02 8.305e+02, threshold=6.494e+02, percent-clipped=1.0 2023-04-28 20:23:27,944 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=74420.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:23:46,739 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6753, 4.7100, 4.9407, 4.8057, 4.7659, 5.3792, 4.9689, 4.5881], device='cuda:4'), covar=tensor([0.1204, 0.1753, 0.1623, 0.1802, 0.2428, 0.0939, 0.1308, 0.2447], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0468, 0.0485, 0.0410, 0.0540, 0.0515, 0.0389, 0.0546], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:23:49,803 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=74436.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:24:11,181 INFO [train.py:904] (4/8) Epoch 8, batch 3400, loss[loss=0.1964, simple_loss=0.2851, pruned_loss=0.05388, over 16613.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2783, pruned_loss=0.05904, over 3309859.50 frames. ], batch size: 62, lr: 8.66e-03, grad_scale: 4.0 2023-04-28 20:24:30,175 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1478, 4.5879, 3.2972, 2.4169, 3.0542, 2.3625, 4.6963, 3.9838], device='cuda:4'), covar=tensor([0.2084, 0.0515, 0.1280, 0.2120, 0.2384, 0.1821, 0.0303, 0.0920], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0256, 0.0277, 0.0265, 0.0284, 0.0212, 0.0256, 0.0287], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:25:14,001 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2133, 5.8175, 5.9239, 5.6059, 5.7131, 6.2264, 5.8754, 5.4873], device='cuda:4'), covar=tensor([0.0704, 0.1654, 0.1481, 0.1921, 0.2440, 0.0927, 0.1050, 0.2066], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0467, 0.0482, 0.0410, 0.0541, 0.0514, 0.0386, 0.0545], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:25:21,715 INFO [train.py:904] (4/8) Epoch 8, batch 3450, loss[loss=0.1809, simple_loss=0.2725, pruned_loss=0.04468, over 17146.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.277, pruned_loss=0.05828, over 3319731.43 frames. ], batch size: 47, lr: 8.65e-03, grad_scale: 4.0 2023-04-28 20:25:22,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.895e+02 2.611e+02 3.016e+02 3.656e+02 6.729e+02, threshold=6.032e+02, percent-clipped=3.0 2023-04-28 20:26:00,331 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7206, 3.4918, 2.8606, 5.1895, 4.4274, 4.7857, 1.6665, 3.4583], device='cuda:4'), covar=tensor([0.1353, 0.0544, 0.1009, 0.0100, 0.0256, 0.0303, 0.1403, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0153, 0.0173, 0.0121, 0.0202, 0.0207, 0.0170, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 20:26:30,776 INFO [train.py:904] (4/8) Epoch 8, batch 3500, loss[loss=0.1744, simple_loss=0.2524, pruned_loss=0.04822, over 16820.00 frames. ], tot_loss[loss=0.1952, simple_loss=0.2751, pruned_loss=0.05763, over 3309113.49 frames. ], batch size: 102, lr: 8.65e-03, grad_scale: 4.0 2023-04-28 20:26:53,201 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8698, 5.2454, 4.9067, 5.0244, 4.6896, 4.6136, 4.6938, 5.2559], device='cuda:4'), covar=tensor([0.0920, 0.0770, 0.0996, 0.0574, 0.0696, 0.0958, 0.0833, 0.0805], device='cuda:4'), in_proj_covar=tensor([0.0495, 0.0630, 0.0523, 0.0426, 0.0396, 0.0412, 0.0522, 0.0466], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:27:32,521 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74595.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:27:33,589 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=74596.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:27:42,853 INFO [train.py:904] (4/8) Epoch 8, batch 3550, loss[loss=0.1989, simple_loss=0.2723, pruned_loss=0.06273, over 16874.00 frames. ], tot_loss[loss=0.1949, simple_loss=0.2745, pruned_loss=0.05768, over 3311249.96 frames. ], batch size: 90, lr: 8.65e-03, grad_scale: 4.0 2023-04-28 20:27:43,961 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.810e+02 2.461e+02 3.024e+02 3.861e+02 7.667e+02, threshold=6.049e+02, percent-clipped=4.0 2023-04-28 20:28:19,263 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-28 20:28:51,888 INFO [train.py:904] (4/8) Epoch 8, batch 3600, loss[loss=0.2336, simple_loss=0.2991, pruned_loss=0.084, over 11859.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2737, pruned_loss=0.05679, over 3312671.70 frames. ], batch size: 246, lr: 8.64e-03, grad_scale: 8.0 2023-04-28 20:28:56,946 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=74656.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:29:21,947 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-28 20:30:00,903 INFO [train.py:904] (4/8) Epoch 8, batch 3650, loss[loss=0.1646, simple_loss=0.2303, pruned_loss=0.04943, over 16805.00 frames. ], tot_loss[loss=0.1928, simple_loss=0.2718, pruned_loss=0.05686, over 3311240.12 frames. ], batch size: 83, lr: 8.64e-03, grad_scale: 8.0 2023-04-28 20:30:02,111 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.583e+02 2.393e+02 2.913e+02 3.939e+02 7.282e+02, threshold=5.826e+02, percent-clipped=2.0 2023-04-28 20:30:27,863 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=74720.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:30:33,665 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7953, 1.2855, 1.6363, 1.5989, 1.7553, 1.8458, 1.4861, 1.6540], device='cuda:4'), covar=tensor([0.0134, 0.0216, 0.0123, 0.0143, 0.0148, 0.0105, 0.0210, 0.0063], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0159, 0.0145, 0.0149, 0.0155, 0.0111, 0.0159, 0.0102], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 20:31:13,893 INFO [train.py:904] (4/8) Epoch 8, batch 3700, loss[loss=0.198, simple_loss=0.2671, pruned_loss=0.06447, over 16282.00 frames. ], tot_loss[loss=0.1949, simple_loss=0.2715, pruned_loss=0.05913, over 3283696.86 frames. ], batch size: 165, lr: 8.64e-03, grad_scale: 8.0 2023-04-28 20:31:27,995 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2648, 4.2208, 4.2671, 3.8366, 4.2508, 1.9066, 4.1087, 4.0347], device='cuda:4'), covar=tensor([0.0083, 0.0069, 0.0098, 0.0228, 0.0068, 0.1920, 0.0100, 0.0146], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0108, 0.0159, 0.0154, 0.0124, 0.0168, 0.0144, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:31:38,734 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=74768.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:31:55,008 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2941, 3.7023, 3.6232, 1.6895, 3.7426, 3.7769, 3.0957, 2.7567], device='cuda:4'), covar=tensor([0.0872, 0.0096, 0.0152, 0.1140, 0.0079, 0.0095, 0.0330, 0.0492], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0096, 0.0085, 0.0138, 0.0070, 0.0093, 0.0119, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 20:31:56,937 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9145, 2.6166, 2.4469, 1.9470, 2.4565, 2.7011, 2.5693, 1.8420], device='cuda:4'), covar=tensor([0.0289, 0.0064, 0.0041, 0.0244, 0.0078, 0.0079, 0.0059, 0.0279], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0066, 0.0064, 0.0118, 0.0069, 0.0079, 0.0070, 0.0112], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:32:29,723 INFO [train.py:904] (4/8) Epoch 8, batch 3750, loss[loss=0.2092, simple_loss=0.2911, pruned_loss=0.06363, over 16773.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.272, pruned_loss=0.06049, over 3260557.15 frames. ], batch size: 39, lr: 8.64e-03, grad_scale: 8.0 2023-04-28 20:32:30,693 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.592e+02 2.639e+02 3.130e+02 3.703e+02 1.006e+03, threshold=6.260e+02, percent-clipped=3.0 2023-04-28 20:32:40,698 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7054, 4.6449, 4.6186, 4.4456, 4.2308, 4.6322, 4.4623, 4.4197], device='cuda:4'), covar=tensor([0.0511, 0.0436, 0.0227, 0.0219, 0.0816, 0.0369, 0.0413, 0.0527], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0277, 0.0274, 0.0244, 0.0306, 0.0278, 0.0187, 0.0312], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:33:03,951 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1568, 4.0766, 4.2522, 4.1614, 4.1904, 4.7103, 4.3425, 3.9789], device='cuda:4'), covar=tensor([0.1654, 0.1863, 0.1663, 0.1793, 0.2489, 0.1075, 0.1255, 0.2283], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0463, 0.0476, 0.0402, 0.0528, 0.0506, 0.0384, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:33:22,781 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 20:33:41,191 INFO [train.py:904] (4/8) Epoch 8, batch 3800, loss[loss=0.2001, simple_loss=0.2807, pruned_loss=0.0597, over 16681.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2745, pruned_loss=0.06224, over 3254170.17 frames. ], batch size: 89, lr: 8.63e-03, grad_scale: 8.0 2023-04-28 20:33:52,210 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8952, 4.8157, 4.7166, 4.1059, 4.8344, 1.7195, 4.5854, 4.5761], device='cuda:4'), covar=tensor([0.0079, 0.0067, 0.0123, 0.0326, 0.0061, 0.2305, 0.0095, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0108, 0.0159, 0.0154, 0.0124, 0.0169, 0.0144, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:34:23,575 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7312, 4.1522, 3.9268, 2.2106, 3.1001, 2.6963, 4.2627, 3.9505], device='cuda:4'), covar=tensor([0.0179, 0.0454, 0.0520, 0.1540, 0.0741, 0.0822, 0.0469, 0.0765], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0142, 0.0155, 0.0140, 0.0135, 0.0125, 0.0136, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 20:34:45,388 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=74896.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:34:51,948 INFO [train.py:904] (4/8) Epoch 8, batch 3850, loss[loss=0.1856, simple_loss=0.2581, pruned_loss=0.05654, over 16489.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2734, pruned_loss=0.06197, over 3268352.64 frames. ], batch size: 68, lr: 8.63e-03, grad_scale: 8.0 2023-04-28 20:34:53,139 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.924e+02 2.503e+02 3.025e+02 3.649e+02 5.657e+02, threshold=6.049e+02, percent-clipped=0.0 2023-04-28 20:35:03,410 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74909.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:35:04,707 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5940, 5.5137, 5.2397, 4.7310, 5.4960, 2.0042, 5.2106, 5.2884], device='cuda:4'), covar=tensor([0.0039, 0.0033, 0.0094, 0.0260, 0.0040, 0.2076, 0.0072, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0108, 0.0159, 0.0154, 0.0124, 0.0169, 0.0144, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:35:05,112 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-28 20:35:17,223 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3284, 3.4620, 3.6983, 2.6943, 3.3004, 3.7775, 3.5696, 2.2170], device='cuda:4'), covar=tensor([0.0319, 0.0075, 0.0032, 0.0214, 0.0055, 0.0056, 0.0048, 0.0278], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0065, 0.0064, 0.0118, 0.0069, 0.0078, 0.0070, 0.0112], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:35:52,592 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=74944.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:35:57,567 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.03 vs. limit=5.0 2023-04-28 20:36:01,484 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=74951.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:36:02,293 INFO [train.py:904] (4/8) Epoch 8, batch 3900, loss[loss=0.1966, simple_loss=0.2685, pruned_loss=0.06232, over 16715.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2734, pruned_loss=0.06252, over 3265800.67 frames. ], batch size: 134, lr: 8.63e-03, grad_scale: 8.0 2023-04-28 20:36:20,347 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74964.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:36:28,138 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=74970.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:37:10,396 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4152, 4.3534, 4.3223, 3.8573, 4.3679, 1.8453, 4.1463, 4.0987], device='cuda:4'), covar=tensor([0.0082, 0.0079, 0.0108, 0.0284, 0.0076, 0.2061, 0.0110, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0107, 0.0158, 0.0152, 0.0123, 0.0168, 0.0143, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:37:12,342 INFO [train.py:904] (4/8) Epoch 8, batch 3950, loss[loss=0.2117, simple_loss=0.2857, pruned_loss=0.06891, over 12800.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2729, pruned_loss=0.06308, over 3253515.14 frames. ], batch size: 247, lr: 8.62e-03, grad_scale: 8.0 2023-04-28 20:37:14,091 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.920e+02 2.648e+02 3.131e+02 3.735e+02 8.073e+02, threshold=6.262e+02, percent-clipped=3.0 2023-04-28 20:37:46,124 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=75025.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:38:23,629 INFO [train.py:904] (4/8) Epoch 8, batch 4000, loss[loss=0.2203, simple_loss=0.2959, pruned_loss=0.07235, over 12418.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2726, pruned_loss=0.063, over 3249341.00 frames. ], batch size: 247, lr: 8.62e-03, grad_scale: 8.0 2023-04-28 20:39:00,488 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7207, 3.0789, 2.4716, 5.0335, 4.0508, 4.4360, 1.8460, 2.8595], device='cuda:4'), covar=tensor([0.1241, 0.0632, 0.1176, 0.0093, 0.0418, 0.0309, 0.1260, 0.0915], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0154, 0.0173, 0.0121, 0.0204, 0.0209, 0.0172, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 20:39:36,987 INFO [train.py:904] (4/8) Epoch 8, batch 4050, loss[loss=0.1925, simple_loss=0.2756, pruned_loss=0.05473, over 15394.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.272, pruned_loss=0.06123, over 3268248.20 frames. ], batch size: 190, lr: 8.62e-03, grad_scale: 8.0 2023-04-28 20:39:38,165 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.540e+02 2.447e+02 2.780e+02 3.424e+02 6.417e+02, threshold=5.561e+02, percent-clipped=2.0 2023-04-28 20:40:49,054 INFO [train.py:904] (4/8) Epoch 8, batch 4100, loss[loss=0.2383, simple_loss=0.316, pruned_loss=0.08031, over 15335.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2738, pruned_loss=0.06083, over 3250707.59 frames. ], batch size: 190, lr: 8.62e-03, grad_scale: 8.0 2023-04-28 20:41:15,970 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-28 20:42:02,417 INFO [train.py:904] (4/8) Epoch 8, batch 4150, loss[loss=0.2153, simple_loss=0.2996, pruned_loss=0.06548, over 16528.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.282, pruned_loss=0.06441, over 3209583.07 frames. ], batch size: 68, lr: 8.61e-03, grad_scale: 8.0 2023-04-28 20:42:04,251 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.780e+02 2.372e+02 3.014e+02 3.829e+02 9.608e+02, threshold=6.029e+02, percent-clipped=8.0 2023-04-28 20:42:08,189 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 20:42:43,276 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=75228.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:43:01,187 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2833, 1.7265, 2.5245, 3.1402, 2.9304, 3.5250, 1.9332, 3.4060], device='cuda:4'), covar=tensor([0.0101, 0.0301, 0.0175, 0.0137, 0.0138, 0.0066, 0.0287, 0.0058], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0158, 0.0142, 0.0146, 0.0151, 0.0108, 0.0157, 0.0099], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 20:43:11,260 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2066, 4.0130, 4.2103, 4.3719, 4.4857, 4.0958, 4.4542, 4.4597], device='cuda:4'), covar=tensor([0.1034, 0.0924, 0.1210, 0.0543, 0.0421, 0.0974, 0.0456, 0.0466], device='cuda:4'), in_proj_covar=tensor([0.0472, 0.0578, 0.0720, 0.0587, 0.0444, 0.0447, 0.0452, 0.0503], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:43:19,258 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75251.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:43:20,233 INFO [train.py:904] (4/8) Epoch 8, batch 4200, loss[loss=0.2602, simple_loss=0.3439, pruned_loss=0.08821, over 15358.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.2898, pruned_loss=0.06673, over 3174289.23 frames. ], batch size: 190, lr: 8.61e-03, grad_scale: 8.0 2023-04-28 20:43:40,419 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=75265.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:43:48,106 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9855, 2.6996, 2.6212, 1.9182, 2.5343, 2.6893, 2.6249, 1.8515], device='cuda:4'), covar=tensor([0.0268, 0.0043, 0.0037, 0.0231, 0.0064, 0.0060, 0.0051, 0.0248], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0064, 0.0064, 0.0118, 0.0068, 0.0078, 0.0070, 0.0112], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:43:48,121 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=75270.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:43:52,332 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4052, 1.3706, 1.8577, 2.3059, 2.2574, 2.4437, 1.6354, 2.3712], device='cuda:4'), covar=tensor([0.0108, 0.0322, 0.0192, 0.0198, 0.0173, 0.0110, 0.0297, 0.0079], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0159, 0.0143, 0.0147, 0.0152, 0.0108, 0.0158, 0.0100], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 20:44:12,631 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3959, 1.9635, 1.6168, 1.8268, 2.3044, 2.0694, 2.5092, 2.5780], device='cuda:4'), covar=tensor([0.0088, 0.0281, 0.0338, 0.0317, 0.0184, 0.0274, 0.0115, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0183, 0.0180, 0.0181, 0.0182, 0.0184, 0.0182, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:44:16,431 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=75289.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:44:29,121 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6120, 4.8047, 4.9267, 4.8077, 4.8618, 5.4291, 4.9489, 4.6675], device='cuda:4'), covar=tensor([0.0879, 0.1548, 0.1318, 0.1669, 0.2339, 0.0842, 0.1267, 0.2208], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0449, 0.0461, 0.0391, 0.0516, 0.0492, 0.0375, 0.0529], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 20:44:30,803 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=75299.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:44:34,932 INFO [train.py:904] (4/8) Epoch 8, batch 4250, loss[loss=0.1839, simple_loss=0.2773, pruned_loss=0.04526, over 15224.00 frames. ], tot_loss[loss=0.2121, simple_loss=0.2926, pruned_loss=0.0658, over 3176737.70 frames. ], batch size: 190, lr: 8.61e-03, grad_scale: 8.0 2023-04-28 20:44:36,197 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.726e+02 2.663e+02 3.246e+02 3.785e+02 9.237e+02, threshold=6.492e+02, percent-clipped=4.0 2023-04-28 20:45:01,373 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=75320.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:45:12,827 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.8562, 6.1326, 5.9645, 6.0510, 5.6184, 5.1113, 5.7461, 6.3126], device='cuda:4'), covar=tensor([0.0720, 0.0633, 0.0688, 0.0436, 0.0635, 0.0545, 0.0558, 0.0594], device='cuda:4'), in_proj_covar=tensor([0.0468, 0.0590, 0.0498, 0.0402, 0.0377, 0.0391, 0.0494, 0.0439], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:45:18,431 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3219, 2.0066, 1.5576, 1.8826, 2.3349, 2.1373, 2.5242, 2.6324], device='cuda:4'), covar=tensor([0.0092, 0.0262, 0.0342, 0.0292, 0.0138, 0.0233, 0.0111, 0.0143], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0181, 0.0178, 0.0178, 0.0180, 0.0182, 0.0179, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:45:18,434 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=75331.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 20:45:29,411 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-28 20:45:48,478 INFO [train.py:904] (4/8) Epoch 8, batch 4300, loss[loss=0.2128, simple_loss=0.295, pruned_loss=0.06532, over 16728.00 frames. ], tot_loss[loss=0.2112, simple_loss=0.2929, pruned_loss=0.06476, over 3179659.27 frames. ], batch size: 124, lr: 8.60e-03, grad_scale: 8.0 2023-04-28 20:45:52,201 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0356, 3.6620, 3.4402, 1.8194, 2.7769, 2.2434, 3.3624, 3.4950], device='cuda:4'), covar=tensor([0.0244, 0.0496, 0.0484, 0.1695, 0.0766, 0.0861, 0.0651, 0.0925], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0135, 0.0152, 0.0137, 0.0132, 0.0121, 0.0133, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 20:46:54,398 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8012, 3.4333, 3.1594, 1.6504, 2.5675, 2.1015, 3.1808, 3.4306], device='cuda:4'), covar=tensor([0.0282, 0.0579, 0.0591, 0.1848, 0.0866, 0.0946, 0.0711, 0.0986], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0136, 0.0153, 0.0138, 0.0132, 0.0121, 0.0133, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 20:47:02,654 INFO [train.py:904] (4/8) Epoch 8, batch 4350, loss[loss=0.2323, simple_loss=0.3087, pruned_loss=0.07795, over 11808.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.2971, pruned_loss=0.06636, over 3177106.37 frames. ], batch size: 248, lr: 8.60e-03, grad_scale: 8.0 2023-04-28 20:47:03,854 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.683e+02 2.615e+02 3.065e+02 3.856e+02 8.729e+02, threshold=6.129e+02, percent-clipped=2.0 2023-04-28 20:47:55,624 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.77 vs. limit=5.0 2023-04-28 20:48:02,570 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6266, 5.5626, 5.2638, 4.8064, 5.6017, 1.9028, 5.3118, 5.3183], device='cuda:4'), covar=tensor([0.0034, 0.0027, 0.0080, 0.0207, 0.0028, 0.1899, 0.0056, 0.0085], device='cuda:4'), in_proj_covar=tensor([0.0114, 0.0101, 0.0150, 0.0146, 0.0117, 0.0161, 0.0136, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:48:17,344 INFO [train.py:904] (4/8) Epoch 8, batch 4400, loss[loss=0.2293, simple_loss=0.3173, pruned_loss=0.07064, over 16481.00 frames. ], tot_loss[loss=0.2164, simple_loss=0.2987, pruned_loss=0.06703, over 3185019.18 frames. ], batch size: 75, lr: 8.60e-03, grad_scale: 8.0 2023-04-28 20:48:37,344 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-28 20:48:59,102 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7759, 2.3151, 2.3151, 4.6623, 2.0210, 2.9052, 2.3741, 2.4674], device='cuda:4'), covar=tensor([0.0631, 0.2522, 0.1624, 0.0238, 0.3261, 0.1542, 0.2130, 0.2677], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0369, 0.0308, 0.0325, 0.0396, 0.0415, 0.0330, 0.0438], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:49:26,865 INFO [train.py:904] (4/8) Epoch 8, batch 4450, loss[loss=0.2243, simple_loss=0.3147, pruned_loss=0.06694, over 16235.00 frames. ], tot_loss[loss=0.2189, simple_loss=0.3016, pruned_loss=0.06804, over 3181721.18 frames. ], batch size: 165, lr: 8.60e-03, grad_scale: 8.0 2023-04-28 20:49:28,913 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.480e+02 2.422e+02 2.913e+02 3.508e+02 6.103e+02, threshold=5.826e+02, percent-clipped=0.0 2023-04-28 20:50:38,195 INFO [train.py:904] (4/8) Epoch 8, batch 4500, loss[loss=0.2039, simple_loss=0.2903, pruned_loss=0.05876, over 16853.00 frames. ], tot_loss[loss=0.2192, simple_loss=0.3019, pruned_loss=0.06828, over 3194476.47 frames. ], batch size: 96, lr: 8.59e-03, grad_scale: 8.0 2023-04-28 20:50:57,678 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75565.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:51:08,161 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.09 vs. limit=5.0 2023-04-28 20:51:24,862 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=75584.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:51:28,661 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3464, 4.6405, 4.3942, 4.4381, 4.1094, 4.0841, 4.2108, 4.6675], device='cuda:4'), covar=tensor([0.0768, 0.0703, 0.0933, 0.0515, 0.0693, 0.1334, 0.0754, 0.0787], device='cuda:4'), in_proj_covar=tensor([0.0474, 0.0590, 0.0499, 0.0403, 0.0376, 0.0391, 0.0493, 0.0440], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 20:51:50,274 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7755, 4.7752, 5.2815, 5.2262, 5.3162, 4.8225, 4.8172, 4.4138], device='cuda:4'), covar=tensor([0.0198, 0.0266, 0.0223, 0.0304, 0.0259, 0.0231, 0.0660, 0.0398], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0286, 0.0287, 0.0275, 0.0325, 0.0303, 0.0407, 0.0247], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 20:51:51,099 INFO [train.py:904] (4/8) Epoch 8, batch 4550, loss[loss=0.218, simple_loss=0.3067, pruned_loss=0.06463, over 17027.00 frames. ], tot_loss[loss=0.2191, simple_loss=0.3017, pruned_loss=0.06823, over 3204477.91 frames. ], batch size: 55, lr: 8.59e-03, grad_scale: 8.0 2023-04-28 20:51:52,274 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.538e+02 2.189e+02 2.672e+02 3.111e+02 5.807e+02, threshold=5.345e+02, percent-clipped=0.0 2023-04-28 20:52:06,459 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=75613.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:52:17,204 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75620.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:52:25,164 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=75626.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 20:53:02,628 INFO [train.py:904] (4/8) Epoch 8, batch 4600, loss[loss=0.207, simple_loss=0.2944, pruned_loss=0.05978, over 16361.00 frames. ], tot_loss[loss=0.2195, simple_loss=0.3028, pruned_loss=0.06808, over 3222846.92 frames. ], batch size: 35, lr: 8.59e-03, grad_scale: 8.0 2023-04-28 20:53:25,888 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=75668.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:53:44,351 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6044, 2.8073, 2.3848, 3.7503, 2.9919, 3.7976, 1.4065, 2.8831], device='cuda:4'), covar=tensor([0.1293, 0.0539, 0.1090, 0.0086, 0.0236, 0.0344, 0.1472, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0152, 0.0173, 0.0117, 0.0203, 0.0203, 0.0172, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 20:54:12,063 INFO [train.py:904] (4/8) Epoch 8, batch 4650, loss[loss=0.2064, simple_loss=0.2894, pruned_loss=0.06174, over 16849.00 frames. ], tot_loss[loss=0.2184, simple_loss=0.3012, pruned_loss=0.06781, over 3230443.07 frames. ], batch size: 42, lr: 8.58e-03, grad_scale: 8.0 2023-04-28 20:54:13,263 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.478e+02 2.294e+02 2.810e+02 3.370e+02 9.861e+02, threshold=5.619e+02, percent-clipped=3.0 2023-04-28 20:54:45,619 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 20:55:23,491 INFO [train.py:904] (4/8) Epoch 8, batch 4700, loss[loss=0.2079, simple_loss=0.2925, pruned_loss=0.06164, over 16972.00 frames. ], tot_loss[loss=0.2158, simple_loss=0.2983, pruned_loss=0.06666, over 3219661.34 frames. ], batch size: 41, lr: 8.58e-03, grad_scale: 8.0 2023-04-28 20:55:25,236 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4902, 3.8030, 3.5722, 1.9148, 2.9502, 2.3617, 3.6556, 3.7793], device='cuda:4'), covar=tensor([0.0206, 0.0610, 0.0557, 0.1774, 0.0765, 0.0864, 0.0609, 0.0791], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0135, 0.0152, 0.0138, 0.0132, 0.0121, 0.0132, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 20:56:31,925 INFO [train.py:904] (4/8) Epoch 8, batch 4750, loss[loss=0.1984, simple_loss=0.2771, pruned_loss=0.05987, over 16494.00 frames. ], tot_loss[loss=0.2112, simple_loss=0.2939, pruned_loss=0.06425, over 3226676.06 frames. ], batch size: 68, lr: 8.58e-03, grad_scale: 8.0 2023-04-28 20:56:33,067 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.310e+02 2.114e+02 2.531e+02 3.127e+02 7.196e+02, threshold=5.061e+02, percent-clipped=1.0 2023-04-28 20:57:27,006 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-28 20:57:44,145 INFO [train.py:904] (4/8) Epoch 8, batch 4800, loss[loss=0.1771, simple_loss=0.2738, pruned_loss=0.04019, over 16869.00 frames. ], tot_loss[loss=0.2079, simple_loss=0.2908, pruned_loss=0.06253, over 3220563.92 frames. ], batch size: 96, lr: 8.58e-03, grad_scale: 8.0 2023-04-28 20:58:19,367 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-28 20:58:31,972 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75884.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 20:58:58,545 INFO [train.py:904] (4/8) Epoch 8, batch 4850, loss[loss=0.2104, simple_loss=0.3093, pruned_loss=0.05572, over 16724.00 frames. ], tot_loss[loss=0.2075, simple_loss=0.2915, pruned_loss=0.06169, over 3214266.14 frames. ], batch size: 83, lr: 8.57e-03, grad_scale: 8.0 2023-04-28 20:59:01,504 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.426e+02 2.327e+02 2.698e+02 3.138e+02 6.949e+02, threshold=5.395e+02, percent-clipped=1.0 2023-04-28 20:59:02,757 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0481, 1.3750, 1.8350, 2.0759, 2.1300, 2.2807, 1.5181, 2.2938], device='cuda:4'), covar=tensor([0.0121, 0.0281, 0.0157, 0.0177, 0.0155, 0.0113, 0.0304, 0.0061], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0159, 0.0143, 0.0145, 0.0153, 0.0108, 0.0161, 0.0100], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 20:59:36,875 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75926.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 20:59:46,581 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=75932.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:00:17,717 INFO [train.py:904] (4/8) Epoch 8, batch 4900, loss[loss=0.2049, simple_loss=0.2849, pruned_loss=0.06242, over 16425.00 frames. ], tot_loss[loss=0.206, simple_loss=0.291, pruned_loss=0.0605, over 3202537.24 frames. ], batch size: 35, lr: 8.57e-03, grad_scale: 8.0 2023-04-28 21:00:49,947 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=75974.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:01:34,233 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-28 21:01:34,529 INFO [train.py:904] (4/8) Epoch 8, batch 4950, loss[loss=0.2436, simple_loss=0.3196, pruned_loss=0.08376, over 12291.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2906, pruned_loss=0.06026, over 3202995.19 frames. ], batch size: 246, lr: 8.57e-03, grad_scale: 8.0 2023-04-28 21:01:36,816 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.340e+02 2.828e+02 3.481e+02 8.052e+02, threshold=5.656e+02, percent-clipped=2.0 2023-04-28 21:01:55,587 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.87 vs. limit=5.0 2023-04-28 21:02:04,351 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7101, 3.7427, 4.0767, 4.0017, 4.0308, 3.7664, 3.7700, 3.7958], device='cuda:4'), covar=tensor([0.0261, 0.0411, 0.0308, 0.0451, 0.0438, 0.0287, 0.0740, 0.0399], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0283, 0.0285, 0.0274, 0.0326, 0.0300, 0.0402, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-28 21:02:45,330 INFO [train.py:904] (4/8) Epoch 8, batch 5000, loss[loss=0.2069, simple_loss=0.2981, pruned_loss=0.05785, over 16860.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2921, pruned_loss=0.06011, over 3222195.60 frames. ], batch size: 116, lr: 8.56e-03, grad_scale: 8.0 2023-04-28 21:03:10,517 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3869, 3.2972, 3.3340, 3.4799, 3.5227, 3.2532, 3.4840, 3.5670], device='cuda:4'), covar=tensor([0.0887, 0.0766, 0.1056, 0.0511, 0.0509, 0.1945, 0.0678, 0.0498], device='cuda:4'), in_proj_covar=tensor([0.0465, 0.0570, 0.0710, 0.0579, 0.0436, 0.0440, 0.0444, 0.0494], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:03:55,754 INFO [train.py:904] (4/8) Epoch 8, batch 5050, loss[loss=0.2141, simple_loss=0.3033, pruned_loss=0.06245, over 16426.00 frames. ], tot_loss[loss=0.2071, simple_loss=0.2933, pruned_loss=0.06042, over 3215820.86 frames. ], batch size: 146, lr: 8.56e-03, grad_scale: 8.0 2023-04-28 21:03:57,926 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.532e+02 2.576e+02 2.969e+02 3.606e+02 8.836e+02, threshold=5.938e+02, percent-clipped=5.0 2023-04-28 21:05:07,314 INFO [train.py:904] (4/8) Epoch 8, batch 5100, loss[loss=0.1845, simple_loss=0.2759, pruned_loss=0.04656, over 16701.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2915, pruned_loss=0.05957, over 3221867.14 frames. ], batch size: 76, lr: 8.56e-03, grad_scale: 8.0 2023-04-28 21:06:20,870 INFO [train.py:904] (4/8) Epoch 8, batch 5150, loss[loss=0.2284, simple_loss=0.3171, pruned_loss=0.0698, over 16399.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2916, pruned_loss=0.05931, over 3196348.56 frames. ], batch size: 146, lr: 8.56e-03, grad_scale: 8.0 2023-04-28 21:06:24,097 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.361e+02 2.714e+02 3.177e+02 5.443e+02, threshold=5.429e+02, percent-clipped=0.0 2023-04-28 21:07:33,713 INFO [train.py:904] (4/8) Epoch 8, batch 5200, loss[loss=0.197, simple_loss=0.2768, pruned_loss=0.05854, over 16382.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2906, pruned_loss=0.05921, over 3183868.97 frames. ], batch size: 146, lr: 8.55e-03, grad_scale: 8.0 2023-04-28 21:07:59,821 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=76269.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 21:08:36,435 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=76294.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:08:48,738 INFO [train.py:904] (4/8) Epoch 8, batch 5250, loss[loss=0.2219, simple_loss=0.3125, pruned_loss=0.06565, over 16246.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2875, pruned_loss=0.05854, over 3199796.45 frames. ], batch size: 165, lr: 8.55e-03, grad_scale: 8.0 2023-04-28 21:08:51,148 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.594e+02 2.354e+02 2.757e+02 3.436e+02 6.643e+02, threshold=5.515e+02, percent-clipped=1.0 2023-04-28 21:08:56,882 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4111, 3.3166, 3.4034, 3.5412, 3.5547, 3.2525, 3.5317, 3.5899], device='cuda:4'), covar=tensor([0.0925, 0.0698, 0.1003, 0.0505, 0.0564, 0.2288, 0.0810, 0.0542], device='cuda:4'), in_proj_covar=tensor([0.0469, 0.0568, 0.0715, 0.0584, 0.0441, 0.0439, 0.0449, 0.0503], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:09:09,048 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-28 21:09:19,358 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=76323.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 21:09:30,950 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=76330.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 21:09:42,562 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-28 21:10:01,828 INFO [train.py:904] (4/8) Epoch 8, batch 5300, loss[loss=0.178, simple_loss=0.2623, pruned_loss=0.0469, over 16647.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.2832, pruned_loss=0.05663, over 3213298.92 frames. ], batch size: 57, lr: 8.55e-03, grad_scale: 8.0 2023-04-28 21:10:06,390 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=76355.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:10:15,562 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9966, 4.2538, 4.5532, 4.4921, 4.4549, 4.2150, 3.9103, 4.0499], device='cuda:4'), covar=tensor([0.0479, 0.0535, 0.0334, 0.0462, 0.0529, 0.0388, 0.1405, 0.0566], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0292, 0.0291, 0.0283, 0.0334, 0.0311, 0.0416, 0.0257], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 21:10:25,516 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-28 21:10:48,026 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=76384.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 21:11:11,098 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.07 vs. limit=5.0 2023-04-28 21:11:13,478 INFO [train.py:904] (4/8) Epoch 8, batch 5350, loss[loss=0.2137, simple_loss=0.2988, pruned_loss=0.0643, over 15262.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2814, pruned_loss=0.05581, over 3217463.23 frames. ], batch size: 191, lr: 8.54e-03, grad_scale: 8.0 2023-04-28 21:11:15,924 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.588e+02 2.303e+02 2.727e+02 3.252e+02 5.747e+02, threshold=5.455e+02, percent-clipped=1.0 2023-04-28 21:12:22,080 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-28 21:12:26,521 INFO [train.py:904] (4/8) Epoch 8, batch 5400, loss[loss=0.1924, simple_loss=0.2795, pruned_loss=0.05268, over 17166.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2843, pruned_loss=0.05673, over 3204582.12 frames. ], batch size: 46, lr: 8.54e-03, grad_scale: 8.0 2023-04-28 21:13:41,160 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1966, 3.4537, 3.6022, 1.5632, 3.8388, 3.8644, 2.8703, 2.7028], device='cuda:4'), covar=tensor([0.0762, 0.0150, 0.0151, 0.1166, 0.0051, 0.0067, 0.0351, 0.0441], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0097, 0.0082, 0.0138, 0.0068, 0.0090, 0.0118, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 21:13:43,643 INFO [train.py:904] (4/8) Epoch 8, batch 5450, loss[loss=0.2418, simple_loss=0.3219, pruned_loss=0.08078, over 16235.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.288, pruned_loss=0.05886, over 3192048.45 frames. ], batch size: 165, lr: 8.54e-03, grad_scale: 8.0 2023-04-28 21:13:46,706 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.612e+02 2.643e+02 3.277e+02 3.879e+02 8.643e+02, threshold=6.553e+02, percent-clipped=9.0 2023-04-28 21:15:01,749 INFO [train.py:904] (4/8) Epoch 8, batch 5500, loss[loss=0.2838, simple_loss=0.3419, pruned_loss=0.1129, over 11925.00 frames. ], tot_loss[loss=0.2132, simple_loss=0.2967, pruned_loss=0.06489, over 3164351.62 frames. ], batch size: 248, lr: 8.54e-03, grad_scale: 8.0 2023-04-28 21:15:55,139 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3742, 3.1818, 2.5786, 2.1708, 2.2395, 2.0379, 3.1262, 3.2380], device='cuda:4'), covar=tensor([0.2315, 0.0780, 0.1480, 0.1910, 0.2153, 0.1769, 0.0514, 0.0904], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0250, 0.0275, 0.0263, 0.0279, 0.0210, 0.0259, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:16:05,359 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-28 21:16:22,240 INFO [train.py:904] (4/8) Epoch 8, batch 5550, loss[loss=0.2362, simple_loss=0.3215, pruned_loss=0.07547, over 16348.00 frames. ], tot_loss[loss=0.2249, simple_loss=0.3057, pruned_loss=0.072, over 3120506.53 frames. ], batch size: 146, lr: 8.53e-03, grad_scale: 8.0 2023-04-28 21:16:26,052 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.250e+02 3.650e+02 4.259e+02 5.295e+02 1.196e+03, threshold=8.517e+02, percent-clipped=11.0 2023-04-28 21:17:00,727 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=76625.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 21:17:22,043 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7525, 1.2562, 1.6304, 1.6167, 1.7288, 1.7835, 1.4402, 1.7222], device='cuda:4'), covar=tensor([0.0128, 0.0199, 0.0111, 0.0150, 0.0132, 0.0080, 0.0209, 0.0059], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0159, 0.0144, 0.0144, 0.0153, 0.0106, 0.0161, 0.0099], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 21:17:40,860 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=76650.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:17:42,921 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-28 21:17:43,813 INFO [train.py:904] (4/8) Epoch 8, batch 5600, loss[loss=0.264, simple_loss=0.3405, pruned_loss=0.09372, over 16778.00 frames. ], tot_loss[loss=0.2329, simple_loss=0.3117, pruned_loss=0.07703, over 3097618.90 frames. ], batch size: 124, lr: 8.53e-03, grad_scale: 8.0 2023-04-28 21:18:28,906 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=76679.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 21:19:06,822 INFO [train.py:904] (4/8) Epoch 8, batch 5650, loss[loss=0.2573, simple_loss=0.3237, pruned_loss=0.09547, over 16670.00 frames. ], tot_loss[loss=0.2423, simple_loss=0.3184, pruned_loss=0.08305, over 3045307.25 frames. ], batch size: 57, lr: 8.53e-03, grad_scale: 8.0 2023-04-28 21:19:10,210 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.391e+02 3.706e+02 4.487e+02 5.532e+02 1.181e+03, threshold=8.975e+02, percent-clipped=2.0 2023-04-28 21:20:27,956 INFO [train.py:904] (4/8) Epoch 8, batch 5700, loss[loss=0.2869, simple_loss=0.3419, pruned_loss=0.1159, over 11234.00 frames. ], tot_loss[loss=0.2466, simple_loss=0.3211, pruned_loss=0.08604, over 3016381.31 frames. ], batch size: 248, lr: 8.53e-03, grad_scale: 8.0 2023-04-28 21:21:15,765 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-28 21:21:49,368 INFO [train.py:904] (4/8) Epoch 8, batch 5750, loss[loss=0.2297, simple_loss=0.3081, pruned_loss=0.07567, over 16389.00 frames. ], tot_loss[loss=0.2471, simple_loss=0.3226, pruned_loss=0.08581, over 3043082.96 frames. ], batch size: 165, lr: 8.52e-03, grad_scale: 4.0 2023-04-28 21:21:54,072 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.644e+02 3.742e+02 4.943e+02 6.173e+02 1.099e+03, threshold=9.887e+02, percent-clipped=1.0 2023-04-28 21:22:23,080 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 21:22:43,804 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3005, 5.2846, 5.1616, 4.9088, 4.6870, 5.2265, 5.1080, 4.8755], device='cuda:4'), covar=tensor([0.0667, 0.0538, 0.0260, 0.0258, 0.1138, 0.0395, 0.0289, 0.0734], device='cuda:4'), in_proj_covar=tensor([0.0214, 0.0252, 0.0248, 0.0218, 0.0276, 0.0253, 0.0169, 0.0285], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:22:52,038 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=76839.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:23:12,544 INFO [train.py:904] (4/8) Epoch 8, batch 5800, loss[loss=0.2425, simple_loss=0.3276, pruned_loss=0.0787, over 16882.00 frames. ], tot_loss[loss=0.2434, simple_loss=0.3207, pruned_loss=0.08302, over 3059444.96 frames. ], batch size: 90, lr: 8.52e-03, grad_scale: 4.0 2023-04-28 21:24:30,545 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=76900.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:24:32,997 INFO [train.py:904] (4/8) Epoch 8, batch 5850, loss[loss=0.2256, simple_loss=0.3121, pruned_loss=0.06952, over 16832.00 frames. ], tot_loss[loss=0.2405, simple_loss=0.3187, pruned_loss=0.08116, over 3061357.01 frames. ], batch size: 116, lr: 8.52e-03, grad_scale: 4.0 2023-04-28 21:24:37,987 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.680e+02 3.434e+02 4.196e+02 5.124e+02 1.086e+03, threshold=8.393e+02, percent-clipped=1.0 2023-04-28 21:25:09,448 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=76925.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 21:25:17,737 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-28 21:25:52,815 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=76950.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:25:56,484 INFO [train.py:904] (4/8) Epoch 8, batch 5900, loss[loss=0.2853, simple_loss=0.3414, pruned_loss=0.1146, over 12062.00 frames. ], tot_loss[loss=0.24, simple_loss=0.3183, pruned_loss=0.08089, over 3076602.08 frames. ], batch size: 247, lr: 8.51e-03, grad_scale: 4.0 2023-04-28 21:26:33,736 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=76973.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 21:26:42,685 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=76979.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 21:27:05,694 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-28 21:27:11,067 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=76998.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:27:17,803 INFO [train.py:904] (4/8) Epoch 8, batch 5950, loss[loss=0.2473, simple_loss=0.3209, pruned_loss=0.08682, over 15267.00 frames. ], tot_loss[loss=0.2396, simple_loss=0.319, pruned_loss=0.08013, over 3071917.91 frames. ], batch size: 190, lr: 8.51e-03, grad_scale: 4.0 2023-04-28 21:27:21,560 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.299e+02 3.417e+02 4.127e+02 4.793e+02 1.224e+03, threshold=8.253e+02, percent-clipped=3.0 2023-04-28 21:27:25,227 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0839, 3.3106, 3.5621, 3.5423, 3.5077, 3.3295, 3.3429, 3.3797], device='cuda:4'), covar=tensor([0.0415, 0.0606, 0.0430, 0.0470, 0.0542, 0.0480, 0.0837, 0.0518], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0295, 0.0297, 0.0284, 0.0336, 0.0317, 0.0418, 0.0259], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 21:27:55,398 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=77027.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 21:27:56,908 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-28 21:28:07,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8371, 2.0734, 2.3953, 3.1815, 2.1486, 2.3530, 2.3009, 2.1381], device='cuda:4'), covar=tensor([0.0786, 0.2408, 0.1305, 0.0451, 0.2896, 0.1620, 0.2108, 0.2478], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0363, 0.0305, 0.0319, 0.0397, 0.0404, 0.0326, 0.0429], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:28:33,572 INFO [train.py:904] (4/8) Epoch 8, batch 6000, loss[loss=0.2159, simple_loss=0.2984, pruned_loss=0.06665, over 16750.00 frames. ], tot_loss[loss=0.2381, simple_loss=0.3177, pruned_loss=0.07927, over 3088255.07 frames. ], batch size: 83, lr: 8.51e-03, grad_scale: 8.0 2023-04-28 21:28:33,572 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 21:28:42,348 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6890, 4.5018, 4.6637, 4.9320, 5.0317, 4.6458, 5.0168, 5.0491], device='cuda:4'), covar=tensor([0.1359, 0.1008, 0.1288, 0.0463, 0.0433, 0.0506, 0.0394, 0.0414], device='cuda:4'), in_proj_covar=tensor([0.0463, 0.0564, 0.0700, 0.0574, 0.0432, 0.0432, 0.0448, 0.0497], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:28:44,114 INFO [train.py:938] (4/8) Epoch 8, validation: loss=0.1707, simple_loss=0.284, pruned_loss=0.02871, over 944034.00 frames. 2023-04-28 21:28:44,115 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 21:29:41,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3732, 3.2767, 3.3218, 3.4570, 3.4804, 3.2119, 3.4412, 3.5301], device='cuda:4'), covar=tensor([0.0856, 0.0711, 0.0994, 0.0515, 0.0552, 0.2307, 0.0792, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0464, 0.0567, 0.0706, 0.0577, 0.0433, 0.0432, 0.0448, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:29:48,601 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77095.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:29:50,085 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-28 21:30:00,319 INFO [train.py:904] (4/8) Epoch 8, batch 6050, loss[loss=0.2107, simple_loss=0.3069, pruned_loss=0.05727, over 16497.00 frames. ], tot_loss[loss=0.2364, simple_loss=0.3163, pruned_loss=0.0783, over 3089443.99 frames. ], batch size: 75, lr: 8.51e-03, grad_scale: 8.0 2023-04-28 21:30:04,227 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.264e+02 3.621e+02 4.460e+02 5.601e+02 1.783e+03, threshold=8.919e+02, percent-clipped=7.0 2023-04-28 21:31:19,254 INFO [train.py:904] (4/8) Epoch 8, batch 6100, loss[loss=0.2299, simple_loss=0.3154, pruned_loss=0.07218, over 16876.00 frames. ], tot_loss[loss=0.2344, simple_loss=0.3151, pruned_loss=0.07683, over 3100267.62 frames. ], batch size: 102, lr: 8.50e-03, grad_scale: 8.0 2023-04-28 21:31:22,584 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3982, 3.2963, 3.3492, 3.4842, 3.5082, 3.1997, 3.4553, 3.5599], device='cuda:4'), covar=tensor([0.0921, 0.0811, 0.1075, 0.0533, 0.0557, 0.2508, 0.0811, 0.0596], device='cuda:4'), in_proj_covar=tensor([0.0463, 0.0566, 0.0705, 0.0576, 0.0435, 0.0432, 0.0449, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:31:26,122 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77156.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 21:32:26,659 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77195.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:32:37,672 INFO [train.py:904] (4/8) Epoch 8, batch 6150, loss[loss=0.2625, simple_loss=0.3254, pruned_loss=0.09974, over 11401.00 frames. ], tot_loss[loss=0.2326, simple_loss=0.3128, pruned_loss=0.07622, over 3105938.74 frames. ], batch size: 248, lr: 8.50e-03, grad_scale: 8.0 2023-04-28 21:32:42,674 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.089e+02 3.253e+02 3.802e+02 4.768e+02 1.072e+03, threshold=7.605e+02, percent-clipped=4.0 2023-04-28 21:32:52,984 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77211.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:33:59,281 INFO [train.py:904] (4/8) Epoch 8, batch 6200, loss[loss=0.1842, simple_loss=0.2713, pruned_loss=0.04853, over 17253.00 frames. ], tot_loss[loss=0.2311, simple_loss=0.3109, pruned_loss=0.0757, over 3109780.06 frames. ], batch size: 52, lr: 8.50e-03, grad_scale: 4.0 2023-04-28 21:34:30,481 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-28 21:34:31,241 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77272.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:34:33,163 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77273.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:34:39,126 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77277.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:35:15,155 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7700, 3.5973, 3.8410, 3.6664, 3.7482, 4.1962, 3.9136, 3.6400], device='cuda:4'), covar=tensor([0.2179, 0.2499, 0.1986, 0.2493, 0.3092, 0.1540, 0.1333, 0.2570], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0434, 0.0458, 0.0385, 0.0511, 0.0483, 0.0372, 0.0518], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 21:35:16,096 INFO [train.py:904] (4/8) Epoch 8, batch 6250, loss[loss=0.2529, simple_loss=0.3188, pruned_loss=0.09353, over 11904.00 frames. ], tot_loss[loss=0.2306, simple_loss=0.3101, pruned_loss=0.0755, over 3101463.18 frames. ], batch size: 247, lr: 8.50e-03, grad_scale: 4.0 2023-04-28 21:35:22,794 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.200e+02 3.064e+02 3.768e+02 4.806e+02 9.942e+02, threshold=7.536e+02, percent-clipped=2.0 2023-04-28 21:35:26,664 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3987, 2.0882, 1.6811, 1.9870, 2.4368, 2.1840, 2.5182, 2.6341], device='cuda:4'), covar=tensor([0.0086, 0.0248, 0.0331, 0.0294, 0.0133, 0.0230, 0.0129, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0109, 0.0180, 0.0178, 0.0178, 0.0175, 0.0179, 0.0176, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:36:07,351 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77334.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:36:12,974 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77338.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:36:30,867 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.46 vs. limit=5.0 2023-04-28 21:36:35,474 INFO [train.py:904] (4/8) Epoch 8, batch 6300, loss[loss=0.2453, simple_loss=0.3112, pruned_loss=0.0897, over 11719.00 frames. ], tot_loss[loss=0.2297, simple_loss=0.3097, pruned_loss=0.0748, over 3109030.32 frames. ], batch size: 246, lr: 8.49e-03, grad_scale: 4.0 2023-04-28 21:37:54,077 INFO [train.py:904] (4/8) Epoch 8, batch 6350, loss[loss=0.3006, simple_loss=0.3491, pruned_loss=0.1261, over 11505.00 frames. ], tot_loss[loss=0.2312, simple_loss=0.3107, pruned_loss=0.0758, over 3111972.35 frames. ], batch size: 246, lr: 8.49e-03, grad_scale: 4.0 2023-04-28 21:38:00,461 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.093e+02 3.385e+02 4.021e+02 5.215e+02 8.857e+02, threshold=8.042e+02, percent-clipped=2.0 2023-04-28 21:39:10,453 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77451.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 21:39:11,216 INFO [train.py:904] (4/8) Epoch 8, batch 6400, loss[loss=0.2091, simple_loss=0.2945, pruned_loss=0.0619, over 16683.00 frames. ], tot_loss[loss=0.2333, simple_loss=0.3118, pruned_loss=0.07744, over 3091499.21 frames. ], batch size: 76, lr: 8.49e-03, grad_scale: 8.0 2023-04-28 21:40:17,231 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77495.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:40:26,236 INFO [train.py:904] (4/8) Epoch 8, batch 6450, loss[loss=0.2161, simple_loss=0.2833, pruned_loss=0.07443, over 11556.00 frames. ], tot_loss[loss=0.2327, simple_loss=0.3114, pruned_loss=0.07696, over 3087301.49 frames. ], batch size: 248, lr: 8.48e-03, grad_scale: 8.0 2023-04-28 21:40:33,076 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.067e+02 3.422e+02 4.688e+02 6.041e+02 9.477e+02, threshold=9.377e+02, percent-clipped=7.0 2023-04-28 21:41:31,682 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=77543.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:41:45,631 INFO [train.py:904] (4/8) Epoch 8, batch 6500, loss[loss=0.2158, simple_loss=0.3012, pruned_loss=0.06521, over 16208.00 frames. ], tot_loss[loss=0.2292, simple_loss=0.3084, pruned_loss=0.07505, over 3108500.20 frames. ], batch size: 165, lr: 8.48e-03, grad_scale: 8.0 2023-04-28 21:42:09,104 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77567.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:42:42,414 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7650, 2.4275, 2.1774, 3.2826, 1.9919, 3.5807, 1.4552, 2.7400], device='cuda:4'), covar=tensor([0.1333, 0.0675, 0.1172, 0.0147, 0.0165, 0.0395, 0.1581, 0.0739], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0149, 0.0173, 0.0117, 0.0200, 0.0202, 0.0173, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 21:43:05,183 INFO [train.py:904] (4/8) Epoch 8, batch 6550, loss[loss=0.2107, simple_loss=0.3083, pruned_loss=0.05653, over 16569.00 frames. ], tot_loss[loss=0.2318, simple_loss=0.3112, pruned_loss=0.07615, over 3110153.18 frames. ], batch size: 62, lr: 8.48e-03, grad_scale: 8.0 2023-04-28 21:43:11,238 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.922e+02 3.133e+02 3.759e+02 4.463e+02 1.371e+03, threshold=7.519e+02, percent-clipped=1.0 2023-04-28 21:43:28,230 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77616.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:43:47,923 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77629.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:43:53,998 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77633.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:44:22,369 INFO [train.py:904] (4/8) Epoch 8, batch 6600, loss[loss=0.2552, simple_loss=0.3346, pruned_loss=0.08793, over 16889.00 frames. ], tot_loss[loss=0.2332, simple_loss=0.3134, pruned_loss=0.07655, over 3115902.24 frames. ], batch size: 116, lr: 8.48e-03, grad_scale: 8.0 2023-04-28 21:44:36,938 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9649, 3.4648, 3.4867, 2.2348, 3.1759, 3.5274, 3.3559, 1.8796], device='cuda:4'), covar=tensor([0.0410, 0.0035, 0.0041, 0.0309, 0.0067, 0.0073, 0.0053, 0.0368], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0061, 0.0063, 0.0119, 0.0067, 0.0078, 0.0069, 0.0112], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 21:45:00,859 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77677.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:45:00,873 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77677.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:45:01,059 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.69 vs. limit=5.0 2023-04-28 21:45:05,890 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-28 21:45:38,845 INFO [train.py:904] (4/8) Epoch 8, batch 6650, loss[loss=0.227, simple_loss=0.3081, pruned_loss=0.0729, over 16831.00 frames. ], tot_loss[loss=0.2341, simple_loss=0.3134, pruned_loss=0.07741, over 3117646.63 frames. ], batch size: 83, lr: 8.47e-03, grad_scale: 8.0 2023-04-28 21:45:45,535 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.195e+02 3.528e+02 4.149e+02 5.029e+02 1.289e+03, threshold=8.299e+02, percent-clipped=5.0 2023-04-28 21:46:19,836 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1706, 1.8361, 2.0945, 3.5929, 1.7569, 2.2462, 1.9685, 1.9050], device='cuda:4'), covar=tensor([0.1012, 0.3379, 0.2017, 0.0529, 0.4371, 0.2450, 0.3077, 0.3648], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0365, 0.0305, 0.0318, 0.0399, 0.0404, 0.0325, 0.0430], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:46:35,365 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77738.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:46:42,616 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3790, 2.1013, 1.6207, 1.8612, 2.3952, 2.0721, 2.5079, 2.6335], device='cuda:4'), covar=tensor([0.0087, 0.0241, 0.0336, 0.0314, 0.0147, 0.0260, 0.0123, 0.0139], device='cuda:4'), in_proj_covar=tensor([0.0110, 0.0183, 0.0180, 0.0181, 0.0177, 0.0183, 0.0178, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:46:53,714 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77751.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 21:46:54,545 INFO [train.py:904] (4/8) Epoch 8, batch 6700, loss[loss=0.2109, simple_loss=0.287, pruned_loss=0.06739, over 16802.00 frames. ], tot_loss[loss=0.2338, simple_loss=0.3124, pruned_loss=0.07761, over 3116885.63 frames. ], batch size: 39, lr: 8.47e-03, grad_scale: 4.0 2023-04-28 21:47:48,760 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-28 21:48:06,137 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=77799.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:48:10,947 INFO [train.py:904] (4/8) Epoch 8, batch 6750, loss[loss=0.2326, simple_loss=0.3105, pruned_loss=0.07731, over 15269.00 frames. ], tot_loss[loss=0.2338, simple_loss=0.3119, pruned_loss=0.07784, over 3109150.37 frames. ], batch size: 190, lr: 8.47e-03, grad_scale: 4.0 2023-04-28 21:48:18,410 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.361e+02 3.344e+02 4.001e+02 5.088e+02 1.053e+03, threshold=8.003e+02, percent-clipped=2.0 2023-04-28 21:49:05,440 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-28 21:49:25,314 INFO [train.py:904] (4/8) Epoch 8, batch 6800, loss[loss=0.2595, simple_loss=0.3198, pruned_loss=0.09956, over 11293.00 frames. ], tot_loss[loss=0.2348, simple_loss=0.3125, pruned_loss=0.07851, over 3073091.39 frames. ], batch size: 248, lr: 8.47e-03, grad_scale: 8.0 2023-04-28 21:49:49,039 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77867.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:50:44,160 INFO [train.py:904] (4/8) Epoch 8, batch 6850, loss[loss=0.2089, simple_loss=0.3145, pruned_loss=0.05159, over 17133.00 frames. ], tot_loss[loss=0.2359, simple_loss=0.3143, pruned_loss=0.07874, over 3076601.93 frames. ], batch size: 49, lr: 8.46e-03, grad_scale: 4.0 2023-04-28 21:50:53,211 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.176e+02 3.270e+02 3.895e+02 4.577e+02 9.421e+02, threshold=7.790e+02, percent-clipped=1.0 2023-04-28 21:51:03,698 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=77915.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:51:24,265 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77929.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:51:27,470 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 21:51:30,534 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77933.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:51:53,809 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7168, 4.6785, 4.5479, 4.3487, 4.1285, 4.5899, 4.5586, 4.2966], device='cuda:4'), covar=tensor([0.0519, 0.0439, 0.0241, 0.0206, 0.0930, 0.0359, 0.0312, 0.0588], device='cuda:4'), in_proj_covar=tensor([0.0216, 0.0255, 0.0251, 0.0221, 0.0278, 0.0256, 0.0173, 0.0290], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:51:59,510 INFO [train.py:904] (4/8) Epoch 8, batch 6900, loss[loss=0.231, simple_loss=0.3187, pruned_loss=0.07165, over 16718.00 frames. ], tot_loss[loss=0.2376, simple_loss=0.3175, pruned_loss=0.07889, over 3092073.61 frames. ], batch size: 89, lr: 8.46e-03, grad_scale: 4.0 2023-04-28 21:52:31,658 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77972.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:52:39,050 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=77977.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:52:45,249 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=77981.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:52:46,817 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-28 21:53:06,061 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3492, 5.6200, 5.3367, 5.4173, 5.0795, 4.7991, 5.1219, 5.7123], device='cuda:4'), covar=tensor([0.0809, 0.0696, 0.0934, 0.0550, 0.0694, 0.0625, 0.0858, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0463, 0.0578, 0.0491, 0.0393, 0.0365, 0.0382, 0.0479, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:53:20,813 INFO [train.py:904] (4/8) Epoch 8, batch 6950, loss[loss=0.2501, simple_loss=0.3296, pruned_loss=0.08527, over 16513.00 frames. ], tot_loss[loss=0.2406, simple_loss=0.3195, pruned_loss=0.08086, over 3085511.93 frames. ], batch size: 75, lr: 8.46e-03, grad_scale: 4.0 2023-04-28 21:53:29,769 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.004e+02 3.295e+02 4.352e+02 5.795e+02 9.816e+02, threshold=8.703e+02, percent-clipped=9.0 2023-04-28 21:54:10,701 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=78033.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:54:14,091 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0278, 2.4199, 2.2677, 2.8260, 2.2260, 3.2570, 1.6604, 2.6854], device='cuda:4'), covar=tensor([0.1050, 0.0492, 0.0963, 0.0116, 0.0121, 0.0371, 0.1267, 0.0629], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0151, 0.0173, 0.0118, 0.0200, 0.0203, 0.0174, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 21:54:38,871 INFO [train.py:904] (4/8) Epoch 8, batch 7000, loss[loss=0.2657, simple_loss=0.3246, pruned_loss=0.1034, over 11588.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.3189, pruned_loss=0.0794, over 3098264.97 frames. ], batch size: 246, lr: 8.45e-03, grad_scale: 4.0 2023-04-28 21:54:48,476 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3239, 5.0601, 5.2862, 5.5589, 5.7360, 4.9522, 5.6286, 5.6496], device='cuda:4'), covar=tensor([0.1492, 0.0928, 0.1363, 0.0518, 0.0471, 0.0661, 0.0507, 0.0430], device='cuda:4'), in_proj_covar=tensor([0.0462, 0.0568, 0.0702, 0.0579, 0.0438, 0.0432, 0.0459, 0.0503], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 21:55:40,807 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 21:55:53,775 INFO [train.py:904] (4/8) Epoch 8, batch 7050, loss[loss=0.2432, simple_loss=0.3238, pruned_loss=0.08134, over 16938.00 frames. ], tot_loss[loss=0.2378, simple_loss=0.3188, pruned_loss=0.07837, over 3117442.80 frames. ], batch size: 109, lr: 8.45e-03, grad_scale: 4.0 2023-04-28 21:56:03,901 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.852e+02 3.078e+02 3.817e+02 4.566e+02 8.440e+02, threshold=7.634e+02, percent-clipped=0.0 2023-04-28 21:56:32,861 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.88 vs. limit=5.0 2023-04-28 21:57:11,210 INFO [train.py:904] (4/8) Epoch 8, batch 7100, loss[loss=0.2251, simple_loss=0.3053, pruned_loss=0.0725, over 16657.00 frames. ], tot_loss[loss=0.2367, simple_loss=0.3172, pruned_loss=0.07815, over 3108737.13 frames. ], batch size: 62, lr: 8.45e-03, grad_scale: 2.0 2023-04-28 21:58:26,606 INFO [train.py:904] (4/8) Epoch 8, batch 7150, loss[loss=0.2054, simple_loss=0.2874, pruned_loss=0.0617, over 16456.00 frames. ], tot_loss[loss=0.2359, simple_loss=0.3154, pruned_loss=0.07823, over 3079560.57 frames. ], batch size: 75, lr: 8.45e-03, grad_scale: 2.0 2023-04-28 21:58:36,168 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.673e+02 3.732e+02 4.325e+02 5.377e+02 1.184e+03, threshold=8.651e+02, percent-clipped=7.0 2023-04-28 21:59:09,690 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=78232.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 21:59:39,847 INFO [train.py:904] (4/8) Epoch 8, batch 7200, loss[loss=0.2091, simple_loss=0.2987, pruned_loss=0.05969, over 16780.00 frames. ], tot_loss[loss=0.233, simple_loss=0.3135, pruned_loss=0.0763, over 3068378.66 frames. ], batch size: 83, lr: 8.44e-03, grad_scale: 4.0 2023-04-28 22:00:06,595 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.10 vs. limit=2.0 2023-04-28 22:00:10,022 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=78272.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:00:42,289 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7622, 2.5912, 2.2669, 3.7782, 2.9665, 3.8561, 1.3366, 2.8723], device='cuda:4'), covar=tensor([0.1224, 0.0644, 0.1177, 0.0136, 0.0254, 0.0355, 0.1556, 0.0727], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0152, 0.0172, 0.0118, 0.0201, 0.0203, 0.0173, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 22:00:44,947 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=78293.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 22:00:59,550 INFO [train.py:904] (4/8) Epoch 8, batch 7250, loss[loss=0.2467, simple_loss=0.3074, pruned_loss=0.09299, over 11495.00 frames. ], tot_loss[loss=0.2301, simple_loss=0.3109, pruned_loss=0.07471, over 3077787.98 frames. ], batch size: 246, lr: 8.44e-03, grad_scale: 4.0 2023-04-28 22:01:10,047 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.993e+02 2.959e+02 3.722e+02 4.557e+02 8.668e+02, threshold=7.444e+02, percent-clipped=1.0 2023-04-28 22:01:26,986 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=78320.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:01:47,627 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=78333.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:02:16,538 INFO [train.py:904] (4/8) Epoch 8, batch 7300, loss[loss=0.2296, simple_loss=0.3116, pruned_loss=0.07378, over 16666.00 frames. ], tot_loss[loss=0.2294, simple_loss=0.3099, pruned_loss=0.07442, over 3083293.21 frames. ], batch size: 134, lr: 8.44e-03, grad_scale: 4.0 2023-04-28 22:02:40,272 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-28 22:03:02,793 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=78381.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:03:13,300 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5579, 2.5132, 2.1354, 3.4654, 2.7791, 3.6787, 1.2618, 2.5862], device='cuda:4'), covar=tensor([0.1496, 0.0731, 0.1409, 0.0155, 0.0270, 0.0421, 0.1779, 0.0974], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0153, 0.0173, 0.0118, 0.0201, 0.0203, 0.0175, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 22:03:34,469 INFO [train.py:904] (4/8) Epoch 8, batch 7350, loss[loss=0.2319, simple_loss=0.3131, pruned_loss=0.07536, over 16437.00 frames. ], tot_loss[loss=0.2308, simple_loss=0.3105, pruned_loss=0.07559, over 3039960.99 frames. ], batch size: 146, lr: 8.44e-03, grad_scale: 4.0 2023-04-28 22:03:45,279 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.445e+02 3.549e+02 4.373e+02 5.467e+02 2.414e+03, threshold=8.746e+02, percent-clipped=10.0 2023-04-28 22:04:17,275 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.78 vs. limit=5.0 2023-04-28 22:04:55,407 INFO [train.py:904] (4/8) Epoch 8, batch 7400, loss[loss=0.2884, simple_loss=0.3349, pruned_loss=0.121, over 11098.00 frames. ], tot_loss[loss=0.233, simple_loss=0.3119, pruned_loss=0.07707, over 3030957.01 frames. ], batch size: 248, lr: 8.43e-03, grad_scale: 4.0 2023-04-28 22:05:08,406 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=78460.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:05:12,362 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6875, 2.1431, 2.3083, 4.5756, 2.0794, 2.7210, 2.3087, 2.4737], device='cuda:4'), covar=tensor([0.0784, 0.3110, 0.1872, 0.0259, 0.3517, 0.1875, 0.2738, 0.2747], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0369, 0.0304, 0.0319, 0.0401, 0.0403, 0.0325, 0.0432], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:05:43,880 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4345, 3.2886, 2.7171, 2.2137, 2.3435, 2.1215, 3.4156, 3.2744], device='cuda:4'), covar=tensor([0.2504, 0.0857, 0.1540, 0.2009, 0.1991, 0.1703, 0.0585, 0.0931], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0255, 0.0277, 0.0264, 0.0281, 0.0213, 0.0261, 0.0277], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:05:50,547 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6965, 4.6542, 4.5017, 4.3139, 4.0948, 4.5933, 4.4560, 4.2777], device='cuda:4'), covar=tensor([0.0512, 0.0364, 0.0275, 0.0242, 0.0953, 0.0353, 0.0375, 0.0605], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0249, 0.0246, 0.0217, 0.0271, 0.0250, 0.0170, 0.0283], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:06:13,356 INFO [train.py:904] (4/8) Epoch 8, batch 7450, loss[loss=0.2014, simple_loss=0.299, pruned_loss=0.05186, over 16843.00 frames. ], tot_loss[loss=0.2337, simple_loss=0.3123, pruned_loss=0.07759, over 3038736.88 frames. ], batch size: 102, lr: 8.43e-03, grad_scale: 4.0 2023-04-28 22:06:26,559 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.920e+02 3.338e+02 4.187e+02 5.103e+02 1.080e+03, threshold=8.375e+02, percent-clipped=2.0 2023-04-28 22:06:40,530 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4584, 3.2221, 2.7557, 2.0338, 2.1928, 2.0819, 3.3475, 3.1731], device='cuda:4'), covar=tensor([0.2361, 0.0702, 0.1356, 0.2075, 0.1878, 0.1771, 0.0439, 0.0883], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0254, 0.0277, 0.0264, 0.0281, 0.0213, 0.0261, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:06:47,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=78521.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:07:34,683 INFO [train.py:904] (4/8) Epoch 8, batch 7500, loss[loss=0.2548, simple_loss=0.3245, pruned_loss=0.0926, over 15266.00 frames. ], tot_loss[loss=0.2343, simple_loss=0.3134, pruned_loss=0.07761, over 3039697.29 frames. ], batch size: 190, lr: 8.43e-03, grad_scale: 4.0 2023-04-28 22:07:42,045 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=78556.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:07:48,635 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8887, 4.7035, 4.9152, 5.1311, 5.2986, 4.7365, 5.2497, 5.2280], device='cuda:4'), covar=tensor([0.1342, 0.0835, 0.1252, 0.0486, 0.0370, 0.0622, 0.0423, 0.0404], device='cuda:4'), in_proj_covar=tensor([0.0460, 0.0565, 0.0697, 0.0575, 0.0440, 0.0428, 0.0457, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:08:33,285 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=78588.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 22:08:55,086 INFO [train.py:904] (4/8) Epoch 8, batch 7550, loss[loss=0.2356, simple_loss=0.3084, pruned_loss=0.08144, over 15412.00 frames. ], tot_loss[loss=0.2345, simple_loss=0.3128, pruned_loss=0.0781, over 3036540.15 frames. ], batch size: 190, lr: 8.42e-03, grad_scale: 4.0 2023-04-28 22:09:00,583 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5847, 3.7808, 1.8460, 4.1770, 2.6054, 4.1192, 2.1436, 2.8092], device='cuda:4'), covar=tensor([0.0210, 0.0297, 0.1750, 0.0071, 0.0812, 0.0357, 0.1441, 0.0671], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0157, 0.0181, 0.0101, 0.0164, 0.0195, 0.0189, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 22:09:05,664 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.406e+02 3.301e+02 4.162e+02 5.431e+02 1.258e+03, threshold=8.325e+02, percent-clipped=7.0 2023-04-28 22:09:20,029 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=78617.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:10:12,495 INFO [train.py:904] (4/8) Epoch 8, batch 7600, loss[loss=0.2707, simple_loss=0.323, pruned_loss=0.1092, over 11484.00 frames. ], tot_loss[loss=0.2331, simple_loss=0.3112, pruned_loss=0.07754, over 3040771.39 frames. ], batch size: 248, lr: 8.42e-03, grad_scale: 8.0 2023-04-28 22:11:30,196 INFO [train.py:904] (4/8) Epoch 8, batch 7650, loss[loss=0.2913, simple_loss=0.355, pruned_loss=0.1139, over 15313.00 frames. ], tot_loss[loss=0.2344, simple_loss=0.312, pruned_loss=0.0784, over 3038922.93 frames. ], batch size: 190, lr: 8.42e-03, grad_scale: 8.0 2023-04-28 22:11:40,445 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.288e+02 3.310e+02 4.230e+02 5.150e+02 8.626e+02, threshold=8.460e+02, percent-clipped=2.0 2023-04-28 22:11:41,050 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1040, 4.0900, 4.0490, 2.8793, 4.0138, 1.4725, 3.7225, 3.6442], device='cuda:4'), covar=tensor([0.0161, 0.0107, 0.0190, 0.0638, 0.0119, 0.3000, 0.0171, 0.0358], device='cuda:4'), in_proj_covar=tensor([0.0108, 0.0095, 0.0141, 0.0138, 0.0113, 0.0160, 0.0127, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:12:45,889 INFO [train.py:904] (4/8) Epoch 8, batch 7700, loss[loss=0.2139, simple_loss=0.2978, pruned_loss=0.06502, over 16772.00 frames. ], tot_loss[loss=0.2342, simple_loss=0.3116, pruned_loss=0.07838, over 3049542.70 frames. ], batch size: 89, lr: 8.42e-03, grad_scale: 4.0 2023-04-28 22:13:45,788 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 22:14:03,977 INFO [train.py:904] (4/8) Epoch 8, batch 7750, loss[loss=0.21, simple_loss=0.3003, pruned_loss=0.05984, over 16964.00 frames. ], tot_loss[loss=0.2335, simple_loss=0.3117, pruned_loss=0.07767, over 3066752.03 frames. ], batch size: 109, lr: 8.41e-03, grad_scale: 2.0 2023-04-28 22:14:17,807 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.129e+02 3.510e+02 4.111e+02 5.372e+02 9.340e+02, threshold=8.221e+02, percent-clipped=3.0 2023-04-28 22:14:23,599 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1868, 5.1653, 4.8969, 4.2523, 5.0011, 1.7389, 4.7623, 4.7880], device='cuda:4'), covar=tensor([0.0053, 0.0039, 0.0104, 0.0282, 0.0056, 0.2020, 0.0086, 0.0117], device='cuda:4'), in_proj_covar=tensor([0.0109, 0.0096, 0.0142, 0.0140, 0.0114, 0.0161, 0.0128, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:14:26,065 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=78816.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:15:19,809 INFO [train.py:904] (4/8) Epoch 8, batch 7800, loss[loss=0.2059, simple_loss=0.293, pruned_loss=0.05935, over 16873.00 frames. ], tot_loss[loss=0.234, simple_loss=0.3121, pruned_loss=0.07797, over 3079183.93 frames. ], batch size: 96, lr: 8.41e-03, grad_scale: 2.0 2023-04-28 22:16:16,611 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=78888.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:16:38,224 INFO [train.py:904] (4/8) Epoch 8, batch 7850, loss[loss=0.2253, simple_loss=0.3112, pruned_loss=0.06965, over 16804.00 frames. ], tot_loss[loss=0.2339, simple_loss=0.3132, pruned_loss=0.07733, over 3092815.00 frames. ], batch size: 124, lr: 8.41e-03, grad_scale: 2.0 2023-04-28 22:16:50,958 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.229e+02 3.386e+02 3.845e+02 4.858e+02 8.286e+02, threshold=7.691e+02, percent-clipped=1.0 2023-04-28 22:16:53,958 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=78912.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:17:01,245 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 22:17:21,328 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-28 22:17:25,109 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3020, 3.2134, 2.3403, 2.0829, 2.3967, 1.9659, 3.2190, 3.1343], device='cuda:4'), covar=tensor([0.2711, 0.1037, 0.1797, 0.1979, 0.2128, 0.1886, 0.0733, 0.1075], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0253, 0.0275, 0.0264, 0.0279, 0.0211, 0.0259, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:17:29,668 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=78936.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:17:44,834 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-28 22:17:54,229 INFO [train.py:904] (4/8) Epoch 8, batch 7900, loss[loss=0.2174, simple_loss=0.3042, pruned_loss=0.06529, over 16810.00 frames. ], tot_loss[loss=0.2341, simple_loss=0.3132, pruned_loss=0.07748, over 3079169.01 frames. ], batch size: 124, lr: 8.41e-03, grad_scale: 2.0 2023-04-28 22:18:15,511 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=78965.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:18:48,199 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1441, 3.3297, 3.5927, 3.5532, 3.5488, 3.3290, 3.3808, 3.4197], device='cuda:4'), covar=tensor([0.0360, 0.0546, 0.0401, 0.0436, 0.0431, 0.0452, 0.0805, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0297, 0.0300, 0.0290, 0.0339, 0.0315, 0.0420, 0.0260], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 22:19:13,438 INFO [train.py:904] (4/8) Epoch 8, batch 7950, loss[loss=0.2221, simple_loss=0.2992, pruned_loss=0.0725, over 16693.00 frames. ], tot_loss[loss=0.2353, simple_loss=0.3139, pruned_loss=0.07837, over 3074201.05 frames. ], batch size: 89, lr: 8.40e-03, grad_scale: 2.0 2023-04-28 22:19:28,064 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.154e+02 3.667e+02 4.128e+02 4.915e+02 9.776e+02, threshold=8.255e+02, percent-clipped=2.0 2023-04-28 22:19:49,336 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0255, 3.6129, 3.2195, 1.8501, 2.8506, 2.2702, 3.3954, 3.6711], device='cuda:4'), covar=tensor([0.0272, 0.0530, 0.0665, 0.1839, 0.0815, 0.0974, 0.0650, 0.0749], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0133, 0.0156, 0.0141, 0.0133, 0.0125, 0.0137, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 22:19:52,382 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79026.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:20:32,280 INFO [train.py:904] (4/8) Epoch 8, batch 8000, loss[loss=0.218, simple_loss=0.3006, pruned_loss=0.06767, over 16624.00 frames. ], tot_loss[loss=0.2368, simple_loss=0.3146, pruned_loss=0.0795, over 3057765.96 frames. ], batch size: 68, lr: 8.40e-03, grad_scale: 4.0 2023-04-28 22:21:01,369 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4734, 3.3933, 2.6986, 2.1521, 2.2728, 2.0955, 3.3829, 3.2227], device='cuda:4'), covar=tensor([0.2389, 0.0662, 0.1410, 0.1920, 0.2129, 0.1692, 0.0475, 0.0905], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0255, 0.0277, 0.0266, 0.0281, 0.0213, 0.0261, 0.0277], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:21:26,648 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6271, 2.7453, 2.2717, 3.9532, 2.8886, 3.7826, 1.4654, 2.7214], device='cuda:4'), covar=tensor([0.1325, 0.0612, 0.1238, 0.0153, 0.0318, 0.0434, 0.1508, 0.0861], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0152, 0.0173, 0.0119, 0.0201, 0.0201, 0.0172, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 22:21:30,953 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8497, 4.2574, 3.1154, 2.4908, 2.9350, 2.5861, 4.3902, 3.8340], device='cuda:4'), covar=tensor([0.2617, 0.0610, 0.1554, 0.2000, 0.2538, 0.1614, 0.0450, 0.0952], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0255, 0.0278, 0.0267, 0.0282, 0.0213, 0.0261, 0.0277], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:21:48,606 INFO [train.py:904] (4/8) Epoch 8, batch 8050, loss[loss=0.2374, simple_loss=0.3211, pruned_loss=0.07682, over 15459.00 frames. ], tot_loss[loss=0.237, simple_loss=0.3153, pruned_loss=0.07932, over 3072406.64 frames. ], batch size: 190, lr: 8.40e-03, grad_scale: 4.0 2023-04-28 22:22:02,033 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.318e+02 3.198e+02 3.767e+02 4.640e+02 1.099e+03, threshold=7.534e+02, percent-clipped=1.0 2023-04-28 22:22:11,143 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79116.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:22:16,986 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.20 vs. limit=5.0 2023-04-28 22:22:24,156 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2201, 3.3875, 1.8323, 3.5839, 2.3954, 3.6070, 1.9826, 2.6297], device='cuda:4'), covar=tensor([0.0230, 0.0321, 0.1534, 0.0102, 0.0828, 0.0427, 0.1475, 0.0688], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0158, 0.0182, 0.0103, 0.0167, 0.0198, 0.0192, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 22:23:05,315 INFO [train.py:904] (4/8) Epoch 8, batch 8100, loss[loss=0.2587, simple_loss=0.3193, pruned_loss=0.09899, over 11755.00 frames. ], tot_loss[loss=0.2359, simple_loss=0.3145, pruned_loss=0.07863, over 3070897.42 frames. ], batch size: 248, lr: 8.40e-03, grad_scale: 4.0 2023-04-28 22:23:18,223 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0702, 3.6945, 3.7238, 2.4641, 3.4683, 3.7366, 3.4815, 2.1692], device='cuda:4'), covar=tensor([0.0402, 0.0030, 0.0037, 0.0272, 0.0057, 0.0066, 0.0050, 0.0300], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0062, 0.0065, 0.0121, 0.0069, 0.0080, 0.0070, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 22:23:23,752 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=79164.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:24:07,061 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.19 vs. limit=5.0 2023-04-28 22:24:22,973 INFO [train.py:904] (4/8) Epoch 8, batch 8150, loss[loss=0.2035, simple_loss=0.2834, pruned_loss=0.06184, over 16915.00 frames. ], tot_loss[loss=0.2323, simple_loss=0.3112, pruned_loss=0.07673, over 3097084.13 frames. ], batch size: 96, lr: 8.39e-03, grad_scale: 4.0 2023-04-28 22:24:36,887 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.139e+02 3.692e+02 4.407e+02 5.311e+02 8.589e+02, threshold=8.814e+02, percent-clipped=3.0 2023-04-28 22:24:38,555 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79212.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:25:21,309 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.62 vs. limit=2.0 2023-04-28 22:25:42,513 INFO [train.py:904] (4/8) Epoch 8, batch 8200, loss[loss=0.1936, simple_loss=0.2834, pruned_loss=0.05194, over 16708.00 frames. ], tot_loss[loss=0.2295, simple_loss=0.308, pruned_loss=0.07547, over 3106096.34 frames. ], batch size: 89, lr: 8.39e-03, grad_scale: 4.0 2023-04-28 22:25:55,506 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=79260.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:26:09,029 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-28 22:26:31,958 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-28 22:26:36,821 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79285.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 22:26:54,883 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-28 22:27:04,307 INFO [train.py:904] (4/8) Epoch 8, batch 8250, loss[loss=0.2399, simple_loss=0.3273, pruned_loss=0.07619, over 16190.00 frames. ], tot_loss[loss=0.2267, simple_loss=0.3069, pruned_loss=0.07329, over 3089721.65 frames. ], batch size: 165, lr: 8.39e-03, grad_scale: 4.0 2023-04-28 22:27:19,443 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.098e+02 3.091e+02 3.915e+02 5.485e+02 9.423e+02, threshold=7.829e+02, percent-clipped=2.0 2023-04-28 22:27:37,112 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79321.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:27:56,957 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79333.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:28:17,681 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79346.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 22:28:26,463 INFO [train.py:904] (4/8) Epoch 8, batch 8300, loss[loss=0.2106, simple_loss=0.2796, pruned_loss=0.07078, over 11899.00 frames. ], tot_loss[loss=0.2214, simple_loss=0.3029, pruned_loss=0.06991, over 3053438.48 frames. ], batch size: 246, lr: 8.39e-03, grad_scale: 4.0 2023-04-28 22:28:29,025 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8560, 4.1827, 3.9836, 4.0438, 3.6920, 3.7677, 3.8453, 4.1495], device='cuda:4'), covar=tensor([0.1067, 0.0981, 0.1014, 0.0648, 0.0739, 0.1513, 0.0857, 0.1004], device='cuda:4'), in_proj_covar=tensor([0.0463, 0.0574, 0.0491, 0.0399, 0.0360, 0.0387, 0.0482, 0.0429], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:28:46,876 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79364.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:29:04,939 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79375.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 22:29:36,133 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79394.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:29:48,237 INFO [train.py:904] (4/8) Epoch 8, batch 8350, loss[loss=0.2184, simple_loss=0.2918, pruned_loss=0.0725, over 12047.00 frames. ], tot_loss[loss=0.2189, simple_loss=0.3021, pruned_loss=0.06785, over 3044407.75 frames. ], batch size: 246, lr: 8.38e-03, grad_scale: 4.0 2023-04-28 22:30:02,871 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.936e+02 2.645e+02 3.311e+02 4.098e+02 6.833e+02, threshold=6.622e+02, percent-clipped=0.0 2023-04-28 22:30:26,126 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79425.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:30:43,664 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79436.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 22:31:09,027 INFO [train.py:904] (4/8) Epoch 8, batch 8400, loss[loss=0.1936, simple_loss=0.2828, pruned_loss=0.05215, over 16756.00 frames. ], tot_loss[loss=0.2157, simple_loss=0.2996, pruned_loss=0.06591, over 3027795.95 frames. ], batch size: 76, lr: 8.38e-03, grad_scale: 8.0 2023-04-28 22:31:20,480 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.01 vs. limit=5.0 2023-04-28 22:32:17,228 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.17 vs. limit=5.0 2023-04-28 22:32:27,063 INFO [train.py:904] (4/8) Epoch 8, batch 8450, loss[loss=0.1739, simple_loss=0.2674, pruned_loss=0.04025, over 16668.00 frames. ], tot_loss[loss=0.2125, simple_loss=0.2972, pruned_loss=0.06386, over 3017482.77 frames. ], batch size: 57, lr: 8.38e-03, grad_scale: 8.0 2023-04-28 22:32:42,129 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.495e+02 2.576e+02 3.232e+02 4.028e+02 7.324e+02, threshold=6.464e+02, percent-clipped=2.0 2023-04-28 22:33:03,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1985, 3.3881, 1.4458, 3.5323, 2.4172, 3.4585, 1.7316, 2.6960], device='cuda:4'), covar=tensor([0.0182, 0.0245, 0.1848, 0.0125, 0.0725, 0.0442, 0.1828, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0152, 0.0177, 0.0100, 0.0160, 0.0189, 0.0188, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 22:33:06,740 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0548, 3.1952, 1.7439, 3.3377, 2.3417, 3.2832, 2.0492, 2.6745], device='cuda:4'), covar=tensor([0.0199, 0.0254, 0.1476, 0.0130, 0.0724, 0.0485, 0.1347, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0152, 0.0177, 0.0100, 0.0160, 0.0189, 0.0188, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 22:33:39,519 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5732, 4.0278, 3.9149, 2.8222, 3.5751, 3.9586, 3.6764, 2.4035], device='cuda:4'), covar=tensor([0.0344, 0.0021, 0.0029, 0.0242, 0.0057, 0.0046, 0.0041, 0.0325], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0059, 0.0062, 0.0116, 0.0066, 0.0076, 0.0067, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 22:33:47,271 INFO [train.py:904] (4/8) Epoch 8, batch 8500, loss[loss=0.1766, simple_loss=0.2658, pruned_loss=0.04371, over 16684.00 frames. ], tot_loss[loss=0.2068, simple_loss=0.2924, pruned_loss=0.06061, over 3020442.69 frames. ], batch size: 83, lr: 8.37e-03, grad_scale: 8.0 2023-04-28 22:33:57,388 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-04-28 22:34:05,984 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-28 22:34:40,416 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79585.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:34:40,505 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6253, 1.4990, 2.0380, 2.5529, 2.3867, 2.6715, 1.8079, 2.6985], device='cuda:4'), covar=tensor([0.0106, 0.0328, 0.0196, 0.0158, 0.0168, 0.0132, 0.0282, 0.0073], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0160, 0.0144, 0.0140, 0.0149, 0.0107, 0.0157, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 22:35:09,510 INFO [train.py:904] (4/8) Epoch 8, batch 8550, loss[loss=0.2209, simple_loss=0.3178, pruned_loss=0.06206, over 16429.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2898, pruned_loss=0.05909, over 3019577.65 frames. ], batch size: 75, lr: 8.37e-03, grad_scale: 8.0 2023-04-28 22:35:26,480 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.911e+02 2.652e+02 3.316e+02 4.199e+02 1.038e+03, threshold=6.632e+02, percent-clipped=3.0 2023-04-28 22:35:47,333 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79621.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:36:25,453 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2835, 4.5220, 4.6084, 4.5222, 4.5004, 4.9785, 4.6006, 4.3892], device='cuda:4'), covar=tensor([0.1287, 0.1613, 0.1619, 0.1796, 0.2213, 0.1039, 0.1272, 0.2291], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0413, 0.0444, 0.0370, 0.0486, 0.0466, 0.0360, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:36:28,517 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79641.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 22:36:37,308 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79646.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:36:48,486 INFO [train.py:904] (4/8) Epoch 8, batch 8600, loss[loss=0.1987, simple_loss=0.2922, pruned_loss=0.05262, over 16904.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2907, pruned_loss=0.05806, over 3036430.08 frames. ], batch size: 102, lr: 8.37e-03, grad_scale: 8.0 2023-04-28 22:36:59,392 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79657.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:37:24,926 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=79669.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:38:02,628 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79689.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:38:20,346 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3014, 4.5088, 4.5839, 4.4603, 4.4022, 4.9671, 4.5253, 4.2008], device='cuda:4'), covar=tensor([0.1092, 0.1517, 0.1484, 0.1732, 0.2641, 0.0868, 0.1245, 0.2302], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0412, 0.0444, 0.0369, 0.0487, 0.0466, 0.0362, 0.0490], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:38:26,205 INFO [train.py:904] (4/8) Epoch 8, batch 8650, loss[loss=0.1806, simple_loss=0.2755, pruned_loss=0.04285, over 16279.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2879, pruned_loss=0.05601, over 3020825.42 frames. ], batch size: 165, lr: 8.37e-03, grad_scale: 4.0 2023-04-28 22:38:50,263 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.014e+02 2.698e+02 3.186e+02 3.919e+02 1.176e+03, threshold=6.372e+02, percent-clipped=4.0 2023-04-28 22:39:04,717 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79718.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:39:09,483 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79720.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:39:32,733 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79731.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 22:40:01,402 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8514, 4.2248, 4.0481, 4.0806, 3.7188, 3.8191, 3.8653, 4.1854], device='cuda:4'), covar=tensor([0.0958, 0.0833, 0.0871, 0.0553, 0.0683, 0.1396, 0.0801, 0.0959], device='cuda:4'), in_proj_covar=tensor([0.0457, 0.0567, 0.0485, 0.0395, 0.0356, 0.0381, 0.0476, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:40:12,066 INFO [train.py:904] (4/8) Epoch 8, batch 8700, loss[loss=0.1947, simple_loss=0.2735, pruned_loss=0.05796, over 12295.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2849, pruned_loss=0.0546, over 3028693.88 frames. ], batch size: 250, lr: 8.36e-03, grad_scale: 4.0 2023-04-28 22:40:32,979 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79762.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:40:49,553 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5456, 3.5949, 2.7820, 2.0904, 2.4031, 2.2855, 3.9570, 3.4165], device='cuda:4'), covar=tensor([0.2528, 0.0741, 0.1520, 0.2239, 0.2053, 0.1678, 0.0404, 0.0869], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0243, 0.0267, 0.0255, 0.0258, 0.0206, 0.0248, 0.0262], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:41:50,473 INFO [train.py:904] (4/8) Epoch 8, batch 8750, loss[loss=0.2046, simple_loss=0.2976, pruned_loss=0.05582, over 16722.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2843, pruned_loss=0.05377, over 3033159.16 frames. ], batch size: 134, lr: 8.36e-03, grad_scale: 4.0 2023-04-28 22:42:15,364 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.735e+02 2.669e+02 3.166e+02 4.113e+02 7.290e+02, threshold=6.332e+02, percent-clipped=2.0 2023-04-28 22:42:43,922 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79823.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:43:14,910 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3509, 4.6376, 4.4304, 4.4604, 4.1487, 4.1400, 4.1529, 4.6649], device='cuda:4'), covar=tensor([0.0737, 0.0755, 0.0846, 0.0527, 0.0632, 0.1156, 0.0812, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0451, 0.0561, 0.0481, 0.0389, 0.0352, 0.0377, 0.0473, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:43:38,850 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9239, 3.0225, 3.1072, 2.0818, 2.9665, 3.1634, 3.0443, 1.8177], device='cuda:4'), covar=tensor([0.0392, 0.0035, 0.0042, 0.0319, 0.0067, 0.0056, 0.0051, 0.0393], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0060, 0.0063, 0.0119, 0.0067, 0.0077, 0.0069, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 22:43:44,531 INFO [train.py:904] (4/8) Epoch 8, batch 8800, loss[loss=0.1849, simple_loss=0.2753, pruned_loss=0.04727, over 16682.00 frames. ], tot_loss[loss=0.1935, simple_loss=0.2823, pruned_loss=0.05234, over 3045097.67 frames. ], batch size: 134, lr: 8.36e-03, grad_scale: 8.0 2023-04-28 22:43:56,220 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2451, 4.0767, 4.3104, 4.4567, 4.5623, 4.0805, 4.5607, 4.5474], device='cuda:4'), covar=tensor([0.1133, 0.0884, 0.1222, 0.0545, 0.0498, 0.0999, 0.0486, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0436, 0.0544, 0.0665, 0.0553, 0.0422, 0.0417, 0.0439, 0.0485], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:45:28,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8654, 2.6187, 2.6230, 1.9382, 2.5736, 2.7197, 2.5551, 1.6819], device='cuda:4'), covar=tensor([0.0290, 0.0048, 0.0048, 0.0261, 0.0071, 0.0051, 0.0071, 0.0383], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0061, 0.0063, 0.0119, 0.0068, 0.0078, 0.0069, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 22:45:31,356 INFO [train.py:904] (4/8) Epoch 8, batch 8850, loss[loss=0.1912, simple_loss=0.2821, pruned_loss=0.05011, over 11975.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.284, pruned_loss=0.05128, over 3038431.61 frames. ], batch size: 246, lr: 8.36e-03, grad_scale: 8.0 2023-04-28 22:45:48,678 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2785, 1.3983, 1.7442, 2.2237, 2.2407, 2.2690, 1.7963, 2.3155], device='cuda:4'), covar=tensor([0.0135, 0.0330, 0.0225, 0.0186, 0.0187, 0.0154, 0.0287, 0.0079], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0157, 0.0141, 0.0138, 0.0145, 0.0104, 0.0154, 0.0094], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 22:45:51,484 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.540e+02 2.661e+02 3.236e+02 3.873e+02 8.211e+02, threshold=6.471e+02, percent-clipped=3.0 2023-04-28 22:46:19,465 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.50 vs. limit=5.0 2023-04-28 22:46:57,808 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79941.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:46:57,902 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79941.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 22:47:12,674 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8150, 2.7249, 2.6651, 1.8944, 2.5389, 2.6615, 2.6279, 1.7770], device='cuda:4'), covar=tensor([0.0318, 0.0036, 0.0038, 0.0258, 0.0058, 0.0052, 0.0046, 0.0311], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0060, 0.0062, 0.0118, 0.0068, 0.0077, 0.0068, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 22:47:20,936 INFO [train.py:904] (4/8) Epoch 8, batch 8900, loss[loss=0.1965, simple_loss=0.2791, pruned_loss=0.05697, over 12851.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2837, pruned_loss=0.0503, over 3036401.44 frames. ], batch size: 248, lr: 8.35e-03, grad_scale: 8.0 2023-04-28 22:47:48,249 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9800, 5.3032, 5.0362, 5.0318, 4.7339, 4.6340, 4.7005, 5.3079], device='cuda:4'), covar=tensor([0.0945, 0.0773, 0.0892, 0.0532, 0.0709, 0.0754, 0.0836, 0.0780], device='cuda:4'), in_proj_covar=tensor([0.0445, 0.0557, 0.0475, 0.0386, 0.0348, 0.0374, 0.0469, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:48:54,504 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=79989.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 22:48:54,545 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79989.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:49:29,483 INFO [train.py:904] (4/8) Epoch 8, batch 8950, loss[loss=0.1836, simple_loss=0.2663, pruned_loss=0.05045, over 12670.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2835, pruned_loss=0.05079, over 3050962.06 frames. ], batch size: 248, lr: 8.35e-03, grad_scale: 8.0 2023-04-28 22:49:50,484 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.741e+02 2.495e+02 3.051e+02 3.831e+02 8.360e+02, threshold=6.103e+02, percent-clipped=2.0 2023-04-28 22:49:54,190 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80013.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:50:08,393 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80020.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:50:32,298 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80031.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 22:50:46,185 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=80037.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:51:17,194 INFO [train.py:904] (4/8) Epoch 8, batch 9000, loss[loss=0.1944, simple_loss=0.2824, pruned_loss=0.05322, over 16190.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2811, pruned_loss=0.0498, over 3045346.30 frames. ], batch size: 165, lr: 8.35e-03, grad_scale: 8.0 2023-04-28 22:51:17,194 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 22:51:25,746 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2529, 4.5627, 4.3487, 4.4046, 4.1872, 4.2179, 4.1597, 4.4967], device='cuda:4'), covar=tensor([0.0833, 0.0598, 0.0720, 0.0444, 0.0547, 0.0447, 0.0769, 0.0729], device='cuda:4'), in_proj_covar=tensor([0.0450, 0.0564, 0.0482, 0.0390, 0.0353, 0.0378, 0.0473, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:51:27,531 INFO [train.py:938] (4/8) Epoch 8, validation: loss=0.1608, simple_loss=0.265, pruned_loss=0.02828, over 944034.00 frames. 2023-04-28 22:51:27,532 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 22:51:31,706 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 22:52:04,336 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=80068.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:52:26,509 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=80079.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 22:53:14,319 INFO [train.py:904] (4/8) Epoch 8, batch 9050, loss[loss=0.2121, simple_loss=0.2918, pruned_loss=0.06625, over 12729.00 frames. ], tot_loss[loss=0.1914, simple_loss=0.2818, pruned_loss=0.05054, over 3042302.20 frames. ], batch size: 246, lr: 8.35e-03, grad_scale: 8.0 2023-04-28 22:53:23,990 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8117, 4.2171, 3.1410, 2.3813, 3.0377, 2.4291, 4.4126, 3.8675], device='cuda:4'), covar=tensor([0.2210, 0.0502, 0.1351, 0.1823, 0.1776, 0.1567, 0.0331, 0.0684], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0242, 0.0266, 0.0253, 0.0249, 0.0204, 0.0247, 0.0259], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:53:35,355 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.717e+02 2.777e+02 3.475e+02 4.225e+02 7.695e+02, threshold=6.949e+02, percent-clipped=5.0 2023-04-28 22:53:45,517 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-28 22:53:48,932 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80118.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:54:59,312 INFO [train.py:904] (4/8) Epoch 8, batch 9100, loss[loss=0.1919, simple_loss=0.2866, pruned_loss=0.04866, over 16702.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2822, pruned_loss=0.05137, over 3048811.03 frames. ], batch size: 134, lr: 8.34e-03, grad_scale: 8.0 2023-04-28 22:56:53,526 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9926, 2.6452, 2.6299, 1.8803, 2.8530, 2.8609, 2.4427, 2.4383], device='cuda:4'), covar=tensor([0.0750, 0.0197, 0.0203, 0.0990, 0.0083, 0.0139, 0.0450, 0.0422], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0094, 0.0080, 0.0137, 0.0064, 0.0086, 0.0116, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 22:56:59,464 INFO [train.py:904] (4/8) Epoch 8, batch 9150, loss[loss=0.1758, simple_loss=0.2637, pruned_loss=0.04399, over 12194.00 frames. ], tot_loss[loss=0.1927, simple_loss=0.2829, pruned_loss=0.05122, over 3041616.39 frames. ], batch size: 250, lr: 8.34e-03, grad_scale: 8.0 2023-04-28 22:57:20,183 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.002e+02 2.760e+02 3.111e+02 3.901e+02 6.426e+02, threshold=6.222e+02, percent-clipped=0.0 2023-04-28 22:58:26,799 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80241.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:58:44,303 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0390, 3.1312, 1.6462, 3.3257, 2.2569, 3.2722, 1.9344, 2.5893], device='cuda:4'), covar=tensor([0.0245, 0.0311, 0.1601, 0.0137, 0.0867, 0.0470, 0.1518, 0.0661], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0152, 0.0180, 0.0100, 0.0160, 0.0187, 0.0190, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-28 22:58:44,954 INFO [train.py:904] (4/8) Epoch 8, batch 9200, loss[loss=0.1848, simple_loss=0.2748, pruned_loss=0.04741, over 15406.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2781, pruned_loss=0.05002, over 3036475.65 frames. ], batch size: 190, lr: 8.34e-03, grad_scale: 8.0 2023-04-28 22:59:36,178 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1482, 4.2036, 4.0015, 3.8025, 3.6968, 4.1411, 3.8849, 3.8254], device='cuda:4'), covar=tensor([0.0514, 0.0487, 0.0245, 0.0212, 0.0638, 0.0384, 0.0540, 0.0470], device='cuda:4'), in_proj_covar=tensor([0.0204, 0.0243, 0.0240, 0.0214, 0.0256, 0.0244, 0.0165, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 22:59:43,154 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80283.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:59:55,917 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=80289.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 22:59:58,523 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4776, 3.3365, 2.6918, 2.1160, 2.1510, 2.1026, 3.3596, 3.0617], device='cuda:4'), covar=tensor([0.2341, 0.0641, 0.1288, 0.2070, 0.2257, 0.1719, 0.0396, 0.0981], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0242, 0.0266, 0.0255, 0.0249, 0.0204, 0.0247, 0.0258], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:00:22,118 INFO [train.py:904] (4/8) Epoch 8, batch 9250, loss[loss=0.2034, simple_loss=0.288, pruned_loss=0.05944, over 16518.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2782, pruned_loss=0.05023, over 3029271.58 frames. ], batch size: 146, lr: 8.34e-03, grad_scale: 4.0 2023-04-28 23:00:42,872 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.717e+02 2.839e+02 3.478e+02 4.215e+02 8.620e+02, threshold=6.955e+02, percent-clipped=3.0 2023-04-28 23:00:44,619 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80313.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:01:09,426 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-28 23:01:45,091 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3858, 3.3214, 3.4306, 3.5070, 3.5395, 3.2233, 3.5030, 3.5716], device='cuda:4'), covar=tensor([0.0980, 0.0737, 0.0926, 0.0512, 0.0534, 0.1865, 0.0788, 0.0593], device='cuda:4'), in_proj_covar=tensor([0.0436, 0.0534, 0.0657, 0.0548, 0.0415, 0.0411, 0.0432, 0.0474], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:01:56,407 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80344.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:02:12,075 INFO [train.py:904] (4/8) Epoch 8, batch 9300, loss[loss=0.1639, simple_loss=0.2523, pruned_loss=0.03778, over 16157.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2766, pruned_loss=0.0496, over 3036852.56 frames. ], batch size: 165, lr: 8.33e-03, grad_scale: 4.0 2023-04-28 23:02:31,762 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=80361.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:03:42,340 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80393.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:03:55,436 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9968, 3.0857, 3.0358, 2.1200, 2.8873, 3.0410, 2.9772, 1.7090], device='cuda:4'), covar=tensor([0.0359, 0.0037, 0.0034, 0.0302, 0.0064, 0.0076, 0.0054, 0.0382], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0060, 0.0062, 0.0118, 0.0068, 0.0077, 0.0068, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 23:03:57,479 INFO [train.py:904] (4/8) Epoch 8, batch 9350, loss[loss=0.1841, simple_loss=0.2711, pruned_loss=0.04853, over 12576.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2762, pruned_loss=0.04926, over 3045701.89 frames. ], batch size: 248, lr: 8.33e-03, grad_scale: 4.0 2023-04-28 23:04:22,274 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.742e+02 2.575e+02 3.007e+02 3.554e+02 5.975e+02, threshold=6.013e+02, percent-clipped=0.0 2023-04-28 23:04:34,131 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80418.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:05:40,898 INFO [train.py:904] (4/8) Epoch 8, batch 9400, loss[loss=0.1809, simple_loss=0.2901, pruned_loss=0.03586, over 16871.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2765, pruned_loss=0.0488, over 3054586.41 frames. ], batch size: 96, lr: 8.33e-03, grad_scale: 4.0 2023-04-28 23:05:46,199 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80454.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:06:05,139 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80463.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:06:10,059 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=80466.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:07:07,324 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80494.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:07:19,861 INFO [train.py:904] (4/8) Epoch 8, batch 9450, loss[loss=0.1961, simple_loss=0.2776, pruned_loss=0.05728, over 12483.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2789, pruned_loss=0.0494, over 3061564.36 frames. ], batch size: 248, lr: 8.33e-03, grad_scale: 4.0 2023-04-28 23:07:38,811 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.685e+02 2.597e+02 3.133e+02 4.086e+02 1.022e+03, threshold=6.266e+02, percent-clipped=6.0 2023-04-28 23:08:04,825 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80524.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:08:58,863 INFO [train.py:904] (4/8) Epoch 8, batch 9500, loss[loss=0.2122, simple_loss=0.3122, pruned_loss=0.05607, over 15314.00 frames. ], tot_loss[loss=0.1872, simple_loss=0.2774, pruned_loss=0.04852, over 3068437.28 frames. ], batch size: 191, lr: 8.32e-03, grad_scale: 4.0 2023-04-28 23:09:08,282 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80555.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:09:39,242 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5811, 1.6258, 2.0873, 2.5318, 2.3724, 2.6282, 1.8287, 2.7091], device='cuda:4'), covar=tensor([0.0107, 0.0310, 0.0202, 0.0166, 0.0170, 0.0113, 0.0316, 0.0086], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0156, 0.0140, 0.0138, 0.0145, 0.0102, 0.0154, 0.0094], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 23:10:07,057 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-28 23:10:36,724 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8388, 2.2547, 2.4632, 4.6237, 2.1537, 2.8757, 2.3272, 2.5287], device='cuda:4'), covar=tensor([0.0654, 0.2981, 0.1738, 0.0246, 0.3493, 0.1845, 0.2787, 0.2956], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0351, 0.0301, 0.0308, 0.0387, 0.0385, 0.0317, 0.0413], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:10:46,570 INFO [train.py:904] (4/8) Epoch 8, batch 9550, loss[loss=0.2331, simple_loss=0.3232, pruned_loss=0.07145, over 16447.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2773, pruned_loss=0.04886, over 3064843.30 frames. ], batch size: 146, lr: 8.32e-03, grad_scale: 4.0 2023-04-28 23:10:51,365 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8681, 2.7115, 2.5684, 1.9610, 2.5238, 2.6330, 2.5504, 1.8436], device='cuda:4'), covar=tensor([0.0282, 0.0035, 0.0039, 0.0249, 0.0071, 0.0059, 0.0051, 0.0312], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0058, 0.0060, 0.0114, 0.0066, 0.0075, 0.0065, 0.0110], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-28 23:11:10,120 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.638e+02 2.584e+02 3.110e+02 3.705e+02 8.746e+02, threshold=6.220e+02, percent-clipped=3.0 2023-04-28 23:11:29,040 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-28 23:11:43,250 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7607, 2.5684, 2.2056, 3.7158, 2.5046, 3.7531, 1.4016, 2.9220], device='cuda:4'), covar=tensor([0.1306, 0.0611, 0.1135, 0.0119, 0.0129, 0.0347, 0.1525, 0.0642], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0149, 0.0170, 0.0114, 0.0178, 0.0197, 0.0170, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 23:11:56,853 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5431, 3.7060, 1.9036, 4.0345, 2.6310, 3.9342, 2.0007, 2.8546], device='cuda:4'), covar=tensor([0.0193, 0.0260, 0.1701, 0.0083, 0.0809, 0.0419, 0.1602, 0.0659], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0150, 0.0181, 0.0101, 0.0161, 0.0186, 0.0190, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 23:12:04,018 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80639.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:12:27,096 INFO [train.py:904] (4/8) Epoch 8, batch 9600, loss[loss=0.2102, simple_loss=0.3103, pruned_loss=0.05503, over 16354.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2783, pruned_loss=0.04974, over 3048413.59 frames. ], batch size: 146, lr: 8.32e-03, grad_scale: 8.0 2023-04-28 23:13:27,698 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3204, 1.3788, 1.8254, 2.2176, 2.1454, 2.2980, 1.6130, 2.3226], device='cuda:4'), covar=tensor([0.0113, 0.0346, 0.0195, 0.0199, 0.0200, 0.0130, 0.0302, 0.0083], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0155, 0.0141, 0.0139, 0.0145, 0.0102, 0.0154, 0.0093], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 23:13:48,249 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6269, 2.7283, 1.8286, 2.8195, 2.1875, 2.8189, 1.9932, 2.4093], device='cuda:4'), covar=tensor([0.0235, 0.0317, 0.1349, 0.0173, 0.0788, 0.0473, 0.1339, 0.0609], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0152, 0.0182, 0.0102, 0.0163, 0.0187, 0.0191, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 23:14:15,049 INFO [train.py:904] (4/8) Epoch 8, batch 9650, loss[loss=0.1801, simple_loss=0.2781, pruned_loss=0.04106, over 16462.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2803, pruned_loss=0.05007, over 3053123.74 frames. ], batch size: 62, lr: 8.31e-03, grad_scale: 8.0 2023-04-28 23:14:42,863 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.647e+02 2.581e+02 3.070e+02 3.886e+02 7.619e+02, threshold=6.139e+02, percent-clipped=2.0 2023-04-28 23:15:29,958 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80736.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:15:59,673 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80749.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:16:02,673 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1374, 3.1650, 3.0988, 1.6968, 3.3024, 3.3919, 2.8497, 2.6401], device='cuda:4'), covar=tensor([0.0789, 0.0156, 0.0181, 0.1150, 0.0074, 0.0099, 0.0328, 0.0400], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0093, 0.0079, 0.0137, 0.0064, 0.0085, 0.0114, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 23:16:03,266 INFO [train.py:904] (4/8) Epoch 8, batch 9700, loss[loss=0.1936, simple_loss=0.2842, pruned_loss=0.05147, over 15314.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2795, pruned_loss=0.04982, over 3055734.38 frames. ], batch size: 190, lr: 8.31e-03, grad_scale: 8.0 2023-04-28 23:16:46,516 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80773.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:17:38,581 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80797.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:17:46,348 INFO [train.py:904] (4/8) Epoch 8, batch 9750, loss[loss=0.1731, simple_loss=0.2751, pruned_loss=0.03553, over 16875.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2791, pruned_loss=0.05005, over 3047049.62 frames. ], batch size: 90, lr: 8.31e-03, grad_scale: 8.0 2023-04-28 23:18:08,281 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.771e+02 2.800e+02 3.435e+02 4.029e+02 7.858e+02, threshold=6.871e+02, percent-clipped=2.0 2023-04-28 23:18:21,145 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80819.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:18:25,928 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3889, 3.5843, 2.4457, 2.0749, 2.3246, 2.0239, 3.7053, 3.2148], device='cuda:4'), covar=tensor([0.2790, 0.0768, 0.1746, 0.2224, 0.2483, 0.2038, 0.0445, 0.0986], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0241, 0.0267, 0.0253, 0.0244, 0.0203, 0.0245, 0.0256], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:18:55,723 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80834.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:19:24,254 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80850.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:19:26,321 INFO [train.py:904] (4/8) Epoch 8, batch 9800, loss[loss=0.1719, simple_loss=0.2741, pruned_loss=0.03482, over 16773.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2789, pruned_loss=0.04893, over 3061378.77 frames. ], batch size: 83, lr: 8.31e-03, grad_scale: 8.0 2023-04-28 23:21:11,973 INFO [train.py:904] (4/8) Epoch 8, batch 9850, loss[loss=0.1813, simple_loss=0.2738, pruned_loss=0.04446, over 16906.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2795, pruned_loss=0.04852, over 3068211.73 frames. ], batch size: 109, lr: 8.30e-03, grad_scale: 8.0 2023-04-28 23:21:33,315 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.375e+02 2.396e+02 3.091e+02 3.765e+02 1.204e+03, threshold=6.182e+02, percent-clipped=3.0 2023-04-28 23:22:37,298 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80939.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:22:43,139 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1010, 2.8614, 2.8193, 1.9093, 2.6109, 2.2427, 2.5625, 2.9654], device='cuda:4'), covar=tensor([0.0328, 0.0671, 0.0460, 0.1559, 0.0663, 0.0784, 0.0785, 0.0709], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0124, 0.0152, 0.0138, 0.0130, 0.0123, 0.0130, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-28 23:23:04,039 INFO [train.py:904] (4/8) Epoch 8, batch 9900, loss[loss=0.2028, simple_loss=0.3014, pruned_loss=0.05209, over 16273.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2797, pruned_loss=0.04846, over 3048672.29 frames. ], batch size: 166, lr: 8.30e-03, grad_scale: 8.0 2023-04-28 23:24:10,403 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2018, 3.3357, 3.5650, 3.5144, 3.5459, 3.3268, 3.4119, 3.4070], device='cuda:4'), covar=tensor([0.0348, 0.0590, 0.0520, 0.0546, 0.0507, 0.0461, 0.0671, 0.0435], device='cuda:4'), in_proj_covar=tensor([0.0269, 0.0268, 0.0274, 0.0269, 0.0309, 0.0289, 0.0373, 0.0234], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-28 23:24:29,639 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=80987.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:25:03,327 INFO [train.py:904] (4/8) Epoch 8, batch 9950, loss[loss=0.1675, simple_loss=0.2645, pruned_loss=0.03524, over 16911.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.282, pruned_loss=0.04889, over 3057342.78 frames. ], batch size: 102, lr: 8.30e-03, grad_scale: 8.0 2023-04-28 23:25:29,444 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.549e+02 2.486e+02 3.063e+02 3.674e+02 7.431e+02, threshold=6.127e+02, percent-clipped=1.0 2023-04-28 23:25:46,971 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0035, 3.2205, 3.3931, 1.7506, 3.5702, 3.5926, 2.7940, 2.6664], device='cuda:4'), covar=tensor([0.0844, 0.0162, 0.0122, 0.1212, 0.0045, 0.0085, 0.0342, 0.0427], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0093, 0.0077, 0.0136, 0.0063, 0.0084, 0.0113, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-28 23:27:01,919 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81049.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:27:07,186 INFO [train.py:904] (4/8) Epoch 8, batch 10000, loss[loss=0.1526, simple_loss=0.2381, pruned_loss=0.03355, over 17178.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2795, pruned_loss=0.04751, over 3087635.11 frames. ], batch size: 46, lr: 8.30e-03, grad_scale: 8.0 2023-04-28 23:27:19,154 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5561, 3.6204, 3.3514, 3.1964, 3.1921, 3.5012, 3.2874, 3.2591], device='cuda:4'), covar=tensor([0.0539, 0.0531, 0.0248, 0.0201, 0.0588, 0.0410, 0.0975, 0.0408], device='cuda:4'), in_proj_covar=tensor([0.0203, 0.0236, 0.0236, 0.0209, 0.0252, 0.0240, 0.0162, 0.0268], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-28 23:27:42,879 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3295, 5.7173, 5.4799, 5.5305, 5.0214, 4.9866, 5.1279, 5.7898], device='cuda:4'), covar=tensor([0.0889, 0.0784, 0.0802, 0.0482, 0.0852, 0.0662, 0.0810, 0.0817], device='cuda:4'), in_proj_covar=tensor([0.0453, 0.0580, 0.0475, 0.0394, 0.0356, 0.0383, 0.0476, 0.0432], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:28:30,671 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81092.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:28:40,810 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=81097.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:28:47,639 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-28 23:28:47,984 INFO [train.py:904] (4/8) Epoch 8, batch 10050, loss[loss=0.1985, simple_loss=0.2991, pruned_loss=0.04898, over 16183.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2788, pruned_loss=0.04712, over 3077949.68 frames. ], batch size: 165, lr: 8.29e-03, grad_scale: 8.0 2023-04-28 23:29:08,273 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.565e+02 2.518e+02 2.989e+02 3.596e+02 8.754e+02, threshold=5.978e+02, percent-clipped=2.0 2023-04-28 23:29:15,172 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81116.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:29:21,111 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81119.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:29:30,191 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 23:29:40,309 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81129.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:29:47,623 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5128, 4.5270, 4.2093, 3.8556, 4.2997, 1.6535, 4.1011, 4.2047], device='cuda:4'), covar=tensor([0.0075, 0.0085, 0.0181, 0.0238, 0.0105, 0.2234, 0.0117, 0.0139], device='cuda:4'), in_proj_covar=tensor([0.0108, 0.0096, 0.0139, 0.0129, 0.0111, 0.0165, 0.0125, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:30:18,756 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81150.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:30:21,529 INFO [train.py:904] (4/8) Epoch 8, batch 10100, loss[loss=0.1936, simple_loss=0.2793, pruned_loss=0.05393, over 16691.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.28, pruned_loss=0.04805, over 3051598.68 frames. ], batch size: 134, lr: 8.29e-03, grad_scale: 8.0 2023-04-28 23:30:26,931 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3568, 4.5063, 4.6053, 4.4871, 4.4575, 4.9790, 4.5530, 4.2423], device='cuda:4'), covar=tensor([0.1117, 0.1518, 0.1302, 0.1642, 0.2403, 0.0889, 0.1261, 0.2092], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0416, 0.0443, 0.0367, 0.0482, 0.0462, 0.0356, 0.0484], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:30:51,849 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=81167.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:31:14,362 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=81177.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:31:23,307 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9583, 3.9885, 4.4077, 4.3165, 4.3467, 4.0167, 4.0527, 3.9758], device='cuda:4'), covar=tensor([0.0273, 0.0588, 0.0341, 0.0450, 0.0405, 0.0362, 0.0805, 0.0456], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0274, 0.0276, 0.0273, 0.0311, 0.0293, 0.0377, 0.0239], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-28 23:31:37,453 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=81198.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:32:08,879 INFO [train.py:904] (4/8) Epoch 9, batch 0, loss[loss=0.312, simple_loss=0.3507, pruned_loss=0.1367, over 16868.00 frames. ], tot_loss[loss=0.312, simple_loss=0.3507, pruned_loss=0.1367, over 16868.00 frames. ], batch size: 116, lr: 7.85e-03, grad_scale: 8.0 2023-04-28 23:32:08,879 INFO [train.py:929] (4/8) Computing validation loss 2023-04-28 23:32:16,262 INFO [train.py:938] (4/8) Epoch 9, validation: loss=0.1602, simple_loss=0.2637, pruned_loss=0.02837, over 944034.00 frames. 2023-04-28 23:32:16,263 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-28 23:32:36,784 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.828e+02 3.571e+02 4.681e+02 1.164e+03, threshold=7.142e+02, percent-clipped=9.0 2023-04-28 23:33:25,079 INFO [train.py:904] (4/8) Epoch 9, batch 50, loss[loss=0.1813, simple_loss=0.2618, pruned_loss=0.05039, over 16837.00 frames. ], tot_loss[loss=0.2204, simple_loss=0.2975, pruned_loss=0.07163, over 751226.85 frames. ], batch size: 39, lr: 7.85e-03, grad_scale: 1.0 2023-04-28 23:33:37,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9508, 3.3734, 2.8876, 5.2017, 4.5375, 4.8044, 1.8444, 3.6446], device='cuda:4'), covar=tensor([0.1328, 0.0591, 0.1089, 0.0110, 0.0246, 0.0315, 0.1427, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0153, 0.0174, 0.0117, 0.0180, 0.0202, 0.0173, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 23:33:57,874 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6302, 1.4950, 2.1379, 2.5025, 2.5733, 2.3873, 1.7046, 2.6156], device='cuda:4'), covar=tensor([0.0113, 0.0310, 0.0189, 0.0160, 0.0139, 0.0144, 0.0289, 0.0066], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0159, 0.0145, 0.0142, 0.0149, 0.0106, 0.0158, 0.0096], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-28 23:34:31,257 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81299.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:34:34,820 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81301.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:34:35,495 INFO [train.py:904] (4/8) Epoch 9, batch 100, loss[loss=0.1873, simple_loss=0.2596, pruned_loss=0.05746, over 16801.00 frames. ], tot_loss[loss=0.2096, simple_loss=0.289, pruned_loss=0.0651, over 1323319.83 frames. ], batch size: 83, lr: 7.84e-03, grad_scale: 1.0 2023-04-28 23:34:54,536 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.676e+02 3.088e+02 4.301e+02 1.054e+03, threshold=6.177e+02, percent-clipped=2.0 2023-04-28 23:35:38,050 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9512, 3.9919, 3.7670, 3.6453, 3.5232, 3.9178, 3.6518, 3.6567], device='cuda:4'), covar=tensor([0.0541, 0.0364, 0.0233, 0.0212, 0.0621, 0.0333, 0.0745, 0.0467], device='cuda:4'), in_proj_covar=tensor([0.0213, 0.0249, 0.0248, 0.0221, 0.0269, 0.0253, 0.0170, 0.0285], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:35:42,947 INFO [train.py:904] (4/8) Epoch 9, batch 150, loss[loss=0.198, simple_loss=0.2892, pruned_loss=0.05336, over 17009.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2872, pruned_loss=0.06398, over 1765559.56 frames. ], batch size: 55, lr: 7.84e-03, grad_scale: 1.0 2023-04-28 23:35:55,871 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=81360.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:35:58,222 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=81362.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:36:40,026 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81392.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:36:54,428 INFO [train.py:904] (4/8) Epoch 9, batch 200, loss[loss=0.2307, simple_loss=0.3134, pruned_loss=0.074, over 16580.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2858, pruned_loss=0.06337, over 2109376.70 frames. ], batch size: 62, lr: 7.84e-03, grad_scale: 1.0 2023-04-28 23:37:01,891 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9336, 4.1299, 2.1997, 4.6116, 2.9862, 4.5817, 2.2315, 3.2670], device='cuda:4'), covar=tensor([0.0206, 0.0330, 0.1529, 0.0148, 0.0796, 0.0386, 0.1492, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0158, 0.0184, 0.0106, 0.0164, 0.0194, 0.0193, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 23:37:13,043 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.798e+02 2.694e+02 3.180e+02 3.951e+02 1.154e+03, threshold=6.361e+02, percent-clipped=2.0 2023-04-28 23:37:31,970 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81429.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:37:45,242 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=81440.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:38:01,972 INFO [train.py:904] (4/8) Epoch 9, batch 250, loss[loss=0.192, simple_loss=0.2759, pruned_loss=0.05401, over 17112.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2839, pruned_loss=0.06277, over 2382558.73 frames. ], batch size: 48, lr: 7.84e-03, grad_scale: 1.0 2023-04-28 23:38:29,687 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81472.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:38:32,232 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7426, 4.6924, 5.3182, 5.2724, 5.2923, 4.9104, 4.8295, 4.6543], device='cuda:4'), covar=tensor([0.0306, 0.0492, 0.0385, 0.0397, 0.0387, 0.0319, 0.0851, 0.0406], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0296, 0.0298, 0.0291, 0.0332, 0.0316, 0.0407, 0.0256], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 23:38:36,871 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=81477.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:38:40,553 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0279, 2.6118, 2.2929, 2.8944, 2.5416, 3.2204, 1.7291, 2.7736], device='cuda:4'), covar=tensor([0.0974, 0.0454, 0.0896, 0.0119, 0.0190, 0.0322, 0.1101, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0150, 0.0172, 0.0119, 0.0183, 0.0202, 0.0170, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 23:39:10,399 INFO [train.py:904] (4/8) Epoch 9, batch 300, loss[loss=0.2224, simple_loss=0.2878, pruned_loss=0.07846, over 16880.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.281, pruned_loss=0.06181, over 2580981.83 frames. ], batch size: 116, lr: 7.83e-03, grad_scale: 1.0 2023-04-28 23:39:15,228 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8468, 3.0104, 2.4599, 4.3846, 3.6967, 4.2478, 1.5384, 2.9542], device='cuda:4'), covar=tensor([0.1182, 0.0519, 0.1035, 0.0121, 0.0231, 0.0323, 0.1268, 0.0720], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0151, 0.0173, 0.0120, 0.0184, 0.0204, 0.0171, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-28 23:39:29,675 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.655e+02 2.475e+02 2.919e+02 3.755e+02 7.155e+02, threshold=5.837e+02, percent-clipped=3.0 2023-04-28 23:40:04,163 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.66 vs. limit=5.0 2023-04-28 23:40:17,833 INFO [train.py:904] (4/8) Epoch 9, batch 350, loss[loss=0.2255, simple_loss=0.2823, pruned_loss=0.08437, over 16799.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2782, pruned_loss=0.06049, over 2737124.23 frames. ], batch size: 83, lr: 7.83e-03, grad_scale: 1.0 2023-04-28 23:41:23,824 INFO [train.py:904] (4/8) Epoch 9, batch 400, loss[loss=0.2268, simple_loss=0.295, pruned_loss=0.07931, over 16318.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2761, pruned_loss=0.0598, over 2868602.39 frames. ], batch size: 165, lr: 7.83e-03, grad_scale: 2.0 2023-04-28 23:41:43,585 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.573e+02 2.386e+02 2.843e+02 3.463e+02 6.249e+02, threshold=5.687e+02, percent-clipped=1.0 2023-04-28 23:41:52,460 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6689, 3.7775, 2.1351, 3.9689, 2.7504, 3.8568, 2.0745, 2.9626], device='cuda:4'), covar=tensor([0.0158, 0.0252, 0.1360, 0.0155, 0.0713, 0.0550, 0.1340, 0.0553], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0160, 0.0183, 0.0109, 0.0164, 0.0196, 0.0192, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 23:42:02,576 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1528, 4.0203, 4.1716, 4.3428, 4.4527, 4.0593, 4.2868, 4.4243], device='cuda:4'), covar=tensor([0.1154, 0.0892, 0.1222, 0.0688, 0.0508, 0.0994, 0.1365, 0.0730], device='cuda:4'), in_proj_covar=tensor([0.0495, 0.0605, 0.0749, 0.0615, 0.0460, 0.0455, 0.0485, 0.0533], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:42:05,615 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9649, 5.0268, 5.5451, 5.4941, 5.5138, 5.0901, 5.0414, 4.8143], device='cuda:4'), covar=tensor([0.0275, 0.0437, 0.0349, 0.0373, 0.0381, 0.0287, 0.0854, 0.0375], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0303, 0.0303, 0.0297, 0.0339, 0.0322, 0.0415, 0.0260], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 23:42:16,454 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7854, 4.8336, 5.4206, 5.3541, 5.4136, 4.9622, 4.8909, 4.6643], device='cuda:4'), covar=tensor([0.0295, 0.0446, 0.0466, 0.0449, 0.0397, 0.0315, 0.0838, 0.0377], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0303, 0.0303, 0.0297, 0.0339, 0.0322, 0.0416, 0.0260], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 23:42:19,700 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6132, 3.9946, 4.2959, 4.2108, 4.2710, 3.8534, 3.5412, 3.8880], device='cuda:4'), covar=tensor([0.0659, 0.0888, 0.0630, 0.0694, 0.0630, 0.0680, 0.1523, 0.0674], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0304, 0.0303, 0.0297, 0.0339, 0.0322, 0.0416, 0.0260], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 23:42:33,320 INFO [train.py:904] (4/8) Epoch 9, batch 450, loss[loss=0.2003, simple_loss=0.2679, pruned_loss=0.06637, over 16715.00 frames. ], tot_loss[loss=0.1962, simple_loss=0.2745, pruned_loss=0.05901, over 2972546.61 frames. ], batch size: 124, lr: 7.83e-03, grad_scale: 2.0 2023-04-28 23:42:35,275 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-28 23:42:37,318 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81655.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:42:40,762 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81657.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:42:46,522 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2806, 3.7291, 3.8664, 2.0815, 3.0640, 2.4403, 3.7204, 3.6517], device='cuda:4'), covar=tensor([0.0289, 0.0673, 0.0431, 0.1680, 0.0726, 0.0908, 0.0619, 0.0944], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0132, 0.0154, 0.0140, 0.0133, 0.0124, 0.0132, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 23:42:58,057 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81670.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:43:42,339 INFO [train.py:904] (4/8) Epoch 9, batch 500, loss[loss=0.1771, simple_loss=0.2554, pruned_loss=0.04942, over 16401.00 frames. ], tot_loss[loss=0.1943, simple_loss=0.2728, pruned_loss=0.05785, over 3036683.08 frames. ], batch size: 75, lr: 7.82e-03, grad_scale: 2.0 2023-04-28 23:44:01,900 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.555e+02 2.446e+02 2.865e+02 3.583e+02 8.926e+02, threshold=5.729e+02, percent-clipped=4.0 2023-04-28 23:44:22,619 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=81731.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 23:44:47,429 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0437, 5.1694, 4.8907, 4.6820, 4.1682, 5.0910, 4.9892, 4.6477], device='cuda:4'), covar=tensor([0.0741, 0.0371, 0.0369, 0.0290, 0.1418, 0.0364, 0.0232, 0.0676], device='cuda:4'), in_proj_covar=tensor([0.0228, 0.0267, 0.0263, 0.0235, 0.0289, 0.0271, 0.0181, 0.0303], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:44:50,228 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4803, 2.9916, 2.8313, 1.9258, 2.5414, 2.0080, 3.0388, 3.0681], device='cuda:4'), covar=tensor([0.0260, 0.0623, 0.0593, 0.1724, 0.0824, 0.1010, 0.0512, 0.0697], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0134, 0.0155, 0.0141, 0.0133, 0.0124, 0.0132, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-28 23:44:50,833 INFO [train.py:904] (4/8) Epoch 9, batch 550, loss[loss=0.2159, simple_loss=0.2844, pruned_loss=0.07366, over 15536.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2725, pruned_loss=0.05718, over 3096914.23 frames. ], batch size: 191, lr: 7.82e-03, grad_scale: 2.0 2023-04-28 23:45:19,931 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81772.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:46:02,608 INFO [train.py:904] (4/8) Epoch 9, batch 600, loss[loss=0.2143, simple_loss=0.2822, pruned_loss=0.07315, over 16528.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.2727, pruned_loss=0.05762, over 3140937.67 frames. ], batch size: 68, lr: 7.82e-03, grad_scale: 2.0 2023-04-28 23:46:21,576 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.753e+02 2.670e+02 3.296e+02 4.052e+02 8.860e+02, threshold=6.592e+02, percent-clipped=6.0 2023-04-28 23:46:26,575 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=81820.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:46:37,386 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 23:47:09,566 INFO [train.py:904] (4/8) Epoch 9, batch 650, loss[loss=0.1475, simple_loss=0.2338, pruned_loss=0.03065, over 17010.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2711, pruned_loss=0.05692, over 3176828.82 frames. ], batch size: 41, lr: 7.82e-03, grad_scale: 2.0 2023-04-28 23:48:18,160 INFO [train.py:904] (4/8) Epoch 9, batch 700, loss[loss=0.2372, simple_loss=0.3001, pruned_loss=0.08709, over 15437.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2714, pruned_loss=0.05691, over 3193331.25 frames. ], batch size: 190, lr: 7.81e-03, grad_scale: 2.0 2023-04-28 23:48:37,201 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.675e+02 2.553e+02 3.037e+02 4.738e+02 2.852e+03, threshold=6.073e+02, percent-clipped=12.0 2023-04-28 23:49:25,153 INFO [train.py:904] (4/8) Epoch 9, batch 750, loss[loss=0.2185, simple_loss=0.2811, pruned_loss=0.07789, over 12347.00 frames. ], tot_loss[loss=0.1919, simple_loss=0.2709, pruned_loss=0.05651, over 3218061.95 frames. ], batch size: 248, lr: 7.81e-03, grad_scale: 2.0 2023-04-28 23:49:29,070 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81955.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:49:31,589 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81957.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:50:14,426 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81987.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:50:37,832 INFO [train.py:904] (4/8) Epoch 9, batch 800, loss[loss=0.2148, simple_loss=0.2865, pruned_loss=0.07161, over 12496.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2708, pruned_loss=0.05561, over 3242174.57 frames. ], batch size: 246, lr: 7.81e-03, grad_scale: 4.0 2023-04-28 23:50:39,916 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=82003.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:50:42,267 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=82005.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:50:56,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.806e+02 2.619e+02 2.968e+02 3.371e+02 5.495e+02, threshold=5.935e+02, percent-clipped=0.0 2023-04-28 23:51:10,868 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82026.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 23:51:25,488 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1826, 5.0518, 4.9823, 4.6166, 4.5219, 4.9155, 5.1066, 4.6034], device='cuda:4'), covar=tensor([0.0574, 0.0403, 0.0270, 0.0261, 0.1139, 0.0456, 0.0254, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0237, 0.0279, 0.0274, 0.0246, 0.0302, 0.0284, 0.0188, 0.0319], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 23:51:40,940 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82048.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:51:45,036 INFO [train.py:904] (4/8) Epoch 9, batch 850, loss[loss=0.2052, simple_loss=0.2897, pruned_loss=0.06036, over 17075.00 frames. ], tot_loss[loss=0.1898, simple_loss=0.2697, pruned_loss=0.0549, over 3266253.65 frames. ], batch size: 55, lr: 7.81e-03, grad_scale: 4.0 2023-04-28 23:51:49,540 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82055.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:51:58,607 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9925, 2.1606, 2.3410, 4.7763, 2.1526, 2.8292, 2.3758, 2.5214], device='cuda:4'), covar=tensor([0.0720, 0.3332, 0.1991, 0.0260, 0.3650, 0.2008, 0.2744, 0.3040], device='cuda:4'), in_proj_covar=tensor([0.0350, 0.0372, 0.0314, 0.0325, 0.0401, 0.0413, 0.0331, 0.0437], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:52:25,207 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1529, 5.4742, 5.1760, 5.2710, 4.8696, 4.7654, 4.9239, 5.5410], device='cuda:4'), covar=tensor([0.0893, 0.0857, 0.1100, 0.0575, 0.0742, 0.0848, 0.0892, 0.0917], device='cuda:4'), in_proj_covar=tensor([0.0496, 0.0641, 0.0534, 0.0437, 0.0396, 0.0412, 0.0530, 0.0471], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 23:52:33,550 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8975, 3.9843, 4.3376, 4.3164, 4.3562, 4.0203, 4.0762, 3.9934], device='cuda:4'), covar=tensor([0.0347, 0.0515, 0.0391, 0.0434, 0.0414, 0.0362, 0.0792, 0.0459], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0318, 0.0319, 0.0310, 0.0362, 0.0338, 0.0438, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-28 23:52:54,644 INFO [train.py:904] (4/8) Epoch 9, batch 900, loss[loss=0.1959, simple_loss=0.2643, pruned_loss=0.06372, over 16721.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2688, pruned_loss=0.05467, over 3270207.13 frames. ], batch size: 134, lr: 7.80e-03, grad_scale: 4.0 2023-04-28 23:53:13,855 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.661e+02 2.604e+02 3.157e+02 3.589e+02 5.638e+02, threshold=6.315e+02, percent-clipped=0.0 2023-04-28 23:53:15,028 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82116.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:54:03,806 INFO [train.py:904] (4/8) Epoch 9, batch 950, loss[loss=0.2553, simple_loss=0.3092, pruned_loss=0.1007, over 12265.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2689, pruned_loss=0.0549, over 3285436.38 frames. ], batch size: 248, lr: 7.80e-03, grad_scale: 4.0 2023-04-28 23:55:11,098 INFO [train.py:904] (4/8) Epoch 9, batch 1000, loss[loss=0.1702, simple_loss=0.2523, pruned_loss=0.04403, over 17206.00 frames. ], tot_loss[loss=0.1894, simple_loss=0.2679, pruned_loss=0.05546, over 3293726.05 frames. ], batch size: 44, lr: 7.80e-03, grad_scale: 4.0 2023-04-28 23:55:31,949 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.413e+02 2.342e+02 2.920e+02 3.464e+02 5.943e+02, threshold=5.839e+02, percent-clipped=0.0 2023-04-28 23:55:38,570 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-28 23:56:20,453 INFO [train.py:904] (4/8) Epoch 9, batch 1050, loss[loss=0.187, simple_loss=0.263, pruned_loss=0.05547, over 16819.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2672, pruned_loss=0.05461, over 3296974.98 frames. ], batch size: 102, lr: 7.80e-03, grad_scale: 4.0 2023-04-28 23:57:28,537 INFO [train.py:904] (4/8) Epoch 9, batch 1100, loss[loss=0.1815, simple_loss=0.2494, pruned_loss=0.05677, over 16414.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2665, pruned_loss=0.05484, over 3295530.17 frames. ], batch size: 75, lr: 7.80e-03, grad_scale: 4.0 2023-04-28 23:57:47,216 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.669e+02 2.514e+02 3.056e+02 3.616e+02 1.290e+03, threshold=6.113e+02, percent-clipped=7.0 2023-04-28 23:58:01,821 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=82326.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:58:24,253 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82343.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:58:28,046 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82346.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:58:35,308 INFO [train.py:904] (4/8) Epoch 9, batch 1150, loss[loss=0.1543, simple_loss=0.2407, pruned_loss=0.03389, over 16866.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2652, pruned_loss=0.05347, over 3300708.66 frames. ], batch size: 42, lr: 7.79e-03, grad_scale: 4.0 2023-04-28 23:59:04,894 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=82374.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:59:44,522 INFO [train.py:904] (4/8) Epoch 9, batch 1200, loss[loss=0.1622, simple_loss=0.2402, pruned_loss=0.04214, over 16963.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.264, pruned_loss=0.05268, over 3308619.09 frames. ], batch size: 41, lr: 7.79e-03, grad_scale: 8.0 2023-04-28 23:59:50,680 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82407.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 23:59:55,202 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82411.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:00:02,690 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.435e+02 2.375e+02 3.047e+02 3.937e+02 1.504e+03, threshold=6.095e+02, percent-clipped=2.0 2023-04-29 00:00:15,769 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7209, 4.5029, 4.6395, 4.9003, 5.0372, 4.5188, 4.9561, 5.0301], device='cuda:4'), covar=tensor([0.1383, 0.1080, 0.1705, 0.0680, 0.0562, 0.1035, 0.0884, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0519, 0.0634, 0.0796, 0.0648, 0.0486, 0.0481, 0.0509, 0.0562], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:00:18,393 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-29 00:00:50,234 INFO [train.py:904] (4/8) Epoch 9, batch 1250, loss[loss=0.193, simple_loss=0.2688, pruned_loss=0.05858, over 16452.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2655, pruned_loss=0.05398, over 3314957.96 frames. ], batch size: 68, lr: 7.79e-03, grad_scale: 4.0 2023-04-29 00:01:47,063 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82493.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 00:01:58,480 INFO [train.py:904] (4/8) Epoch 9, batch 1300, loss[loss=0.196, simple_loss=0.2877, pruned_loss=0.05219, over 17219.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2668, pruned_loss=0.05434, over 3319819.43 frames. ], batch size: 45, lr: 7.79e-03, grad_scale: 4.0 2023-04-29 00:02:18,042 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.645e+02 2.404e+02 2.965e+02 3.987e+02 6.881e+02, threshold=5.930e+02, percent-clipped=4.0 2023-04-29 00:03:05,222 INFO [train.py:904] (4/8) Epoch 9, batch 1350, loss[loss=0.2174, simple_loss=0.2966, pruned_loss=0.06914, over 16310.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2667, pruned_loss=0.05402, over 3305407.56 frames. ], batch size: 165, lr: 7.78e-03, grad_scale: 4.0 2023-04-29 00:03:08,002 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82554.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 00:03:29,861 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82569.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:04:12,575 INFO [train.py:904] (4/8) Epoch 9, batch 1400, loss[loss=0.1985, simple_loss=0.2751, pruned_loss=0.06094, over 16855.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2664, pruned_loss=0.05338, over 3314483.74 frames. ], batch size: 96, lr: 7.78e-03, grad_scale: 4.0 2023-04-29 00:04:33,568 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.521e+02 2.260e+02 2.970e+02 4.048e+02 1.598e+03, threshold=5.940e+02, percent-clipped=4.0 2023-04-29 00:04:53,180 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82630.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:05:09,542 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=82643.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:05:22,654 INFO [train.py:904] (4/8) Epoch 9, batch 1450, loss[loss=0.2116, simple_loss=0.2746, pruned_loss=0.07436, over 16864.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2665, pruned_loss=0.05385, over 3310536.10 frames. ], batch size: 116, lr: 7.78e-03, grad_scale: 4.0 2023-04-29 00:06:04,702 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7949, 3.8968, 2.0485, 4.1303, 2.9009, 4.0464, 2.2525, 2.9994], device='cuda:4'), covar=tensor([0.0150, 0.0220, 0.1398, 0.0132, 0.0628, 0.0481, 0.1201, 0.0592], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0161, 0.0183, 0.0115, 0.0164, 0.0202, 0.0193, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 00:06:09,919 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6644, 3.0428, 2.7005, 4.6714, 3.8402, 4.5273, 1.8674, 3.0504], device='cuda:4'), covar=tensor([0.1525, 0.0670, 0.1186, 0.0190, 0.0371, 0.0359, 0.1463, 0.0842], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0154, 0.0174, 0.0127, 0.0196, 0.0211, 0.0173, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 00:06:12,923 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2558, 5.1986, 5.1099, 4.6619, 4.6247, 5.1316, 5.2420, 4.7046], device='cuda:4'), covar=tensor([0.0544, 0.0450, 0.0253, 0.0290, 0.1174, 0.0341, 0.0238, 0.0687], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0286, 0.0280, 0.0253, 0.0307, 0.0287, 0.0192, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:06:16,062 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=82691.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:06:29,714 INFO [train.py:904] (4/8) Epoch 9, batch 1500, loss[loss=0.1639, simple_loss=0.2416, pruned_loss=0.04313, over 15909.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2649, pruned_loss=0.05351, over 3313157.39 frames. ], batch size: 35, lr: 7.78e-03, grad_scale: 4.0 2023-04-29 00:06:29,979 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82702.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:06:42,792 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=82711.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:06:51,612 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.560e+02 2.540e+02 3.048e+02 3.562e+02 8.559e+02, threshold=6.096e+02, percent-clipped=4.0 2023-04-29 00:07:39,172 INFO [train.py:904] (4/8) Epoch 9, batch 1550, loss[loss=0.2347, simple_loss=0.2955, pruned_loss=0.08695, over 16879.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2677, pruned_loss=0.05559, over 3320929.37 frames. ], batch size: 109, lr: 7.77e-03, grad_scale: 4.0 2023-04-29 00:07:49,786 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=82759.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:08:48,744 INFO [train.py:904] (4/8) Epoch 9, batch 1600, loss[loss=0.2335, simple_loss=0.3146, pruned_loss=0.07621, over 15402.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2704, pruned_loss=0.05642, over 3299536.02 frames. ], batch size: 190, lr: 7.77e-03, grad_scale: 8.0 2023-04-29 00:09:09,709 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.618e+02 2.591e+02 3.276e+02 4.037e+02 8.145e+02, threshold=6.551e+02, percent-clipped=5.0 2023-04-29 00:09:52,306 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82849.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 00:09:56,036 INFO [train.py:904] (4/8) Epoch 9, batch 1650, loss[loss=0.2259, simple_loss=0.3023, pruned_loss=0.0747, over 16814.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2721, pruned_loss=0.0569, over 3306354.09 frames. ], batch size: 90, lr: 7.77e-03, grad_scale: 8.0 2023-04-29 00:11:05,716 INFO [train.py:904] (4/8) Epoch 9, batch 1700, loss[loss=0.1964, simple_loss=0.2888, pruned_loss=0.05195, over 17078.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2723, pruned_loss=0.0565, over 3321065.23 frames. ], batch size: 55, lr: 7.77e-03, grad_scale: 8.0 2023-04-29 00:11:24,536 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.504e+02 2.790e+02 3.377e+02 4.170e+02 9.157e+02, threshold=6.754e+02, percent-clipped=3.0 2023-04-29 00:11:37,133 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82925.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:12:13,189 INFO [train.py:904] (4/8) Epoch 9, batch 1750, loss[loss=0.1857, simple_loss=0.2626, pruned_loss=0.05442, over 16906.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2724, pruned_loss=0.05588, over 3323194.32 frames. ], batch size: 90, lr: 7.76e-03, grad_scale: 8.0 2023-04-29 00:12:28,844 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82963.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:12:34,856 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0799, 4.5589, 4.6740, 3.3835, 4.0871, 4.6712, 4.2248, 2.9950], device='cuda:4'), covar=tensor([0.0294, 0.0051, 0.0026, 0.0219, 0.0046, 0.0051, 0.0040, 0.0261], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0067, 0.0066, 0.0121, 0.0071, 0.0080, 0.0071, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:13:19,613 INFO [train.py:904] (4/8) Epoch 9, batch 1800, loss[loss=0.2241, simple_loss=0.2993, pruned_loss=0.07441, over 15562.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2739, pruned_loss=0.0564, over 3312406.61 frames. ], batch size: 191, lr: 7.76e-03, grad_scale: 8.0 2023-04-29 00:13:20,533 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83002.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:13:40,257 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.771e+02 2.500e+02 2.970e+02 3.668e+02 6.624e+02, threshold=5.940e+02, percent-clipped=0.0 2023-04-29 00:13:50,155 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=83024.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:14:26,664 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=83050.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:14:28,691 INFO [train.py:904] (4/8) Epoch 9, batch 1850, loss[loss=0.1919, simple_loss=0.29, pruned_loss=0.04693, over 17111.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2745, pruned_loss=0.05637, over 3308303.76 frames. ], batch size: 48, lr: 7.76e-03, grad_scale: 8.0 2023-04-29 00:15:11,084 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-29 00:15:37,693 INFO [train.py:904] (4/8) Epoch 9, batch 1900, loss[loss=0.1924, simple_loss=0.2681, pruned_loss=0.05835, over 16773.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2737, pruned_loss=0.0551, over 3309194.13 frames. ], batch size: 124, lr: 7.76e-03, grad_scale: 8.0 2023-04-29 00:15:56,126 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1985, 4.8705, 5.1381, 5.3605, 5.5792, 4.8316, 5.5115, 5.5012], device='cuda:4'), covar=tensor([0.1261, 0.1001, 0.1388, 0.0561, 0.0447, 0.0647, 0.0426, 0.0446], device='cuda:4'), in_proj_covar=tensor([0.0519, 0.0633, 0.0798, 0.0647, 0.0486, 0.0487, 0.0504, 0.0561], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:15:59,151 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.589e+02 2.476e+02 2.876e+02 3.479e+02 6.930e+02, threshold=5.751e+02, percent-clipped=2.0 2023-04-29 00:16:42,381 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83149.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 00:16:46,826 INFO [train.py:904] (4/8) Epoch 9, batch 1950, loss[loss=0.1659, simple_loss=0.2498, pruned_loss=0.04103, over 17224.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2747, pruned_loss=0.05521, over 3298297.62 frames. ], batch size: 44, lr: 7.76e-03, grad_scale: 8.0 2023-04-29 00:17:08,558 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-29 00:17:47,198 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4822, 5.9509, 5.6935, 5.7447, 5.2812, 5.2371, 5.4831, 6.0764], device='cuda:4'), covar=tensor([0.1036, 0.0913, 0.0987, 0.0746, 0.0825, 0.0661, 0.0844, 0.0760], device='cuda:4'), in_proj_covar=tensor([0.0514, 0.0655, 0.0546, 0.0448, 0.0406, 0.0420, 0.0540, 0.0482], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:17:48,399 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83196.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:17:49,289 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=83197.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 00:17:55,080 INFO [train.py:904] (4/8) Epoch 9, batch 2000, loss[loss=0.2453, simple_loss=0.3167, pruned_loss=0.08696, over 15513.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.274, pruned_loss=0.05482, over 3307296.53 frames. ], batch size: 190, lr: 7.75e-03, grad_scale: 8.0 2023-04-29 00:18:02,727 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8149, 1.8948, 2.3036, 2.8169, 2.4996, 3.3381, 2.2005, 3.0981], device='cuda:4'), covar=tensor([0.0152, 0.0307, 0.0223, 0.0171, 0.0195, 0.0115, 0.0277, 0.0126], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0166, 0.0151, 0.0152, 0.0157, 0.0114, 0.0164, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 00:18:17,512 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.705e+02 2.480e+02 2.857e+02 3.443e+02 6.900e+02, threshold=5.715e+02, percent-clipped=1.0 2023-04-29 00:18:26,839 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83225.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:19:04,146 INFO [train.py:904] (4/8) Epoch 9, batch 2050, loss[loss=0.2018, simple_loss=0.2754, pruned_loss=0.06411, over 16511.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2741, pruned_loss=0.05477, over 3307771.25 frames. ], batch size: 75, lr: 7.75e-03, grad_scale: 8.0 2023-04-29 00:19:12,740 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=83257.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:19:33,894 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=83273.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:19:48,980 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83284.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:20:05,352 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9452, 4.2794, 4.4615, 2.0089, 4.7081, 4.7512, 3.3237, 3.7197], device='cuda:4'), covar=tensor([0.0700, 0.0123, 0.0134, 0.1095, 0.0047, 0.0093, 0.0367, 0.0307], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0097, 0.0087, 0.0141, 0.0070, 0.0098, 0.0121, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 00:20:13,196 INFO [train.py:904] (4/8) Epoch 9, batch 2100, loss[loss=0.1698, simple_loss=0.2467, pruned_loss=0.04644, over 16764.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2746, pruned_loss=0.05502, over 3316047.57 frames. ], batch size: 39, lr: 7.75e-03, grad_scale: 8.0 2023-04-29 00:20:17,732 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0147, 5.4210, 5.4633, 5.3444, 5.3583, 5.9288, 5.4609, 5.1754], device='cuda:4'), covar=tensor([0.0852, 0.1528, 0.1627, 0.1774, 0.2474, 0.0894, 0.1324, 0.2422], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0476, 0.0507, 0.0422, 0.0558, 0.0536, 0.0406, 0.0563], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:20:33,165 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2580, 3.2314, 1.9093, 3.4568, 2.4412, 3.4413, 1.8507, 2.6707], device='cuda:4'), covar=tensor([0.0202, 0.0405, 0.1473, 0.0193, 0.0785, 0.0483, 0.1312, 0.0602], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0166, 0.0186, 0.0120, 0.0167, 0.0208, 0.0195, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 00:20:35,021 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.834e+02 2.645e+02 3.321e+02 4.089e+02 1.048e+03, threshold=6.642e+02, percent-clipped=7.0 2023-04-29 00:20:37,089 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=83319.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:21:05,998 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8995, 4.5451, 4.8438, 5.0827, 5.3201, 4.5176, 5.2444, 5.2102], device='cuda:4'), covar=tensor([0.1402, 0.1185, 0.1764, 0.0718, 0.0490, 0.0844, 0.0533, 0.0499], device='cuda:4'), in_proj_covar=tensor([0.0520, 0.0635, 0.0802, 0.0649, 0.0485, 0.0484, 0.0504, 0.0560], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:21:11,872 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=83345.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:21:20,735 INFO [train.py:904] (4/8) Epoch 9, batch 2150, loss[loss=0.1915, simple_loss=0.2807, pruned_loss=0.05113, over 17252.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2756, pruned_loss=0.0555, over 3304403.32 frames. ], batch size: 52, lr: 7.75e-03, grad_scale: 8.0 2023-04-29 00:21:48,253 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7598, 5.1393, 4.8821, 4.8764, 4.5760, 4.5765, 4.5957, 5.1659], device='cuda:4'), covar=tensor([0.0948, 0.0814, 0.0930, 0.0639, 0.0673, 0.0890, 0.0864, 0.0852], device='cuda:4'), in_proj_covar=tensor([0.0511, 0.0655, 0.0542, 0.0448, 0.0404, 0.0417, 0.0539, 0.0486], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:22:07,838 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0185, 2.6094, 2.6151, 1.7798, 2.7977, 2.7768, 2.3845, 2.3509], device='cuda:4'), covar=tensor([0.0788, 0.0207, 0.0245, 0.1032, 0.0103, 0.0201, 0.0456, 0.0487], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0098, 0.0087, 0.0142, 0.0070, 0.0098, 0.0121, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 00:22:31,367 INFO [train.py:904] (4/8) Epoch 9, batch 2200, loss[loss=0.2321, simple_loss=0.3085, pruned_loss=0.07782, over 11979.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2764, pruned_loss=0.05642, over 3298816.97 frames. ], batch size: 246, lr: 7.74e-03, grad_scale: 8.0 2023-04-29 00:22:54,054 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.989e+02 2.797e+02 3.398e+02 4.283e+02 9.169e+02, threshold=6.797e+02, percent-clipped=3.0 2023-04-29 00:23:41,053 INFO [train.py:904] (4/8) Epoch 9, batch 2250, loss[loss=0.2007, simple_loss=0.2784, pruned_loss=0.06147, over 16842.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2778, pruned_loss=0.05773, over 3296944.37 frames. ], batch size: 102, lr: 7.74e-03, grad_scale: 8.0 2023-04-29 00:23:49,783 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8927, 1.7809, 2.3320, 2.8570, 2.6516, 2.9752, 1.9676, 3.0549], device='cuda:4'), covar=tensor([0.0125, 0.0303, 0.0210, 0.0162, 0.0172, 0.0136, 0.0308, 0.0083], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0165, 0.0150, 0.0152, 0.0158, 0.0115, 0.0164, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 00:24:49,197 INFO [train.py:904] (4/8) Epoch 9, batch 2300, loss[loss=0.1879, simple_loss=0.2593, pruned_loss=0.05818, over 16874.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2764, pruned_loss=0.05711, over 3307146.05 frames. ], batch size: 90, lr: 7.74e-03, grad_scale: 8.0 2023-04-29 00:25:12,015 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.453e+02 2.674e+02 3.137e+02 4.026e+02 1.366e+03, threshold=6.274e+02, percent-clipped=5.0 2023-04-29 00:25:59,051 INFO [train.py:904] (4/8) Epoch 9, batch 2350, loss[loss=0.1814, simple_loss=0.272, pruned_loss=0.04542, over 17115.00 frames. ], tot_loss[loss=0.1963, simple_loss=0.2769, pruned_loss=0.05782, over 3308235.74 frames. ], batch size: 49, lr: 7.74e-03, grad_scale: 8.0 2023-04-29 00:25:59,382 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=83552.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:26:11,706 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7012, 4.6296, 4.8511, 4.7798, 4.6559, 5.3854, 4.9341, 4.6263], device='cuda:4'), covar=tensor([0.1273, 0.1839, 0.1780, 0.1898, 0.3217, 0.0999, 0.1358, 0.2497], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0473, 0.0502, 0.0418, 0.0553, 0.0528, 0.0401, 0.0559], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:26:51,075 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4754, 4.8602, 4.6002, 4.6811, 4.3280, 4.3854, 4.3271, 4.9140], device='cuda:4'), covar=tensor([0.0970, 0.0796, 0.0933, 0.0611, 0.0756, 0.1072, 0.0905, 0.0800], device='cuda:4'), in_proj_covar=tensor([0.0514, 0.0660, 0.0550, 0.0450, 0.0407, 0.0419, 0.0545, 0.0486], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:27:06,437 INFO [train.py:904] (4/8) Epoch 9, batch 2400, loss[loss=0.1597, simple_loss=0.245, pruned_loss=0.03721, over 17253.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.2781, pruned_loss=0.05813, over 3311370.96 frames. ], batch size: 45, lr: 7.73e-03, grad_scale: 8.0 2023-04-29 00:27:29,672 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.715e+02 2.500e+02 2.959e+02 3.579e+02 6.970e+02, threshold=5.919e+02, percent-clipped=2.0 2023-04-29 00:27:31,274 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83619.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:27:49,039 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-04-29 00:27:59,099 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=83640.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:28:14,956 INFO [train.py:904] (4/8) Epoch 9, batch 2450, loss[loss=0.2283, simple_loss=0.3099, pruned_loss=0.07336, over 16692.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2787, pruned_loss=0.05799, over 3305589.96 frames. ], batch size: 62, lr: 7.73e-03, grad_scale: 8.0 2023-04-29 00:28:35,098 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=83667.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:28:58,748 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-29 00:29:22,512 INFO [train.py:904] (4/8) Epoch 9, batch 2500, loss[loss=0.171, simple_loss=0.2533, pruned_loss=0.04437, over 16782.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2782, pruned_loss=0.05797, over 3296450.78 frames. ], batch size: 39, lr: 7.73e-03, grad_scale: 8.0 2023-04-29 00:29:44,586 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.720e+02 2.505e+02 2.999e+02 3.639e+02 7.354e+02, threshold=5.999e+02, percent-clipped=3.0 2023-04-29 00:30:28,782 INFO [train.py:904] (4/8) Epoch 9, batch 2550, loss[loss=0.1886, simple_loss=0.2822, pruned_loss=0.0475, over 17120.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.2786, pruned_loss=0.0576, over 3298488.03 frames. ], batch size: 49, lr: 7.73e-03, grad_scale: 8.0 2023-04-29 00:30:41,862 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83760.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:31:09,249 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0100, 4.2725, 2.3315, 4.7665, 3.0697, 4.7076, 2.4450, 3.2838], device='cuda:4'), covar=tensor([0.0203, 0.0286, 0.1478, 0.0138, 0.0695, 0.0336, 0.1390, 0.0581], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0165, 0.0184, 0.0119, 0.0164, 0.0206, 0.0193, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 00:31:38,614 INFO [train.py:904] (4/8) Epoch 9, batch 2600, loss[loss=0.1985, simple_loss=0.2886, pruned_loss=0.0542, over 17006.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2787, pruned_loss=0.05744, over 3300996.86 frames. ], batch size: 50, lr: 7.73e-03, grad_scale: 8.0 2023-04-29 00:31:59,379 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.575e+02 2.494e+02 3.064e+02 3.833e+02 6.863e+02, threshold=6.129e+02, percent-clipped=2.0 2023-04-29 00:32:04,249 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=83821.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:32:04,374 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-29 00:32:45,547 INFO [train.py:904] (4/8) Epoch 9, batch 2650, loss[loss=0.1719, simple_loss=0.2704, pruned_loss=0.0367, over 17043.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2785, pruned_loss=0.0566, over 3312548.71 frames. ], batch size: 53, lr: 7.72e-03, grad_scale: 8.0 2023-04-29 00:32:45,900 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83852.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:32:53,098 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7308, 3.8540, 2.9393, 2.2654, 2.6318, 2.3145, 3.9439, 3.5398], device='cuda:4'), covar=tensor([0.2145, 0.0509, 0.1298, 0.2103, 0.2220, 0.1606, 0.0442, 0.0969], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0255, 0.0277, 0.0268, 0.0280, 0.0215, 0.0261, 0.0288], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:33:16,604 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.61 vs. limit=5.0 2023-04-29 00:33:51,888 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=83900.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:33:53,983 INFO [train.py:904] (4/8) Epoch 9, batch 2700, loss[loss=0.2495, simple_loss=0.3133, pruned_loss=0.09279, over 12316.00 frames. ], tot_loss[loss=0.1948, simple_loss=0.2781, pruned_loss=0.05576, over 3313938.42 frames. ], batch size: 246, lr: 7.72e-03, grad_scale: 8.0 2023-04-29 00:34:17,451 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.417e+02 2.501e+02 2.856e+02 3.411e+02 7.585e+02, threshold=5.711e+02, percent-clipped=1.0 2023-04-29 00:34:47,077 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83940.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:34:49,600 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-29 00:34:58,219 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83949.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:35:02,460 INFO [train.py:904] (4/8) Epoch 9, batch 2750, loss[loss=0.2109, simple_loss=0.2825, pruned_loss=0.06969, over 16725.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2779, pruned_loss=0.05551, over 3322701.20 frames. ], batch size: 124, lr: 7.72e-03, grad_scale: 8.0 2023-04-29 00:35:41,685 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0708, 5.4520, 5.5638, 5.3129, 5.3689, 5.9633, 5.4678, 5.2534], device='cuda:4'), covar=tensor([0.0796, 0.1574, 0.1564, 0.1959, 0.2376, 0.0876, 0.1260, 0.2152], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0471, 0.0500, 0.0416, 0.0548, 0.0522, 0.0398, 0.0556], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:35:48,328 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83985.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:35:49,548 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7795, 4.6570, 4.6656, 4.4714, 4.3077, 4.7020, 4.5602, 4.4239], device='cuda:4'), covar=tensor([0.0500, 0.0412, 0.0233, 0.0211, 0.0815, 0.0333, 0.0317, 0.0572], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0298, 0.0288, 0.0262, 0.0314, 0.0294, 0.0197, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:35:51,695 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=83988.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:35:52,244 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-29 00:36:15,269 INFO [train.py:904] (4/8) Epoch 9, batch 2800, loss[loss=0.1689, simple_loss=0.2468, pruned_loss=0.04552, over 16778.00 frames. ], tot_loss[loss=0.1954, simple_loss=0.2786, pruned_loss=0.05608, over 3324055.51 frames. ], batch size: 83, lr: 7.72e-03, grad_scale: 8.0 2023-04-29 00:36:25,614 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84010.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 00:36:36,295 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.816e+02 2.751e+02 3.290e+02 3.928e+02 7.223e+02, threshold=6.581e+02, percent-clipped=1.0 2023-04-29 00:37:16,011 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84046.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:37:23,281 INFO [train.py:904] (4/8) Epoch 9, batch 2850, loss[loss=0.3061, simple_loss=0.354, pruned_loss=0.1291, over 11957.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2777, pruned_loss=0.05615, over 3317418.81 frames. ], batch size: 246, lr: 7.71e-03, grad_scale: 8.0 2023-04-29 00:38:25,982 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84097.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:38:32,592 INFO [train.py:904] (4/8) Epoch 9, batch 2900, loss[loss=0.1971, simple_loss=0.2605, pruned_loss=0.0669, over 16913.00 frames. ], tot_loss[loss=0.1935, simple_loss=0.2761, pruned_loss=0.05549, over 3328831.69 frames. ], batch size: 109, lr: 7.71e-03, grad_scale: 8.0 2023-04-29 00:38:52,442 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84116.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:38:54,442 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.514e+02 2.515e+02 3.057e+02 3.611e+02 6.139e+02, threshold=6.114e+02, percent-clipped=0.0 2023-04-29 00:39:23,271 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9762, 3.9797, 3.8204, 3.6441, 3.5875, 3.9392, 3.5959, 3.7240], device='cuda:4'), covar=tensor([0.0627, 0.0490, 0.0276, 0.0259, 0.0766, 0.0426, 0.0949, 0.0559], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0296, 0.0285, 0.0260, 0.0312, 0.0293, 0.0195, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:39:27,539 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-29 00:39:27,654 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-04-29 00:39:43,177 INFO [train.py:904] (4/8) Epoch 9, batch 2950, loss[loss=0.1786, simple_loss=0.2633, pruned_loss=0.04696, over 16840.00 frames. ], tot_loss[loss=0.1935, simple_loss=0.2753, pruned_loss=0.05589, over 3333017.14 frames. ], batch size: 42, lr: 7.71e-03, grad_scale: 8.0 2023-04-29 00:39:51,105 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84158.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:40:47,983 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84199.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:40:52,809 INFO [train.py:904] (4/8) Epoch 9, batch 3000, loss[loss=0.1796, simple_loss=0.2703, pruned_loss=0.04447, over 17115.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2752, pruned_loss=0.05612, over 3333966.55 frames. ], batch size: 48, lr: 7.71e-03, grad_scale: 8.0 2023-04-29 00:40:52,809 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 00:41:02,059 INFO [train.py:938] (4/8) Epoch 9, validation: loss=0.1444, simple_loss=0.2507, pruned_loss=0.019, over 944034.00 frames. 2023-04-29 00:41:02,060 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-29 00:41:23,147 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.086e+02 2.665e+02 3.010e+02 4.048e+02 8.692e+02, threshold=6.019e+02, percent-clipped=1.0 2023-04-29 00:41:40,119 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-04-29 00:41:47,945 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84236.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:42:09,948 INFO [train.py:904] (4/8) Epoch 9, batch 3050, loss[loss=0.1645, simple_loss=0.2488, pruned_loss=0.04009, over 15830.00 frames. ], tot_loss[loss=0.1938, simple_loss=0.2749, pruned_loss=0.05632, over 3335887.12 frames. ], batch size: 35, lr: 7.70e-03, grad_scale: 8.0 2023-04-29 00:42:20,929 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84260.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:42:59,853 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6689, 3.7391, 4.0751, 2.7307, 3.6781, 4.0790, 3.7702, 2.4581], device='cuda:4'), covar=tensor([0.0341, 0.0187, 0.0031, 0.0255, 0.0066, 0.0056, 0.0053, 0.0294], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0069, 0.0068, 0.0125, 0.0075, 0.0084, 0.0076, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:43:06,686 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9936, 3.1156, 3.1754, 1.9870, 2.9492, 3.2444, 2.9399, 1.9379], device='cuda:4'), covar=tensor([0.0359, 0.0076, 0.0033, 0.0300, 0.0069, 0.0061, 0.0063, 0.0296], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0069, 0.0068, 0.0125, 0.0075, 0.0084, 0.0076, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:43:11,781 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84297.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:43:17,164 INFO [train.py:904] (4/8) Epoch 9, batch 3100, loss[loss=0.1824, simple_loss=0.2711, pruned_loss=0.04688, over 17244.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2733, pruned_loss=0.05554, over 3337515.11 frames. ], batch size: 45, lr: 7.70e-03, grad_scale: 8.0 2023-04-29 00:43:22,115 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84305.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 00:43:39,988 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.846e+02 2.519e+02 3.109e+02 3.888e+02 1.112e+03, threshold=6.217e+02, percent-clipped=5.0 2023-04-29 00:43:57,203 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6022, 3.7951, 2.9610, 2.1528, 2.5721, 2.2151, 3.7808, 3.5212], device='cuda:4'), covar=tensor([0.2375, 0.0556, 0.1335, 0.2226, 0.2255, 0.1716, 0.0431, 0.1006], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0257, 0.0279, 0.0269, 0.0282, 0.0216, 0.0263, 0.0292], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:44:06,117 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84336.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:44:12,239 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84341.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:44:28,373 INFO [train.py:904] (4/8) Epoch 9, batch 3150, loss[loss=0.1856, simple_loss=0.2768, pruned_loss=0.04725, over 17123.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2734, pruned_loss=0.05539, over 3324195.20 frames. ], batch size: 47, lr: 7.70e-03, grad_scale: 8.0 2023-04-29 00:44:40,099 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0194, 2.4169, 1.7959, 2.0989, 2.8851, 2.5941, 3.0768, 3.0531], device='cuda:4'), covar=tensor([0.0097, 0.0261, 0.0381, 0.0336, 0.0132, 0.0223, 0.0165, 0.0146], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0193, 0.0187, 0.0190, 0.0190, 0.0193, 0.0200, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:44:47,604 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-29 00:44:59,655 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7028, 4.2039, 4.3279, 3.0397, 3.8523, 4.2637, 3.8178, 2.1169], device='cuda:4'), covar=tensor([0.0372, 0.0049, 0.0036, 0.0260, 0.0066, 0.0076, 0.0084, 0.0394], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0068, 0.0067, 0.0122, 0.0074, 0.0083, 0.0075, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:45:30,314 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84397.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:45:36,485 INFO [train.py:904] (4/8) Epoch 9, batch 3200, loss[loss=0.2296, simple_loss=0.3029, pruned_loss=0.07819, over 12241.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2725, pruned_loss=0.05505, over 3324129.04 frames. ], batch size: 246, lr: 7.70e-03, grad_scale: 8.0 2023-04-29 00:45:56,355 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84416.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:45:59,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.715e+02 2.601e+02 3.159e+02 4.189e+02 1.012e+03, threshold=6.318e+02, percent-clipped=3.0 2023-04-29 00:46:45,526 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-29 00:46:45,891 INFO [train.py:904] (4/8) Epoch 9, batch 3250, loss[loss=0.1744, simple_loss=0.2642, pruned_loss=0.04233, over 16830.00 frames. ], tot_loss[loss=0.1914, simple_loss=0.2725, pruned_loss=0.05518, over 3326073.38 frames. ], batch size: 96, lr: 7.70e-03, grad_scale: 8.0 2023-04-29 00:46:47,424 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84453.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:47:03,267 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=84464.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:47:32,644 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.78 vs. limit=5.0 2023-04-29 00:47:55,748 INFO [train.py:904] (4/8) Epoch 9, batch 3300, loss[loss=0.2061, simple_loss=0.2874, pruned_loss=0.06244, over 16814.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2732, pruned_loss=0.05585, over 3316658.55 frames. ], batch size: 102, lr: 7.69e-03, grad_scale: 8.0 2023-04-29 00:48:14,269 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-29 00:48:18,089 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.545e+02 2.660e+02 3.024e+02 3.740e+02 8.260e+02, threshold=6.049e+02, percent-clipped=1.0 2023-04-29 00:48:27,856 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84525.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:49:04,553 INFO [train.py:904] (4/8) Epoch 9, batch 3350, loss[loss=0.1681, simple_loss=0.2659, pruned_loss=0.03521, over 17098.00 frames. ], tot_loss[loss=0.1932, simple_loss=0.2744, pruned_loss=0.05603, over 3317356.11 frames. ], batch size: 49, lr: 7.69e-03, grad_scale: 8.0 2023-04-29 00:49:10,250 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84555.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:49:53,267 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84586.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:49:56,527 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-29 00:50:01,510 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84592.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:50:16,095 INFO [train.py:904] (4/8) Epoch 9, batch 3400, loss[loss=0.2042, simple_loss=0.2748, pruned_loss=0.06676, over 15577.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2735, pruned_loss=0.05506, over 3318799.36 frames. ], batch size: 191, lr: 7.69e-03, grad_scale: 8.0 2023-04-29 00:50:20,126 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84605.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:50:38,674 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.593e+02 2.352e+02 2.993e+02 3.807e+02 8.102e+02, threshold=5.985e+02, percent-clipped=2.0 2023-04-29 00:50:59,428 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84633.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:51:00,466 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5150, 3.5601, 3.1945, 3.0591, 3.1238, 3.4176, 3.2314, 3.2398], device='cuda:4'), covar=tensor([0.0478, 0.0424, 0.0254, 0.0242, 0.0580, 0.0311, 0.1335, 0.0428], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0304, 0.0294, 0.0266, 0.0320, 0.0303, 0.0201, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 00:51:10,771 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84641.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:51:24,587 INFO [train.py:904] (4/8) Epoch 9, batch 3450, loss[loss=0.1825, simple_loss=0.2705, pruned_loss=0.04724, over 17238.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2727, pruned_loss=0.05463, over 3320048.84 frames. ], batch size: 52, lr: 7.69e-03, grad_scale: 8.0 2023-04-29 00:51:26,678 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=84653.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:52:16,069 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=84689.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:52:17,343 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2746, 3.2289, 3.3026, 1.7429, 3.4818, 3.5257, 2.8657, 2.6045], device='cuda:4'), covar=tensor([0.0717, 0.0159, 0.0184, 0.1142, 0.0077, 0.0126, 0.0369, 0.0415], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0096, 0.0087, 0.0139, 0.0069, 0.0100, 0.0121, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 00:52:20,946 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84692.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:52:24,262 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84694.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:52:35,106 INFO [train.py:904] (4/8) Epoch 9, batch 3500, loss[loss=0.1975, simple_loss=0.2704, pruned_loss=0.06234, over 16824.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2714, pruned_loss=0.05432, over 3311738.59 frames. ], batch size: 102, lr: 7.68e-03, grad_scale: 8.0 2023-04-29 00:52:56,975 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.766e+02 2.287e+02 2.792e+02 3.603e+02 8.093e+02, threshold=5.583e+02, percent-clipped=2.0 2023-04-29 00:53:44,891 INFO [train.py:904] (4/8) Epoch 9, batch 3550, loss[loss=0.1925, simple_loss=0.267, pruned_loss=0.05902, over 16625.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2703, pruned_loss=0.05433, over 3318006.82 frames. ], batch size: 89, lr: 7.68e-03, grad_scale: 8.0 2023-04-29 00:53:47,071 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84753.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:53:59,012 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8453, 2.1418, 2.2567, 4.6027, 2.0900, 2.7164, 2.2380, 2.3753], device='cuda:4'), covar=tensor([0.0775, 0.3238, 0.2025, 0.0327, 0.3629, 0.2154, 0.2914, 0.3148], device='cuda:4'), in_proj_covar=tensor([0.0360, 0.0380, 0.0319, 0.0330, 0.0405, 0.0431, 0.0340, 0.0448], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:54:53,077 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84800.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:54:54,906 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=84801.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:54:55,837 INFO [train.py:904] (4/8) Epoch 9, batch 3600, loss[loss=0.202, simple_loss=0.2686, pruned_loss=0.06769, over 16708.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.27, pruned_loss=0.05429, over 3302801.17 frames. ], batch size: 124, lr: 7.68e-03, grad_scale: 8.0 2023-04-29 00:55:06,039 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3899, 5.8785, 5.5592, 5.6974, 5.1645, 4.9845, 5.2823, 5.9895], device='cuda:4'), covar=tensor([0.1199, 0.1028, 0.1122, 0.0640, 0.0870, 0.0709, 0.1008, 0.0937], device='cuda:4'), in_proj_covar=tensor([0.0524, 0.0668, 0.0548, 0.0451, 0.0411, 0.0422, 0.0550, 0.0494], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:55:17,925 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.706e+02 2.544e+02 2.999e+02 3.772e+02 8.043e+02, threshold=5.998e+02, percent-clipped=5.0 2023-04-29 00:55:50,655 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-29 00:56:07,144 INFO [train.py:904] (4/8) Epoch 9, batch 3650, loss[loss=0.1936, simple_loss=0.2622, pruned_loss=0.06251, over 16666.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2711, pruned_loss=0.05515, over 3310607.56 frames. ], batch size: 89, lr: 7.68e-03, grad_scale: 8.0 2023-04-29 00:56:12,816 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84855.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:56:21,165 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84861.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:56:50,511 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84881.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:57:07,438 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84892.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:57:21,437 INFO [train.py:904] (4/8) Epoch 9, batch 3700, loss[loss=0.1963, simple_loss=0.2622, pruned_loss=0.06524, over 16768.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2704, pruned_loss=0.05697, over 3305188.69 frames. ], batch size: 83, lr: 7.68e-03, grad_scale: 8.0 2023-04-29 00:57:23,506 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=84903.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:57:38,128 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-29 00:57:46,406 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.577e+02 2.579e+02 2.947e+02 3.773e+02 6.988e+02, threshold=5.894e+02, percent-clipped=1.0 2023-04-29 00:57:56,329 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-29 00:58:05,452 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.65 vs. limit=5.0 2023-04-29 00:58:19,041 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=84940.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:58:24,298 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6310, 2.2174, 2.3905, 4.3824, 2.1991, 2.6569, 2.2768, 2.4487], device='cuda:4'), covar=tensor([0.0825, 0.3147, 0.1815, 0.0312, 0.3282, 0.2047, 0.2900, 0.2603], device='cuda:4'), in_proj_covar=tensor([0.0362, 0.0384, 0.0322, 0.0332, 0.0407, 0.0435, 0.0343, 0.0452], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 00:58:34,936 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84951.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:58:35,654 INFO [train.py:904] (4/8) Epoch 9, batch 3750, loss[loss=0.2066, simple_loss=0.2737, pruned_loss=0.06971, over 16253.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2709, pruned_loss=0.05825, over 3279032.64 frames. ], batch size: 165, lr: 7.67e-03, grad_scale: 8.0 2023-04-29 00:59:28,297 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84989.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:59:34,010 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84992.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 00:59:47,554 INFO [train.py:904] (4/8) Epoch 9, batch 3800, loss[loss=0.2207, simple_loss=0.2908, pruned_loss=0.07523, over 16862.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2719, pruned_loss=0.05994, over 3277702.09 frames. ], batch size: 109, lr: 7.67e-03, grad_scale: 8.0 2023-04-29 01:00:03,361 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85012.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:00:11,014 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.825e+02 2.585e+02 2.931e+02 3.484e+02 5.921e+02, threshold=5.862e+02, percent-clipped=1.0 2023-04-29 01:00:43,934 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=85040.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:01:00,353 INFO [train.py:904] (4/8) Epoch 9, batch 3850, loss[loss=0.2194, simple_loss=0.2769, pruned_loss=0.08095, over 16722.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.2716, pruned_loss=0.06058, over 3282760.04 frames. ], batch size: 134, lr: 7.67e-03, grad_scale: 8.0 2023-04-29 01:01:19,053 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9634, 1.8225, 2.4506, 2.8408, 2.7873, 2.9677, 1.8414, 2.9974], device='cuda:4'), covar=tensor([0.0089, 0.0283, 0.0177, 0.0173, 0.0150, 0.0104, 0.0280, 0.0086], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0164, 0.0151, 0.0153, 0.0158, 0.0115, 0.0162, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 01:01:37,101 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-04-29 01:02:02,840 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8203, 3.9859, 3.1419, 2.4170, 2.8529, 2.4201, 4.1881, 3.7296], device='cuda:4'), covar=tensor([0.2209, 0.0530, 0.1298, 0.1978, 0.2219, 0.1598, 0.0365, 0.0901], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0256, 0.0279, 0.0271, 0.0284, 0.0217, 0.0262, 0.0292], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:02:13,000 INFO [train.py:904] (4/8) Epoch 9, batch 3900, loss[loss=0.1912, simple_loss=0.2651, pruned_loss=0.05868, over 16884.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2711, pruned_loss=0.06117, over 3292173.83 frames. ], batch size: 109, lr: 7.67e-03, grad_scale: 8.0 2023-04-29 01:02:21,679 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85107.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:02:32,537 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85115.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:02:36,232 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.667e+02 2.478e+02 2.844e+02 3.462e+02 6.399e+02, threshold=5.687e+02, percent-clipped=3.0 2023-04-29 01:03:09,964 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1530, 4.0920, 4.5046, 4.5167, 4.5292, 4.1565, 4.2796, 4.1169], device='cuda:4'), covar=tensor([0.0302, 0.0580, 0.0338, 0.0388, 0.0450, 0.0347, 0.0756, 0.0518], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0324, 0.0326, 0.0313, 0.0369, 0.0341, 0.0447, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 01:03:22,416 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1761, 4.1654, 4.0843, 3.9379, 3.8344, 4.1377, 3.8138, 3.9552], device='cuda:4'), covar=tensor([0.0583, 0.0434, 0.0252, 0.0244, 0.0724, 0.0389, 0.0836, 0.0566], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0293, 0.0286, 0.0258, 0.0310, 0.0295, 0.0194, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 01:03:24,357 INFO [train.py:904] (4/8) Epoch 9, batch 3950, loss[loss=0.2055, simple_loss=0.2736, pruned_loss=0.06867, over 16461.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2706, pruned_loss=0.06172, over 3295209.28 frames. ], batch size: 68, lr: 7.66e-03, grad_scale: 8.0 2023-04-29 01:03:32,139 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85156.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:03:50,097 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85168.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 01:04:00,701 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85176.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:04:07,243 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85181.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:04:28,178 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.10 vs. limit=5.0 2023-04-29 01:04:37,098 INFO [train.py:904] (4/8) Epoch 9, batch 4000, loss[loss=0.1876, simple_loss=0.268, pruned_loss=0.05358, over 16313.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2707, pruned_loss=0.06163, over 3286928.09 frames. ], batch size: 165, lr: 7.66e-03, grad_scale: 16.0 2023-04-29 01:05:00,771 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.397e+02 2.490e+02 3.071e+02 3.606e+02 5.108e+02, threshold=6.141e+02, percent-clipped=0.0 2023-04-29 01:05:13,690 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.00 vs. limit=5.0 2023-04-29 01:05:17,722 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=85229.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:05:51,193 INFO [train.py:904] (4/8) Epoch 9, batch 4050, loss[loss=0.2048, simple_loss=0.2909, pruned_loss=0.05939, over 16832.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2704, pruned_loss=0.05991, over 3279737.85 frames. ], batch size: 83, lr: 7.66e-03, grad_scale: 16.0 2023-04-29 01:06:45,647 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85289.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:07:04,621 INFO [train.py:904] (4/8) Epoch 9, batch 4100, loss[loss=0.2051, simple_loss=0.2794, pruned_loss=0.0654, over 12005.00 frames. ], tot_loss[loss=0.1943, simple_loss=0.2711, pruned_loss=0.05874, over 3279359.39 frames. ], batch size: 247, lr: 7.66e-03, grad_scale: 16.0 2023-04-29 01:07:12,416 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85307.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:07:28,042 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 2.051e+02 2.369e+02 2.790e+02 6.834e+02, threshold=4.737e+02, percent-clipped=1.0 2023-04-29 01:07:34,513 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7874, 3.9770, 2.2443, 4.4572, 2.9181, 4.3923, 2.2724, 3.1007], device='cuda:4'), covar=tensor([0.0185, 0.0287, 0.1458, 0.0067, 0.0661, 0.0253, 0.1330, 0.0567], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0164, 0.0183, 0.0117, 0.0164, 0.0205, 0.0190, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 01:07:57,981 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=85337.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:08:20,246 INFO [train.py:904] (4/8) Epoch 9, batch 4150, loss[loss=0.1999, simple_loss=0.2974, pruned_loss=0.05117, over 16736.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2791, pruned_loss=0.06214, over 3219987.12 frames. ], batch size: 134, lr: 7.66e-03, grad_scale: 16.0 2023-04-29 01:08:57,421 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-29 01:09:37,254 INFO [train.py:904] (4/8) Epoch 9, batch 4200, loss[loss=0.2456, simple_loss=0.326, pruned_loss=0.08265, over 15219.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2873, pruned_loss=0.06454, over 3197894.79 frames. ], batch size: 190, lr: 7.65e-03, grad_scale: 16.0 2023-04-29 01:10:02,492 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.568e+02 2.639e+02 3.175e+02 3.973e+02 9.081e+02, threshold=6.349e+02, percent-clipped=14.0 2023-04-29 01:10:17,691 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85428.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:10:45,447 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4033, 4.0297, 4.0030, 2.7052, 3.5263, 3.8856, 3.6277, 2.3363], device='cuda:4'), covar=tensor([0.0348, 0.0033, 0.0039, 0.0248, 0.0049, 0.0084, 0.0051, 0.0296], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0065, 0.0066, 0.0120, 0.0072, 0.0082, 0.0072, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 01:10:52,280 INFO [train.py:904] (4/8) Epoch 9, batch 4250, loss[loss=0.2055, simple_loss=0.2943, pruned_loss=0.05831, over 16313.00 frames. ], tot_loss[loss=0.2101, simple_loss=0.2906, pruned_loss=0.06476, over 3178309.34 frames. ], batch size: 165, lr: 7.65e-03, grad_scale: 16.0 2023-04-29 01:10:59,123 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85456.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:11:08,338 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85463.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 01:11:10,232 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6005, 4.8949, 4.7084, 4.6871, 4.3234, 4.3460, 4.3310, 4.9884], device='cuda:4'), covar=tensor([0.0877, 0.0761, 0.0843, 0.0577, 0.0712, 0.0866, 0.0864, 0.0713], device='cuda:4'), in_proj_covar=tensor([0.0489, 0.0620, 0.0515, 0.0424, 0.0386, 0.0399, 0.0512, 0.0463], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:11:20,279 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85471.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:11:27,760 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-29 01:11:39,072 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85483.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:11:48,326 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85489.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:12:07,380 INFO [train.py:904] (4/8) Epoch 9, batch 4300, loss[loss=0.2191, simple_loss=0.3064, pruned_loss=0.06592, over 16551.00 frames. ], tot_loss[loss=0.2084, simple_loss=0.2907, pruned_loss=0.06302, over 3195387.37 frames. ], batch size: 75, lr: 7.65e-03, grad_scale: 16.0 2023-04-29 01:12:11,407 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=85504.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:12:30,789 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.243e+02 2.412e+02 2.949e+02 3.511e+02 9.601e+02, threshold=5.898e+02, percent-clipped=2.0 2023-04-29 01:12:49,224 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85530.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:13:10,300 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85544.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:13:21,362 INFO [train.py:904] (4/8) Epoch 9, batch 4350, loss[loss=0.233, simple_loss=0.3159, pruned_loss=0.07506, over 16803.00 frames. ], tot_loss[loss=0.2107, simple_loss=0.2937, pruned_loss=0.06381, over 3195166.46 frames. ], batch size: 116, lr: 7.65e-03, grad_scale: 16.0 2023-04-29 01:13:24,230 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2773, 5.6053, 5.2707, 5.3672, 4.9914, 4.7321, 4.9371, 5.6702], device='cuda:4'), covar=tensor([0.0723, 0.0621, 0.0901, 0.0548, 0.0665, 0.0740, 0.0818, 0.0705], device='cuda:4'), in_proj_covar=tensor([0.0489, 0.0616, 0.0514, 0.0424, 0.0385, 0.0400, 0.0513, 0.0463], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:14:18,980 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85591.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:14:30,676 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8537, 3.3837, 3.3494, 1.9899, 3.0593, 3.2558, 3.1074, 1.6934], device='cuda:4'), covar=tensor([0.0425, 0.0033, 0.0046, 0.0360, 0.0071, 0.0082, 0.0060, 0.0412], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0064, 0.0064, 0.0119, 0.0071, 0.0080, 0.0071, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 01:14:34,737 INFO [train.py:904] (4/8) Epoch 9, batch 4400, loss[loss=0.2479, simple_loss=0.3162, pruned_loss=0.08978, over 11637.00 frames. ], tot_loss[loss=0.2132, simple_loss=0.296, pruned_loss=0.06514, over 3189230.01 frames. ], batch size: 247, lr: 7.64e-03, grad_scale: 16.0 2023-04-29 01:14:41,659 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85607.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:14:56,972 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.792e+02 2.712e+02 3.095e+02 3.599e+02 6.298e+02, threshold=6.190e+02, percent-clipped=2.0 2023-04-29 01:15:06,905 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85625.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:15:46,021 INFO [train.py:904] (4/8) Epoch 9, batch 4450, loss[loss=0.211, simple_loss=0.2928, pruned_loss=0.06467, over 16398.00 frames. ], tot_loss[loss=0.2147, simple_loss=0.2987, pruned_loss=0.06538, over 3214976.83 frames. ], batch size: 35, lr: 7.64e-03, grad_scale: 16.0 2023-04-29 01:15:50,416 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=85655.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:15:58,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2504, 4.2162, 4.0727, 3.5075, 4.1385, 1.6564, 3.9243, 3.7757], device='cuda:4'), covar=tensor([0.0064, 0.0056, 0.0103, 0.0303, 0.0066, 0.2327, 0.0104, 0.0149], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0108, 0.0156, 0.0149, 0.0126, 0.0169, 0.0142, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:16:33,299 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85686.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 01:16:44,352 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4812, 3.5596, 1.8501, 4.0144, 2.5713, 3.8732, 1.9860, 2.7334], device='cuda:4'), covar=tensor([0.0175, 0.0294, 0.1698, 0.0060, 0.0729, 0.0387, 0.1437, 0.0664], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0161, 0.0181, 0.0112, 0.0162, 0.0200, 0.0190, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 01:16:56,347 INFO [train.py:904] (4/8) Epoch 9, batch 4500, loss[loss=0.2073, simple_loss=0.2951, pruned_loss=0.05974, over 16831.00 frames. ], tot_loss[loss=0.2154, simple_loss=0.299, pruned_loss=0.06594, over 3213145.39 frames. ], batch size: 83, lr: 7.64e-03, grad_scale: 16.0 2023-04-29 01:17:20,314 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.479e+02 2.131e+02 2.570e+02 3.015e+02 5.229e+02, threshold=5.140e+02, percent-clipped=0.0 2023-04-29 01:18:07,071 INFO [train.py:904] (4/8) Epoch 9, batch 4550, loss[loss=0.2215, simple_loss=0.3085, pruned_loss=0.06718, over 17117.00 frames. ], tot_loss[loss=0.2162, simple_loss=0.2993, pruned_loss=0.06656, over 3224392.20 frames. ], batch size: 48, lr: 7.64e-03, grad_scale: 16.0 2023-04-29 01:18:23,727 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85763.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:18:34,972 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85771.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:18:54,374 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85784.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:19:20,645 INFO [train.py:904] (4/8) Epoch 9, batch 4600, loss[loss=0.2178, simple_loss=0.3005, pruned_loss=0.06762, over 16727.00 frames. ], tot_loss[loss=0.2166, simple_loss=0.3002, pruned_loss=0.06648, over 3233768.13 frames. ], batch size: 124, lr: 7.64e-03, grad_scale: 16.0 2023-04-29 01:19:32,653 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=85811.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:19:42,422 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.567e+02 2.222e+02 2.575e+02 3.036e+02 5.036e+02, threshold=5.150e+02, percent-clipped=0.0 2023-04-29 01:19:44,975 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=85819.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:20:12,955 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85839.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:20:30,003 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0954, 1.8750, 2.0390, 3.6431, 1.8866, 2.2932, 2.0578, 2.0437], device='cuda:4'), covar=tensor([0.0953, 0.3084, 0.2036, 0.0401, 0.3604, 0.2095, 0.2801, 0.2985], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0377, 0.0314, 0.0316, 0.0404, 0.0429, 0.0335, 0.0444], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:20:30,858 INFO [train.py:904] (4/8) Epoch 9, batch 4650, loss[loss=0.2079, simple_loss=0.2895, pruned_loss=0.06313, over 16299.00 frames. ], tot_loss[loss=0.2162, simple_loss=0.2992, pruned_loss=0.06657, over 3231953.47 frames. ], batch size: 165, lr: 7.63e-03, grad_scale: 16.0 2023-04-29 01:20:37,228 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1904, 5.2113, 4.8370, 4.5451, 5.1339, 1.8378, 4.8652, 4.8228], device='cuda:4'), covar=tensor([0.0039, 0.0030, 0.0104, 0.0225, 0.0044, 0.2067, 0.0063, 0.0115], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0105, 0.0154, 0.0145, 0.0123, 0.0166, 0.0139, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:20:52,695 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4355, 3.4786, 1.6119, 3.8451, 2.3997, 3.7788, 1.8720, 2.6803], device='cuda:4'), covar=tensor([0.0179, 0.0268, 0.1854, 0.0074, 0.0743, 0.0370, 0.1534, 0.0620], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0161, 0.0184, 0.0112, 0.0164, 0.0201, 0.0191, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 01:21:04,755 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85875.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:21:20,348 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85886.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:21:33,856 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.46 vs. limit=5.0 2023-04-29 01:21:42,720 INFO [train.py:904] (4/8) Epoch 9, batch 4700, loss[loss=0.2025, simple_loss=0.2833, pruned_loss=0.06084, over 16890.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2966, pruned_loss=0.06559, over 3220258.93 frames. ], batch size: 109, lr: 7.63e-03, grad_scale: 16.0 2023-04-29 01:21:43,515 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-29 01:21:50,619 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4634, 3.4374, 1.6466, 3.7745, 2.3046, 3.7223, 1.8595, 2.6423], device='cuda:4'), covar=tensor([0.0163, 0.0290, 0.1752, 0.0083, 0.0814, 0.0386, 0.1492, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0163, 0.0185, 0.0113, 0.0165, 0.0202, 0.0193, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 01:22:06,316 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.424e+02 2.169e+02 2.505e+02 2.903e+02 6.033e+02, threshold=5.010e+02, percent-clipped=1.0 2023-04-29 01:22:10,879 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-29 01:22:25,131 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85930.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 01:22:32,876 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85936.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:22:55,504 INFO [train.py:904] (4/8) Epoch 9, batch 4750, loss[loss=0.1839, simple_loss=0.2697, pruned_loss=0.04909, over 16822.00 frames. ], tot_loss[loss=0.2097, simple_loss=0.2925, pruned_loss=0.06346, over 3224668.52 frames. ], batch size: 102, lr: 7.63e-03, grad_scale: 16.0 2023-04-29 01:22:59,903 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2342, 4.2605, 4.4000, 4.2730, 4.2894, 4.7686, 4.3926, 4.0522], device='cuda:4'), covar=tensor([0.1403, 0.1516, 0.1444, 0.1770, 0.2387, 0.0963, 0.1189, 0.2371], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0447, 0.0473, 0.0395, 0.0523, 0.0506, 0.0385, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 01:23:04,401 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85957.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:23:39,897 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85981.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 01:23:55,326 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85991.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 01:24:13,773 INFO [train.py:904] (4/8) Epoch 9, batch 4800, loss[loss=0.182, simple_loss=0.2642, pruned_loss=0.0499, over 17045.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2895, pruned_loss=0.06138, over 3221103.37 frames. ], batch size: 53, lr: 7.63e-03, grad_scale: 16.0 2023-04-29 01:24:14,236 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86002.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:24:30,382 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1659, 3.3654, 3.5603, 3.5359, 3.5483, 3.3130, 3.3766, 3.4040], device='cuda:4'), covar=tensor([0.0361, 0.0568, 0.0434, 0.0424, 0.0469, 0.0422, 0.0817, 0.0428], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0303, 0.0314, 0.0300, 0.0353, 0.0326, 0.0432, 0.0260], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 01:24:37,221 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.418e+02 2.079e+02 2.389e+02 2.964e+02 6.421e+02, threshold=4.777e+02, percent-clipped=3.0 2023-04-29 01:24:38,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86018.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:25:00,985 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5367, 4.5590, 4.3456, 4.1382, 3.9428, 4.4729, 4.3310, 4.1447], device='cuda:4'), covar=tensor([0.0505, 0.0305, 0.0258, 0.0250, 0.0973, 0.0349, 0.0384, 0.0629], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0271, 0.0264, 0.0239, 0.0288, 0.0271, 0.0181, 0.0304], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:25:28,113 INFO [train.py:904] (4/8) Epoch 9, batch 4850, loss[loss=0.2378, simple_loss=0.324, pruned_loss=0.07578, over 15396.00 frames. ], tot_loss[loss=0.2068, simple_loss=0.2909, pruned_loss=0.06128, over 3202065.69 frames. ], batch size: 190, lr: 7.62e-03, grad_scale: 16.0 2023-04-29 01:25:28,606 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86052.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:25:45,099 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86063.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:25:54,705 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86069.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:26:16,456 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86084.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:26:42,283 INFO [train.py:904] (4/8) Epoch 9, batch 4900, loss[loss=0.1904, simple_loss=0.2678, pruned_loss=0.05647, over 17129.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.2898, pruned_loss=0.06029, over 3175615.06 frames. ], batch size: 49, lr: 7.62e-03, grad_scale: 16.0 2023-04-29 01:26:59,812 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86113.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:27:05,786 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.491e+02 2.278e+02 2.647e+02 3.027e+02 5.480e+02, threshold=5.294e+02, percent-clipped=1.0 2023-04-29 01:27:23,798 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86130.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:27:26,528 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86132.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:27:37,271 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86139.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:27:55,680 INFO [train.py:904] (4/8) Epoch 9, batch 4950, loss[loss=0.2333, simple_loss=0.3129, pruned_loss=0.07684, over 12128.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2893, pruned_loss=0.05943, over 3187662.44 frames. ], batch size: 247, lr: 7.62e-03, grad_scale: 16.0 2023-04-29 01:28:46,015 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86186.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:28:46,995 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86187.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:29:08,679 INFO [train.py:904] (4/8) Epoch 9, batch 5000, loss[loss=0.1857, simple_loss=0.2802, pruned_loss=0.04558, over 16859.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2912, pruned_loss=0.05999, over 3180216.55 frames. ], batch size: 96, lr: 7.62e-03, grad_scale: 16.0 2023-04-29 01:29:28,493 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.50 vs. limit=5.0 2023-04-29 01:29:32,025 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.770e+02 2.365e+02 2.931e+02 3.414e+02 6.733e+02, threshold=5.862e+02, percent-clipped=2.0 2023-04-29 01:29:50,440 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86231.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:29:54,567 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86234.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:30:11,832 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86246.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:30:19,715 INFO [train.py:904] (4/8) Epoch 9, batch 5050, loss[loss=0.2227, simple_loss=0.3091, pruned_loss=0.06817, over 15255.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.2911, pruned_loss=0.05963, over 3180887.43 frames. ], batch size: 190, lr: 7.62e-03, grad_scale: 16.0 2023-04-29 01:31:01,251 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86281.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:31:07,941 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86286.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 01:31:29,532 INFO [train.py:904] (4/8) Epoch 9, batch 5100, loss[loss=0.1701, simple_loss=0.2589, pruned_loss=0.04068, over 16470.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2883, pruned_loss=0.0582, over 3186873.00 frames. ], batch size: 68, lr: 7.61e-03, grad_scale: 16.0 2023-04-29 01:31:37,433 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86307.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:31:45,608 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86313.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:31:52,760 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.623e+02 2.422e+02 2.762e+02 3.223e+02 5.276e+02, threshold=5.525e+02, percent-clipped=0.0 2023-04-29 01:32:07,223 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-29 01:32:08,053 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86329.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:32:23,756 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-29 01:32:41,540 INFO [train.py:904] (4/8) Epoch 9, batch 5150, loss[loss=0.2032, simple_loss=0.2924, pruned_loss=0.05704, over 16928.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2883, pruned_loss=0.05743, over 3179673.38 frames. ], batch size: 109, lr: 7.61e-03, grad_scale: 16.0 2023-04-29 01:32:50,285 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86358.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:33:11,104 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6103, 4.5223, 4.5120, 3.7875, 4.4345, 1.5507, 4.1839, 4.3642], device='cuda:4'), covar=tensor([0.0079, 0.0071, 0.0112, 0.0402, 0.0082, 0.2257, 0.0126, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0106, 0.0153, 0.0146, 0.0122, 0.0168, 0.0137, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:33:54,430 INFO [train.py:904] (4/8) Epoch 9, batch 5200, loss[loss=0.182, simple_loss=0.2735, pruned_loss=0.04524, over 16822.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2874, pruned_loss=0.05722, over 3181534.91 frames. ], batch size: 83, lr: 7.61e-03, grad_scale: 16.0 2023-04-29 01:34:02,829 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86408.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:34:04,459 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.27 vs. limit=5.0 2023-04-29 01:34:16,833 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.23 vs. limit=5.0 2023-04-29 01:34:17,297 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.533e+02 2.298e+02 2.670e+02 3.115e+02 5.719e+02, threshold=5.340e+02, percent-clipped=1.0 2023-04-29 01:34:27,688 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86425.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:34:50,071 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86440.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:34:57,838 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86445.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:35:06,327 INFO [train.py:904] (4/8) Epoch 9, batch 5250, loss[loss=0.2172, simple_loss=0.3071, pruned_loss=0.06362, over 16877.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2846, pruned_loss=0.05684, over 3185992.40 frames. ], batch size: 116, lr: 7.61e-03, grad_scale: 16.0 2023-04-29 01:35:41,805 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6333, 3.7721, 2.9792, 2.2854, 2.6619, 2.3527, 3.9817, 3.6714], device='cuda:4'), covar=tensor([0.2402, 0.0676, 0.1347, 0.1988, 0.1984, 0.1539, 0.0431, 0.0782], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0252, 0.0274, 0.0268, 0.0274, 0.0212, 0.0261, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:36:16,647 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-29 01:36:17,584 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86501.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:36:19,004 INFO [train.py:904] (4/8) Epoch 9, batch 5300, loss[loss=0.1836, simple_loss=0.2673, pruned_loss=0.04997, over 15537.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2813, pruned_loss=0.0558, over 3177679.49 frames. ], batch size: 191, lr: 7.60e-03, grad_scale: 16.0 2023-04-29 01:36:25,587 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86506.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 01:36:42,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.402e+02 2.448e+02 2.767e+02 3.292e+02 5.111e+02, threshold=5.534e+02, percent-clipped=0.0 2023-04-29 01:36:47,599 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5411, 3.8711, 4.1004, 1.9478, 4.4052, 4.3663, 3.1144, 3.2346], device='cuda:4'), covar=tensor([0.0735, 0.0136, 0.0128, 0.1090, 0.0031, 0.0055, 0.0352, 0.0357], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0096, 0.0084, 0.0137, 0.0067, 0.0094, 0.0118, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-29 01:36:58,303 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8432, 3.3589, 3.0858, 1.9116, 2.7677, 2.2737, 3.3950, 3.3471], device='cuda:4'), covar=tensor([0.0229, 0.0534, 0.0576, 0.1653, 0.0745, 0.0890, 0.0580, 0.0718], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0138, 0.0156, 0.0140, 0.0134, 0.0124, 0.0135, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 01:37:01,904 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86531.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:37:32,903 INFO [train.py:904] (4/8) Epoch 9, batch 5350, loss[loss=0.211, simple_loss=0.2921, pruned_loss=0.06495, over 16567.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2791, pruned_loss=0.05448, over 3186234.72 frames. ], batch size: 68, lr: 7.60e-03, grad_scale: 16.0 2023-04-29 01:38:11,780 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86579.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:38:13,776 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86580.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:38:23,807 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86586.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 01:38:45,867 INFO [train.py:904] (4/8) Epoch 9, batch 5400, loss[loss=0.2086, simple_loss=0.2885, pruned_loss=0.0643, over 12005.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2823, pruned_loss=0.05548, over 3186660.98 frames. ], batch size: 247, lr: 7.60e-03, grad_scale: 16.0 2023-04-29 01:38:46,249 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86602.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:39:01,684 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86613.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:39:09,033 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.647e+02 2.152e+02 2.423e+02 2.907e+02 4.710e+02, threshold=4.847e+02, percent-clipped=0.0 2023-04-29 01:39:12,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6805, 3.5757, 3.5942, 3.9125, 3.9533, 3.6455, 3.9155, 4.0172], device='cuda:4'), covar=tensor([0.1369, 0.1126, 0.1841, 0.0779, 0.0714, 0.1714, 0.0816, 0.0674], device='cuda:4'), in_proj_covar=tensor([0.0499, 0.0606, 0.0752, 0.0614, 0.0465, 0.0466, 0.0483, 0.0536], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 01:39:32,123 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86634.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 01:39:41,814 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86641.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:39:54,409 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-29 01:40:00,485 INFO [train.py:904] (4/8) Epoch 9, batch 5450, loss[loss=0.254, simple_loss=0.329, pruned_loss=0.08955, over 16677.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2852, pruned_loss=0.05674, over 3184534.53 frames. ], batch size: 134, lr: 7.60e-03, grad_scale: 16.0 2023-04-29 01:40:03,998 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-29 01:40:11,008 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86658.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:40:15,979 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86661.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:41:18,953 INFO [train.py:904] (4/8) Epoch 9, batch 5500, loss[loss=0.2484, simple_loss=0.3242, pruned_loss=0.08629, over 16257.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.2944, pruned_loss=0.06336, over 3147876.54 frames. ], batch size: 165, lr: 7.60e-03, grad_scale: 16.0 2023-04-29 01:41:24,318 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86706.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:41:27,604 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86708.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:41:43,593 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.927e+02 3.483e+02 4.212e+02 5.366e+02 1.166e+03, threshold=8.424e+02, percent-clipped=36.0 2023-04-29 01:41:54,542 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86725.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:42:36,876 INFO [train.py:904] (4/8) Epoch 9, batch 5550, loss[loss=0.2324, simple_loss=0.3084, pruned_loss=0.07823, over 16382.00 frames. ], tot_loss[loss=0.2202, simple_loss=0.3021, pruned_loss=0.06912, over 3126546.86 frames. ], batch size: 146, lr: 7.59e-03, grad_scale: 16.0 2023-04-29 01:42:43,066 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86756.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:43:09,698 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86773.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:43:44,627 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86796.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:43:53,985 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86801.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 01:43:54,665 INFO [train.py:904] (4/8) Epoch 9, batch 5600, loss[loss=0.2522, simple_loss=0.3283, pruned_loss=0.08799, over 16763.00 frames. ], tot_loss[loss=0.2276, simple_loss=0.3074, pruned_loss=0.07388, over 3110463.53 frames. ], batch size: 124, lr: 7.59e-03, grad_scale: 8.0 2023-04-29 01:44:14,697 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86814.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:44:21,510 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.583e+02 3.990e+02 4.962e+02 6.279e+02 1.585e+03, threshold=9.923e+02, percent-clipped=6.0 2023-04-29 01:45:17,290 INFO [train.py:904] (4/8) Epoch 9, batch 5650, loss[loss=0.2267, simple_loss=0.3025, pruned_loss=0.07547, over 16903.00 frames. ], tot_loss[loss=0.2338, simple_loss=0.3122, pruned_loss=0.0777, over 3096959.18 frames. ], batch size: 116, lr: 7.59e-03, grad_scale: 4.0 2023-04-29 01:45:53,442 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86875.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:45:58,526 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86878.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:46:33,109 INFO [train.py:904] (4/8) Epoch 9, batch 5700, loss[loss=0.25, simple_loss=0.3339, pruned_loss=0.08302, over 15259.00 frames. ], tot_loss[loss=0.2366, simple_loss=0.3142, pruned_loss=0.07955, over 3091304.37 frames. ], batch size: 190, lr: 7.59e-03, grad_scale: 4.0 2023-04-29 01:46:33,461 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86902.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:46:37,214 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7962, 2.7149, 2.6020, 1.9305, 2.5356, 2.7108, 2.5392, 1.7993], device='cuda:4'), covar=tensor([0.0306, 0.0039, 0.0058, 0.0250, 0.0079, 0.0073, 0.0059, 0.0287], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0064, 0.0064, 0.0120, 0.0071, 0.0081, 0.0070, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 01:46:59,880 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.388e+02 3.612e+02 4.435e+02 5.746e+02 9.391e+02, threshold=8.870e+02, percent-clipped=0.0 2023-04-29 01:47:25,188 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86936.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:47:31,052 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86939.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:47:39,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0176, 2.3758, 2.2466, 2.7328, 2.1930, 3.1594, 1.6831, 2.7209], device='cuda:4'), covar=tensor([0.1172, 0.0501, 0.1020, 0.0144, 0.0159, 0.0396, 0.1333, 0.0638], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0152, 0.0176, 0.0126, 0.0199, 0.0205, 0.0174, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 01:47:48,291 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=86950.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:47:50,854 INFO [train.py:904] (4/8) Epoch 9, batch 5750, loss[loss=0.2112, simple_loss=0.2977, pruned_loss=0.06232, over 16786.00 frames. ], tot_loss[loss=0.2399, simple_loss=0.3169, pruned_loss=0.08146, over 3061060.11 frames. ], batch size: 83, lr: 7.58e-03, grad_scale: 4.0 2023-04-29 01:49:10,134 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9388, 3.4853, 3.5124, 2.2501, 3.2511, 3.4895, 3.3428, 1.8366], device='cuda:4'), covar=tensor([0.0419, 0.0040, 0.0038, 0.0298, 0.0068, 0.0087, 0.0052, 0.0375], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0065, 0.0065, 0.0121, 0.0072, 0.0082, 0.0070, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 01:49:12,516 INFO [train.py:904] (4/8) Epoch 9, batch 5800, loss[loss=0.226, simple_loss=0.3137, pruned_loss=0.06912, over 16933.00 frames. ], tot_loss[loss=0.2386, simple_loss=0.3165, pruned_loss=0.0804, over 3045703.07 frames. ], batch size: 83, lr: 7.58e-03, grad_scale: 4.0 2023-04-29 01:49:13,807 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87002.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 01:49:40,721 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.964e+02 3.181e+02 3.841e+02 4.855e+02 7.782e+02, threshold=7.681e+02, percent-clipped=0.0 2023-04-29 01:50:30,438 INFO [train.py:904] (4/8) Epoch 9, batch 5850, loss[loss=0.2479, simple_loss=0.3052, pruned_loss=0.09525, over 11288.00 frames. ], tot_loss[loss=0.2363, simple_loss=0.3147, pruned_loss=0.07892, over 3055439.54 frames. ], batch size: 248, lr: 7.58e-03, grad_scale: 4.0 2023-04-29 01:50:47,484 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87063.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 01:51:42,081 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87096.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:51:50,120 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87101.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 01:51:50,871 INFO [train.py:904] (4/8) Epoch 9, batch 5900, loss[loss=0.2119, simple_loss=0.3024, pruned_loss=0.0607, over 16674.00 frames. ], tot_loss[loss=0.2365, simple_loss=0.3148, pruned_loss=0.07912, over 3039920.16 frames. ], batch size: 57, lr: 7.58e-03, grad_scale: 4.0 2023-04-29 01:52:22,906 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.063e+02 3.002e+02 3.675e+02 4.500e+02 8.851e+02, threshold=7.350e+02, percent-clipped=1.0 2023-04-29 01:53:01,034 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=87144.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:53:09,063 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=87149.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:53:13,003 INFO [train.py:904] (4/8) Epoch 9, batch 5950, loss[loss=0.2537, simple_loss=0.328, pruned_loss=0.08969, over 11875.00 frames. ], tot_loss[loss=0.2342, simple_loss=0.3148, pruned_loss=0.07675, over 3064355.53 frames. ], batch size: 246, lr: 7.58e-03, grad_scale: 4.0 2023-04-29 01:53:42,054 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87170.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:53:48,765 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6888, 3.2642, 2.9962, 1.8368, 2.7223, 2.1023, 3.1728, 3.3356], device='cuda:4'), covar=tensor([0.0290, 0.0531, 0.0624, 0.1671, 0.0772, 0.0919, 0.0630, 0.0778], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0139, 0.0158, 0.0142, 0.0134, 0.0125, 0.0136, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 01:54:02,370 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-29 01:54:33,104 INFO [train.py:904] (4/8) Epoch 9, batch 6000, loss[loss=0.2213, simple_loss=0.3008, pruned_loss=0.07093, over 16627.00 frames. ], tot_loss[loss=0.2324, simple_loss=0.3131, pruned_loss=0.07583, over 3071470.57 frames. ], batch size: 62, lr: 7.57e-03, grad_scale: 8.0 2023-04-29 01:54:33,104 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 01:54:44,302 INFO [train.py:938] (4/8) Epoch 9, validation: loss=0.1674, simple_loss=0.2809, pruned_loss=0.02692, over 944034.00 frames. 2023-04-29 01:54:44,302 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-29 01:55:11,800 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.951e+02 3.247e+02 3.795e+02 4.870e+02 1.523e+03, threshold=7.589e+02, percent-clipped=3.0 2023-04-29 01:55:28,590 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87230.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:55:34,793 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87234.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:55:37,454 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87236.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:56:03,542 INFO [train.py:904] (4/8) Epoch 9, batch 6050, loss[loss=0.2326, simple_loss=0.3212, pruned_loss=0.07193, over 16632.00 frames. ], tot_loss[loss=0.2301, simple_loss=0.3111, pruned_loss=0.07451, over 3092455.26 frames. ], batch size: 68, lr: 7.57e-03, grad_scale: 8.0 2023-04-29 01:56:44,147 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9678, 3.3749, 3.3729, 2.2297, 3.1360, 3.3845, 3.2609, 1.7519], device='cuda:4'), covar=tensor([0.0397, 0.0033, 0.0037, 0.0298, 0.0064, 0.0073, 0.0050, 0.0384], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0064, 0.0065, 0.0122, 0.0072, 0.0082, 0.0071, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 01:56:56,178 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=87284.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:57:06,960 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87291.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 01:57:23,335 INFO [train.py:904] (4/8) Epoch 9, batch 6100, loss[loss=0.2295, simple_loss=0.3104, pruned_loss=0.07433, over 17182.00 frames. ], tot_loss[loss=0.2286, simple_loss=0.3106, pruned_loss=0.07335, over 3108429.93 frames. ], batch size: 46, lr: 7.57e-03, grad_scale: 8.0 2023-04-29 01:57:52,501 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.809e+02 2.938e+02 3.700e+02 4.608e+02 9.243e+02, threshold=7.400e+02, percent-clipped=2.0 2023-04-29 01:58:40,619 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5736, 2.6944, 2.2966, 3.8770, 2.8241, 3.8344, 1.2520, 2.7264], device='cuda:4'), covar=tensor([0.1425, 0.0647, 0.1241, 0.0138, 0.0294, 0.0403, 0.1748, 0.0849], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0153, 0.0178, 0.0127, 0.0203, 0.0206, 0.0175, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 01:58:42,527 INFO [train.py:904] (4/8) Epoch 9, batch 6150, loss[loss=0.2836, simple_loss=0.3479, pruned_loss=0.1097, over 11273.00 frames. ], tot_loss[loss=0.2276, simple_loss=0.3091, pruned_loss=0.07304, over 3087330.91 frames. ], batch size: 248, lr: 7.57e-03, grad_scale: 8.0 2023-04-29 01:58:52,854 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87358.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 01:58:59,193 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87362.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:00:00,627 INFO [train.py:904] (4/8) Epoch 9, batch 6200, loss[loss=0.2388, simple_loss=0.3053, pruned_loss=0.08614, over 11868.00 frames. ], tot_loss[loss=0.2271, simple_loss=0.3079, pruned_loss=0.07314, over 3085765.18 frames. ], batch size: 248, lr: 7.57e-03, grad_scale: 8.0 2023-04-29 02:00:01,242 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0446, 1.8429, 2.0712, 3.4161, 1.7990, 2.1691, 2.0224, 1.9426], device='cuda:4'), covar=tensor([0.0981, 0.3416, 0.2083, 0.0556, 0.4316, 0.2479, 0.3101, 0.3530], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0371, 0.0311, 0.0316, 0.0400, 0.0420, 0.0333, 0.0439], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:00:28,300 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.203e+02 3.670e+02 4.311e+02 5.768e+02 9.493e+02, threshold=8.622e+02, percent-clipped=7.0 2023-04-29 02:00:34,905 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87423.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:01:07,123 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9328, 1.9995, 2.3788, 3.1653, 2.1857, 2.2936, 2.2640, 2.1175], device='cuda:4'), covar=tensor([0.0847, 0.2828, 0.1580, 0.0494, 0.2988, 0.1846, 0.2183, 0.2820], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0372, 0.0311, 0.0316, 0.0401, 0.0421, 0.0333, 0.0440], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:01:18,183 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2771, 3.4982, 3.7148, 1.7444, 3.9200, 3.9460, 2.9672, 2.7778], device='cuda:4'), covar=tensor([0.0761, 0.0157, 0.0134, 0.1107, 0.0049, 0.0092, 0.0326, 0.0444], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0098, 0.0084, 0.0137, 0.0066, 0.0094, 0.0118, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-29 02:01:18,220 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7047, 2.1581, 1.7783, 1.9487, 2.5502, 2.2575, 2.6662, 2.7741], device='cuda:4'), covar=tensor([0.0097, 0.0289, 0.0361, 0.0322, 0.0165, 0.0279, 0.0138, 0.0167], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0188, 0.0187, 0.0185, 0.0186, 0.0190, 0.0188, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:01:18,823 INFO [train.py:904] (4/8) Epoch 9, batch 6250, loss[loss=0.2453, simple_loss=0.3285, pruned_loss=0.08102, over 16852.00 frames. ], tot_loss[loss=0.2265, simple_loss=0.3073, pruned_loss=0.07288, over 3088596.98 frames. ], batch size: 116, lr: 7.56e-03, grad_scale: 4.0 2023-04-29 02:01:21,190 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.4665, 2.7969, 2.2642, 4.3946, 2.8824, 4.1941, 1.4410, 2.6283], device='cuda:4'), covar=tensor([0.1528, 0.0702, 0.1428, 0.0131, 0.0321, 0.0408, 0.1696, 0.1053], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0152, 0.0176, 0.0126, 0.0201, 0.0205, 0.0174, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 02:01:32,678 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87461.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:01:38,832 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2911, 1.9026, 1.6329, 1.7468, 2.2214, 1.9916, 2.1851, 2.3828], device='cuda:4'), covar=tensor([0.0088, 0.0241, 0.0316, 0.0303, 0.0157, 0.0244, 0.0144, 0.0183], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0188, 0.0186, 0.0185, 0.0186, 0.0189, 0.0188, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:01:45,456 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87470.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:02:33,518 INFO [train.py:904] (4/8) Epoch 9, batch 6300, loss[loss=0.2424, simple_loss=0.3201, pruned_loss=0.08236, over 16758.00 frames. ], tot_loss[loss=0.2252, simple_loss=0.3069, pruned_loss=0.07174, over 3119578.61 frames. ], batch size: 124, lr: 7.56e-03, grad_scale: 4.0 2023-04-29 02:02:38,942 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87505.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:02:54,510 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1774, 4.1696, 4.6247, 4.5854, 4.5573, 4.2641, 4.2831, 4.1171], device='cuda:4'), covar=tensor([0.0288, 0.0483, 0.0345, 0.0401, 0.0460, 0.0325, 0.0880, 0.0457], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0308, 0.0320, 0.0306, 0.0361, 0.0329, 0.0441, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 02:02:59,938 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=87518.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:03:04,115 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.747e+02 3.209e+02 3.966e+02 4.812e+02 1.231e+03, threshold=7.932e+02, percent-clipped=2.0 2023-04-29 02:03:06,783 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87522.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:03:24,773 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87534.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:03:50,409 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7580, 3.6147, 3.7999, 3.6063, 3.7289, 4.1524, 3.8782, 3.6281], device='cuda:4'), covar=tensor([0.2163, 0.2343, 0.2302, 0.3111, 0.3209, 0.1871, 0.1617, 0.2684], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0453, 0.0485, 0.0401, 0.0526, 0.0507, 0.0391, 0.0539], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:03:51,224 INFO [train.py:904] (4/8) Epoch 9, batch 6350, loss[loss=0.2728, simple_loss=0.3317, pruned_loss=0.1069, over 11522.00 frames. ], tot_loss[loss=0.2277, simple_loss=0.3081, pruned_loss=0.07364, over 3087685.64 frames. ], batch size: 247, lr: 7.56e-03, grad_scale: 4.0 2023-04-29 02:04:13,030 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87566.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 02:04:37,475 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=87582.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:04:40,649 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87584.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:04:44,420 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87586.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:04:50,132 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87589.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:05:08,846 INFO [train.py:904] (4/8) Epoch 9, batch 6400, loss[loss=0.1937, simple_loss=0.2814, pruned_loss=0.05296, over 17274.00 frames. ], tot_loss[loss=0.2284, simple_loss=0.3081, pruned_loss=0.07436, over 3092199.94 frames. ], batch size: 52, lr: 7.56e-03, grad_scale: 8.0 2023-04-29 02:05:13,474 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87605.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:05:37,567 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.591e+02 3.429e+02 4.257e+02 5.158e+02 9.236e+02, threshold=8.515e+02, percent-clipped=3.0 2023-04-29 02:05:51,891 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8838, 4.1499, 3.9337, 4.0276, 3.6789, 3.7496, 3.8244, 4.1241], device='cuda:4'), covar=tensor([0.0881, 0.0832, 0.0953, 0.0603, 0.0702, 0.1354, 0.0762, 0.0888], device='cuda:4'), in_proj_covar=tensor([0.0497, 0.0626, 0.0525, 0.0426, 0.0388, 0.0405, 0.0523, 0.0466], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:06:02,401 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3663, 3.1744, 2.9940, 1.8397, 2.6940, 2.0966, 3.0494, 3.2787], device='cuda:4'), covar=tensor([0.0253, 0.0520, 0.0482, 0.1770, 0.0783, 0.0914, 0.0513, 0.0489], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0138, 0.0157, 0.0141, 0.0134, 0.0124, 0.0135, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 02:06:05,197 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2473, 5.1729, 5.0622, 4.3435, 5.0531, 1.7487, 4.7941, 4.8985], device='cuda:4'), covar=tensor([0.0076, 0.0079, 0.0127, 0.0394, 0.0093, 0.2295, 0.0139, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0104, 0.0152, 0.0145, 0.0121, 0.0166, 0.0136, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:06:14,992 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87645.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 02:06:22,235 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87650.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:06:24,600 INFO [train.py:904] (4/8) Epoch 9, batch 6450, loss[loss=0.2188, simple_loss=0.3008, pruned_loss=0.0684, over 16459.00 frames. ], tot_loss[loss=0.2276, simple_loss=0.3079, pruned_loss=0.07361, over 3091705.52 frames. ], batch size: 68, lr: 7.55e-03, grad_scale: 4.0 2023-04-29 02:06:27,895 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-29 02:06:34,947 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87658.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 02:06:38,896 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8949, 4.1548, 3.9686, 3.9986, 3.6950, 3.7616, 3.8815, 4.1234], device='cuda:4'), covar=tensor([0.0978, 0.1035, 0.0987, 0.0732, 0.0796, 0.1626, 0.0863, 0.1016], device='cuda:4'), in_proj_covar=tensor([0.0494, 0.0622, 0.0522, 0.0425, 0.0387, 0.0402, 0.0519, 0.0463], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:06:46,405 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87666.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 02:06:58,103 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9212, 2.7983, 2.6678, 1.9928, 2.5650, 2.1330, 2.7446, 2.8958], device='cuda:4'), covar=tensor([0.0292, 0.0636, 0.0461, 0.1533, 0.0734, 0.0874, 0.0502, 0.0561], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0139, 0.0157, 0.0141, 0.0134, 0.0124, 0.0135, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 02:07:41,885 INFO [train.py:904] (4/8) Epoch 9, batch 6500, loss[loss=0.2458, simple_loss=0.3187, pruned_loss=0.08642, over 15504.00 frames. ], tot_loss[loss=0.2263, simple_loss=0.3065, pruned_loss=0.07308, over 3089924.03 frames. ], batch size: 190, lr: 7.55e-03, grad_scale: 2.0 2023-04-29 02:07:48,426 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=87706.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 02:07:51,454 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8462, 3.2362, 3.1673, 2.0702, 2.9205, 3.1253, 3.0643, 1.8523], device='cuda:4'), covar=tensor([0.0395, 0.0033, 0.0039, 0.0290, 0.0069, 0.0080, 0.0055, 0.0337], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0064, 0.0065, 0.0121, 0.0071, 0.0083, 0.0071, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:07:59,485 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5463, 4.5840, 4.3919, 4.1910, 3.9941, 4.4539, 4.3096, 4.0895], device='cuda:4'), covar=tensor([0.0490, 0.0348, 0.0230, 0.0213, 0.0888, 0.0309, 0.0399, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0225, 0.0272, 0.0261, 0.0237, 0.0285, 0.0275, 0.0179, 0.0308], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:08:05,641 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.59 vs. limit=5.0 2023-04-29 02:08:06,441 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87718.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:08:13,250 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.348e+02 3.236e+02 4.021e+02 5.179e+02 1.078e+03, threshold=8.043e+02, percent-clipped=2.0 2023-04-29 02:08:34,364 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-29 02:09:00,842 INFO [train.py:904] (4/8) Epoch 9, batch 6550, loss[loss=0.2516, simple_loss=0.3396, pruned_loss=0.08184, over 15292.00 frames. ], tot_loss[loss=0.2297, simple_loss=0.3098, pruned_loss=0.07479, over 3092320.82 frames. ], batch size: 190, lr: 7.55e-03, grad_scale: 2.0 2023-04-29 02:09:35,356 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9524, 3.3237, 3.3552, 2.0495, 3.0563, 3.2942, 3.2614, 1.7576], device='cuda:4'), covar=tensor([0.0395, 0.0041, 0.0032, 0.0311, 0.0068, 0.0078, 0.0050, 0.0373], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0063, 0.0063, 0.0119, 0.0070, 0.0082, 0.0070, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:09:49,904 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0608, 3.6286, 3.2175, 1.8559, 2.7134, 2.4409, 3.2794, 3.5554], device='cuda:4'), covar=tensor([0.0296, 0.0613, 0.0676, 0.2006, 0.0944, 0.0912, 0.0789, 0.0912], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0137, 0.0155, 0.0139, 0.0132, 0.0123, 0.0134, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 02:10:18,518 INFO [train.py:904] (4/8) Epoch 9, batch 6600, loss[loss=0.2245, simple_loss=0.3078, pruned_loss=0.07065, over 16780.00 frames. ], tot_loss[loss=0.2312, simple_loss=0.3118, pruned_loss=0.07531, over 3088273.26 frames. ], batch size: 83, lr: 7.55e-03, grad_scale: 2.0 2023-04-29 02:10:42,559 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87817.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:10:51,645 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.077e+02 3.184e+02 4.033e+02 5.145e+02 1.254e+03, threshold=8.065e+02, percent-clipped=5.0 2023-04-29 02:11:21,067 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9044, 4.8451, 4.7203, 4.4738, 4.2681, 4.7749, 4.5872, 4.4380], device='cuda:4'), covar=tensor([0.0403, 0.0286, 0.0214, 0.0209, 0.0903, 0.0295, 0.0315, 0.0560], device='cuda:4'), in_proj_covar=tensor([0.0225, 0.0272, 0.0261, 0.0237, 0.0285, 0.0276, 0.0180, 0.0308], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:11:36,981 INFO [train.py:904] (4/8) Epoch 9, batch 6650, loss[loss=0.3162, simple_loss=0.3639, pruned_loss=0.1343, over 11597.00 frames. ], tot_loss[loss=0.2325, simple_loss=0.3123, pruned_loss=0.07632, over 3079576.72 frames. ], batch size: 250, lr: 7.55e-03, grad_scale: 2.0 2023-04-29 02:11:41,503 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-29 02:11:51,506 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87861.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 02:12:28,594 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87886.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:12:53,898 INFO [train.py:904] (4/8) Epoch 9, batch 6700, loss[loss=0.2714, simple_loss=0.3298, pruned_loss=0.1065, over 11201.00 frames. ], tot_loss[loss=0.2317, simple_loss=0.3109, pruned_loss=0.07623, over 3072243.84 frames. ], batch size: 248, lr: 7.54e-03, grad_scale: 2.0 2023-04-29 02:13:26,701 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.105e+02 3.383e+02 4.194e+02 5.361e+02 9.838e+02, threshold=8.388e+02, percent-clipped=4.0 2023-04-29 02:13:44,326 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=87934.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:13:52,511 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87940.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 02:14:00,206 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87945.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:14:10,405 INFO [train.py:904] (4/8) Epoch 9, batch 6750, loss[loss=0.3187, simple_loss=0.3603, pruned_loss=0.1386, over 11560.00 frames. ], tot_loss[loss=0.232, simple_loss=0.3104, pruned_loss=0.07675, over 3072084.01 frames. ], batch size: 248, lr: 7.54e-03, grad_scale: 2.0 2023-04-29 02:14:24,885 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87961.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 02:15:29,294 INFO [train.py:904] (4/8) Epoch 9, batch 6800, loss[loss=0.2239, simple_loss=0.3124, pruned_loss=0.06766, over 16723.00 frames. ], tot_loss[loss=0.2308, simple_loss=0.3097, pruned_loss=0.07598, over 3081675.31 frames. ], batch size: 124, lr: 7.54e-03, grad_scale: 4.0 2023-04-29 02:15:54,190 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88018.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:15:55,600 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-29 02:15:59,031 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7587, 4.9192, 4.8577, 3.0145, 4.3139, 4.7946, 4.1424, 2.7604], device='cuda:4'), covar=tensor([0.0354, 0.0012, 0.0019, 0.0271, 0.0044, 0.0052, 0.0038, 0.0286], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0063, 0.0063, 0.0120, 0.0070, 0.0082, 0.0071, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:16:01,653 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-29 02:16:02,234 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.936e+02 3.136e+02 3.707e+02 4.804e+02 8.151e+02, threshold=7.414e+02, percent-clipped=0.0 2023-04-29 02:16:45,306 INFO [train.py:904] (4/8) Epoch 9, batch 6850, loss[loss=0.2138, simple_loss=0.3034, pruned_loss=0.06207, over 16367.00 frames. ], tot_loss[loss=0.2311, simple_loss=0.3102, pruned_loss=0.07605, over 3089086.08 frames. ], batch size: 165, lr: 7.54e-03, grad_scale: 4.0 2023-04-29 02:17:06,648 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=88066.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:17:18,895 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3319, 3.2846, 3.3383, 3.4750, 3.5082, 3.2291, 3.4302, 3.5326], device='cuda:4'), covar=tensor([0.1120, 0.0907, 0.1173, 0.0636, 0.0671, 0.2147, 0.1032, 0.0720], device='cuda:4'), in_proj_covar=tensor([0.0479, 0.0596, 0.0737, 0.0607, 0.0467, 0.0461, 0.0486, 0.0536], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:17:46,559 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7933, 3.5074, 2.8697, 1.6025, 2.4559, 2.0694, 3.2218, 3.4836], device='cuda:4'), covar=tensor([0.0269, 0.0507, 0.0820, 0.2257, 0.1082, 0.1148, 0.0684, 0.0702], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0139, 0.0157, 0.0142, 0.0134, 0.0125, 0.0136, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 02:17:59,039 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=88100.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:18:01,000 INFO [train.py:904] (4/8) Epoch 9, batch 6900, loss[loss=0.2481, simple_loss=0.3309, pruned_loss=0.08262, over 16773.00 frames. ], tot_loss[loss=0.2332, simple_loss=0.3133, pruned_loss=0.07654, over 3078603.97 frames. ], batch size: 124, lr: 7.54e-03, grad_scale: 4.0 2023-04-29 02:18:24,581 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88117.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:18:33,102 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.937e+02 3.211e+02 3.844e+02 4.463e+02 9.504e+02, threshold=7.687e+02, percent-clipped=4.0 2023-04-29 02:19:05,052 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6937, 4.4911, 4.6969, 4.9268, 5.0650, 4.5837, 5.0427, 5.0259], device='cuda:4'), covar=tensor([0.1315, 0.0983, 0.1341, 0.0620, 0.0477, 0.0695, 0.0504, 0.0494], device='cuda:4'), in_proj_covar=tensor([0.0479, 0.0599, 0.0740, 0.0607, 0.0469, 0.0464, 0.0487, 0.0540], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:19:10,689 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2352, 4.0814, 4.2555, 4.4705, 4.5773, 4.1530, 4.4917, 4.5611], device='cuda:4'), covar=tensor([0.1352, 0.0966, 0.1335, 0.0563, 0.0512, 0.1074, 0.0674, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0480, 0.0601, 0.0742, 0.0609, 0.0470, 0.0465, 0.0488, 0.0541], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:19:18,084 INFO [train.py:904] (4/8) Epoch 9, batch 6950, loss[loss=0.2249, simple_loss=0.3031, pruned_loss=0.0733, over 16685.00 frames. ], tot_loss[loss=0.2348, simple_loss=0.3146, pruned_loss=0.07752, over 3068031.80 frames. ], batch size: 134, lr: 7.53e-03, grad_scale: 4.0 2023-04-29 02:19:32,103 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88161.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 02:19:32,201 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=88161.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:19:38,178 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=88165.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:20:23,299 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-29 02:20:31,313 INFO [train.py:904] (4/8) Epoch 9, batch 7000, loss[loss=0.2387, simple_loss=0.3243, pruned_loss=0.07654, over 16683.00 frames. ], tot_loss[loss=0.2328, simple_loss=0.3138, pruned_loss=0.07592, over 3078073.23 frames. ], batch size: 134, lr: 7.53e-03, grad_scale: 4.0 2023-04-29 02:20:43,257 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=88209.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:20:45,889 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7629, 4.5952, 4.7660, 5.0029, 5.1118, 4.5423, 5.0949, 5.0657], device='cuda:4'), covar=tensor([0.1302, 0.0919, 0.1226, 0.0495, 0.0478, 0.0693, 0.0506, 0.0448], device='cuda:4'), in_proj_covar=tensor([0.0478, 0.0599, 0.0738, 0.0605, 0.0468, 0.0461, 0.0487, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:21:03,406 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.948e+02 3.320e+02 4.209e+02 4.909e+02 8.638e+02, threshold=8.417e+02, percent-clipped=2.0 2023-04-29 02:21:29,020 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88240.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:21:33,406 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1115, 5.4479, 5.1180, 5.1760, 4.8577, 4.7838, 4.8545, 5.5384], device='cuda:4'), covar=tensor([0.1000, 0.0887, 0.1055, 0.0688, 0.0759, 0.0790, 0.0991, 0.0793], device='cuda:4'), in_proj_covar=tensor([0.0497, 0.0629, 0.0525, 0.0426, 0.0386, 0.0408, 0.0522, 0.0469], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:21:35,935 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88245.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:21:47,209 INFO [train.py:904] (4/8) Epoch 9, batch 7050, loss[loss=0.2049, simple_loss=0.2891, pruned_loss=0.06031, over 16797.00 frames. ], tot_loss[loss=0.232, simple_loss=0.3142, pruned_loss=0.07487, over 3109936.38 frames. ], batch size: 39, lr: 7.53e-03, grad_scale: 4.0 2023-04-29 02:22:00,335 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88261.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 02:22:41,149 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=88288.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:22:45,420 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=88291.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 02:22:47,885 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=88293.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:23:02,082 INFO [train.py:904] (4/8) Epoch 9, batch 7100, loss[loss=0.2111, simple_loss=0.2913, pruned_loss=0.06546, over 16699.00 frames. ], tot_loss[loss=0.2312, simple_loss=0.3128, pruned_loss=0.07479, over 3110576.37 frames. ], batch size: 57, lr: 7.53e-03, grad_scale: 4.0 2023-04-29 02:23:12,559 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=88309.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:23:19,269 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9866, 3.2483, 2.9481, 5.2388, 4.2669, 4.5355, 1.7802, 3.4537], device='cuda:4'), covar=tensor([0.1191, 0.0603, 0.1032, 0.0136, 0.0381, 0.0360, 0.1349, 0.0717], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0154, 0.0178, 0.0128, 0.0202, 0.0207, 0.0175, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 02:23:33,102 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.153e+02 3.325e+02 4.085e+02 5.101e+02 9.859e+02, threshold=8.169e+02, percent-clipped=1.0 2023-04-29 02:24:15,978 INFO [train.py:904] (4/8) Epoch 9, batch 7150, loss[loss=0.2152, simple_loss=0.2999, pruned_loss=0.0653, over 17053.00 frames. ], tot_loss[loss=0.2293, simple_loss=0.3104, pruned_loss=0.07407, over 3105410.44 frames. ], batch size: 55, lr: 7.52e-03, grad_scale: 4.0 2023-04-29 02:24:17,146 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=88352.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 02:24:23,126 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6497, 2.6056, 1.8113, 2.7568, 2.1809, 2.7191, 2.0025, 2.3577], device='cuda:4'), covar=tensor([0.0231, 0.0384, 0.1239, 0.0143, 0.0607, 0.0562, 0.1140, 0.0547], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0160, 0.0185, 0.0112, 0.0164, 0.0201, 0.0193, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 02:24:38,728 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3888, 3.8456, 4.1485, 1.7339, 4.4073, 4.5359, 3.0793, 2.9319], device='cuda:4'), covar=tensor([0.1023, 0.0149, 0.0179, 0.1299, 0.0061, 0.0070, 0.0395, 0.0566], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0100, 0.0087, 0.0142, 0.0068, 0.0097, 0.0121, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 02:25:28,109 INFO [train.py:904] (4/8) Epoch 9, batch 7200, loss[loss=0.1836, simple_loss=0.2657, pruned_loss=0.05072, over 17020.00 frames. ], tot_loss[loss=0.2266, simple_loss=0.3078, pruned_loss=0.07266, over 3094775.20 frames. ], batch size: 53, lr: 7.52e-03, grad_scale: 8.0 2023-04-29 02:26:00,148 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.048e+02 3.005e+02 3.439e+02 4.228e+02 8.504e+02, threshold=6.879e+02, percent-clipped=1.0 2023-04-29 02:26:11,227 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8256, 5.1369, 4.8771, 4.8805, 4.6032, 4.6124, 4.6131, 5.2387], device='cuda:4'), covar=tensor([0.0836, 0.0739, 0.0901, 0.0611, 0.0721, 0.0780, 0.0891, 0.0684], device='cuda:4'), in_proj_covar=tensor([0.0493, 0.0623, 0.0521, 0.0422, 0.0383, 0.0405, 0.0516, 0.0467], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:26:47,218 INFO [train.py:904] (4/8) Epoch 9, batch 7250, loss[loss=0.1987, simple_loss=0.2787, pruned_loss=0.05931, over 17134.00 frames. ], tot_loss[loss=0.223, simple_loss=0.3047, pruned_loss=0.0707, over 3090662.04 frames. ], batch size: 47, lr: 7.52e-03, grad_scale: 8.0 2023-04-29 02:26:53,241 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=88456.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:27:38,138 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8982, 3.5788, 3.3475, 1.8308, 2.9954, 2.2397, 3.4740, 3.8074], device='cuda:4'), covar=tensor([0.0291, 0.0590, 0.0670, 0.2032, 0.0829, 0.1076, 0.0659, 0.0711], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0139, 0.0158, 0.0143, 0.0135, 0.0126, 0.0137, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 02:28:00,985 INFO [train.py:904] (4/8) Epoch 9, batch 7300, loss[loss=0.2179, simple_loss=0.3117, pruned_loss=0.06209, over 16856.00 frames. ], tot_loss[loss=0.2224, simple_loss=0.3041, pruned_loss=0.07035, over 3105404.66 frames. ], batch size: 116, lr: 7.52e-03, grad_scale: 8.0 2023-04-29 02:28:02,189 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-29 02:28:33,453 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.978e+02 3.255e+02 4.092e+02 5.788e+02 1.345e+03, threshold=8.184e+02, percent-clipped=12.0 2023-04-29 02:28:34,733 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-29 02:29:04,909 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6473, 4.6647, 4.4753, 4.2847, 4.0431, 4.6102, 4.4594, 4.2705], device='cuda:4'), covar=tensor([0.0500, 0.0274, 0.0262, 0.0234, 0.0952, 0.0321, 0.0337, 0.0530], device='cuda:4'), in_proj_covar=tensor([0.0219, 0.0264, 0.0253, 0.0232, 0.0276, 0.0265, 0.0177, 0.0297], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:29:14,111 INFO [train.py:904] (4/8) Epoch 9, batch 7350, loss[loss=0.2283, simple_loss=0.3083, pruned_loss=0.07418, over 16931.00 frames. ], tot_loss[loss=0.2235, simple_loss=0.3048, pruned_loss=0.07117, over 3085025.85 frames. ], batch size: 109, lr: 7.52e-03, grad_scale: 4.0 2023-04-29 02:30:17,940 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9976, 4.9497, 4.7515, 4.1585, 4.7763, 1.8780, 4.5349, 4.6678], device='cuda:4'), covar=tensor([0.0059, 0.0056, 0.0142, 0.0315, 0.0077, 0.2164, 0.0102, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0103, 0.0150, 0.0144, 0.0118, 0.0166, 0.0134, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:30:27,925 INFO [train.py:904] (4/8) Epoch 9, batch 7400, loss[loss=0.2198, simple_loss=0.2998, pruned_loss=0.06987, over 16648.00 frames. ], tot_loss[loss=0.2259, simple_loss=0.3066, pruned_loss=0.07258, over 3069243.34 frames. ], batch size: 62, lr: 7.51e-03, grad_scale: 4.0 2023-04-29 02:31:01,751 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.901e+02 3.181e+02 3.584e+02 4.494e+02 7.659e+02, threshold=7.169e+02, percent-clipped=0.0 2023-04-29 02:31:34,846 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1943, 4.1831, 4.0791, 3.4936, 4.1068, 1.5265, 3.8879, 3.8358], device='cuda:4'), covar=tensor([0.0093, 0.0069, 0.0128, 0.0285, 0.0080, 0.2414, 0.0107, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0103, 0.0151, 0.0145, 0.0119, 0.0167, 0.0135, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:31:38,444 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=88647.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 02:31:44,621 INFO [train.py:904] (4/8) Epoch 9, batch 7450, loss[loss=0.2165, simple_loss=0.3148, pruned_loss=0.05912, over 16861.00 frames. ], tot_loss[loss=0.2285, simple_loss=0.3086, pruned_loss=0.07417, over 3065479.33 frames. ], batch size: 96, lr: 7.51e-03, grad_scale: 4.0 2023-04-29 02:31:48,077 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5037, 2.1226, 2.2464, 4.1281, 2.0881, 2.5899, 2.2126, 2.3072], device='cuda:4'), covar=tensor([0.0885, 0.3058, 0.2038, 0.0400, 0.3609, 0.1982, 0.2801, 0.2941], device='cuda:4'), in_proj_covar=tensor([0.0350, 0.0376, 0.0314, 0.0318, 0.0405, 0.0427, 0.0336, 0.0439], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:33:03,515 INFO [train.py:904] (4/8) Epoch 9, batch 7500, loss[loss=0.2231, simple_loss=0.3053, pruned_loss=0.07043, over 16747.00 frames. ], tot_loss[loss=0.2282, simple_loss=0.309, pruned_loss=0.07365, over 3062079.09 frames. ], batch size: 124, lr: 7.51e-03, grad_scale: 4.0 2023-04-29 02:33:36,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.091e+02 3.551e+02 4.142e+02 5.326e+02 1.060e+03, threshold=8.283e+02, percent-clipped=5.0 2023-04-29 02:34:18,158 INFO [train.py:904] (4/8) Epoch 9, batch 7550, loss[loss=0.2218, simple_loss=0.2999, pruned_loss=0.07183, over 16275.00 frames. ], tot_loss[loss=0.2306, simple_loss=0.3098, pruned_loss=0.07569, over 3009217.62 frames. ], batch size: 165, lr: 7.51e-03, grad_scale: 4.0 2023-04-29 02:34:23,590 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88756.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:34:52,141 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-29 02:35:06,682 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5196, 4.4472, 4.3118, 4.1336, 3.9845, 4.3972, 4.2512, 4.1215], device='cuda:4'), covar=tensor([0.0431, 0.0324, 0.0250, 0.0244, 0.0844, 0.0342, 0.0383, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0268, 0.0255, 0.0235, 0.0276, 0.0269, 0.0179, 0.0299], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:35:31,492 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-29 02:35:33,596 INFO [train.py:904] (4/8) Epoch 9, batch 7600, loss[loss=0.2427, simple_loss=0.3064, pruned_loss=0.08951, over 11403.00 frames. ], tot_loss[loss=0.2295, simple_loss=0.3082, pruned_loss=0.07536, over 3021595.38 frames. ], batch size: 246, lr: 7.51e-03, grad_scale: 8.0 2023-04-29 02:35:37,865 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=88804.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:36:05,551 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.836e+02 3.313e+02 3.789e+02 4.639e+02 1.052e+03, threshold=7.577e+02, percent-clipped=2.0 2023-04-29 02:36:45,919 INFO [train.py:904] (4/8) Epoch 9, batch 7650, loss[loss=0.2349, simple_loss=0.3111, pruned_loss=0.07933, over 15386.00 frames. ], tot_loss[loss=0.231, simple_loss=0.3091, pruned_loss=0.07646, over 3020107.07 frames. ], batch size: 191, lr: 7.50e-03, grad_scale: 8.0 2023-04-29 02:38:01,877 INFO [train.py:904] (4/8) Epoch 9, batch 7700, loss[loss=0.2275, simple_loss=0.3067, pruned_loss=0.07415, over 16383.00 frames. ], tot_loss[loss=0.2311, simple_loss=0.309, pruned_loss=0.07661, over 3029511.98 frames. ], batch size: 146, lr: 7.50e-03, grad_scale: 8.0 2023-04-29 02:38:24,544 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4630, 3.4552, 2.6618, 2.1261, 2.4080, 2.2485, 3.5457, 3.3250], device='cuda:4'), covar=tensor([0.2702, 0.0866, 0.1717, 0.2138, 0.2070, 0.1785, 0.0576, 0.0952], device='cuda:4'), in_proj_covar=tensor([0.0300, 0.0258, 0.0282, 0.0271, 0.0281, 0.0216, 0.0266, 0.0284], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:38:34,565 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.958e+02 3.502e+02 4.245e+02 5.386e+02 9.706e+02, threshold=8.489e+02, percent-clipped=2.0 2023-04-29 02:38:35,597 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1010, 4.1243, 3.9437, 3.8021, 3.5608, 4.0704, 3.8280, 3.7724], device='cuda:4'), covar=tensor([0.0605, 0.0540, 0.0321, 0.0282, 0.1063, 0.0476, 0.0702, 0.0618], device='cuda:4'), in_proj_covar=tensor([0.0221, 0.0268, 0.0255, 0.0235, 0.0278, 0.0268, 0.0179, 0.0299], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:39:08,846 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88947.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 02:39:11,691 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7903, 5.2284, 5.4298, 5.2107, 5.2010, 5.8204, 5.3537, 5.1408], device='cuda:4'), covar=tensor([0.0845, 0.1607, 0.1601, 0.1553, 0.2141, 0.0820, 0.1291, 0.2036], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0453, 0.0488, 0.0404, 0.0529, 0.0517, 0.0395, 0.0544], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:39:16,828 INFO [train.py:904] (4/8) Epoch 9, batch 7750, loss[loss=0.2297, simple_loss=0.3131, pruned_loss=0.0732, over 16414.00 frames. ], tot_loss[loss=0.2316, simple_loss=0.3098, pruned_loss=0.07667, over 3028321.48 frames. ], batch size: 146, lr: 7.50e-03, grad_scale: 8.0 2023-04-29 02:39:31,143 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7371, 4.1248, 3.1586, 2.3058, 3.1374, 2.5624, 4.3620, 3.9132], device='cuda:4'), covar=tensor([0.2503, 0.0596, 0.1403, 0.1922, 0.2054, 0.1538, 0.0433, 0.0832], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0256, 0.0281, 0.0270, 0.0278, 0.0215, 0.0264, 0.0283], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:39:45,485 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0468, 5.3384, 5.1066, 5.0928, 4.7750, 4.7065, 4.7659, 5.4351], device='cuda:4'), covar=tensor([0.0900, 0.0753, 0.0995, 0.0643, 0.0743, 0.0767, 0.1006, 0.0850], device='cuda:4'), in_proj_covar=tensor([0.0491, 0.0624, 0.0526, 0.0423, 0.0385, 0.0406, 0.0517, 0.0470], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:40:20,776 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=88995.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 02:40:27,361 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2247, 4.2069, 4.6975, 4.6716, 4.6486, 4.3038, 4.3317, 4.1459], device='cuda:4'), covar=tensor([0.0296, 0.0514, 0.0295, 0.0375, 0.0396, 0.0333, 0.0898, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0304, 0.0310, 0.0297, 0.0345, 0.0324, 0.0428, 0.0261], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-29 02:40:29,342 INFO [train.py:904] (4/8) Epoch 9, batch 7800, loss[loss=0.2285, simple_loss=0.3086, pruned_loss=0.07415, over 16428.00 frames. ], tot_loss[loss=0.2319, simple_loss=0.3102, pruned_loss=0.07683, over 3039922.50 frames. ], batch size: 146, lr: 7.50e-03, grad_scale: 8.0 2023-04-29 02:41:02,671 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.440e+02 3.396e+02 4.306e+02 5.337e+02 1.534e+03, threshold=8.611e+02, percent-clipped=4.0 2023-04-29 02:41:07,763 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89027.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:41:19,801 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.35 vs. limit=5.0 2023-04-29 02:41:28,844 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5198, 3.6554, 2.6131, 2.0899, 2.6777, 2.2634, 3.6781, 3.4106], device='cuda:4'), covar=tensor([0.2687, 0.0739, 0.1821, 0.2280, 0.2252, 0.1814, 0.0596, 0.0983], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0255, 0.0279, 0.0269, 0.0277, 0.0213, 0.0262, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:41:42,475 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1186, 1.7163, 2.4658, 3.0453, 2.8152, 3.5236, 1.9180, 3.3586], device='cuda:4'), covar=tensor([0.0120, 0.0353, 0.0208, 0.0159, 0.0195, 0.0094, 0.0362, 0.0072], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0159, 0.0143, 0.0144, 0.0154, 0.0112, 0.0162, 0.0102], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 02:41:44,457 INFO [train.py:904] (4/8) Epoch 9, batch 7850, loss[loss=0.2144, simple_loss=0.3085, pruned_loss=0.06017, over 16821.00 frames. ], tot_loss[loss=0.2313, simple_loss=0.3106, pruned_loss=0.07594, over 3042366.02 frames. ], batch size: 83, lr: 7.49e-03, grad_scale: 8.0 2023-04-29 02:41:47,315 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-29 02:42:38,975 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89088.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:42:48,612 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-29 02:42:57,227 INFO [train.py:904] (4/8) Epoch 9, batch 7900, loss[loss=0.281, simple_loss=0.347, pruned_loss=0.1075, over 11263.00 frames. ], tot_loss[loss=0.2292, simple_loss=0.3093, pruned_loss=0.07448, over 3088943.07 frames. ], batch size: 246, lr: 7.49e-03, grad_scale: 8.0 2023-04-29 02:43:05,416 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.80 vs. limit=5.0 2023-04-29 02:43:22,246 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9695, 1.8639, 2.1167, 3.4762, 1.9122, 2.3095, 2.0235, 2.0111], device='cuda:4'), covar=tensor([0.0988, 0.3328, 0.2043, 0.0470, 0.3850, 0.2182, 0.2996, 0.3087], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0378, 0.0314, 0.0318, 0.0407, 0.0426, 0.0337, 0.0439], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:43:28,442 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.998e+02 3.084e+02 3.809e+02 4.466e+02 8.463e+02, threshold=7.617e+02, percent-clipped=0.0 2023-04-29 02:43:34,592 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9818, 2.2434, 1.7055, 2.0495, 2.6494, 2.4224, 2.9504, 2.9487], device='cuda:4'), covar=tensor([0.0076, 0.0254, 0.0409, 0.0336, 0.0164, 0.0259, 0.0150, 0.0151], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0189, 0.0188, 0.0186, 0.0188, 0.0190, 0.0190, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:44:12,924 INFO [train.py:904] (4/8) Epoch 9, batch 7950, loss[loss=0.2187, simple_loss=0.2987, pruned_loss=0.06936, over 16817.00 frames. ], tot_loss[loss=0.2297, simple_loss=0.3094, pruned_loss=0.07498, over 3079810.30 frames. ], batch size: 83, lr: 7.49e-03, grad_scale: 8.0 2023-04-29 02:44:39,093 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89169.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:45:00,709 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5758, 2.5889, 2.3177, 3.8532, 2.7289, 3.8354, 1.3729, 2.5915], device='cuda:4'), covar=tensor([0.1393, 0.0715, 0.1292, 0.0194, 0.0283, 0.0414, 0.1615, 0.0978], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0155, 0.0179, 0.0129, 0.0203, 0.0205, 0.0176, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 02:45:26,762 INFO [train.py:904] (4/8) Epoch 9, batch 8000, loss[loss=0.2405, simple_loss=0.3196, pruned_loss=0.08072, over 16700.00 frames. ], tot_loss[loss=0.2315, simple_loss=0.3105, pruned_loss=0.07621, over 3068894.44 frames. ], batch size: 134, lr: 7.49e-03, grad_scale: 8.0 2023-04-29 02:45:34,238 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0276, 3.5512, 3.4993, 2.2035, 3.1767, 3.4758, 3.3488, 1.9653], device='cuda:4'), covar=tensor([0.0393, 0.0031, 0.0034, 0.0301, 0.0077, 0.0088, 0.0060, 0.0353], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0063, 0.0065, 0.0124, 0.0072, 0.0084, 0.0072, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:45:59,520 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.399e+02 3.668e+02 4.137e+02 4.605e+02 8.803e+02, threshold=8.275e+02, percent-clipped=3.0 2023-04-29 02:46:09,109 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89230.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:46:40,073 INFO [train.py:904] (4/8) Epoch 9, batch 8050, loss[loss=0.2294, simple_loss=0.308, pruned_loss=0.07542, over 16950.00 frames. ], tot_loss[loss=0.2325, simple_loss=0.3112, pruned_loss=0.07687, over 3054870.76 frames. ], batch size: 41, lr: 7.49e-03, grad_scale: 4.0 2023-04-29 02:47:14,567 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6601, 5.0004, 5.2386, 5.0080, 4.9966, 5.6019, 5.1462, 4.8813], device='cuda:4'), covar=tensor([0.0912, 0.1570, 0.1616, 0.1686, 0.2364, 0.0872, 0.1288, 0.2343], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0454, 0.0488, 0.0407, 0.0528, 0.0514, 0.0394, 0.0543], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:47:24,441 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3710, 3.3167, 3.3880, 3.4922, 3.5327, 3.2474, 3.5065, 3.5683], device='cuda:4'), covar=tensor([0.1038, 0.0882, 0.0969, 0.0536, 0.0653, 0.2037, 0.0750, 0.0660], device='cuda:4'), in_proj_covar=tensor([0.0482, 0.0599, 0.0727, 0.0598, 0.0464, 0.0459, 0.0485, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:47:35,736 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-29 02:47:55,802 INFO [train.py:904] (4/8) Epoch 9, batch 8100, loss[loss=0.2124, simple_loss=0.3058, pruned_loss=0.0595, over 16753.00 frames. ], tot_loss[loss=0.23, simple_loss=0.3096, pruned_loss=0.07522, over 3067895.25 frames. ], batch size: 83, lr: 7.48e-03, grad_scale: 4.0 2023-04-29 02:48:01,463 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6645, 4.9660, 5.1434, 5.0498, 5.0344, 5.5642, 5.1085, 4.8555], device='cuda:4'), covar=tensor([0.0948, 0.1665, 0.1712, 0.1511, 0.2279, 0.0871, 0.1313, 0.2148], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0455, 0.0489, 0.0407, 0.0526, 0.0514, 0.0395, 0.0542], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:48:29,433 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.308e+02 2.948e+02 3.621e+02 4.670e+02 8.254e+02, threshold=7.243e+02, percent-clipped=0.0 2023-04-29 02:49:10,984 INFO [train.py:904] (4/8) Epoch 9, batch 8150, loss[loss=0.2123, simple_loss=0.2879, pruned_loss=0.06832, over 15403.00 frames. ], tot_loss[loss=0.228, simple_loss=0.3072, pruned_loss=0.07442, over 3063024.65 frames. ], batch size: 191, lr: 7.48e-03, grad_scale: 4.0 2023-04-29 02:49:45,589 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89374.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:49:59,365 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89383.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:50:28,007 INFO [train.py:904] (4/8) Epoch 9, batch 8200, loss[loss=0.2296, simple_loss=0.3092, pruned_loss=0.07504, over 16243.00 frames. ], tot_loss[loss=0.2255, simple_loss=0.3045, pruned_loss=0.07328, over 3092261.12 frames. ], batch size: 165, lr: 7.48e-03, grad_scale: 4.0 2023-04-29 02:51:05,467 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89425.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:51:06,155 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.992e+02 3.313e+02 4.005e+02 4.576e+02 8.683e+02, threshold=8.011e+02, percent-clipped=3.0 2023-04-29 02:51:20,848 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89435.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:51:27,463 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89439.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 02:51:47,090 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-29 02:51:47,428 INFO [train.py:904] (4/8) Epoch 9, batch 8250, loss[loss=0.1991, simple_loss=0.2893, pruned_loss=0.05446, over 15175.00 frames. ], tot_loss[loss=0.2229, simple_loss=0.3037, pruned_loss=0.07104, over 3065306.51 frames. ], batch size: 190, lr: 7.48e-03, grad_scale: 2.0 2023-04-29 02:51:58,052 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4523, 1.9593, 1.6931, 1.6420, 2.2113, 1.9813, 2.2658, 2.3880], device='cuda:4'), covar=tensor([0.0081, 0.0236, 0.0306, 0.0309, 0.0142, 0.0210, 0.0129, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0188, 0.0187, 0.0185, 0.0186, 0.0188, 0.0188, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:52:43,703 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89486.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:52:49,163 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.53 vs. limit=5.0 2023-04-29 02:53:06,351 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89500.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 02:53:08,807 INFO [train.py:904] (4/8) Epoch 9, batch 8300, loss[loss=0.1984, simple_loss=0.2983, pruned_loss=0.04932, over 16847.00 frames. ], tot_loss[loss=0.2182, simple_loss=0.3009, pruned_loss=0.06777, over 3058578.10 frames. ], batch size: 102, lr: 7.48e-03, grad_scale: 2.0 2023-04-29 02:53:21,165 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89509.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:53:31,525 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3709, 4.4410, 4.5777, 4.4541, 4.4782, 5.0084, 4.6663, 4.3493], device='cuda:4'), covar=tensor([0.1251, 0.1593, 0.1790, 0.1818, 0.2620, 0.0942, 0.1166, 0.2344], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0442, 0.0474, 0.0391, 0.0509, 0.0504, 0.0383, 0.0524], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 02:53:42,898 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6141, 3.7088, 2.9208, 2.1015, 2.6610, 2.2875, 3.9600, 3.5445], device='cuda:4'), covar=tensor([0.2573, 0.0684, 0.1409, 0.2381, 0.2341, 0.1859, 0.0418, 0.0859], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0253, 0.0279, 0.0269, 0.0274, 0.0214, 0.0260, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:53:47,062 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89525.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:53:48,528 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.763e+02 2.745e+02 3.164e+02 3.682e+02 6.265e+02, threshold=6.329e+02, percent-clipped=0.0 2023-04-29 02:54:31,537 INFO [train.py:904] (4/8) Epoch 9, batch 8350, loss[loss=0.1959, simple_loss=0.2919, pruned_loss=0.04994, over 16893.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.2996, pruned_loss=0.0651, over 3070659.60 frames. ], batch size: 96, lr: 7.47e-03, grad_scale: 2.0 2023-04-29 02:54:32,328 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2492, 1.5052, 1.8860, 2.2992, 2.3756, 2.4967, 1.5962, 2.4478], device='cuda:4'), covar=tensor([0.0133, 0.0333, 0.0228, 0.0183, 0.0175, 0.0147, 0.0359, 0.0081], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0163, 0.0146, 0.0148, 0.0158, 0.0115, 0.0165, 0.0103], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 02:54:58,150 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9237, 4.3363, 3.7082, 2.5609, 3.1077, 2.6352, 4.6722, 3.9607], device='cuda:4'), covar=tensor([0.2190, 0.0439, 0.1011, 0.2029, 0.2173, 0.1567, 0.0255, 0.0683], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0250, 0.0277, 0.0266, 0.0270, 0.0212, 0.0257, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:55:01,213 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89570.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:55:01,467 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-29 02:55:52,201 INFO [train.py:904] (4/8) Epoch 9, batch 8400, loss[loss=0.1916, simple_loss=0.2722, pruned_loss=0.05549, over 12046.00 frames. ], tot_loss[loss=0.2119, simple_loss=0.2974, pruned_loss=0.06323, over 3052488.09 frames. ], batch size: 247, lr: 7.47e-03, grad_scale: 4.0 2023-04-29 02:56:31,315 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.203e+02 2.904e+02 3.369e+02 3.930e+02 8.032e+02, threshold=6.737e+02, percent-clipped=2.0 2023-04-29 02:57:13,775 INFO [train.py:904] (4/8) Epoch 9, batch 8450, loss[loss=0.1954, simple_loss=0.2875, pruned_loss=0.05169, over 15278.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2948, pruned_loss=0.06117, over 3044814.99 frames. ], batch size: 190, lr: 7.47e-03, grad_scale: 4.0 2023-04-29 02:57:53,871 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8790, 2.2311, 1.8548, 1.9840, 2.5618, 2.3037, 2.8070, 2.8224], device='cuda:4'), covar=tensor([0.0083, 0.0251, 0.0331, 0.0312, 0.0157, 0.0240, 0.0126, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0187, 0.0185, 0.0186, 0.0186, 0.0185, 0.0185, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 02:58:00,151 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89680.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:58:04,764 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=89683.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:58:34,964 INFO [train.py:904] (4/8) Epoch 9, batch 8500, loss[loss=0.1867, simple_loss=0.2741, pruned_loss=0.04965, over 16690.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2905, pruned_loss=0.05833, over 3040405.31 frames. ], batch size: 124, lr: 7.47e-03, grad_scale: 4.0 2023-04-29 02:58:59,625 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.42 vs. limit=5.0 2023-04-29 02:59:14,096 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.548e+02 2.444e+02 3.113e+02 3.731e+02 7.658e+02, threshold=6.225e+02, percent-clipped=1.0 2023-04-29 02:59:20,981 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89730.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:59:22,203 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=89731.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:59:39,437 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89741.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 02:59:48,823 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.28 vs. limit=5.0 2023-04-29 02:59:58,037 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-29 02:59:58,592 INFO [train.py:904] (4/8) Epoch 9, batch 8550, loss[loss=0.2134, simple_loss=0.2837, pruned_loss=0.07152, over 11604.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2875, pruned_loss=0.05708, over 3006522.17 frames. ], batch size: 247, lr: 7.47e-03, grad_scale: 4.0 2023-04-29 03:00:55,135 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89781.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:01:24,797 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89795.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 03:01:37,952 INFO [train.py:904] (4/8) Epoch 9, batch 8600, loss[loss=0.1924, simple_loss=0.283, pruned_loss=0.05092, over 16144.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2876, pruned_loss=0.05628, over 3004214.81 frames. ], batch size: 165, lr: 7.46e-03, grad_scale: 4.0 2023-04-29 03:01:39,742 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4585, 3.3730, 3.4777, 3.5734, 3.5987, 3.2669, 3.5842, 3.6170], device='cuda:4'), covar=tensor([0.1048, 0.0837, 0.1005, 0.0558, 0.0593, 0.2079, 0.0663, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0467, 0.0577, 0.0698, 0.0583, 0.0452, 0.0451, 0.0468, 0.0526], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:02:25,763 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=89825.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:02:26,423 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.775e+02 2.695e+02 3.313e+02 4.176e+02 8.025e+02, threshold=6.625e+02, percent-clipped=2.0 2023-04-29 03:03:15,780 INFO [train.py:904] (4/8) Epoch 9, batch 8650, loss[loss=0.1735, simple_loss=0.2596, pruned_loss=0.04374, over 12267.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2844, pruned_loss=0.0545, over 2972950.47 frames. ], batch size: 247, lr: 7.46e-03, grad_scale: 4.0 2023-04-29 03:03:47,700 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89865.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:03:47,807 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89865.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:04:05,356 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=89873.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:04:31,937 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8045, 3.7201, 3.9282, 3.7323, 3.8635, 4.2838, 3.9709, 3.6123], device='cuda:4'), covar=tensor([0.1904, 0.2192, 0.1973, 0.2309, 0.3017, 0.1610, 0.1383, 0.3094], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0439, 0.0469, 0.0384, 0.0505, 0.0494, 0.0383, 0.0516], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 03:05:02,347 INFO [train.py:904] (4/8) Epoch 9, batch 8700, loss[loss=0.1959, simple_loss=0.2797, pruned_loss=0.05608, over 16818.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2813, pruned_loss=0.05268, over 3004983.62 frames. ], batch size: 124, lr: 7.46e-03, grad_scale: 4.0 2023-04-29 03:05:15,399 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.26 vs. limit=5.0 2023-04-29 03:05:45,060 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.721e+02 2.547e+02 3.087e+02 3.551e+02 5.785e+02, threshold=6.175e+02, percent-clipped=0.0 2023-04-29 03:05:45,906 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2642, 2.2914, 1.9778, 1.9965, 2.7698, 2.4313, 3.1027, 2.8743], device='cuda:4'), covar=tensor([0.0076, 0.0275, 0.0343, 0.0319, 0.0169, 0.0252, 0.0105, 0.0178], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0186, 0.0184, 0.0184, 0.0183, 0.0185, 0.0180, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:05:45,909 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89926.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 03:06:36,263 INFO [train.py:904] (4/8) Epoch 9, batch 8750, loss[loss=0.218, simple_loss=0.3057, pruned_loss=0.06521, over 16676.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2814, pruned_loss=0.05234, over 3007034.19 frames. ], batch size: 134, lr: 7.46e-03, grad_scale: 4.0 2023-04-29 03:07:05,551 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6515, 3.2410, 3.1593, 1.7214, 2.7046, 2.2284, 3.0899, 3.2026], device='cuda:4'), covar=tensor([0.0278, 0.0581, 0.0528, 0.1920, 0.0738, 0.0903, 0.0755, 0.0810], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0131, 0.0152, 0.0138, 0.0129, 0.0122, 0.0132, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 03:08:32,637 INFO [train.py:904] (4/8) Epoch 9, batch 8800, loss[loss=0.2204, simple_loss=0.2962, pruned_loss=0.07235, over 12881.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.28, pruned_loss=0.05148, over 2992599.14 frames. ], batch size: 249, lr: 7.46e-03, grad_scale: 8.0 2023-04-29 03:09:21,968 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.576e+02 2.546e+02 3.019e+02 3.613e+02 7.330e+02, threshold=6.037e+02, percent-clipped=3.0 2023-04-29 03:09:31,828 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90030.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:09:44,355 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90036.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:10:17,856 INFO [train.py:904] (4/8) Epoch 9, batch 8850, loss[loss=0.2075, simple_loss=0.3065, pruned_loss=0.05421, over 16360.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2829, pruned_loss=0.05081, over 3005934.02 frames. ], batch size: 146, lr: 7.45e-03, grad_scale: 8.0 2023-04-29 03:10:32,663 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0334, 4.0174, 3.9054, 3.4632, 3.9541, 1.5438, 3.7517, 3.6043], device='cuda:4'), covar=tensor([0.0074, 0.0065, 0.0116, 0.0210, 0.0074, 0.2283, 0.0104, 0.0187], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0101, 0.0147, 0.0139, 0.0118, 0.0167, 0.0133, 0.0137], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:11:12,094 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=90078.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:11:18,197 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90081.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:11:50,041 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90095.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 03:12:02,853 INFO [train.py:904] (4/8) Epoch 9, batch 8900, loss[loss=0.1855, simple_loss=0.2752, pruned_loss=0.04792, over 12553.00 frames. ], tot_loss[loss=0.1917, simple_loss=0.2833, pruned_loss=0.05003, over 3031222.52 frames. ], batch size: 246, lr: 7.45e-03, grad_scale: 4.0 2023-04-29 03:12:57,513 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.859e+02 2.641e+02 3.095e+02 3.699e+02 6.742e+02, threshold=6.190e+02, percent-clipped=4.0 2023-04-29 03:13:07,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=90129.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:13:10,470 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90130.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:13:47,019 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-29 03:13:48,655 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=90143.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 03:14:08,704 INFO [train.py:904] (4/8) Epoch 9, batch 8950, loss[loss=0.1764, simple_loss=0.2664, pruned_loss=0.04322, over 16314.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2824, pruned_loss=0.04986, over 3063224.07 frames. ], batch size: 166, lr: 7.45e-03, grad_scale: 4.0 2023-04-29 03:14:12,263 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-29 03:14:38,339 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90165.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:15:35,030 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90191.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:15:50,886 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-29 03:15:57,359 INFO [train.py:904] (4/8) Epoch 9, batch 9000, loss[loss=0.1815, simple_loss=0.2721, pruned_loss=0.04541, over 16798.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2794, pruned_loss=0.0484, over 3060611.16 frames. ], batch size: 124, lr: 7.45e-03, grad_scale: 4.0 2023-04-29 03:15:57,360 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 03:16:07,533 INFO [train.py:938] (4/8) Epoch 9, validation: loss=0.1581, simple_loss=0.2623, pruned_loss=0.02697, over 944034.00 frames. 2023-04-29 03:16:07,533 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-29 03:16:31,272 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=90213.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:16:40,086 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4082, 3.3396, 3.4353, 3.5275, 3.5683, 3.2526, 3.5341, 3.5869], device='cuda:4'), covar=tensor([0.1032, 0.0875, 0.1033, 0.0609, 0.0558, 0.2603, 0.0896, 0.0657], device='cuda:4'), in_proj_covar=tensor([0.0458, 0.0566, 0.0687, 0.0573, 0.0440, 0.0441, 0.0460, 0.0516], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:16:47,981 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90221.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 03:16:51,506 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-29 03:16:58,710 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.701e+02 2.410e+02 2.881e+02 3.616e+02 7.746e+02, threshold=5.761e+02, percent-clipped=2.0 2023-04-29 03:17:08,224 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90232.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:17:33,945 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6636, 1.6260, 2.1154, 2.7048, 2.4756, 2.9283, 1.8398, 2.8222], device='cuda:4'), covar=tensor([0.0117, 0.0348, 0.0225, 0.0163, 0.0196, 0.0121, 0.0324, 0.0095], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0159, 0.0144, 0.0143, 0.0153, 0.0110, 0.0161, 0.0099], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 03:17:51,878 INFO [train.py:904] (4/8) Epoch 9, batch 9050, loss[loss=0.2101, simple_loss=0.2878, pruned_loss=0.0662, over 13007.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2817, pruned_loss=0.04948, over 3067364.31 frames. ], batch size: 248, lr: 7.45e-03, grad_scale: 4.0 2023-04-29 03:18:12,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7220, 3.8197, 2.1230, 4.3196, 2.6897, 4.1866, 2.2703, 2.9929], device='cuda:4'), covar=tensor([0.0207, 0.0298, 0.1613, 0.0105, 0.0875, 0.0416, 0.1480, 0.0651], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0151, 0.0178, 0.0106, 0.0158, 0.0188, 0.0185, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-29 03:18:31,400 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.72 vs. limit=5.0 2023-04-29 03:19:19,149 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90293.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:19:36,840 INFO [train.py:904] (4/8) Epoch 9, batch 9100, loss[loss=0.1886, simple_loss=0.2697, pruned_loss=0.05375, over 12335.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.281, pruned_loss=0.04964, over 3067192.81 frames. ], batch size: 250, lr: 7.44e-03, grad_scale: 4.0 2023-04-29 03:20:34,108 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.733e+02 2.617e+02 3.274e+02 4.462e+02 7.495e+02, threshold=6.548e+02, percent-clipped=8.0 2023-04-29 03:20:58,699 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90336.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:21:33,859 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90351.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:21:35,357 INFO [train.py:904] (4/8) Epoch 9, batch 9150, loss[loss=0.1723, simple_loss=0.2597, pruned_loss=0.04247, over 12150.00 frames. ], tot_loss[loss=0.1898, simple_loss=0.2808, pruned_loss=0.04942, over 3054703.49 frames. ], batch size: 248, lr: 7.44e-03, grad_scale: 4.0 2023-04-29 03:22:45,425 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=90384.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:22:59,595 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0828, 2.8231, 2.8239, 2.0576, 2.6409, 2.1459, 2.7112, 2.9068], device='cuda:4'), covar=tensor([0.0268, 0.0689, 0.0477, 0.1490, 0.0671, 0.0850, 0.0676, 0.0699], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0130, 0.0151, 0.0138, 0.0130, 0.0122, 0.0132, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 03:23:22,749 INFO [train.py:904] (4/8) Epoch 9, batch 9200, loss[loss=0.1816, simple_loss=0.275, pruned_loss=0.04407, over 16864.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2766, pruned_loss=0.04832, over 3051989.93 frames. ], batch size: 96, lr: 7.44e-03, grad_scale: 8.0 2023-04-29 03:23:34,027 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-29 03:23:42,315 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90412.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:24:07,624 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.818e+02 2.608e+02 2.952e+02 3.728e+02 7.219e+02, threshold=5.904e+02, percent-clipped=3.0 2023-04-29 03:24:58,588 INFO [train.py:904] (4/8) Epoch 9, batch 9250, loss[loss=0.1556, simple_loss=0.2377, pruned_loss=0.03677, over 11980.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2765, pruned_loss=0.04806, over 3058666.47 frames. ], batch size: 246, lr: 7.44e-03, grad_scale: 8.0 2023-04-29 03:25:14,805 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90459.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:25:41,608 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-29 03:26:08,323 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-29 03:26:13,012 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90486.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:26:38,482 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90496.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:26:49,729 INFO [train.py:904] (4/8) Epoch 9, batch 9300, loss[loss=0.1756, simple_loss=0.2563, pruned_loss=0.04744, over 12288.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2739, pruned_loss=0.04714, over 3037060.68 frames. ], batch size: 246, lr: 7.43e-03, grad_scale: 8.0 2023-04-29 03:27:33,267 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9342, 3.8070, 3.9796, 4.1385, 4.2358, 3.7933, 4.2097, 4.2255], device='cuda:4'), covar=tensor([0.1291, 0.0979, 0.1248, 0.0613, 0.0492, 0.1438, 0.0629, 0.0569], device='cuda:4'), in_proj_covar=tensor([0.0456, 0.0561, 0.0679, 0.0568, 0.0438, 0.0436, 0.0455, 0.0511], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:27:35,089 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90520.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:27:36,965 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90521.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 03:27:46,916 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.664e+02 2.420e+02 2.895e+02 3.675e+02 7.033e+02, threshold=5.790e+02, percent-clipped=1.0 2023-04-29 03:28:36,079 INFO [train.py:904] (4/8) Epoch 9, batch 9350, loss[loss=0.1936, simple_loss=0.2808, pruned_loss=0.05324, over 16961.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2741, pruned_loss=0.04683, over 3080976.54 frames. ], batch size: 109, lr: 7.43e-03, grad_scale: 8.0 2023-04-29 03:28:48,572 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90557.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:29:14,438 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=90569.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:29:42,428 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-29 03:29:50,781 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90588.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:30:18,678 INFO [train.py:904] (4/8) Epoch 9, batch 9400, loss[loss=0.2021, simple_loss=0.2954, pruned_loss=0.05439, over 16297.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2738, pruned_loss=0.04655, over 3072279.02 frames. ], batch size: 146, lr: 7.43e-03, grad_scale: 8.0 2023-04-29 03:30:46,280 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7408, 1.2465, 1.6470, 1.6586, 1.7665, 1.8312, 1.5415, 1.8319], device='cuda:4'), covar=tensor([0.0168, 0.0281, 0.0155, 0.0190, 0.0207, 0.0141, 0.0286, 0.0076], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0160, 0.0144, 0.0143, 0.0154, 0.0110, 0.0161, 0.0100], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 03:30:47,752 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-29 03:31:09,565 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.896e+02 2.474e+02 2.959e+02 3.714e+02 8.908e+02, threshold=5.918e+02, percent-clipped=5.0 2023-04-29 03:32:00,895 INFO [train.py:904] (4/8) Epoch 9, batch 9450, loss[loss=0.1909, simple_loss=0.2808, pruned_loss=0.05052, over 16961.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2756, pruned_loss=0.04682, over 3063384.74 frames. ], batch size: 109, lr: 7.43e-03, grad_scale: 8.0 2023-04-29 03:33:25,598 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2910, 1.9675, 2.1108, 3.9151, 1.9780, 2.4591, 2.1426, 2.1179], device='cuda:4'), covar=tensor([0.0836, 0.3440, 0.2163, 0.0345, 0.3852, 0.2173, 0.3151, 0.3332], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0362, 0.0308, 0.0305, 0.0395, 0.0404, 0.0327, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:33:30,920 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0185, 4.0048, 3.8841, 3.4420, 3.9412, 1.7314, 3.7326, 3.6616], device='cuda:4'), covar=tensor([0.0073, 0.0070, 0.0126, 0.0215, 0.0074, 0.2265, 0.0106, 0.0173], device='cuda:4'), in_proj_covar=tensor([0.0112, 0.0099, 0.0144, 0.0133, 0.0115, 0.0166, 0.0130, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-29 03:33:43,967 INFO [train.py:904] (4/8) Epoch 9, batch 9500, loss[loss=0.167, simple_loss=0.2522, pruned_loss=0.04088, over 12762.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2743, pruned_loss=0.04619, over 3061126.02 frames. ], batch size: 246, lr: 7.43e-03, grad_scale: 8.0 2023-04-29 03:33:58,715 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90707.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:34:35,316 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.646e+02 2.298e+02 2.954e+02 3.660e+02 6.291e+02, threshold=5.908e+02, percent-clipped=3.0 2023-04-29 03:35:24,990 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6553, 4.9074, 5.0413, 4.9275, 4.9605, 5.4766, 5.0223, 4.7784], device='cuda:4'), covar=tensor([0.0940, 0.1896, 0.1705, 0.1707, 0.2553, 0.1098, 0.1489, 0.2147], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0433, 0.0456, 0.0376, 0.0496, 0.0483, 0.0377, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:35:30,036 INFO [train.py:904] (4/8) Epoch 9, batch 9550, loss[loss=0.1855, simple_loss=0.2718, pruned_loss=0.04956, over 12685.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2743, pruned_loss=0.04641, over 3060638.83 frames. ], batch size: 248, lr: 7.42e-03, grad_scale: 8.0 2023-04-29 03:36:43,165 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90786.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:37:11,178 INFO [train.py:904] (4/8) Epoch 9, batch 9600, loss[loss=0.2374, simple_loss=0.3273, pruned_loss=0.07381, over 16224.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2767, pruned_loss=0.04768, over 3062766.07 frames. ], batch size: 165, lr: 7.42e-03, grad_scale: 8.0 2023-04-29 03:37:37,099 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90815.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:37:59,117 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.470e+02 2.542e+02 3.036e+02 4.023e+02 8.440e+02, threshold=6.073e+02, percent-clipped=4.0 2023-04-29 03:38:17,844 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=90834.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:38:55,249 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90849.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:39:01,490 INFO [train.py:904] (4/8) Epoch 9, batch 9650, loss[loss=0.1641, simple_loss=0.2601, pruned_loss=0.03403, over 16536.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2779, pruned_loss=0.04749, over 3074449.02 frames. ], batch size: 68, lr: 7.42e-03, grad_scale: 8.0 2023-04-29 03:39:03,619 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90852.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:39:03,778 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9637, 2.3908, 2.3277, 3.0673, 2.1273, 3.2928, 1.6153, 2.7821], device='cuda:4'), covar=tensor([0.1216, 0.0510, 0.1017, 0.0131, 0.0120, 0.0403, 0.1414, 0.0662], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0150, 0.0176, 0.0123, 0.0181, 0.0202, 0.0175, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 03:40:18,928 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90888.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:40:48,113 INFO [train.py:904] (4/8) Epoch 9, batch 9700, loss[loss=0.1699, simple_loss=0.2558, pruned_loss=0.04201, over 12331.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2772, pruned_loss=0.04767, over 3069810.45 frames. ], batch size: 248, lr: 7.42e-03, grad_scale: 8.0 2023-04-29 03:41:04,597 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90910.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:41:18,447 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4820, 3.3395, 2.7736, 2.1163, 2.3517, 2.1428, 3.4125, 3.1882], device='cuda:4'), covar=tensor([0.2369, 0.0650, 0.1355, 0.2237, 0.2159, 0.1777, 0.0404, 0.0904], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0243, 0.0268, 0.0260, 0.0247, 0.0206, 0.0251, 0.0264], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:41:40,521 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.598e+02 2.339e+02 3.062e+02 3.716e+02 7.920e+02, threshold=6.123e+02, percent-clipped=1.0 2023-04-29 03:41:59,659 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=90936.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:42:31,486 INFO [train.py:904] (4/8) Epoch 9, batch 9750, loss[loss=0.1798, simple_loss=0.2626, pruned_loss=0.04848, over 12597.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2755, pruned_loss=0.0478, over 3040848.67 frames. ], batch size: 248, lr: 7.42e-03, grad_scale: 8.0 2023-04-29 03:44:00,347 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1857, 3.3626, 3.6266, 3.6022, 3.6080, 3.3908, 3.4192, 3.4467], device='cuda:4'), covar=tensor([0.0392, 0.0632, 0.0384, 0.0443, 0.0502, 0.0438, 0.0775, 0.0400], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0291, 0.0296, 0.0285, 0.0333, 0.0312, 0.0401, 0.0252], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-29 03:44:10,866 INFO [train.py:904] (4/8) Epoch 9, batch 9800, loss[loss=0.194, simple_loss=0.2922, pruned_loss=0.04789, over 15471.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2753, pruned_loss=0.0464, over 3044928.56 frames. ], batch size: 190, lr: 7.41e-03, grad_scale: 8.0 2023-04-29 03:44:21,925 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91007.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:44:44,572 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-29 03:44:57,913 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.670e+02 2.313e+02 2.751e+02 3.431e+02 5.847e+02, threshold=5.502e+02, percent-clipped=0.0 2023-04-29 03:45:23,243 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1588, 3.6247, 3.6352, 1.9604, 2.9757, 2.4340, 3.5715, 3.5336], device='cuda:4'), covar=tensor([0.0256, 0.0540, 0.0473, 0.1716, 0.0689, 0.0866, 0.0658, 0.0891], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0130, 0.0153, 0.0140, 0.0131, 0.0123, 0.0133, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 03:45:57,961 INFO [train.py:904] (4/8) Epoch 9, batch 9850, loss[loss=0.1846, simple_loss=0.2808, pruned_loss=0.04417, over 16847.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2765, pruned_loss=0.04573, over 3067804.64 frames. ], batch size: 124, lr: 7.41e-03, grad_scale: 8.0 2023-04-29 03:46:04,900 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=91055.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:47:48,711 INFO [train.py:904] (4/8) Epoch 9, batch 9900, loss[loss=0.1949, simple_loss=0.2752, pruned_loss=0.05736, over 12536.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.2776, pruned_loss=0.04602, over 3074696.68 frames. ], batch size: 250, lr: 7.41e-03, grad_scale: 8.0 2023-04-29 03:48:20,561 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91115.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:48:47,932 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.713e+02 2.434e+02 3.232e+02 4.008e+02 9.044e+02, threshold=6.464e+02, percent-clipped=5.0 2023-04-29 03:49:47,990 INFO [train.py:904] (4/8) Epoch 9, batch 9950, loss[loss=0.1772, simple_loss=0.2712, pruned_loss=0.04159, over 16661.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2798, pruned_loss=0.0465, over 3087469.31 frames. ], batch size: 134, lr: 7.41e-03, grad_scale: 8.0 2023-04-29 03:49:49,171 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91152.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:50:14,538 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=91163.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:51:44,349 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=91200.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:51:47,235 INFO [train.py:904] (4/8) Epoch 9, batch 10000, loss[loss=0.1778, simple_loss=0.2826, pruned_loss=0.03654, over 17032.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.2775, pruned_loss=0.04549, over 3107058.09 frames. ], batch size: 55, lr: 7.41e-03, grad_scale: 8.0 2023-04-29 03:51:53,872 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=91205.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:52:17,636 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6573, 2.6952, 2.5616, 4.2179, 2.8178, 4.0944, 1.4882, 3.0657], device='cuda:4'), covar=tensor([0.1378, 0.0644, 0.1046, 0.0090, 0.0152, 0.0316, 0.1441, 0.0627], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0149, 0.0173, 0.0120, 0.0176, 0.0198, 0.0173, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 03:52:35,744 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.638e+02 2.574e+02 2.985e+02 3.622e+02 7.238e+02, threshold=5.969e+02, percent-clipped=2.0 2023-04-29 03:53:27,231 INFO [train.py:904] (4/8) Epoch 9, batch 10050, loss[loss=0.2037, simple_loss=0.2958, pruned_loss=0.05583, over 16949.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2774, pruned_loss=0.04533, over 3104588.41 frames. ], batch size: 109, lr: 7.40e-03, grad_scale: 8.0 2023-04-29 03:54:30,782 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91285.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:54:48,359 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9209, 1.7347, 1.5600, 1.4651, 1.9294, 1.5555, 1.7361, 1.9325], device='cuda:4'), covar=tensor([0.0073, 0.0219, 0.0274, 0.0279, 0.0153, 0.0223, 0.0135, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0190, 0.0184, 0.0184, 0.0183, 0.0186, 0.0178, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:55:01,277 INFO [train.py:904] (4/8) Epoch 9, batch 10100, loss[loss=0.1828, simple_loss=0.2621, pruned_loss=0.05178, over 12529.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2776, pruned_loss=0.04576, over 3090193.94 frames. ], batch size: 248, lr: 7.40e-03, grad_scale: 8.0 2023-04-29 03:55:52,059 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.735e+02 2.558e+02 3.237e+02 3.853e+02 8.628e+02, threshold=6.474e+02, percent-clipped=2.0 2023-04-29 03:56:08,640 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0725, 2.7979, 2.7793, 2.0325, 2.5792, 2.1414, 2.7541, 2.8848], device='cuda:4'), covar=tensor([0.0323, 0.0749, 0.0484, 0.1616, 0.0729, 0.0934, 0.0639, 0.0832], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0129, 0.0154, 0.0140, 0.0130, 0.0123, 0.0131, 0.0140], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 03:56:14,317 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=91346.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 03:56:45,259 INFO [train.py:904] (4/8) Epoch 10, batch 0, loss[loss=0.3131, simple_loss=0.3575, pruned_loss=0.1344, over 16690.00 frames. ], tot_loss[loss=0.3131, simple_loss=0.3575, pruned_loss=0.1344, over 16690.00 frames. ], batch size: 134, lr: 7.04e-03, grad_scale: 8.0 2023-04-29 03:56:45,259 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 03:56:50,560 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2280, 4.5371, 4.4055, 4.4511, 4.1411, 4.3173, 4.1318, 4.5513], device='cuda:4'), covar=tensor([0.0810, 0.0706, 0.0648, 0.0473, 0.0723, 0.0375, 0.0755, 0.0638], device='cuda:4'), in_proj_covar=tensor([0.0467, 0.0596, 0.0485, 0.0407, 0.0373, 0.0388, 0.0496, 0.0443], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:56:52,894 INFO [train.py:938] (4/8) Epoch 10, validation: loss=0.158, simple_loss=0.2614, pruned_loss=0.02732, over 944034.00 frames. 2023-04-29 03:56:52,895 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17726MB 2023-04-29 03:57:33,842 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2100, 1.8831, 1.3981, 1.6064, 2.2771, 2.1061, 2.4316, 2.4123], device='cuda:4'), covar=tensor([0.0103, 0.0319, 0.0398, 0.0391, 0.0161, 0.0274, 0.0127, 0.0199], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0189, 0.0184, 0.0184, 0.0183, 0.0187, 0.0178, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 03:58:02,502 INFO [train.py:904] (4/8) Epoch 10, batch 50, loss[loss=0.2373, simple_loss=0.3083, pruned_loss=0.08312, over 16723.00 frames. ], tot_loss[loss=0.2174, simple_loss=0.2935, pruned_loss=0.07059, over 753467.77 frames. ], batch size: 134, lr: 7.04e-03, grad_scale: 2.0 2023-04-29 03:58:39,941 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.612e+02 2.984e+02 3.603e+02 4.569e+02 8.591e+02, threshold=7.207e+02, percent-clipped=1.0 2023-04-29 03:59:08,801 INFO [train.py:904] (4/8) Epoch 10, batch 100, loss[loss=0.1994, simple_loss=0.2846, pruned_loss=0.05712, over 17121.00 frames. ], tot_loss[loss=0.2079, simple_loss=0.2863, pruned_loss=0.06469, over 1325889.86 frames. ], batch size: 47, lr: 7.03e-03, grad_scale: 1.0 2023-04-29 03:59:38,019 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-04-29 03:59:58,501 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0558, 3.9691, 4.5111, 4.4746, 4.5036, 4.0801, 4.2277, 4.0870], device='cuda:4'), covar=tensor([0.0351, 0.0566, 0.0347, 0.0435, 0.0419, 0.0393, 0.0798, 0.0484], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0303, 0.0307, 0.0293, 0.0342, 0.0325, 0.0417, 0.0261], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-29 04:00:15,269 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-29 04:00:16,885 INFO [train.py:904] (4/8) Epoch 10, batch 150, loss[loss=0.2523, simple_loss=0.3258, pruned_loss=0.08935, over 15428.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2838, pruned_loss=0.06177, over 1776996.49 frames. ], batch size: 190, lr: 7.03e-03, grad_scale: 1.0 2023-04-29 04:00:22,948 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91505.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:00:56,393 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.669e+02 2.683e+02 3.349e+02 4.064e+02 6.042e+02, threshold=6.698e+02, percent-clipped=0.0 2023-04-29 04:01:12,392 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91542.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 04:01:26,563 INFO [train.py:904] (4/8) Epoch 10, batch 200, loss[loss=0.1705, simple_loss=0.253, pruned_loss=0.04399, over 16807.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2848, pruned_loss=0.06243, over 2112936.09 frames. ], batch size: 39, lr: 7.03e-03, grad_scale: 1.0 2023-04-29 04:01:28,065 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=91553.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:01:48,902 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5991, 4.1428, 4.4352, 3.1824, 3.7977, 4.2768, 4.0388, 2.5831], device='cuda:4'), covar=tensor([0.0380, 0.0040, 0.0020, 0.0245, 0.0068, 0.0052, 0.0051, 0.0322], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0067, 0.0066, 0.0124, 0.0075, 0.0081, 0.0074, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:02:34,674 INFO [train.py:904] (4/8) Epoch 10, batch 250, loss[loss=0.1821, simple_loss=0.2756, pruned_loss=0.04428, over 17136.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2815, pruned_loss=0.06035, over 2381943.49 frames. ], batch size: 48, lr: 7.03e-03, grad_scale: 1.0 2023-04-29 04:02:36,446 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=91603.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 04:03:11,344 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.913e+02 2.504e+02 3.085e+02 3.691e+02 6.376e+02, threshold=6.169e+02, percent-clipped=0.0 2023-04-29 04:03:28,030 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=91641.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:03:42,030 INFO [train.py:904] (4/8) Epoch 10, batch 300, loss[loss=0.1897, simple_loss=0.2674, pruned_loss=0.05596, over 16551.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2788, pruned_loss=0.05944, over 2591597.11 frames. ], batch size: 68, lr: 7.03e-03, grad_scale: 1.0 2023-04-29 04:04:51,290 INFO [train.py:904] (4/8) Epoch 10, batch 350, loss[loss=0.188, simple_loss=0.2662, pruned_loss=0.05487, over 16861.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2748, pruned_loss=0.05725, over 2748368.24 frames. ], batch size: 96, lr: 7.02e-03, grad_scale: 1.0 2023-04-29 04:05:25,755 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.71 vs. limit=5.0 2023-04-29 04:05:28,619 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.553e+02 2.384e+02 2.940e+02 3.585e+02 5.710e+02, threshold=5.881e+02, percent-clipped=0.0 2023-04-29 04:05:59,422 INFO [train.py:904] (4/8) Epoch 10, batch 400, loss[loss=0.1994, simple_loss=0.2724, pruned_loss=0.0632, over 16545.00 frames. ], tot_loss[loss=0.1928, simple_loss=0.2727, pruned_loss=0.05647, over 2880420.82 frames. ], batch size: 68, lr: 7.02e-03, grad_scale: 2.0 2023-04-29 04:06:07,111 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5691, 3.8125, 4.0427, 1.9523, 4.0752, 4.2059, 3.2346, 2.9330], device='cuda:4'), covar=tensor([0.0788, 0.0174, 0.0143, 0.1194, 0.0081, 0.0120, 0.0385, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0098, 0.0084, 0.0141, 0.0069, 0.0098, 0.0121, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 04:07:11,422 INFO [train.py:904] (4/8) Epoch 10, batch 450, loss[loss=0.2296, simple_loss=0.2961, pruned_loss=0.08155, over 12472.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2709, pruned_loss=0.05525, over 2984336.31 frames. ], batch size: 246, lr: 7.02e-03, grad_scale: 2.0 2023-04-29 04:07:50,757 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.774e+02 2.266e+02 2.904e+02 3.599e+02 6.333e+02, threshold=5.808e+02, percent-clipped=1.0 2023-04-29 04:08:20,248 INFO [train.py:904] (4/8) Epoch 10, batch 500, loss[loss=0.1822, simple_loss=0.265, pruned_loss=0.04964, over 16524.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2689, pruned_loss=0.05433, over 3054040.47 frames. ], batch size: 68, lr: 7.02e-03, grad_scale: 2.0 2023-04-29 04:09:15,098 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91891.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:09:23,881 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=91898.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 04:09:29,464 INFO [train.py:904] (4/8) Epoch 10, batch 550, loss[loss=0.2114, simple_loss=0.2804, pruned_loss=0.07121, over 16284.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2689, pruned_loss=0.05412, over 3115424.81 frames. ], batch size: 165, lr: 7.02e-03, grad_scale: 2.0 2023-04-29 04:09:32,790 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-29 04:09:46,663 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-29 04:10:07,828 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.776e+02 2.517e+02 3.166e+02 3.975e+02 1.037e+03, threshold=6.333e+02, percent-clipped=5.0 2023-04-29 04:10:08,372 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2609, 1.8520, 2.6395, 3.1129, 2.9705, 3.5207, 1.7592, 3.4267], device='cuda:4'), covar=tensor([0.0110, 0.0379, 0.0182, 0.0172, 0.0156, 0.0111, 0.0457, 0.0091], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0169, 0.0154, 0.0155, 0.0164, 0.0120, 0.0170, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 04:10:22,770 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91941.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:10:38,342 INFO [train.py:904] (4/8) Epoch 10, batch 600, loss[loss=0.2027, simple_loss=0.2898, pruned_loss=0.05777, over 17078.00 frames. ], tot_loss[loss=0.1885, simple_loss=0.2685, pruned_loss=0.05421, over 3168466.19 frames. ], batch size: 55, lr: 7.01e-03, grad_scale: 2.0 2023-04-29 04:10:38,844 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=91952.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:10:42,863 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91955.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:11:18,679 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91981.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 04:11:30,057 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=91989.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:11:51,293 INFO [train.py:904] (4/8) Epoch 10, batch 650, loss[loss=0.149, simple_loss=0.2323, pruned_loss=0.03289, over 16938.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2666, pruned_loss=0.05339, over 3203981.08 frames. ], batch size: 41, lr: 7.01e-03, grad_scale: 2.0 2023-04-29 04:12:12,135 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92016.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:12:25,119 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6461, 2.5882, 1.8201, 2.7299, 2.1931, 2.7315, 1.9970, 2.3574], device='cuda:4'), covar=tensor([0.0233, 0.0343, 0.1249, 0.0246, 0.0601, 0.0488, 0.1195, 0.0573], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0164, 0.0188, 0.0124, 0.0167, 0.0203, 0.0194, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 04:12:31,800 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.739e+02 2.417e+02 3.017e+02 3.569e+02 8.138e+02, threshold=6.033e+02, percent-clipped=1.0 2023-04-29 04:12:49,482 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92042.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 04:13:02,284 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9459, 4.9156, 4.7649, 4.3924, 4.2735, 4.8801, 4.8159, 4.4596], device='cuda:4'), covar=tensor([0.0622, 0.0401, 0.0319, 0.0313, 0.1266, 0.0397, 0.0336, 0.0679], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0298, 0.0284, 0.0262, 0.0310, 0.0299, 0.0193, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:13:03,065 INFO [train.py:904] (4/8) Epoch 10, batch 700, loss[loss=0.197, simple_loss=0.2857, pruned_loss=0.05418, over 16674.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2671, pruned_loss=0.05317, over 3216026.11 frames. ], batch size: 62, lr: 7.01e-03, grad_scale: 2.0 2023-04-29 04:14:12,201 INFO [train.py:904] (4/8) Epoch 10, batch 750, loss[loss=0.1833, simple_loss=0.2594, pruned_loss=0.05359, over 12685.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2672, pruned_loss=0.05302, over 3237299.35 frames. ], batch size: 247, lr: 7.01e-03, grad_scale: 2.0 2023-04-29 04:14:50,129 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8434, 3.2261, 3.0216, 5.0884, 4.2951, 4.7133, 1.5911, 3.3200], device='cuda:4'), covar=tensor([0.1258, 0.0596, 0.0987, 0.0114, 0.0259, 0.0314, 0.1448, 0.0665], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0152, 0.0177, 0.0132, 0.0190, 0.0208, 0.0176, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 04:14:52,014 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.724e+02 2.681e+02 3.065e+02 4.087e+02 6.685e+02, threshold=6.130e+02, percent-clipped=5.0 2023-04-29 04:15:22,699 INFO [train.py:904] (4/8) Epoch 10, batch 800, loss[loss=0.1495, simple_loss=0.2276, pruned_loss=0.03571, over 16814.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2672, pruned_loss=0.05316, over 3255402.57 frames. ], batch size: 39, lr: 7.01e-03, grad_scale: 4.0 2023-04-29 04:15:43,442 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0490, 1.7227, 2.3339, 2.7882, 2.7017, 3.3319, 1.8353, 3.1148], device='cuda:4'), covar=tensor([0.0138, 0.0350, 0.0212, 0.0184, 0.0186, 0.0122, 0.0357, 0.0111], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0166, 0.0151, 0.0154, 0.0163, 0.0119, 0.0168, 0.0107], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 04:16:27,559 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=92198.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 04:16:32,819 INFO [train.py:904] (4/8) Epoch 10, batch 850, loss[loss=0.1585, simple_loss=0.2507, pruned_loss=0.03312, over 17219.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.266, pruned_loss=0.05228, over 3271630.75 frames. ], batch size: 44, lr: 7.00e-03, grad_scale: 4.0 2023-04-29 04:17:06,707 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.85 vs. limit=5.0 2023-04-29 04:17:10,140 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.539e+02 2.453e+02 2.911e+02 3.766e+02 9.676e+02, threshold=5.821e+02, percent-clipped=2.0 2023-04-29 04:17:29,405 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8752, 2.4467, 1.8233, 2.2095, 2.8849, 2.7101, 3.0453, 3.0258], device='cuda:4'), covar=tensor([0.0144, 0.0259, 0.0408, 0.0360, 0.0145, 0.0223, 0.0163, 0.0151], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0197, 0.0192, 0.0192, 0.0194, 0.0197, 0.0199, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:17:33,334 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=92246.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 04:17:34,334 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=92247.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:17:41,160 INFO [train.py:904] (4/8) Epoch 10, batch 900, loss[loss=0.1739, simple_loss=0.2693, pruned_loss=0.03923, over 17031.00 frames. ], tot_loss[loss=0.184, simple_loss=0.265, pruned_loss=0.05145, over 3288556.97 frames. ], batch size: 50, lr: 7.00e-03, grad_scale: 4.0 2023-04-29 04:17:45,180 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-29 04:17:59,388 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3200, 5.6656, 5.4718, 5.5264, 5.0235, 5.0082, 5.1181, 5.8725], device='cuda:4'), covar=tensor([0.1269, 0.0915, 0.0938, 0.0657, 0.0878, 0.0620, 0.1053, 0.0837], device='cuda:4'), in_proj_covar=tensor([0.0526, 0.0667, 0.0547, 0.0456, 0.0419, 0.0424, 0.0560, 0.0500], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:18:50,838 INFO [train.py:904] (4/8) Epoch 10, batch 950, loss[loss=0.1654, simple_loss=0.2424, pruned_loss=0.04421, over 12225.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2654, pruned_loss=0.05145, over 3298722.78 frames. ], batch size: 246, lr: 7.00e-03, grad_scale: 4.0 2023-04-29 04:19:04,204 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=92311.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:19:29,762 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.546e+02 2.335e+02 2.688e+02 3.318e+02 6.424e+02, threshold=5.375e+02, percent-clipped=3.0 2023-04-29 04:19:39,509 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=92337.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 04:19:58,791 INFO [train.py:904] (4/8) Epoch 10, batch 1000, loss[loss=0.2254, simple_loss=0.2897, pruned_loss=0.0806, over 16550.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2646, pruned_loss=0.05132, over 3310618.40 frames. ], batch size: 68, lr: 7.00e-03, grad_scale: 4.0 2023-04-29 04:21:09,132 INFO [train.py:904] (4/8) Epoch 10, batch 1050, loss[loss=0.1807, simple_loss=0.2572, pruned_loss=0.05209, over 12198.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.2639, pruned_loss=0.05068, over 3316735.99 frames. ], batch size: 248, lr: 7.00e-03, grad_scale: 4.0 2023-04-29 04:21:24,633 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5230, 2.5897, 2.0896, 2.2774, 2.8894, 2.6263, 3.3719, 3.1352], device='cuda:4'), covar=tensor([0.0105, 0.0266, 0.0375, 0.0359, 0.0217, 0.0280, 0.0184, 0.0201], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0198, 0.0194, 0.0193, 0.0197, 0.0198, 0.0201, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:21:48,359 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.641e+02 2.612e+02 3.041e+02 3.664e+02 7.677e+02, threshold=6.083e+02, percent-clipped=4.0 2023-04-29 04:21:54,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0711, 5.5998, 5.6889, 5.5447, 5.6292, 6.0637, 5.7111, 5.4910], device='cuda:4'), covar=tensor([0.0811, 0.1681, 0.2287, 0.1810, 0.2447, 0.0996, 0.1334, 0.2110], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0493, 0.0518, 0.0422, 0.0565, 0.0539, 0.0419, 0.0569], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:22:17,257 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-29 04:22:18,533 INFO [train.py:904] (4/8) Epoch 10, batch 1100, loss[loss=0.1812, simple_loss=0.2595, pruned_loss=0.05142, over 16949.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2635, pruned_loss=0.05041, over 3325384.71 frames. ], batch size: 41, lr: 7.00e-03, grad_scale: 4.0 2023-04-29 04:23:28,218 INFO [train.py:904] (4/8) Epoch 10, batch 1150, loss[loss=0.168, simple_loss=0.2452, pruned_loss=0.04534, over 16857.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2627, pruned_loss=0.04999, over 3327229.36 frames. ], batch size: 90, lr: 6.99e-03, grad_scale: 4.0 2023-04-29 04:24:08,393 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.423e+02 2.186e+02 2.590e+02 3.516e+02 6.121e+02, threshold=5.179e+02, percent-clipped=1.0 2023-04-29 04:24:25,990 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-29 04:24:32,772 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=92547.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:24:38,818 INFO [train.py:904] (4/8) Epoch 10, batch 1200, loss[loss=0.1694, simple_loss=0.2451, pruned_loss=0.04684, over 16796.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.262, pruned_loss=0.05017, over 3326528.14 frames. ], batch size: 39, lr: 6.99e-03, grad_scale: 8.0 2023-04-29 04:25:39,054 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=92595.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:25:47,774 INFO [train.py:904] (4/8) Epoch 10, batch 1250, loss[loss=0.1574, simple_loss=0.2427, pruned_loss=0.0361, over 16849.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2617, pruned_loss=0.05053, over 3329464.27 frames. ], batch size: 42, lr: 6.99e-03, grad_scale: 8.0 2023-04-29 04:26:01,835 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=92611.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:26:20,494 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92625.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:26:27,271 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.538e+02 2.554e+02 2.910e+02 3.549e+02 5.849e+02, threshold=5.821e+02, percent-clipped=4.0 2023-04-29 04:26:37,459 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=92637.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 04:26:58,797 INFO [train.py:904] (4/8) Epoch 10, batch 1300, loss[loss=0.1951, simple_loss=0.2638, pruned_loss=0.06322, over 16702.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2618, pruned_loss=0.05083, over 3321541.47 frames. ], batch size: 134, lr: 6.99e-03, grad_scale: 8.0 2023-04-29 04:27:07,401 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=92659.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:27:44,063 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=92685.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 04:27:45,947 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92686.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:27:57,791 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5833, 3.6286, 2.0511, 3.8683, 2.7339, 3.8331, 2.1393, 2.8184], device='cuda:4'), covar=tensor([0.0187, 0.0293, 0.1349, 0.0178, 0.0645, 0.0539, 0.1227, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0165, 0.0188, 0.0125, 0.0166, 0.0204, 0.0193, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 04:28:07,580 INFO [train.py:904] (4/8) Epoch 10, batch 1350, loss[loss=0.1609, simple_loss=0.2434, pruned_loss=0.03918, over 16669.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2629, pruned_loss=0.05083, over 3328366.66 frames. ], batch size: 37, lr: 6.99e-03, grad_scale: 8.0 2023-04-29 04:28:22,790 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0125, 5.3979, 5.5761, 5.3639, 5.3946, 5.9326, 5.5853, 5.2342], device='cuda:4'), covar=tensor([0.0861, 0.1796, 0.1621, 0.1785, 0.2539, 0.0935, 0.1166, 0.2023], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0497, 0.0523, 0.0430, 0.0570, 0.0548, 0.0423, 0.0573], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:28:34,049 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92720.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:28:47,042 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.588e+02 2.580e+02 3.033e+02 3.471e+02 5.594e+02, threshold=6.066e+02, percent-clipped=0.0 2023-04-29 04:28:59,123 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7535, 3.8750, 4.2161, 2.9035, 3.7091, 4.0806, 3.8395, 2.3621], device='cuda:4'), covar=tensor([0.0317, 0.0109, 0.0027, 0.0263, 0.0081, 0.0068, 0.0055, 0.0335], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0070, 0.0068, 0.0126, 0.0077, 0.0085, 0.0076, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:29:04,534 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6736, 4.1339, 4.3823, 3.0270, 3.7011, 4.1628, 3.9017, 2.2878], device='cuda:4'), covar=tensor([0.0370, 0.0050, 0.0024, 0.0257, 0.0077, 0.0063, 0.0058, 0.0387], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0070, 0.0068, 0.0126, 0.0077, 0.0085, 0.0076, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:29:18,974 INFO [train.py:904] (4/8) Epoch 10, batch 1400, loss[loss=0.1546, simple_loss=0.2369, pruned_loss=0.03615, over 16455.00 frames. ], tot_loss[loss=0.182, simple_loss=0.2626, pruned_loss=0.05074, over 3323536.04 frames. ], batch size: 75, lr: 6.98e-03, grad_scale: 8.0 2023-04-29 04:30:00,624 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92781.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:30:28,850 INFO [train.py:904] (4/8) Epoch 10, batch 1450, loss[loss=0.1994, simple_loss=0.2691, pruned_loss=0.06481, over 16544.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2618, pruned_loss=0.0509, over 3324958.88 frames. ], batch size: 75, lr: 6.98e-03, grad_scale: 8.0 2023-04-29 04:30:39,508 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2765, 3.8076, 3.6188, 1.9395, 3.0291, 2.3628, 3.6785, 3.6920], device='cuda:4'), covar=tensor([0.0306, 0.0666, 0.0520, 0.1669, 0.0764, 0.0896, 0.0621, 0.0969], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0141, 0.0156, 0.0141, 0.0134, 0.0123, 0.0134, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 04:31:07,980 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.451e+02 2.441e+02 3.023e+02 3.567e+02 6.383e+02, threshold=6.046e+02, percent-clipped=1.0 2023-04-29 04:31:09,656 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92831.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:31:36,462 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5070, 3.2590, 2.7340, 2.1574, 2.2952, 2.1668, 3.3118, 3.0868], device='cuda:4'), covar=tensor([0.2299, 0.0668, 0.1386, 0.2241, 0.2268, 0.1771, 0.0477, 0.1199], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0258, 0.0282, 0.0274, 0.0276, 0.0219, 0.0266, 0.0294], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:31:38,220 INFO [train.py:904] (4/8) Epoch 10, batch 1500, loss[loss=0.1869, simple_loss=0.2821, pruned_loss=0.04588, over 17149.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.2627, pruned_loss=0.05195, over 3313718.68 frames. ], batch size: 47, lr: 6.98e-03, grad_scale: 8.0 2023-04-29 04:32:30,070 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7634, 2.1228, 2.2731, 4.4241, 2.0725, 2.7610, 2.2982, 2.4492], device='cuda:4'), covar=tensor([0.0827, 0.3453, 0.2228, 0.0372, 0.3761, 0.2169, 0.2919, 0.3288], device='cuda:4'), in_proj_covar=tensor([0.0359, 0.0385, 0.0326, 0.0324, 0.0409, 0.0436, 0.0345, 0.0452], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:32:34,081 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92892.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:32:45,536 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92899.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:32:48,539 INFO [train.py:904] (4/8) Epoch 10, batch 1550, loss[loss=0.1561, simple_loss=0.2374, pruned_loss=0.03737, over 17011.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2639, pruned_loss=0.05236, over 3324081.06 frames. ], batch size: 41, lr: 6.98e-03, grad_scale: 8.0 2023-04-29 04:33:21,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3134, 5.6605, 5.3940, 5.4620, 5.0246, 4.9776, 5.1430, 5.7826], device='cuda:4'), covar=tensor([0.1011, 0.0822, 0.0997, 0.0621, 0.0829, 0.0672, 0.0898, 0.0822], device='cuda:4'), in_proj_covar=tensor([0.0527, 0.0673, 0.0556, 0.0462, 0.0423, 0.0430, 0.0560, 0.0506], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:33:26,143 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.809e+02 2.944e+02 3.490e+02 4.097e+02 8.318e+02, threshold=6.980e+02, percent-clipped=5.0 2023-04-29 04:33:28,351 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92931.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:33:54,935 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-04-29 04:33:56,444 INFO [train.py:904] (4/8) Epoch 10, batch 1600, loss[loss=0.1818, simple_loss=0.2705, pruned_loss=0.04651, over 15962.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2666, pruned_loss=0.05415, over 3320590.89 frames. ], batch size: 35, lr: 6.98e-03, grad_scale: 8.0 2023-04-29 04:34:07,743 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92960.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:34:29,060 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92975.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:34:37,308 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=92981.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:34:52,610 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92992.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:35:06,236 INFO [train.py:904] (4/8) Epoch 10, batch 1650, loss[loss=0.1606, simple_loss=0.2451, pruned_loss=0.03807, over 16879.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2662, pruned_loss=0.05338, over 3331689.81 frames. ], batch size: 42, lr: 6.97e-03, grad_scale: 8.0 2023-04-29 04:35:18,232 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93010.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:35:45,199 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.722e+02 2.314e+02 2.824e+02 3.373e+02 5.658e+02, threshold=5.648e+02, percent-clipped=0.0 2023-04-29 04:35:53,102 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93036.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:35:53,150 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93036.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:36:15,631 INFO [train.py:904] (4/8) Epoch 10, batch 1700, loss[loss=0.1994, simple_loss=0.2888, pruned_loss=0.05494, over 17038.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.2687, pruned_loss=0.05436, over 3315771.14 frames. ], batch size: 55, lr: 6.97e-03, grad_scale: 8.0 2023-04-29 04:36:29,526 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-29 04:36:42,583 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93071.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:36:48,264 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93076.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:36:53,642 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2970, 3.2225, 3.4660, 2.4478, 3.2024, 3.5229, 3.3181, 1.9733], device='cuda:4'), covar=tensor([0.0331, 0.0086, 0.0033, 0.0257, 0.0069, 0.0050, 0.0055, 0.0352], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0070, 0.0067, 0.0124, 0.0076, 0.0083, 0.0075, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:37:17,801 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93097.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 04:37:23,690 INFO [train.py:904] (4/8) Epoch 10, batch 1750, loss[loss=0.1729, simple_loss=0.2631, pruned_loss=0.04136, over 16790.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2701, pruned_loss=0.05406, over 3319327.09 frames. ], batch size: 39, lr: 6.97e-03, grad_scale: 8.0 2023-04-29 04:37:42,873 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93115.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:37:51,247 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9637, 4.9541, 4.7299, 4.2585, 4.8412, 1.8894, 4.5946, 4.6628], device='cuda:4'), covar=tensor([0.0067, 0.0058, 0.0129, 0.0284, 0.0067, 0.2199, 0.0105, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0115, 0.0165, 0.0155, 0.0134, 0.0178, 0.0152, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:38:01,414 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.528e+02 2.469e+02 2.889e+02 3.646e+02 7.131e+02, threshold=5.778e+02, percent-clipped=4.0 2023-04-29 04:38:23,183 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0054, 4.0423, 2.9935, 2.3451, 2.8452, 2.4632, 4.2745, 3.7693], device='cuda:4'), covar=tensor([0.2209, 0.0687, 0.1593, 0.2216, 0.2104, 0.1689, 0.0445, 0.0957], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0256, 0.0280, 0.0272, 0.0274, 0.0218, 0.0264, 0.0294], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:38:32,562 INFO [train.py:904] (4/8) Epoch 10, batch 1800, loss[loss=0.1988, simple_loss=0.2793, pruned_loss=0.05914, over 16509.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.271, pruned_loss=0.05398, over 3328479.58 frames. ], batch size: 68, lr: 6.97e-03, grad_scale: 8.0 2023-04-29 04:39:06,710 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93176.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:39:20,436 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93187.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:39:42,563 INFO [train.py:904] (4/8) Epoch 10, batch 1850, loss[loss=0.1836, simple_loss=0.2777, pruned_loss=0.04474, over 16633.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2726, pruned_loss=0.05411, over 3327431.86 frames. ], batch size: 62, lr: 6.97e-03, grad_scale: 8.0 2023-04-29 04:39:47,692 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-29 04:39:48,168 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3463, 5.7332, 5.5269, 5.5276, 5.1292, 5.0275, 5.1479, 5.9037], device='cuda:4'), covar=tensor([0.1064, 0.0875, 0.0929, 0.0649, 0.0837, 0.0632, 0.1021, 0.0753], device='cuda:4'), in_proj_covar=tensor([0.0527, 0.0672, 0.0553, 0.0463, 0.0422, 0.0429, 0.0562, 0.0505], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:40:21,084 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.739e+02 2.500e+02 2.909e+02 3.446e+02 8.007e+02, threshold=5.817e+02, percent-clipped=2.0 2023-04-29 04:40:52,059 INFO [train.py:904] (4/8) Epoch 10, batch 1900, loss[loss=0.211, simple_loss=0.2772, pruned_loss=0.07238, over 16855.00 frames. ], tot_loss[loss=0.189, simple_loss=0.2711, pruned_loss=0.05344, over 3315782.76 frames. ], batch size: 96, lr: 6.97e-03, grad_scale: 8.0 2023-04-29 04:40:56,692 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93255.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:41:33,395 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93281.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:41:41,654 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93287.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:42:02,723 INFO [train.py:904] (4/8) Epoch 10, batch 1950, loss[loss=0.2132, simple_loss=0.2873, pruned_loss=0.06954, over 16433.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.272, pruned_loss=0.05329, over 3316357.09 frames. ], batch size: 146, lr: 6.96e-03, grad_scale: 8.0 2023-04-29 04:42:40,591 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93329.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:42:41,416 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.729e+02 2.458e+02 3.015e+02 3.561e+02 8.313e+02, threshold=6.031e+02, percent-clipped=4.0 2023-04-29 04:42:43,567 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93331.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:43:02,735 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7796, 3.6749, 2.8904, 2.2887, 2.5693, 2.2538, 3.7188, 3.4069], device='cuda:4'), covar=tensor([0.2141, 0.0557, 0.1423, 0.2136, 0.2167, 0.1652, 0.0507, 0.1088], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0258, 0.0281, 0.0273, 0.0277, 0.0219, 0.0265, 0.0295], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:43:12,556 INFO [train.py:904] (4/8) Epoch 10, batch 2000, loss[loss=0.2162, simple_loss=0.2973, pruned_loss=0.06757, over 12574.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.2729, pruned_loss=0.05359, over 3307590.77 frames. ], batch size: 248, lr: 6.96e-03, grad_scale: 8.0 2023-04-29 04:43:19,780 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9451, 4.2306, 4.3296, 3.3553, 3.6591, 4.2266, 3.8849, 2.7464], device='cuda:4'), covar=tensor([0.0304, 0.0049, 0.0028, 0.0222, 0.0083, 0.0064, 0.0065, 0.0315], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0070, 0.0068, 0.0126, 0.0077, 0.0085, 0.0076, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:43:31,505 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93366.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:43:44,496 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93376.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:44:08,368 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93392.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 04:44:21,491 INFO [train.py:904] (4/8) Epoch 10, batch 2050, loss[loss=0.2071, simple_loss=0.2802, pruned_loss=0.06702, over 16489.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2725, pruned_loss=0.05379, over 3314064.18 frames. ], batch size: 75, lr: 6.96e-03, grad_scale: 4.0 2023-04-29 04:44:28,956 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3624, 2.0518, 2.3008, 4.0132, 2.1107, 2.5356, 2.1440, 2.2800], device='cuda:4'), covar=tensor([0.0924, 0.2947, 0.1870, 0.0387, 0.3068, 0.1899, 0.2753, 0.2449], device='cuda:4'), in_proj_covar=tensor([0.0360, 0.0384, 0.0324, 0.0324, 0.0407, 0.0436, 0.0345, 0.0452], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:44:51,420 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93424.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:45:00,992 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.855e+02 2.531e+02 2.855e+02 3.295e+02 5.942e+02, threshold=5.709e+02, percent-clipped=0.0 2023-04-29 04:45:26,127 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4671, 4.5097, 4.6743, 4.5393, 4.5293, 5.1667, 4.7589, 4.4060], device='cuda:4'), covar=tensor([0.1425, 0.2034, 0.2030, 0.2069, 0.2841, 0.1096, 0.1485, 0.2632], device='cuda:4'), in_proj_covar=tensor([0.0348, 0.0501, 0.0524, 0.0433, 0.0578, 0.0550, 0.0422, 0.0580], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:45:29,945 INFO [train.py:904] (4/8) Epoch 10, batch 2100, loss[loss=0.19, simple_loss=0.2848, pruned_loss=0.04761, over 16614.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2731, pruned_loss=0.05437, over 3316392.19 frames. ], batch size: 62, lr: 6.96e-03, grad_scale: 4.0 2023-04-29 04:45:56,795 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93471.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:46:07,614 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-29 04:46:18,623 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93487.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:46:40,127 INFO [train.py:904] (4/8) Epoch 10, batch 2150, loss[loss=0.1768, simple_loss=0.2619, pruned_loss=0.0458, over 16846.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2738, pruned_loss=0.05518, over 3322667.48 frames. ], batch size: 42, lr: 6.96e-03, grad_scale: 4.0 2023-04-29 04:47:18,311 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.855e+02 2.473e+02 3.061e+02 3.473e+02 5.653e+02, threshold=6.122e+02, percent-clipped=0.0 2023-04-29 04:47:24,634 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93535.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:47:46,702 INFO [train.py:904] (4/8) Epoch 10, batch 2200, loss[loss=0.2, simple_loss=0.2782, pruned_loss=0.06089, over 16857.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2746, pruned_loss=0.05607, over 3299747.51 frames. ], batch size: 102, lr: 6.95e-03, grad_scale: 4.0 2023-04-29 04:47:52,004 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93555.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:48:10,862 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8124, 2.7894, 2.5100, 4.1139, 3.4445, 4.1571, 1.5608, 2.8473], device='cuda:4'), covar=tensor([0.1289, 0.0573, 0.1002, 0.0139, 0.0173, 0.0316, 0.1360, 0.0734], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0156, 0.0178, 0.0137, 0.0198, 0.0212, 0.0176, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 04:48:35,506 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93587.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:48:54,025 INFO [train.py:904] (4/8) Epoch 10, batch 2250, loss[loss=0.1706, simple_loss=0.2549, pruned_loss=0.04312, over 16807.00 frames. ], tot_loss[loss=0.1927, simple_loss=0.2741, pruned_loss=0.05565, over 3305045.72 frames. ], batch size: 39, lr: 6.95e-03, grad_scale: 4.0 2023-04-29 04:48:55,410 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93603.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:49:33,831 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.688e+02 2.560e+02 3.200e+02 3.958e+02 7.230e+02, threshold=6.400e+02, percent-clipped=2.0 2023-04-29 04:49:34,890 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93631.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:49:35,052 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5991, 2.2820, 2.2901, 4.3959, 2.1009, 2.7155, 2.3146, 2.4433], device='cuda:4'), covar=tensor([0.0860, 0.2972, 0.2068, 0.0356, 0.3525, 0.2067, 0.2659, 0.2959], device='cuda:4'), in_proj_covar=tensor([0.0363, 0.0388, 0.0326, 0.0326, 0.0410, 0.0442, 0.0348, 0.0456], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:49:39,246 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93635.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:50:04,001 INFO [train.py:904] (4/8) Epoch 10, batch 2300, loss[loss=0.1888, simple_loss=0.2671, pruned_loss=0.05527, over 16456.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.274, pruned_loss=0.05546, over 3307027.29 frames. ], batch size: 146, lr: 6.95e-03, grad_scale: 4.0 2023-04-29 04:50:05,727 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9158, 3.2632, 2.9633, 1.9673, 2.6872, 2.1881, 3.5108, 3.4517], device='cuda:4'), covar=tensor([0.0256, 0.0758, 0.0679, 0.1666, 0.0824, 0.0953, 0.0498, 0.0764], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0144, 0.0158, 0.0143, 0.0136, 0.0126, 0.0137, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 04:50:22,476 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93666.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:50:23,908 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7973, 2.6201, 2.1333, 2.3243, 3.0252, 2.8649, 3.6942, 3.2994], device='cuda:4'), covar=tensor([0.0076, 0.0345, 0.0488, 0.0403, 0.0221, 0.0289, 0.0225, 0.0182], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0201, 0.0197, 0.0199, 0.0200, 0.0201, 0.0210, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:50:39,710 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93679.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:50:57,949 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93692.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:51:09,768 INFO [train.py:904] (4/8) Epoch 10, batch 2350, loss[loss=0.1734, simple_loss=0.2532, pruned_loss=0.04678, over 16837.00 frames. ], tot_loss[loss=0.1917, simple_loss=0.2731, pruned_loss=0.05511, over 3318930.38 frames. ], batch size: 39, lr: 6.95e-03, grad_scale: 4.0 2023-04-29 04:51:27,701 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93714.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:51:49,907 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.552e+02 2.363e+02 2.805e+02 3.305e+02 9.718e+02, threshold=5.610e+02, percent-clipped=1.0 2023-04-29 04:52:03,151 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93740.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:52:17,489 INFO [train.py:904] (4/8) Epoch 10, batch 2400, loss[loss=0.2036, simple_loss=0.288, pruned_loss=0.05954, over 17098.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.274, pruned_loss=0.05521, over 3323971.91 frames. ], batch size: 53, lr: 6.95e-03, grad_scale: 8.0 2023-04-29 04:52:25,292 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8901, 4.1510, 3.1792, 2.2422, 2.7714, 2.4442, 4.5435, 3.8052], device='cuda:4'), covar=tensor([0.2349, 0.0644, 0.1425, 0.2277, 0.2669, 0.1715, 0.0347, 0.0936], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0259, 0.0283, 0.0274, 0.0281, 0.0220, 0.0266, 0.0297], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:52:41,453 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93770.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:52:43,110 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93771.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:53:01,000 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9803, 2.4925, 2.4954, 4.8860, 2.2974, 3.0057, 2.5502, 2.7483], device='cuda:4'), covar=tensor([0.0805, 0.3117, 0.2043, 0.0282, 0.3519, 0.2058, 0.2600, 0.2900], device='cuda:4'), in_proj_covar=tensor([0.0359, 0.0384, 0.0323, 0.0323, 0.0406, 0.0438, 0.0344, 0.0452], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 04:53:26,813 INFO [train.py:904] (4/8) Epoch 10, batch 2450, loss[loss=0.2122, simple_loss=0.2826, pruned_loss=0.07095, over 16767.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.2751, pruned_loss=0.05488, over 3331253.44 frames. ], batch size: 134, lr: 6.94e-03, grad_scale: 8.0 2023-04-29 04:53:49,967 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=93819.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:54:05,084 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.668e+02 2.505e+02 3.029e+02 3.825e+02 7.236e+02, threshold=6.058e+02, percent-clipped=4.0 2023-04-29 04:54:05,518 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93831.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:54:34,536 INFO [train.py:904] (4/8) Epoch 10, batch 2500, loss[loss=0.1874, simple_loss=0.2637, pruned_loss=0.05561, over 16796.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.2741, pruned_loss=0.05441, over 3320385.16 frames. ], batch size: 102, lr: 6.94e-03, grad_scale: 8.0 2023-04-29 04:55:25,744 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7067, 2.6567, 2.3404, 3.9941, 3.3109, 3.9690, 1.4373, 2.7991], device='cuda:4'), covar=tensor([0.1231, 0.0563, 0.1041, 0.0144, 0.0121, 0.0345, 0.1303, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0156, 0.0179, 0.0139, 0.0199, 0.0212, 0.0176, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 04:55:42,142 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8764, 5.2878, 5.3372, 5.2041, 5.1428, 5.8056, 5.2794, 5.0005], device='cuda:4'), covar=tensor([0.0985, 0.1568, 0.1592, 0.1564, 0.2638, 0.0885, 0.1232, 0.2108], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0497, 0.0520, 0.0426, 0.0573, 0.0549, 0.0422, 0.0576], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:55:43,655 INFO [train.py:904] (4/8) Epoch 10, batch 2550, loss[loss=0.2209, simple_loss=0.3003, pruned_loss=0.07079, over 16536.00 frames. ], tot_loss[loss=0.1928, simple_loss=0.2751, pruned_loss=0.05526, over 3305722.72 frames. ], batch size: 68, lr: 6.94e-03, grad_scale: 8.0 2023-04-29 04:56:23,962 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.543e+02 2.408e+02 2.927e+02 3.582e+02 7.180e+02, threshold=5.854e+02, percent-clipped=2.0 2023-04-29 04:56:48,066 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0428, 4.8884, 4.8465, 4.5910, 4.4734, 4.9225, 4.8358, 4.5272], device='cuda:4'), covar=tensor([0.0532, 0.0497, 0.0251, 0.0263, 0.1020, 0.0373, 0.0298, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0251, 0.0310, 0.0295, 0.0272, 0.0320, 0.0308, 0.0204, 0.0345], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 04:56:52,858 INFO [train.py:904] (4/8) Epoch 10, batch 2600, loss[loss=0.1684, simple_loss=0.2584, pruned_loss=0.03916, over 17265.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2743, pruned_loss=0.05463, over 3310635.15 frames. ], batch size: 45, lr: 6.94e-03, grad_scale: 8.0 2023-04-29 04:57:09,882 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-29 04:57:32,289 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93981.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:57:42,498 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0326, 4.0122, 4.3448, 1.9994, 4.6143, 4.6870, 3.2712, 3.4245], device='cuda:4'), covar=tensor([0.0646, 0.0205, 0.0208, 0.1161, 0.0062, 0.0123, 0.0386, 0.0393], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0098, 0.0087, 0.0138, 0.0070, 0.0101, 0.0121, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 04:57:56,755 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1612, 4.1572, 4.6313, 4.5948, 4.6012, 4.2425, 4.3061, 4.1488], device='cuda:4'), covar=tensor([0.0347, 0.0532, 0.0366, 0.0479, 0.0464, 0.0392, 0.0858, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0344, 0.0347, 0.0330, 0.0388, 0.0362, 0.0469, 0.0291], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 04:58:03,852 INFO [train.py:904] (4/8) Epoch 10, batch 2650, loss[loss=0.225, simple_loss=0.2909, pruned_loss=0.07957, over 16880.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2753, pruned_loss=0.05444, over 3323753.26 frames. ], batch size: 116, lr: 6.94e-03, grad_scale: 8.0 2023-04-29 04:58:42,888 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.48 vs. limit=5.0 2023-04-29 04:58:43,799 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.633e+02 2.229e+02 2.744e+02 3.271e+02 8.724e+02, threshold=5.488e+02, percent-clipped=1.0 2023-04-29 04:59:00,257 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=94042.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 04:59:13,552 INFO [train.py:904] (4/8) Epoch 10, batch 2700, loss[loss=0.2176, simple_loss=0.2978, pruned_loss=0.06869, over 16263.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2755, pruned_loss=0.05399, over 3332814.82 frames. ], batch size: 165, lr: 6.94e-03, grad_scale: 8.0 2023-04-29 05:00:23,323 INFO [train.py:904] (4/8) Epoch 10, batch 2750, loss[loss=0.2075, simple_loss=0.2774, pruned_loss=0.06884, over 16477.00 frames. ], tot_loss[loss=0.1906, simple_loss=0.275, pruned_loss=0.05314, over 3339801.73 frames. ], batch size: 146, lr: 6.93e-03, grad_scale: 8.0 2023-04-29 05:00:55,428 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=94126.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:01:01,088 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.410e+02 2.958e+02 3.419e+02 5.641e+02, threshold=5.917e+02, percent-clipped=1.0 2023-04-29 05:01:29,685 INFO [train.py:904] (4/8) Epoch 10, batch 2800, loss[loss=0.1709, simple_loss=0.2523, pruned_loss=0.04471, over 16982.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2742, pruned_loss=0.0528, over 3330896.98 frames. ], batch size: 41, lr: 6.93e-03, grad_scale: 8.0 2023-04-29 05:02:39,399 INFO [train.py:904] (4/8) Epoch 10, batch 2850, loss[loss=0.2272, simple_loss=0.2996, pruned_loss=0.07734, over 15383.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2735, pruned_loss=0.05282, over 3328779.12 frames. ], batch size: 191, lr: 6.93e-03, grad_scale: 8.0 2023-04-29 05:03:09,004 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=94223.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 05:03:18,385 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-29 05:03:20,127 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.572e+02 2.429e+02 2.851e+02 3.333e+02 6.061e+02, threshold=5.703e+02, percent-clipped=1.0 2023-04-29 05:03:49,048 INFO [train.py:904] (4/8) Epoch 10, batch 2900, loss[loss=0.1904, simple_loss=0.2918, pruned_loss=0.04447, over 17112.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2725, pruned_loss=0.05272, over 3332507.42 frames. ], batch size: 49, lr: 6.93e-03, grad_scale: 8.0 2023-04-29 05:04:14,914 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0236, 3.9883, 4.4051, 4.4187, 4.4579, 4.1251, 4.1604, 4.0362], device='cuda:4'), covar=tensor([0.0385, 0.0590, 0.0445, 0.0432, 0.0436, 0.0385, 0.0839, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0346, 0.0354, 0.0358, 0.0340, 0.0395, 0.0371, 0.0483, 0.0297], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 05:04:33,897 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=94284.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 05:04:58,055 INFO [train.py:904] (4/8) Epoch 10, batch 2950, loss[loss=0.2159, simple_loss=0.2837, pruned_loss=0.07407, over 16884.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2728, pruned_loss=0.05434, over 3321918.84 frames. ], batch size: 116, lr: 6.93e-03, grad_scale: 8.0 2023-04-29 05:05:39,543 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.726e+02 2.713e+02 3.214e+02 3.938e+02 7.856e+02, threshold=6.427e+02, percent-clipped=3.0 2023-04-29 05:05:49,212 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=94337.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:06:03,092 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=94348.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:06:08,004 INFO [train.py:904] (4/8) Epoch 10, batch 3000, loss[loss=0.1716, simple_loss=0.2674, pruned_loss=0.03792, over 17137.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2722, pruned_loss=0.0541, over 3326942.80 frames. ], batch size: 49, lr: 6.92e-03, grad_scale: 8.0 2023-04-29 05:06:08,004 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 05:06:17,138 INFO [train.py:938] (4/8) Epoch 10, validation: loss=0.1426, simple_loss=0.2488, pruned_loss=0.01818, over 944034.00 frames. 2023-04-29 05:06:17,139 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 05:06:40,741 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5901, 5.9645, 5.7033, 5.7406, 5.3121, 5.1554, 5.4203, 6.0763], device='cuda:4'), covar=tensor([0.1042, 0.0857, 0.0913, 0.0680, 0.0781, 0.0628, 0.0943, 0.0831], device='cuda:4'), in_proj_covar=tensor([0.0528, 0.0674, 0.0554, 0.0458, 0.0420, 0.0428, 0.0558, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:07:26,695 INFO [train.py:904] (4/8) Epoch 10, batch 3050, loss[loss=0.1926, simple_loss=0.2822, pruned_loss=0.05153, over 17057.00 frames. ], tot_loss[loss=0.1894, simple_loss=0.2716, pruned_loss=0.05361, over 3326416.09 frames. ], batch size: 55, lr: 6.92e-03, grad_scale: 8.0 2023-04-29 05:07:30,853 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-29 05:07:36,828 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=94409.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 05:07:56,496 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4540, 4.3559, 4.5581, 4.4006, 4.4319, 5.0445, 4.6409, 4.3481], device='cuda:4'), covar=tensor([0.1597, 0.2166, 0.1747, 0.2378, 0.2931, 0.1175, 0.1306, 0.2745], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0497, 0.0520, 0.0432, 0.0574, 0.0548, 0.0419, 0.0581], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 05:07:57,737 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=94426.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:08:05,351 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.835e+02 2.667e+02 3.252e+02 4.083e+02 7.974e+02, threshold=6.505e+02, percent-clipped=3.0 2023-04-29 05:08:10,100 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6914, 3.2161, 2.7767, 5.1195, 4.4254, 4.6380, 1.6691, 3.3261], device='cuda:4'), covar=tensor([0.1330, 0.0594, 0.1101, 0.0177, 0.0265, 0.0342, 0.1391, 0.0721], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0157, 0.0179, 0.0139, 0.0200, 0.0214, 0.0176, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 05:08:33,247 INFO [train.py:904] (4/8) Epoch 10, batch 3100, loss[loss=0.1903, simple_loss=0.2773, pruned_loss=0.05161, over 16452.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2709, pruned_loss=0.05297, over 3336133.89 frames. ], batch size: 68, lr: 6.92e-03, grad_scale: 8.0 2023-04-29 05:09:04,482 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=94474.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:09:43,505 INFO [train.py:904] (4/8) Epoch 10, batch 3150, loss[loss=0.2247, simple_loss=0.291, pruned_loss=0.07918, over 16925.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2686, pruned_loss=0.05213, over 3343121.09 frames. ], batch size: 116, lr: 6.92e-03, grad_scale: 8.0 2023-04-29 05:09:48,920 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.91 vs. limit=5.0 2023-04-29 05:09:59,523 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-29 05:10:23,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.666e+02 2.494e+02 3.047e+02 3.554e+02 8.509e+02, threshold=6.093e+02, percent-clipped=1.0 2023-04-29 05:10:28,141 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5943, 6.0410, 5.7465, 5.8100, 5.3101, 5.1715, 5.4889, 6.1610], device='cuda:4'), covar=tensor([0.1000, 0.0764, 0.1047, 0.0653, 0.0805, 0.0655, 0.0969, 0.0838], device='cuda:4'), in_proj_covar=tensor([0.0540, 0.0686, 0.0567, 0.0469, 0.0429, 0.0435, 0.0569, 0.0519], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:10:36,147 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.06 vs. limit=5.0 2023-04-29 05:10:52,169 INFO [train.py:904] (4/8) Epoch 10, batch 3200, loss[loss=0.1578, simple_loss=0.2485, pruned_loss=0.03355, over 17234.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2679, pruned_loss=0.05168, over 3343714.68 frames. ], batch size: 45, lr: 6.92e-03, grad_scale: 8.0 2023-04-29 05:10:56,092 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9900, 5.5317, 5.6874, 5.4691, 5.5542, 6.0865, 5.6794, 5.4295], device='cuda:4'), covar=tensor([0.0863, 0.1718, 0.1751, 0.1883, 0.2608, 0.0862, 0.1097, 0.2080], device='cuda:4'), in_proj_covar=tensor([0.0346, 0.0498, 0.0522, 0.0431, 0.0573, 0.0548, 0.0423, 0.0579], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 05:11:32,215 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=94579.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 05:12:04,562 INFO [train.py:904] (4/8) Epoch 10, batch 3250, loss[loss=0.1584, simple_loss=0.2479, pruned_loss=0.03441, over 17257.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2679, pruned_loss=0.05234, over 3335616.86 frames. ], batch size: 43, lr: 6.92e-03, grad_scale: 8.0 2023-04-29 05:12:44,900 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.625e+02 2.346e+02 2.940e+02 3.507e+02 9.203e+02, threshold=5.881e+02, percent-clipped=1.0 2023-04-29 05:12:53,011 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=94637.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:13:14,625 INFO [train.py:904] (4/8) Epoch 10, batch 3300, loss[loss=0.1851, simple_loss=0.2804, pruned_loss=0.04486, over 17061.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2698, pruned_loss=0.05326, over 3327262.36 frames. ], batch size: 50, lr: 6.91e-03, grad_scale: 8.0 2023-04-29 05:14:02,317 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=94685.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:14:20,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6201, 3.7632, 2.1705, 4.0158, 2.6730, 4.0158, 2.1057, 2.9092], device='cuda:4'), covar=tensor([0.0203, 0.0308, 0.1328, 0.0202, 0.0749, 0.0515, 0.1363, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0167, 0.0187, 0.0132, 0.0168, 0.0210, 0.0196, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 05:14:24,568 INFO [train.py:904] (4/8) Epoch 10, batch 3350, loss[loss=0.2012, simple_loss=0.2785, pruned_loss=0.06193, over 16488.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2718, pruned_loss=0.05434, over 3326729.03 frames. ], batch size: 146, lr: 6.91e-03, grad_scale: 8.0 2023-04-29 05:14:28,542 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=94704.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 05:14:31,043 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7745, 4.9006, 5.3654, 5.3292, 5.3781, 5.0444, 4.8884, 4.7038], device='cuda:4'), covar=tensor([0.0506, 0.0609, 0.0511, 0.0614, 0.0563, 0.0489, 0.1250, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0348, 0.0354, 0.0361, 0.0340, 0.0400, 0.0375, 0.0484, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 05:14:51,044 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-29 05:14:54,055 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-29 05:15:05,006 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.683e+02 2.488e+02 2.932e+02 3.912e+02 8.438e+02, threshold=5.863e+02, percent-clipped=4.0 2023-04-29 05:15:08,202 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5088, 4.0149, 4.1821, 2.1051, 3.3046, 2.4199, 4.1564, 4.0321], device='cuda:4'), covar=tensor([0.0219, 0.0657, 0.0406, 0.1723, 0.0711, 0.0966, 0.0459, 0.0773], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0146, 0.0158, 0.0143, 0.0136, 0.0125, 0.0138, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 05:15:35,790 INFO [train.py:904] (4/8) Epoch 10, batch 3400, loss[loss=0.1638, simple_loss=0.2517, pruned_loss=0.03793, over 17106.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2712, pruned_loss=0.05327, over 3325819.39 frames. ], batch size: 47, lr: 6.91e-03, grad_scale: 8.0 2023-04-29 05:16:44,906 INFO [train.py:904] (4/8) Epoch 10, batch 3450, loss[loss=0.1743, simple_loss=0.2565, pruned_loss=0.04604, over 17222.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2706, pruned_loss=0.05332, over 3324840.99 frames. ], batch size: 44, lr: 6.91e-03, grad_scale: 8.0 2023-04-29 05:17:22,610 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-29 05:17:26,306 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.690e+02 2.564e+02 2.957e+02 3.532e+02 6.981e+02, threshold=5.915e+02, percent-clipped=3.0 2023-04-29 05:17:56,544 INFO [train.py:904] (4/8) Epoch 10, batch 3500, loss[loss=0.1861, simple_loss=0.2676, pruned_loss=0.05225, over 16744.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2689, pruned_loss=0.05303, over 3323542.90 frames. ], batch size: 89, lr: 6.91e-03, grad_scale: 8.0 2023-04-29 05:18:35,686 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=94879.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 05:19:06,951 INFO [train.py:904] (4/8) Epoch 10, batch 3550, loss[loss=0.1623, simple_loss=0.2523, pruned_loss=0.03615, over 17249.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2684, pruned_loss=0.05227, over 3328961.41 frames. ], batch size: 45, lr: 6.90e-03, grad_scale: 8.0 2023-04-29 05:19:42,074 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=94927.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 05:19:47,556 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.652e+02 2.230e+02 2.675e+02 3.251e+02 5.912e+02, threshold=5.350e+02, percent-clipped=0.0 2023-04-29 05:20:17,531 INFO [train.py:904] (4/8) Epoch 10, batch 3600, loss[loss=0.1934, simple_loss=0.2677, pruned_loss=0.05953, over 16353.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2675, pruned_loss=0.05203, over 3325201.41 frames. ], batch size: 165, lr: 6.90e-03, grad_scale: 8.0 2023-04-29 05:21:28,700 INFO [train.py:904] (4/8) Epoch 10, batch 3650, loss[loss=0.2061, simple_loss=0.2787, pruned_loss=0.06673, over 11260.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2669, pruned_loss=0.05265, over 3315267.62 frames. ], batch size: 246, lr: 6.90e-03, grad_scale: 8.0 2023-04-29 05:21:32,970 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95004.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:22:01,908 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6668, 4.8069, 4.8411, 4.8845, 4.7699, 5.3544, 4.9914, 4.7335], device='cuda:4'), covar=tensor([0.1236, 0.1618, 0.1707, 0.1642, 0.2564, 0.0939, 0.1131, 0.2066], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0495, 0.0521, 0.0430, 0.0568, 0.0547, 0.0420, 0.0581], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 05:22:10,213 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.931e+02 2.260e+02 2.711e+02 3.366e+02 9.321e+02, threshold=5.423e+02, percent-clipped=5.0 2023-04-29 05:22:43,033 INFO [train.py:904] (4/8) Epoch 10, batch 3700, loss[loss=0.1975, simple_loss=0.2671, pruned_loss=0.06393, over 16851.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2661, pruned_loss=0.05427, over 3303309.47 frames. ], batch size: 109, lr: 6.90e-03, grad_scale: 8.0 2023-04-29 05:22:43,301 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=95052.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:23:50,024 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4272, 3.4082, 3.6343, 1.9781, 3.7682, 3.7826, 2.9436, 2.8613], device='cuda:4'), covar=tensor([0.0686, 0.0185, 0.0112, 0.0977, 0.0057, 0.0108, 0.0347, 0.0396], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0100, 0.0087, 0.0138, 0.0070, 0.0102, 0.0121, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 05:23:56,092 INFO [train.py:904] (4/8) Epoch 10, batch 3750, loss[loss=0.1966, simple_loss=0.272, pruned_loss=0.06065, over 15475.00 frames. ], tot_loss[loss=0.189, simple_loss=0.2667, pruned_loss=0.05567, over 3278175.68 frames. ], batch size: 190, lr: 6.90e-03, grad_scale: 8.0 2023-04-29 05:24:00,753 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95105.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 05:24:38,276 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.833e+02 2.465e+02 2.786e+02 3.363e+02 5.307e+02, threshold=5.572e+02, percent-clipped=0.0 2023-04-29 05:25:07,896 INFO [train.py:904] (4/8) Epoch 10, batch 3800, loss[loss=0.1867, simple_loss=0.2629, pruned_loss=0.05525, over 16887.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2689, pruned_loss=0.05759, over 3273155.29 frames. ], batch size: 96, lr: 6.90e-03, grad_scale: 8.0 2023-04-29 05:25:28,641 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95166.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 05:26:20,912 INFO [train.py:904] (4/8) Epoch 10, batch 3850, loss[loss=0.1789, simple_loss=0.257, pruned_loss=0.05044, over 16850.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.269, pruned_loss=0.05806, over 3269662.78 frames. ], batch size: 42, lr: 6.89e-03, grad_scale: 8.0 2023-04-29 05:27:00,982 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.608e+02 2.452e+02 2.920e+02 3.414e+02 5.310e+02, threshold=5.839e+02, percent-clipped=0.0 2023-04-29 05:27:31,950 INFO [train.py:904] (4/8) Epoch 10, batch 3900, loss[loss=0.1856, simple_loss=0.271, pruned_loss=0.05008, over 17017.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.268, pruned_loss=0.05821, over 3270801.54 frames. ], batch size: 50, lr: 6.89e-03, grad_scale: 8.0 2023-04-29 05:27:36,769 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1667, 1.9728, 2.1215, 3.7623, 1.9877, 2.3512, 2.0620, 2.1117], device='cuda:4'), covar=tensor([0.1042, 0.3280, 0.2162, 0.0475, 0.3454, 0.2189, 0.3167, 0.3015], device='cuda:4'), in_proj_covar=tensor([0.0363, 0.0390, 0.0325, 0.0325, 0.0409, 0.0447, 0.0353, 0.0462], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:27:43,912 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95260.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:27:56,104 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95268.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:28:17,599 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-29 05:28:45,286 INFO [train.py:904] (4/8) Epoch 10, batch 3950, loss[loss=0.1985, simple_loss=0.2683, pruned_loss=0.06439, over 16871.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2667, pruned_loss=0.05822, over 3269196.56 frames. ], batch size: 116, lr: 6.89e-03, grad_scale: 8.0 2023-04-29 05:29:02,508 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6469, 2.7035, 2.3290, 3.8081, 3.1668, 3.9209, 1.4243, 2.7222], device='cuda:4'), covar=tensor([0.1273, 0.0587, 0.1124, 0.0162, 0.0194, 0.0353, 0.1371, 0.0786], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0158, 0.0180, 0.0141, 0.0203, 0.0211, 0.0175, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 05:29:12,939 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95321.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:29:23,108 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95329.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:29:23,199 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1202, 3.3359, 3.2329, 1.9498, 2.7998, 2.3772, 3.4845, 3.5035], device='cuda:4'), covar=tensor([0.0223, 0.0671, 0.0568, 0.1643, 0.0754, 0.0890, 0.0475, 0.0725], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0145, 0.0155, 0.0142, 0.0135, 0.0124, 0.0136, 0.0156], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 05:29:25,657 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.566e+02 2.393e+02 2.876e+02 3.494e+02 7.568e+02, threshold=5.751e+02, percent-clipped=4.0 2023-04-29 05:29:38,153 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95339.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:29:56,313 INFO [train.py:904] (4/8) Epoch 10, batch 4000, loss[loss=0.1914, simple_loss=0.2675, pruned_loss=0.05771, over 16736.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2667, pruned_loss=0.05874, over 3278468.11 frames. ], batch size: 124, lr: 6.89e-03, grad_scale: 8.0 2023-04-29 05:30:31,438 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7090, 1.7076, 2.1499, 2.5280, 2.6211, 2.8737, 1.6458, 2.6939], device='cuda:4'), covar=tensor([0.0135, 0.0338, 0.0213, 0.0186, 0.0170, 0.0104, 0.0336, 0.0068], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0168, 0.0153, 0.0158, 0.0164, 0.0122, 0.0167, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 05:31:05,728 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95400.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:31:07,588 INFO [train.py:904] (4/8) Epoch 10, batch 4050, loss[loss=0.2237, simple_loss=0.2939, pruned_loss=0.07676, over 11884.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.267, pruned_loss=0.05774, over 3280100.78 frames. ], batch size: 247, lr: 6.89e-03, grad_scale: 16.0 2023-04-29 05:31:49,145 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.375e+02 1.964e+02 2.335e+02 2.702e+02 4.319e+02, threshold=4.671e+02, percent-clipped=0.0 2023-04-29 05:32:20,029 INFO [train.py:904] (4/8) Epoch 10, batch 4100, loss[loss=0.1883, simple_loss=0.2698, pruned_loss=0.0534, over 16414.00 frames. ], tot_loss[loss=0.1906, simple_loss=0.2678, pruned_loss=0.05676, over 3283444.74 frames. ], batch size: 68, lr: 6.88e-03, grad_scale: 16.0 2023-04-29 05:32:34,908 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95461.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 05:32:37,305 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-29 05:33:33,931 INFO [train.py:904] (4/8) Epoch 10, batch 4150, loss[loss=0.2335, simple_loss=0.3144, pruned_loss=0.07625, over 16626.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2764, pruned_loss=0.0607, over 3221874.32 frames. ], batch size: 62, lr: 6.88e-03, grad_scale: 16.0 2023-04-29 05:34:17,107 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.022e+02 2.709e+02 3.150e+02 3.936e+02 7.135e+02, threshold=6.300e+02, percent-clipped=10.0 2023-04-29 05:34:49,630 INFO [train.py:904] (4/8) Epoch 10, batch 4200, loss[loss=0.265, simple_loss=0.3299, pruned_loss=0.1, over 11099.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.284, pruned_loss=0.06278, over 3185832.13 frames. ], batch size: 248, lr: 6.88e-03, grad_scale: 16.0 2023-04-29 05:36:04,061 INFO [train.py:904] (4/8) Epoch 10, batch 4250, loss[loss=0.2067, simple_loss=0.3005, pruned_loss=0.05642, over 16764.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2861, pruned_loss=0.06152, over 3191367.42 frames. ], batch size: 39, lr: 6.88e-03, grad_scale: 8.0 2023-04-29 05:36:24,739 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95616.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:36:26,057 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95617.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:36:28,245 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.62 vs. limit=2.0 2023-04-29 05:36:37,903 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95624.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:36:48,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7477, 4.3649, 4.3277, 2.9609, 3.8564, 4.1802, 3.9734, 2.6840], device='cuda:4'), covar=tensor([0.0323, 0.0018, 0.0024, 0.0253, 0.0053, 0.0057, 0.0039, 0.0269], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0066, 0.0067, 0.0121, 0.0076, 0.0083, 0.0073, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 05:36:49,130 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.632e+02 2.325e+02 2.809e+02 3.314e+02 5.860e+02, threshold=5.619e+02, percent-clipped=0.0 2023-04-29 05:37:19,468 INFO [train.py:904] (4/8) Epoch 10, batch 4300, loss[loss=0.2163, simple_loss=0.303, pruned_loss=0.06477, over 16646.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2872, pruned_loss=0.06022, over 3195336.83 frames. ], batch size: 134, lr: 6.88e-03, grad_scale: 8.0 2023-04-29 05:37:34,496 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-29 05:37:59,685 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95678.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:38:24,123 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95695.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:38:33,665 INFO [train.py:904] (4/8) Epoch 10, batch 4350, loss[loss=0.2078, simple_loss=0.2928, pruned_loss=0.0614, over 16747.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.29, pruned_loss=0.06112, over 3175143.28 frames. ], batch size: 39, lr: 6.88e-03, grad_scale: 8.0 2023-04-29 05:38:57,024 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2193, 4.0679, 4.0182, 2.3297, 3.6783, 3.8904, 3.7468, 2.1739], device='cuda:4'), covar=tensor([0.0394, 0.0016, 0.0022, 0.0337, 0.0049, 0.0059, 0.0042, 0.0314], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0067, 0.0067, 0.0122, 0.0076, 0.0084, 0.0073, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 05:39:04,045 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-29 05:39:18,685 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.762e+02 2.523e+02 3.025e+02 3.679e+02 8.417e+02, threshold=6.050e+02, percent-clipped=3.0 2023-04-29 05:39:20,880 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1680, 4.1905, 4.5686, 4.4922, 4.5210, 4.1949, 4.2266, 4.0435], device='cuda:4'), covar=tensor([0.0254, 0.0389, 0.0261, 0.0402, 0.0389, 0.0305, 0.0699, 0.0447], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0327, 0.0330, 0.0316, 0.0375, 0.0351, 0.0453, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 05:39:49,467 INFO [train.py:904] (4/8) Epoch 10, batch 4400, loss[loss=0.2364, simple_loss=0.3205, pruned_loss=0.07616, over 12004.00 frames. ], tot_loss[loss=0.2087, simple_loss=0.2924, pruned_loss=0.0625, over 3166715.94 frames. ], batch size: 248, lr: 6.87e-03, grad_scale: 8.0 2023-04-29 05:40:02,632 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95761.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 05:40:06,630 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-29 05:41:01,592 INFO [train.py:904] (4/8) Epoch 10, batch 4450, loss[loss=0.2059, simple_loss=0.296, pruned_loss=0.05783, over 16941.00 frames. ], tot_loss[loss=0.2115, simple_loss=0.2959, pruned_loss=0.06357, over 3192889.89 frames. ], batch size: 109, lr: 6.87e-03, grad_scale: 8.0 2023-04-29 05:41:13,573 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=95809.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 05:41:13,810 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8529, 3.9987, 3.0977, 2.4240, 2.9238, 2.3087, 4.3536, 3.8105], device='cuda:4'), covar=tensor([0.2300, 0.0617, 0.1422, 0.1967, 0.2155, 0.1801, 0.0393, 0.0772], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0258, 0.0282, 0.0276, 0.0287, 0.0220, 0.0270, 0.0297], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:41:17,726 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95812.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 05:41:46,154 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.454e+02 2.254e+02 2.672e+02 3.297e+02 5.015e+02, threshold=5.344e+02, percent-clipped=0.0 2023-04-29 05:42:08,839 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95847.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:42:15,010 INFO [train.py:904] (4/8) Epoch 10, batch 4500, loss[loss=0.2026, simple_loss=0.2812, pruned_loss=0.06203, over 16602.00 frames. ], tot_loss[loss=0.2124, simple_loss=0.2961, pruned_loss=0.06432, over 3183243.35 frames. ], batch size: 35, lr: 6.87e-03, grad_scale: 8.0 2023-04-29 05:42:45,927 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95873.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 05:42:54,380 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6832, 3.0258, 2.7644, 4.7965, 3.7917, 4.3379, 1.5823, 3.1611], device='cuda:4'), covar=tensor([0.1275, 0.0603, 0.0989, 0.0122, 0.0313, 0.0277, 0.1428, 0.0721], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0157, 0.0178, 0.0138, 0.0202, 0.0206, 0.0177, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 05:43:03,684 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2759, 5.2663, 5.1302, 4.9175, 4.7057, 5.2258, 5.1203, 4.7302], device='cuda:4'), covar=tensor([0.0554, 0.0222, 0.0217, 0.0226, 0.0974, 0.0261, 0.0206, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0290, 0.0279, 0.0256, 0.0302, 0.0289, 0.0191, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:43:27,208 INFO [train.py:904] (4/8) Epoch 10, batch 4550, loss[loss=0.2124, simple_loss=0.3066, pruned_loss=0.0591, over 17030.00 frames. ], tot_loss[loss=0.2137, simple_loss=0.2969, pruned_loss=0.06528, over 3194053.61 frames. ], batch size: 50, lr: 6.87e-03, grad_scale: 8.0 2023-04-29 05:43:35,515 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95908.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 05:43:47,800 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95916.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:43:59,515 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95924.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:44:06,426 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.39 vs. limit=5.0 2023-04-29 05:44:10,340 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.676e+02 2.125e+02 2.355e+02 2.832e+02 4.769e+02, threshold=4.710e+02, percent-clipped=0.0 2023-04-29 05:44:39,213 INFO [train.py:904] (4/8) Epoch 10, batch 4600, loss[loss=0.196, simple_loss=0.2845, pruned_loss=0.05372, over 16503.00 frames. ], tot_loss[loss=0.2138, simple_loss=0.2978, pruned_loss=0.06488, over 3202347.05 frames. ], batch size: 68, lr: 6.87e-03, grad_scale: 8.0 2023-04-29 05:44:42,552 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4437, 4.4497, 4.8493, 4.8246, 4.8255, 4.4650, 4.5172, 4.2380], device='cuda:4'), covar=tensor([0.0280, 0.0377, 0.0273, 0.0324, 0.0381, 0.0327, 0.0790, 0.0426], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0319, 0.0324, 0.0308, 0.0367, 0.0342, 0.0443, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 05:44:57,860 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=95964.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:45:09,544 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=95972.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:45:11,969 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95973.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:45:19,776 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5840, 4.5404, 4.2924, 3.4327, 4.4698, 1.4514, 4.1440, 4.0235], device='cuda:4'), covar=tensor([0.0060, 0.0054, 0.0128, 0.0404, 0.0068, 0.2857, 0.0105, 0.0201], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0113, 0.0159, 0.0153, 0.0130, 0.0172, 0.0148, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:45:20,965 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95980.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:45:40,799 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-29 05:45:43,244 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95995.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:45:55,175 INFO [train.py:904] (4/8) Epoch 10, batch 4650, loss[loss=0.2001, simple_loss=0.2868, pruned_loss=0.05672, over 16501.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.2969, pruned_loss=0.0648, over 3213450.25 frames. ], batch size: 68, lr: 6.87e-03, grad_scale: 8.0 2023-04-29 05:45:57,320 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1217, 4.2431, 4.3994, 4.1999, 4.2272, 4.7734, 4.3281, 4.0515], device='cuda:4'), covar=tensor([0.1690, 0.1548, 0.1475, 0.1943, 0.2560, 0.1001, 0.1272, 0.2314], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0468, 0.0494, 0.0406, 0.0539, 0.0521, 0.0401, 0.0553], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 05:46:40,627 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.329e+02 2.110e+02 2.403e+02 2.925e+02 6.538e+02, threshold=4.805e+02, percent-clipped=1.0 2023-04-29 05:46:55,346 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96041.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:46:58,683 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=96043.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:47:06,143 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0920, 4.0171, 1.7789, 4.8495, 3.0659, 4.5731, 1.9857, 2.9087], device='cuda:4'), covar=tensor([0.0147, 0.0291, 0.1824, 0.0055, 0.0656, 0.0253, 0.1460, 0.0647], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0164, 0.0186, 0.0121, 0.0165, 0.0202, 0.0190, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 05:47:10,300 INFO [train.py:904] (4/8) Epoch 10, batch 4700, loss[loss=0.2106, simple_loss=0.29, pruned_loss=0.06556, over 16931.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.2941, pruned_loss=0.06356, over 3217089.97 frames. ], batch size: 109, lr: 6.86e-03, grad_scale: 8.0 2023-04-29 05:47:52,224 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96080.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:48:23,998 INFO [train.py:904] (4/8) Epoch 10, batch 4750, loss[loss=0.1864, simple_loss=0.2704, pruned_loss=0.05119, over 16465.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2898, pruned_loss=0.0611, over 3225740.72 frames. ], batch size: 35, lr: 6.86e-03, grad_scale: 8.0 2023-04-29 05:49:08,952 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.464e+02 2.162e+02 2.549e+02 3.399e+02 6.347e+02, threshold=5.097e+02, percent-clipped=5.0 2023-04-29 05:49:23,233 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96141.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:49:38,023 INFO [train.py:904] (4/8) Epoch 10, batch 4800, loss[loss=0.2154, simple_loss=0.2924, pruned_loss=0.06919, over 11664.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.2861, pruned_loss=0.059, over 3213610.29 frames. ], batch size: 246, lr: 6.86e-03, grad_scale: 8.0 2023-04-29 05:49:38,367 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9674, 5.3031, 5.0637, 5.0464, 4.8211, 4.6031, 4.7498, 5.3695], device='cuda:4'), covar=tensor([0.1109, 0.0684, 0.0822, 0.0636, 0.0662, 0.0843, 0.0943, 0.0759], device='cuda:4'), in_proj_covar=tensor([0.0506, 0.0641, 0.0531, 0.0439, 0.0401, 0.0414, 0.0530, 0.0488], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:49:58,814 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96165.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:50:03,011 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96168.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 05:50:19,740 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7378, 3.6587, 3.8310, 3.9541, 4.0303, 3.6181, 3.9794, 4.0678], device='cuda:4'), covar=tensor([0.1169, 0.0922, 0.1071, 0.0499, 0.0457, 0.1684, 0.0628, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0494, 0.0612, 0.0747, 0.0623, 0.0465, 0.0479, 0.0488, 0.0551], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:50:54,658 INFO [train.py:904] (4/8) Epoch 10, batch 4850, loss[loss=0.213, simple_loss=0.3069, pruned_loss=0.05952, over 15361.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2866, pruned_loss=0.05824, over 3203995.51 frames. ], batch size: 190, lr: 6.86e-03, grad_scale: 8.0 2023-04-29 05:50:56,264 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96203.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 05:51:29,413 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.38 vs. limit=5.0 2023-04-29 05:51:32,220 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96226.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:51:40,157 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.515e+02 2.214e+02 2.693e+02 3.150e+02 7.967e+02, threshold=5.386e+02, percent-clipped=5.0 2023-04-29 05:51:48,802 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-29 05:52:10,085 INFO [train.py:904] (4/8) Epoch 10, batch 4900, loss[loss=0.1657, simple_loss=0.2609, pruned_loss=0.03519, over 16818.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2858, pruned_loss=0.05708, over 3192959.94 frames. ], batch size: 96, lr: 6.86e-03, grad_scale: 8.0 2023-04-29 05:52:40,252 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-29 05:52:42,917 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96273.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:53:24,547 INFO [train.py:904] (4/8) Epoch 10, batch 4950, loss[loss=0.1997, simple_loss=0.2938, pruned_loss=0.05281, over 16652.00 frames. ], tot_loss[loss=0.1994, simple_loss=0.2855, pruned_loss=0.05661, over 3193314.81 frames. ], batch size: 134, lr: 6.85e-03, grad_scale: 8.0 2023-04-29 05:53:52,125 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=96321.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:53:56,275 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96324.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:54:05,871 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.526e+02 2.230e+02 2.596e+02 3.100e+02 4.863e+02, threshold=5.193e+02, percent-clipped=0.0 2023-04-29 05:54:07,805 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-29 05:54:12,440 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96336.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:54:19,819 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2277, 5.1136, 5.0705, 4.3651, 5.1334, 1.7371, 4.8105, 5.0383], device='cuda:4'), covar=tensor([0.0056, 0.0059, 0.0085, 0.0353, 0.0063, 0.2128, 0.0090, 0.0124], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0109, 0.0154, 0.0150, 0.0126, 0.0168, 0.0144, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 05:54:33,261 INFO [train.py:904] (4/8) Epoch 10, batch 5000, loss[loss=0.2271, simple_loss=0.3183, pruned_loss=0.06798, over 16944.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2872, pruned_loss=0.05681, over 3198563.58 frames. ], batch size: 109, lr: 6.85e-03, grad_scale: 8.0 2023-04-29 05:55:06,916 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9364, 5.3619, 5.6442, 5.2784, 5.4607, 5.9888, 5.5378, 5.2102], device='cuda:4'), covar=tensor([0.0804, 0.1451, 0.1500, 0.1792, 0.2214, 0.0903, 0.1114, 0.2117], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0457, 0.0484, 0.0402, 0.0532, 0.0519, 0.0396, 0.0547], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 05:55:20,962 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96385.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:55:28,056 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96390.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:55:44,796 INFO [train.py:904] (4/8) Epoch 10, batch 5050, loss[loss=0.2062, simple_loss=0.2814, pruned_loss=0.0655, over 16676.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.288, pruned_loss=0.05712, over 3211607.87 frames. ], batch size: 57, lr: 6.85e-03, grad_scale: 8.0 2023-04-29 05:56:18,211 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2013, 1.8809, 2.5637, 3.0329, 2.9871, 3.6681, 2.1713, 3.4811], device='cuda:4'), covar=tensor([0.0141, 0.0375, 0.0254, 0.0203, 0.0193, 0.0088, 0.0353, 0.0075], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0168, 0.0153, 0.0157, 0.0165, 0.0120, 0.0171, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 05:56:27,890 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.480e+02 2.386e+02 2.799e+02 3.376e+02 6.474e+02, threshold=5.598e+02, percent-clipped=3.0 2023-04-29 05:56:33,453 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96436.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:56:55,126 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96451.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:56:55,806 INFO [train.py:904] (4/8) Epoch 10, batch 5100, loss[loss=0.1676, simple_loss=0.2494, pruned_loss=0.0429, over 16858.00 frames. ], tot_loss[loss=0.2003, simple_loss=0.287, pruned_loss=0.05681, over 3203170.10 frames. ], batch size: 42, lr: 6.85e-03, grad_scale: 8.0 2023-04-29 05:57:20,108 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96468.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 05:58:08,610 INFO [train.py:904] (4/8) Epoch 10, batch 5150, loss[loss=0.2278, simple_loss=0.3203, pruned_loss=0.06766, over 16319.00 frames. ], tot_loss[loss=0.2003, simple_loss=0.2877, pruned_loss=0.05644, over 3161984.05 frames. ], batch size: 165, lr: 6.85e-03, grad_scale: 8.0 2023-04-29 05:58:11,196 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96503.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:58:30,253 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=96516.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 05:58:31,549 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9154, 1.7745, 2.3200, 2.7851, 2.6298, 3.2204, 1.8610, 3.1622], device='cuda:4'), covar=tensor([0.0132, 0.0335, 0.0230, 0.0177, 0.0209, 0.0089, 0.0367, 0.0067], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0166, 0.0151, 0.0156, 0.0163, 0.0119, 0.0170, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 05:58:36,670 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96521.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:58:50,306 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.78 vs. limit=5.0 2023-04-29 05:58:52,021 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.417e+02 2.088e+02 2.479e+02 2.937e+02 7.130e+02, threshold=4.958e+02, percent-clipped=1.0 2023-04-29 05:59:15,976 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-29 05:59:19,730 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6861, 1.7876, 2.2338, 2.6685, 2.6336, 2.9877, 1.8238, 2.9308], device='cuda:4'), covar=tensor([0.0137, 0.0370, 0.0232, 0.0197, 0.0198, 0.0112, 0.0371, 0.0085], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0167, 0.0152, 0.0157, 0.0164, 0.0119, 0.0171, 0.0110], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 05:59:21,922 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=96551.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 05:59:22,751 INFO [train.py:904] (4/8) Epoch 10, batch 5200, loss[loss=0.1831, simple_loss=0.266, pruned_loss=0.05006, over 16462.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2862, pruned_loss=0.05607, over 3172603.22 frames. ], batch size: 75, lr: 6.85e-03, grad_scale: 8.0 2023-04-29 05:59:32,131 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0574, 3.1663, 1.6569, 3.2555, 2.3040, 3.3231, 1.9609, 2.5206], device='cuda:4'), covar=tensor([0.0182, 0.0283, 0.1558, 0.0113, 0.0803, 0.0388, 0.1391, 0.0626], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0162, 0.0186, 0.0119, 0.0166, 0.0200, 0.0192, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 05:59:33,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5588, 4.4792, 4.4367, 3.7562, 4.4675, 1.6755, 4.1389, 4.2655], device='cuda:4'), covar=tensor([0.0068, 0.0065, 0.0111, 0.0352, 0.0067, 0.2266, 0.0110, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0110, 0.0155, 0.0153, 0.0127, 0.0170, 0.0144, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:00:35,342 INFO [train.py:904] (4/8) Epoch 10, batch 5250, loss[loss=0.1761, simple_loss=0.2704, pruned_loss=0.04087, over 16310.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2835, pruned_loss=0.05535, over 3168941.81 frames. ], batch size: 146, lr: 6.84e-03, grad_scale: 4.0 2023-04-29 06:00:40,785 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8080, 4.0071, 3.1542, 2.3224, 2.7785, 2.4821, 4.2250, 3.5778], device='cuda:4'), covar=tensor([0.2412, 0.0577, 0.1333, 0.2180, 0.2164, 0.1603, 0.0422, 0.0884], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0255, 0.0276, 0.0273, 0.0280, 0.0216, 0.0267, 0.0290], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:01:21,023 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.540e+02 2.301e+02 2.656e+02 3.144e+02 5.435e+02, threshold=5.311e+02, percent-clipped=2.0 2023-04-29 06:01:26,453 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96636.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:01:31,125 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3879, 3.7392, 3.6220, 2.1960, 3.2339, 2.6992, 3.7733, 3.8977], device='cuda:4'), covar=tensor([0.0196, 0.0561, 0.0504, 0.1618, 0.0659, 0.0743, 0.0496, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0141, 0.0156, 0.0141, 0.0134, 0.0123, 0.0135, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 06:01:48,984 INFO [train.py:904] (4/8) Epoch 10, batch 5300, loss[loss=0.1566, simple_loss=0.2394, pruned_loss=0.03691, over 16787.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2793, pruned_loss=0.05365, over 3176088.31 frames. ], batch size: 83, lr: 6.84e-03, grad_scale: 4.0 2023-04-29 06:02:17,488 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5929, 3.6547, 4.0748, 1.9115, 4.3046, 4.3089, 2.9946, 3.0778], device='cuda:4'), covar=tensor([0.0747, 0.0195, 0.0112, 0.1164, 0.0032, 0.0054, 0.0369, 0.0436], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0098, 0.0085, 0.0137, 0.0069, 0.0095, 0.0119, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 06:02:30,154 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96680.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:02:36,966 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=96684.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:03:03,535 INFO [train.py:904] (4/8) Epoch 10, batch 5350, loss[loss=0.2226, simple_loss=0.317, pruned_loss=0.06407, over 16298.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2776, pruned_loss=0.05305, over 3187865.73 frames. ], batch size: 165, lr: 6.84e-03, grad_scale: 4.0 2023-04-29 06:03:27,984 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-29 06:03:32,264 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.85 vs. limit=5.0 2023-04-29 06:03:48,671 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.632e+02 2.305e+02 2.858e+02 3.473e+02 6.088e+02, threshold=5.716e+02, percent-clipped=3.0 2023-04-29 06:03:53,470 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96736.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:04:08,473 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96746.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:04:15,625 INFO [train.py:904] (4/8) Epoch 10, batch 5400, loss[loss=0.1967, simple_loss=0.2837, pruned_loss=0.05488, over 16590.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2809, pruned_loss=0.05365, over 3196632.96 frames. ], batch size: 62, lr: 6.84e-03, grad_scale: 4.0 2023-04-29 06:04:52,065 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96776.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:05:02,543 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=96784.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:05:31,404 INFO [train.py:904] (4/8) Epoch 10, batch 5450, loss[loss=0.2611, simple_loss=0.3215, pruned_loss=0.1003, over 11753.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2841, pruned_loss=0.05554, over 3193678.56 frames. ], batch size: 247, lr: 6.84e-03, grad_scale: 4.0 2023-04-29 06:05:41,670 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-29 06:05:50,160 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-29 06:06:02,205 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96821.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:06:20,032 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.555e+02 2.545e+02 3.377e+02 4.195e+02 1.255e+03, threshold=6.754e+02, percent-clipped=10.0 2023-04-29 06:06:27,197 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96837.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:06:49,547 INFO [train.py:904] (4/8) Epoch 10, batch 5500, loss[loss=0.2857, simple_loss=0.3396, pruned_loss=0.1159, over 11514.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2922, pruned_loss=0.0612, over 3168479.84 frames. ], batch size: 248, lr: 6.83e-03, grad_scale: 4.0 2023-04-29 06:06:55,196 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0860, 2.0156, 2.7159, 3.0076, 3.0436, 3.8250, 2.2345, 3.6453], device='cuda:4'), covar=tensor([0.0152, 0.0300, 0.0200, 0.0173, 0.0167, 0.0063, 0.0308, 0.0061], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0166, 0.0152, 0.0157, 0.0163, 0.0118, 0.0169, 0.0110], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 06:07:17,928 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=96869.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:08:08,493 INFO [train.py:904] (4/8) Epoch 10, batch 5550, loss[loss=0.2313, simple_loss=0.3109, pruned_loss=0.07584, over 16423.00 frames. ], tot_loss[loss=0.2165, simple_loss=0.2999, pruned_loss=0.0666, over 3156956.87 frames. ], batch size: 146, lr: 6.83e-03, grad_scale: 4.0 2023-04-29 06:08:50,673 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-29 06:09:01,353 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.332e+02 3.697e+02 4.404e+02 5.267e+02 9.227e+02, threshold=8.809e+02, percent-clipped=8.0 2023-04-29 06:09:28,004 INFO [train.py:904] (4/8) Epoch 10, batch 5600, loss[loss=0.3304, simple_loss=0.3704, pruned_loss=0.1452, over 11344.00 frames. ], tot_loss[loss=0.2255, simple_loss=0.3064, pruned_loss=0.07235, over 3110368.31 frames. ], batch size: 247, lr: 6.83e-03, grad_scale: 2.0 2023-04-29 06:09:48,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4537, 2.5970, 2.0926, 2.2683, 2.9940, 2.6038, 3.2596, 3.2110], device='cuda:4'), covar=tensor([0.0061, 0.0248, 0.0350, 0.0301, 0.0159, 0.0260, 0.0145, 0.0146], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0194, 0.0192, 0.0191, 0.0193, 0.0195, 0.0197, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:10:15,943 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96980.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:10:51,167 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97000.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:10:53,952 INFO [train.py:904] (4/8) Epoch 10, batch 5650, loss[loss=0.2288, simple_loss=0.3127, pruned_loss=0.07243, over 16383.00 frames. ], tot_loss[loss=0.2323, simple_loss=0.3115, pruned_loss=0.07658, over 3081335.52 frames. ], batch size: 146, lr: 6.83e-03, grad_scale: 2.0 2023-04-29 06:11:03,481 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0594, 2.5094, 2.7393, 4.8660, 2.3924, 2.9071, 2.5746, 2.8572], device='cuda:4'), covar=tensor([0.0735, 0.2707, 0.1742, 0.0255, 0.3218, 0.1815, 0.2371, 0.2415], device='cuda:4'), in_proj_covar=tensor([0.0356, 0.0383, 0.0320, 0.0318, 0.0403, 0.0436, 0.0343, 0.0448], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:11:31,348 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-29 06:11:34,447 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=97028.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:11:43,932 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.521e+02 3.909e+02 4.849e+02 5.778e+02 1.289e+03, threshold=9.698e+02, percent-clipped=3.0 2023-04-29 06:12:02,684 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97046.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:12:11,111 INFO [train.py:904] (4/8) Epoch 10, batch 5700, loss[loss=0.2219, simple_loss=0.3079, pruned_loss=0.06799, over 16906.00 frames. ], tot_loss[loss=0.2361, simple_loss=0.314, pruned_loss=0.07904, over 3066169.63 frames. ], batch size: 109, lr: 6.83e-03, grad_scale: 2.0 2023-04-29 06:12:25,240 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97061.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:12:47,376 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-29 06:13:15,309 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4367, 1.5919, 2.0113, 2.3147, 2.4539, 2.7325, 1.8097, 2.5606], device='cuda:4'), covar=tensor([0.0147, 0.0336, 0.0217, 0.0209, 0.0185, 0.0121, 0.0313, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0165, 0.0150, 0.0154, 0.0161, 0.0117, 0.0167, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 06:13:16,960 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=97094.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:13:29,501 INFO [train.py:904] (4/8) Epoch 10, batch 5750, loss[loss=0.2598, simple_loss=0.3163, pruned_loss=0.1017, over 11287.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.3162, pruned_loss=0.08081, over 3026935.68 frames. ], batch size: 246, lr: 6.83e-03, grad_scale: 2.0 2023-04-29 06:14:17,794 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97132.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:14:22,108 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.407e+02 3.370e+02 4.173e+02 4.986e+02 1.197e+03, threshold=8.346e+02, percent-clipped=2.0 2023-04-29 06:14:49,724 INFO [train.py:904] (4/8) Epoch 10, batch 5800, loss[loss=0.2197, simple_loss=0.3067, pruned_loss=0.06631, over 16671.00 frames. ], tot_loss[loss=0.2359, simple_loss=0.3148, pruned_loss=0.07851, over 3034312.92 frames. ], batch size: 76, lr: 6.82e-03, grad_scale: 2.0 2023-04-29 06:14:58,043 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7741, 3.8482, 2.9988, 2.2701, 2.8128, 2.4224, 4.0665, 3.5384], device='cuda:4'), covar=tensor([0.2362, 0.0684, 0.1501, 0.2126, 0.2131, 0.1678, 0.0436, 0.0933], device='cuda:4'), in_proj_covar=tensor([0.0300, 0.0256, 0.0278, 0.0274, 0.0280, 0.0216, 0.0264, 0.0289], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:15:08,754 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2962, 3.2794, 1.6576, 3.6253, 2.4101, 3.5338, 1.7300, 2.5089], device='cuda:4'), covar=tensor([0.0232, 0.0361, 0.1916, 0.0153, 0.0856, 0.0514, 0.1929, 0.0820], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0162, 0.0185, 0.0119, 0.0166, 0.0201, 0.0189, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 06:15:58,999 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6485, 2.7730, 2.6698, 4.0075, 3.1001, 4.0706, 1.3084, 2.7901], device='cuda:4'), covar=tensor([0.1358, 0.0634, 0.0984, 0.0141, 0.0295, 0.0360, 0.1625, 0.0823], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0156, 0.0177, 0.0136, 0.0198, 0.0206, 0.0178, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 06:16:02,421 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.73 vs. limit=5.0 2023-04-29 06:16:07,864 INFO [train.py:904] (4/8) Epoch 10, batch 5850, loss[loss=0.2359, simple_loss=0.3257, pruned_loss=0.07304, over 16749.00 frames. ], tot_loss[loss=0.2332, simple_loss=0.3128, pruned_loss=0.07677, over 3042615.01 frames. ], batch size: 83, lr: 6.82e-03, grad_scale: 2.0 2023-04-29 06:16:31,848 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-29 06:16:56,894 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-29 06:17:00,834 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.062e+02 2.965e+02 3.707e+02 4.622e+02 9.015e+02, threshold=7.415e+02, percent-clipped=1.0 2023-04-29 06:17:28,942 INFO [train.py:904] (4/8) Epoch 10, batch 5900, loss[loss=0.2364, simple_loss=0.3245, pruned_loss=0.07412, over 17042.00 frames. ], tot_loss[loss=0.2322, simple_loss=0.3123, pruned_loss=0.07606, over 3072641.26 frames. ], batch size: 41, lr: 6.82e-03, grad_scale: 2.0 2023-04-29 06:17:45,043 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97260.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:18:40,311 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6859, 4.6890, 5.1401, 5.1102, 5.0879, 4.8078, 4.7774, 4.5255], device='cuda:4'), covar=tensor([0.0292, 0.0506, 0.0343, 0.0341, 0.0412, 0.0390, 0.0873, 0.0413], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0319, 0.0322, 0.0309, 0.0373, 0.0347, 0.0442, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 06:18:49,538 INFO [train.py:904] (4/8) Epoch 10, batch 5950, loss[loss=0.2655, simple_loss=0.337, pruned_loss=0.09707, over 11919.00 frames. ], tot_loss[loss=0.2312, simple_loss=0.3127, pruned_loss=0.07483, over 3064757.05 frames. ], batch size: 247, lr: 6.82e-03, grad_scale: 2.0 2023-04-29 06:19:00,771 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.74 vs. limit=2.0 2023-04-29 06:19:11,245 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.10 vs. limit=5.0 2023-04-29 06:19:20,359 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97321.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:19:41,580 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.885e+02 2.897e+02 3.264e+02 4.239e+02 7.603e+02, threshold=6.529e+02, percent-clipped=1.0 2023-04-29 06:19:52,723 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97342.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:20:01,441 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97347.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:20:09,122 INFO [train.py:904] (4/8) Epoch 10, batch 6000, loss[loss=0.1993, simple_loss=0.289, pruned_loss=0.05485, over 16864.00 frames. ], tot_loss[loss=0.2288, simple_loss=0.3108, pruned_loss=0.07342, over 3083387.71 frames. ], batch size: 96, lr: 6.82e-03, grad_scale: 4.0 2023-04-29 06:20:09,122 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 06:20:23,728 INFO [train.py:938] (4/8) Epoch 10, validation: loss=0.165, simple_loss=0.2783, pruned_loss=0.02583, over 944034.00 frames. 2023-04-29 06:20:23,729 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 06:20:30,440 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97356.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:20:38,048 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7842, 3.8007, 2.2580, 4.3741, 2.8118, 4.3241, 2.3358, 2.9714], device='cuda:4'), covar=tensor([0.0203, 0.0356, 0.1631, 0.0103, 0.0786, 0.0434, 0.1526, 0.0661], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0163, 0.0186, 0.0120, 0.0166, 0.0201, 0.0191, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 06:20:58,023 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97374.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:21:42,480 INFO [train.py:904] (4/8) Epoch 10, batch 6050, loss[loss=0.2523, simple_loss=0.3128, pruned_loss=0.09587, over 11651.00 frames. ], tot_loss[loss=0.2274, simple_loss=0.3091, pruned_loss=0.07287, over 3087888.34 frames. ], batch size: 248, lr: 6.82e-03, grad_scale: 2.0 2023-04-29 06:21:45,428 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97403.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 06:21:52,798 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97408.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:22:14,247 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2327, 4.0750, 4.2950, 4.4780, 4.5863, 4.1895, 4.5463, 4.5909], device='cuda:4'), covar=tensor([0.1503, 0.1047, 0.1312, 0.0584, 0.0536, 0.0888, 0.0671, 0.0510], device='cuda:4'), in_proj_covar=tensor([0.0505, 0.0624, 0.0759, 0.0636, 0.0484, 0.0484, 0.0500, 0.0567], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:22:30,429 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97432.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:22:35,050 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97435.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:22:35,671 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.881e+02 3.012e+02 3.614e+02 4.338e+02 6.849e+02, threshold=7.229e+02, percent-clipped=1.0 2023-04-29 06:23:02,109 INFO [train.py:904] (4/8) Epoch 10, batch 6100, loss[loss=0.2134, simple_loss=0.2967, pruned_loss=0.06505, over 17163.00 frames. ], tot_loss[loss=0.2257, simple_loss=0.3085, pruned_loss=0.0715, over 3104959.78 frames. ], batch size: 46, lr: 6.81e-03, grad_scale: 2.0 2023-04-29 06:23:42,854 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-29 06:23:49,208 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=97480.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:24:18,287 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-29 06:24:23,702 INFO [train.py:904] (4/8) Epoch 10, batch 6150, loss[loss=0.1943, simple_loss=0.2779, pruned_loss=0.05536, over 17257.00 frames. ], tot_loss[loss=0.2241, simple_loss=0.3063, pruned_loss=0.07089, over 3111535.19 frames. ], batch size: 52, lr: 6.81e-03, grad_scale: 2.0 2023-04-29 06:25:17,520 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.161e+02 3.229e+02 3.939e+02 5.020e+02 8.476e+02, threshold=7.879e+02, percent-clipped=2.0 2023-04-29 06:25:41,181 INFO [train.py:904] (4/8) Epoch 10, batch 6200, loss[loss=0.1855, simple_loss=0.2753, pruned_loss=0.04786, over 16505.00 frames. ], tot_loss[loss=0.2218, simple_loss=0.3039, pruned_loss=0.06988, over 3129024.31 frames. ], batch size: 68, lr: 6.81e-03, grad_scale: 2.0 2023-04-29 06:25:46,673 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97555.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:26:57,937 INFO [train.py:904] (4/8) Epoch 10, batch 6250, loss[loss=0.2628, simple_loss=0.3168, pruned_loss=0.1044, over 11869.00 frames. ], tot_loss[loss=0.2223, simple_loss=0.3039, pruned_loss=0.07039, over 3099284.10 frames. ], batch size: 248, lr: 6.81e-03, grad_scale: 2.0 2023-04-29 06:27:18,723 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97616.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:27:18,857 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97616.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:27:47,811 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.754e+02 3.287e+02 3.846e+02 4.993e+02 1.247e+03, threshold=7.692e+02, percent-clipped=7.0 2023-04-29 06:28:00,099 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-29 06:28:11,794 INFO [train.py:904] (4/8) Epoch 10, batch 6300, loss[loss=0.1905, simple_loss=0.2734, pruned_loss=0.05383, over 17112.00 frames. ], tot_loss[loss=0.2215, simple_loss=0.3036, pruned_loss=0.06972, over 3115303.75 frames. ], batch size: 49, lr: 6.81e-03, grad_scale: 2.0 2023-04-29 06:28:18,569 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97656.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:28:48,989 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97674.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 06:29:24,485 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97698.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 06:29:30,238 INFO [train.py:904] (4/8) Epoch 10, batch 6350, loss[loss=0.2549, simple_loss=0.3254, pruned_loss=0.09224, over 15445.00 frames. ], tot_loss[loss=0.2235, simple_loss=0.3047, pruned_loss=0.07121, over 3099675.12 frames. ], batch size: 190, lr: 6.81e-03, grad_scale: 2.0 2023-04-29 06:29:31,921 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97703.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:29:33,046 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=97704.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:30:01,349 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-04-29 06:30:13,648 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97730.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:30:22,252 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97735.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 06:30:22,844 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.978e+02 3.422e+02 4.156e+02 5.083e+02 8.318e+02, threshold=8.312e+02, percent-clipped=1.0 2023-04-29 06:30:46,248 INFO [train.py:904] (4/8) Epoch 10, batch 6400, loss[loss=0.2341, simple_loss=0.3067, pruned_loss=0.08071, over 16898.00 frames. ], tot_loss[loss=0.2245, simple_loss=0.3044, pruned_loss=0.07228, over 3094966.59 frames. ], batch size: 116, lr: 6.80e-03, grad_scale: 4.0 2023-04-29 06:30:50,610 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-29 06:32:00,788 INFO [train.py:904] (4/8) Epoch 10, batch 6450, loss[loss=0.1852, simple_loss=0.2812, pruned_loss=0.04459, over 16840.00 frames. ], tot_loss[loss=0.2234, simple_loss=0.304, pruned_loss=0.07139, over 3093199.47 frames. ], batch size: 102, lr: 6.80e-03, grad_scale: 4.0 2023-04-29 06:32:16,490 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-29 06:32:57,254 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.110e+02 2.867e+02 3.528e+02 4.342e+02 1.041e+03, threshold=7.056e+02, percent-clipped=1.0 2023-04-29 06:33:21,801 INFO [train.py:904] (4/8) Epoch 10, batch 6500, loss[loss=0.1942, simple_loss=0.2747, pruned_loss=0.05691, over 17118.00 frames. ], tot_loss[loss=0.2214, simple_loss=0.3016, pruned_loss=0.07057, over 3078584.27 frames. ], batch size: 47, lr: 6.80e-03, grad_scale: 4.0 2023-04-29 06:34:04,180 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-29 06:34:17,538 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97888.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:34:41,150 INFO [train.py:904] (4/8) Epoch 10, batch 6550, loss[loss=0.2152, simple_loss=0.3148, pruned_loss=0.05786, over 16463.00 frames. ], tot_loss[loss=0.2244, simple_loss=0.3053, pruned_loss=0.07179, over 3077299.96 frames. ], batch size: 146, lr: 6.80e-03, grad_scale: 4.0 2023-04-29 06:34:50,145 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0867, 5.3976, 4.9016, 5.2430, 4.8435, 4.6269, 4.9703, 5.4541], device='cuda:4'), covar=tensor([0.1840, 0.1323, 0.2108, 0.1184, 0.1447, 0.1370, 0.1848, 0.1521], device='cuda:4'), in_proj_covar=tensor([0.0511, 0.0637, 0.0534, 0.0446, 0.0400, 0.0417, 0.0532, 0.0483], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:34:55,678 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97911.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:35:04,946 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97916.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:35:35,377 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.197e+02 3.236e+02 3.952e+02 5.112e+02 9.277e+02, threshold=7.905e+02, percent-clipped=2.0 2023-04-29 06:35:56,856 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97949.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:36:00,716 INFO [train.py:904] (4/8) Epoch 10, batch 6600, loss[loss=0.2218, simple_loss=0.3032, pruned_loss=0.07019, over 16927.00 frames. ], tot_loss[loss=0.2261, simple_loss=0.3074, pruned_loss=0.07239, over 3065579.56 frames. ], batch size: 116, lr: 6.80e-03, grad_scale: 4.0 2023-04-29 06:36:19,345 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=97964.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:37:14,714 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97998.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:37:22,483 INFO [train.py:904] (4/8) Epoch 10, batch 6650, loss[loss=0.2756, simple_loss=0.3323, pruned_loss=0.1094, over 11296.00 frames. ], tot_loss[loss=0.2281, simple_loss=0.3085, pruned_loss=0.07384, over 3058457.53 frames. ], batch size: 246, lr: 6.79e-03, grad_scale: 4.0 2023-04-29 06:37:24,759 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98003.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:37:27,361 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4964, 3.5690, 3.3312, 3.1039, 3.1684, 3.4713, 3.2974, 3.2387], device='cuda:4'), covar=tensor([0.0572, 0.0469, 0.0244, 0.0211, 0.0618, 0.0367, 0.0953, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0234, 0.0293, 0.0275, 0.0253, 0.0298, 0.0287, 0.0188, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:38:05,317 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98030.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 06:38:05,337 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98030.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:38:14,887 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.646e+02 3.281e+02 3.856e+02 4.909e+02 8.809e+02, threshold=7.713e+02, percent-clipped=1.0 2023-04-29 06:38:16,868 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6899, 2.6845, 2.0804, 2.4707, 3.1643, 2.7943, 3.5413, 3.3812], device='cuda:4'), covar=tensor([0.0048, 0.0256, 0.0354, 0.0301, 0.0149, 0.0251, 0.0117, 0.0142], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0194, 0.0192, 0.0193, 0.0192, 0.0197, 0.0198, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:38:31,003 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=98046.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:38:39,002 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=98051.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:38:39,792 INFO [train.py:904] (4/8) Epoch 10, batch 6700, loss[loss=0.2705, simple_loss=0.3214, pruned_loss=0.1098, over 11491.00 frames. ], tot_loss[loss=0.2284, simple_loss=0.308, pruned_loss=0.07443, over 3036591.82 frames. ], batch size: 246, lr: 6.79e-03, grad_scale: 4.0 2023-04-29 06:39:20,941 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=98078.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:39:57,931 INFO [train.py:904] (4/8) Epoch 10, batch 6750, loss[loss=0.218, simple_loss=0.2963, pruned_loss=0.06986, over 16750.00 frames. ], tot_loss[loss=0.227, simple_loss=0.3067, pruned_loss=0.0736, over 3056273.60 frames. ], batch size: 124, lr: 6.79e-03, grad_scale: 4.0 2023-04-29 06:40:04,412 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98106.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:40:49,800 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.169e+02 3.584e+02 4.478e+02 5.468e+02 7.454e+02, threshold=8.956e+02, percent-clipped=0.0 2023-04-29 06:41:00,740 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2816, 1.9440, 2.5831, 3.0787, 3.1363, 3.6290, 2.0926, 3.5320], device='cuda:4'), covar=tensor([0.0113, 0.0325, 0.0203, 0.0164, 0.0152, 0.0085, 0.0316, 0.0075], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0165, 0.0148, 0.0153, 0.0161, 0.0116, 0.0168, 0.0108], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 06:41:15,032 INFO [train.py:904] (4/8) Epoch 10, batch 6800, loss[loss=0.1928, simple_loss=0.2819, pruned_loss=0.05189, over 16848.00 frames. ], tot_loss[loss=0.2269, simple_loss=0.307, pruned_loss=0.07342, over 3062619.44 frames. ], batch size: 102, lr: 6.79e-03, grad_scale: 8.0 2023-04-29 06:41:39,164 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98167.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:42:33,969 INFO [train.py:904] (4/8) Epoch 10, batch 6850, loss[loss=0.1999, simple_loss=0.3084, pruned_loss=0.04571, over 16704.00 frames. ], tot_loss[loss=0.2283, simple_loss=0.3083, pruned_loss=0.07419, over 3058491.91 frames. ], batch size: 89, lr: 6.79e-03, grad_scale: 8.0 2023-04-29 06:42:47,564 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98211.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:42:58,587 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1073, 2.9908, 3.1455, 1.6744, 3.3377, 3.3896, 2.6784, 2.6168], device='cuda:4'), covar=tensor([0.0731, 0.0198, 0.0166, 0.1111, 0.0062, 0.0130, 0.0367, 0.0424], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0098, 0.0087, 0.0139, 0.0068, 0.0096, 0.0120, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 06:43:24,606 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.168e+02 2.932e+02 3.586e+02 4.673e+02 8.284e+02, threshold=7.173e+02, percent-clipped=0.0 2023-04-29 06:43:37,282 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98244.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:43:49,037 INFO [train.py:904] (4/8) Epoch 10, batch 6900, loss[loss=0.2187, simple_loss=0.3031, pruned_loss=0.06716, over 16603.00 frames. ], tot_loss[loss=0.229, simple_loss=0.3106, pruned_loss=0.07368, over 3069929.56 frames. ], batch size: 57, lr: 6.79e-03, grad_scale: 8.0 2023-04-29 06:44:01,149 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=98259.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:44:24,007 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9683, 2.2562, 2.2827, 2.8258, 2.1506, 3.1741, 1.7562, 2.6575], device='cuda:4'), covar=tensor([0.1022, 0.0518, 0.0913, 0.0113, 0.0118, 0.0311, 0.1211, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0155, 0.0178, 0.0136, 0.0201, 0.0204, 0.0178, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 06:44:40,574 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4740, 5.8511, 5.5051, 5.6468, 5.1666, 5.1526, 5.3222, 5.9166], device='cuda:4'), covar=tensor([0.0908, 0.0691, 0.0945, 0.0626, 0.0772, 0.0564, 0.0909, 0.0725], device='cuda:4'), in_proj_covar=tensor([0.0511, 0.0642, 0.0535, 0.0446, 0.0401, 0.0417, 0.0538, 0.0487], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 06:45:09,157 INFO [train.py:904] (4/8) Epoch 10, batch 6950, loss[loss=0.2145, simple_loss=0.2984, pruned_loss=0.06533, over 16263.00 frames. ], tot_loss[loss=0.2308, simple_loss=0.3119, pruned_loss=0.07486, over 3069575.01 frames. ], batch size: 165, lr: 6.78e-03, grad_scale: 8.0 2023-04-29 06:45:44,685 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2587, 1.5334, 1.9094, 2.2359, 2.3626, 2.4052, 1.5260, 2.2766], device='cuda:4'), covar=tensor([0.0149, 0.0376, 0.0218, 0.0227, 0.0201, 0.0143, 0.0397, 0.0118], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0166, 0.0149, 0.0155, 0.0164, 0.0118, 0.0170, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 06:45:54,243 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98330.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 06:46:01,761 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.874e+02 3.204e+02 3.949e+02 4.845e+02 8.099e+02, threshold=7.898e+02, percent-clipped=3.0 2023-04-29 06:46:27,403 INFO [train.py:904] (4/8) Epoch 10, batch 7000, loss[loss=0.2277, simple_loss=0.3218, pruned_loss=0.06676, over 16841.00 frames. ], tot_loss[loss=0.2293, simple_loss=0.3112, pruned_loss=0.07363, over 3074446.62 frames. ], batch size: 116, lr: 6.78e-03, grad_scale: 8.0 2023-04-29 06:47:08,045 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=98378.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 06:47:43,136 INFO [train.py:904] (4/8) Epoch 10, batch 7050, loss[loss=0.2422, simple_loss=0.3178, pruned_loss=0.08327, over 16311.00 frames. ], tot_loss[loss=0.2286, simple_loss=0.3114, pruned_loss=0.07287, over 3098118.19 frames. ], batch size: 165, lr: 6.78e-03, grad_scale: 8.0 2023-04-29 06:47:57,162 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98410.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:48:34,410 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.770e+02 3.125e+02 3.922e+02 4.883e+02 1.063e+03, threshold=7.844e+02, percent-clipped=4.0 2023-04-29 06:48:59,553 INFO [train.py:904] (4/8) Epoch 10, batch 7100, loss[loss=0.2146, simple_loss=0.3012, pruned_loss=0.06406, over 16483.00 frames. ], tot_loss[loss=0.2271, simple_loss=0.3096, pruned_loss=0.07234, over 3103276.79 frames. ], batch size: 75, lr: 6.78e-03, grad_scale: 8.0 2023-04-29 06:49:14,092 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98462.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:49:27,947 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98471.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:50:00,828 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9816, 5.8059, 6.0913, 5.7329, 5.9090, 6.3564, 5.7828, 5.5671], device='cuda:4'), covar=tensor([0.0871, 0.1635, 0.1650, 0.1649, 0.2032, 0.0813, 0.1512, 0.2473], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0477, 0.0512, 0.0415, 0.0548, 0.0541, 0.0417, 0.0566], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 06:50:06,171 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5402, 4.1000, 4.3708, 1.6550, 4.5535, 4.6133, 3.2341, 3.3821], device='cuda:4'), covar=tensor([0.0959, 0.0123, 0.0136, 0.1372, 0.0049, 0.0063, 0.0343, 0.0456], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0099, 0.0086, 0.0138, 0.0068, 0.0096, 0.0120, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 06:50:11,985 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1782, 4.1697, 4.4035, 4.2124, 4.3131, 4.7954, 4.3582, 4.1482], device='cuda:4'), covar=tensor([0.1614, 0.1966, 0.1623, 0.1996, 0.2676, 0.1083, 0.1485, 0.2590], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0477, 0.0512, 0.0415, 0.0548, 0.0541, 0.0417, 0.0566], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 06:50:12,805 INFO [train.py:904] (4/8) Epoch 10, batch 7150, loss[loss=0.2613, simple_loss=0.3385, pruned_loss=0.09204, over 15518.00 frames. ], tot_loss[loss=0.2264, simple_loss=0.3083, pruned_loss=0.07227, over 3115320.87 frames. ], batch size: 191, lr: 6.78e-03, grad_scale: 8.0 2023-04-29 06:50:32,384 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-29 06:51:03,785 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.197e+02 3.117e+02 3.720e+02 4.761e+02 8.896e+02, threshold=7.439e+02, percent-clipped=3.0 2023-04-29 06:51:15,747 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98543.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:51:16,705 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98544.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:51:29,368 INFO [train.py:904] (4/8) Epoch 10, batch 7200, loss[loss=0.2152, simple_loss=0.2954, pruned_loss=0.06752, over 12055.00 frames. ], tot_loss[loss=0.225, simple_loss=0.3067, pruned_loss=0.07161, over 3074402.69 frames. ], batch size: 248, lr: 6.78e-03, grad_scale: 8.0 2023-04-29 06:52:32,650 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=98592.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:52:49,206 INFO [train.py:904] (4/8) Epoch 10, batch 7250, loss[loss=0.2131, simple_loss=0.2929, pruned_loss=0.06669, over 16411.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.3044, pruned_loss=0.07045, over 3075240.91 frames. ], batch size: 146, lr: 6.77e-03, grad_scale: 2.0 2023-04-29 06:52:53,569 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98604.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:53:45,147 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.674e+02 3.052e+02 3.513e+02 4.472e+02 1.147e+03, threshold=7.026e+02, percent-clipped=3.0 2023-04-29 06:54:05,880 INFO [train.py:904] (4/8) Epoch 10, batch 7300, loss[loss=0.2394, simple_loss=0.3156, pruned_loss=0.08155, over 16313.00 frames. ], tot_loss[loss=0.2223, simple_loss=0.3036, pruned_loss=0.07056, over 3066108.34 frames. ], batch size: 165, lr: 6.77e-03, grad_scale: 2.0 2023-04-29 06:55:23,864 INFO [train.py:904] (4/8) Epoch 10, batch 7350, loss[loss=0.1995, simple_loss=0.2895, pruned_loss=0.05475, over 16847.00 frames. ], tot_loss[loss=0.2231, simple_loss=0.3041, pruned_loss=0.07101, over 3055859.16 frames. ], batch size: 102, lr: 6.77e-03, grad_scale: 2.0 2023-04-29 06:56:19,128 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.271e+02 3.294e+02 3.832e+02 4.622e+02 6.263e+02, threshold=7.664e+02, percent-clipped=0.0 2023-04-29 06:56:34,294 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-29 06:56:41,001 INFO [train.py:904] (4/8) Epoch 10, batch 7400, loss[loss=0.2242, simple_loss=0.2986, pruned_loss=0.0749, over 16719.00 frames. ], tot_loss[loss=0.2239, simple_loss=0.3049, pruned_loss=0.07139, over 3065883.29 frames. ], batch size: 57, lr: 6.77e-03, grad_scale: 2.0 2023-04-29 06:56:57,600 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98762.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:57:03,124 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98766.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:57:53,572 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-29 06:57:59,093 INFO [train.py:904] (4/8) Epoch 10, batch 7450, loss[loss=0.2642, simple_loss=0.3415, pruned_loss=0.09344, over 15266.00 frames. ], tot_loss[loss=0.2247, simple_loss=0.3058, pruned_loss=0.07181, over 3079435.79 frames. ], batch size: 190, lr: 6.77e-03, grad_scale: 2.0 2023-04-29 06:58:04,817 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5532, 3.6220, 1.8884, 3.9875, 2.5195, 3.9969, 2.0465, 2.6967], device='cuda:4'), covar=tensor([0.0188, 0.0315, 0.1642, 0.0146, 0.0838, 0.0405, 0.1536, 0.0717], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0161, 0.0186, 0.0117, 0.0164, 0.0202, 0.0191, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 06:58:14,249 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=98810.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 06:58:17,616 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.84 vs. limit=5.0 2023-04-29 06:58:57,290 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.965e+02 3.303e+02 3.942e+02 4.934e+02 1.157e+03, threshold=7.883e+02, percent-clipped=3.0 2023-04-29 06:59:19,899 INFO [train.py:904] (4/8) Epoch 10, batch 7500, loss[loss=0.2197, simple_loss=0.29, pruned_loss=0.07474, over 17205.00 frames. ], tot_loss[loss=0.2244, simple_loss=0.3061, pruned_loss=0.07139, over 3076616.09 frames. ], batch size: 46, lr: 6.77e-03, grad_scale: 2.0 2023-04-29 07:00:20,507 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98890.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:00:27,988 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98895.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:00:34,223 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98899.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:00:38,282 INFO [train.py:904] (4/8) Epoch 10, batch 7550, loss[loss=0.2475, simple_loss=0.3111, pruned_loss=0.09195, over 11220.00 frames. ], tot_loss[loss=0.2236, simple_loss=0.3047, pruned_loss=0.07126, over 3078829.22 frames. ], batch size: 248, lr: 6.76e-03, grad_scale: 2.0 2023-04-29 07:01:16,802 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8577, 3.1753, 3.1816, 2.0282, 2.9919, 3.1293, 3.0633, 1.7671], device='cuda:4'), covar=tensor([0.0463, 0.0037, 0.0043, 0.0365, 0.0071, 0.0088, 0.0064, 0.0376], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0066, 0.0068, 0.0126, 0.0076, 0.0087, 0.0075, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 07:01:32,281 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.425e+02 3.280e+02 3.927e+02 5.001e+02 7.830e+02, threshold=7.854e+02, percent-clipped=0.0 2023-04-29 07:01:36,704 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2642, 4.3031, 4.0994, 3.8800, 3.7940, 4.2006, 4.0179, 3.8913], device='cuda:4'), covar=tensor([0.0576, 0.0394, 0.0267, 0.0275, 0.0862, 0.0438, 0.0509, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0232, 0.0292, 0.0269, 0.0250, 0.0292, 0.0285, 0.0186, 0.0315], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 07:01:51,282 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-29 07:01:53,321 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98951.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:01:53,973 INFO [train.py:904] (4/8) Epoch 10, batch 7600, loss[loss=0.1915, simple_loss=0.2787, pruned_loss=0.05215, over 16695.00 frames. ], tot_loss[loss=0.2227, simple_loss=0.3036, pruned_loss=0.07093, over 3071901.58 frames. ], batch size: 89, lr: 6.76e-03, grad_scale: 4.0 2023-04-29 07:02:01,647 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98956.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:02:13,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5781, 3.5938, 4.0036, 1.9236, 4.2543, 4.2779, 3.0042, 3.1651], device='cuda:4'), covar=tensor([0.0765, 0.0183, 0.0160, 0.1152, 0.0041, 0.0090, 0.0373, 0.0411], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0099, 0.0085, 0.0136, 0.0067, 0.0096, 0.0119, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 07:02:24,328 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98971.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:03:11,998 INFO [train.py:904] (4/8) Epoch 10, batch 7650, loss[loss=0.274, simple_loss=0.3351, pruned_loss=0.1065, over 11313.00 frames. ], tot_loss[loss=0.2247, simple_loss=0.305, pruned_loss=0.07221, over 3063833.31 frames. ], batch size: 247, lr: 6.76e-03, grad_scale: 4.0 2023-04-29 07:03:35,368 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8458, 2.6894, 2.6082, 1.8794, 2.5073, 2.5984, 2.5776, 1.8608], device='cuda:4'), covar=tensor([0.0354, 0.0049, 0.0050, 0.0290, 0.0094, 0.0082, 0.0074, 0.0307], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0066, 0.0069, 0.0126, 0.0076, 0.0087, 0.0075, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 07:03:59,239 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99032.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:04:08,570 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.981e+02 3.374e+02 4.235e+02 5.349e+02 9.496e+02, threshold=8.471e+02, percent-clipped=6.0 2023-04-29 07:04:20,658 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4294, 4.1954, 4.2448, 2.7604, 3.6521, 4.1227, 3.8707, 2.2851], device='cuda:4'), covar=tensor([0.0403, 0.0025, 0.0025, 0.0298, 0.0069, 0.0077, 0.0046, 0.0342], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0066, 0.0068, 0.0125, 0.0076, 0.0087, 0.0074, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 07:04:28,313 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99051.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:04:29,003 INFO [train.py:904] (4/8) Epoch 10, batch 7700, loss[loss=0.2271, simple_loss=0.3206, pruned_loss=0.06682, over 16768.00 frames. ], tot_loss[loss=0.2258, simple_loss=0.3056, pruned_loss=0.07302, over 3054269.21 frames. ], batch size: 89, lr: 6.76e-03, grad_scale: 4.0 2023-04-29 07:04:50,118 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99066.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:05:37,131 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.40 vs. limit=5.0 2023-04-29 07:05:43,695 INFO [train.py:904] (4/8) Epoch 10, batch 7750, loss[loss=0.2672, simple_loss=0.3372, pruned_loss=0.0986, over 11753.00 frames. ], tot_loss[loss=0.2258, simple_loss=0.306, pruned_loss=0.07275, over 3079987.65 frames. ], batch size: 248, lr: 6.76e-03, grad_scale: 4.0 2023-04-29 07:06:00,276 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99112.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:06:02,449 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=99114.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:06:14,843 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7623, 1.7281, 1.5167, 1.4832, 1.8748, 1.6141, 1.7319, 1.9709], device='cuda:4'), covar=tensor([0.0110, 0.0210, 0.0282, 0.0243, 0.0141, 0.0190, 0.0134, 0.0139], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0194, 0.0194, 0.0192, 0.0193, 0.0197, 0.0196, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 07:06:38,055 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.181e+02 3.321e+02 3.885e+02 5.264e+02 1.269e+03, threshold=7.770e+02, percent-clipped=1.0 2023-04-29 07:06:59,327 INFO [train.py:904] (4/8) Epoch 10, batch 7800, loss[loss=0.2646, simple_loss=0.3244, pruned_loss=0.1024, over 11338.00 frames. ], tot_loss[loss=0.2281, simple_loss=0.3075, pruned_loss=0.07437, over 3058666.69 frames. ], batch size: 247, lr: 6.76e-03, grad_scale: 4.0 2023-04-29 07:07:34,829 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.31 vs. limit=5.0 2023-04-29 07:07:44,906 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-29 07:08:03,005 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-29 07:08:12,875 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99199.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:08:16,573 INFO [train.py:904] (4/8) Epoch 10, batch 7850, loss[loss=0.2159, simple_loss=0.3044, pruned_loss=0.06368, over 16692.00 frames. ], tot_loss[loss=0.2269, simple_loss=0.3075, pruned_loss=0.07318, over 3068705.55 frames. ], batch size: 124, lr: 6.75e-03, grad_scale: 4.0 2023-04-29 07:08:57,343 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99229.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:09:10,220 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.079e+02 2.953e+02 3.751e+02 4.670e+02 9.934e+02, threshold=7.502e+02, percent-clipped=3.0 2023-04-29 07:09:22,457 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99246.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:09:24,228 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=99247.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:09:29,661 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99251.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:09:30,436 INFO [train.py:904] (4/8) Epoch 10, batch 7900, loss[loss=0.2061, simple_loss=0.2943, pruned_loss=0.05898, over 16847.00 frames. ], tot_loss[loss=0.2264, simple_loss=0.3069, pruned_loss=0.07291, over 3065651.62 frames. ], batch size: 96, lr: 6.75e-03, grad_scale: 4.0 2023-04-29 07:09:55,563 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99268.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:10:31,501 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99290.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:10:49,021 INFO [train.py:904] (4/8) Epoch 10, batch 7950, loss[loss=0.2251, simple_loss=0.3041, pruned_loss=0.07306, over 15474.00 frames. ], tot_loss[loss=0.2271, simple_loss=0.3077, pruned_loss=0.07326, over 3066703.47 frames. ], batch size: 191, lr: 6.75e-03, grad_scale: 4.0 2023-04-29 07:11:19,077 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-29 07:11:22,413 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99324.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:11:27,241 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99327.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:11:30,499 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99329.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:11:43,785 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.817e+02 3.138e+02 3.909e+02 4.608e+02 7.916e+02, threshold=7.818e+02, percent-clipped=1.0 2023-04-29 07:11:51,037 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2901, 1.9561, 2.0837, 3.8007, 1.9191, 2.4376, 2.0625, 2.1889], device='cuda:4'), covar=tensor([0.0958, 0.3490, 0.2269, 0.0448, 0.4025, 0.2186, 0.3011, 0.3121], device='cuda:4'), in_proj_covar=tensor([0.0350, 0.0382, 0.0318, 0.0318, 0.0409, 0.0432, 0.0343, 0.0446], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 07:12:06,027 INFO [train.py:904] (4/8) Epoch 10, batch 8000, loss[loss=0.2161, simple_loss=0.3084, pruned_loss=0.06187, over 16762.00 frames. ], tot_loss[loss=0.2281, simple_loss=0.3082, pruned_loss=0.07401, over 3074487.86 frames. ], batch size: 89, lr: 6.75e-03, grad_scale: 8.0 2023-04-29 07:12:57,311 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99385.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:13:21,633 INFO [train.py:904] (4/8) Epoch 10, batch 8050, loss[loss=0.2338, simple_loss=0.305, pruned_loss=0.0813, over 11631.00 frames. ], tot_loss[loss=0.2279, simple_loss=0.3083, pruned_loss=0.07382, over 3066704.38 frames. ], batch size: 247, lr: 6.75e-03, grad_scale: 8.0 2023-04-29 07:13:29,868 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99407.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:13:50,420 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99420.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:14:18,153 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.046e+02 3.114e+02 3.682e+02 4.495e+02 9.204e+02, threshold=7.364e+02, percent-clipped=1.0 2023-04-29 07:14:39,575 INFO [train.py:904] (4/8) Epoch 10, batch 8100, loss[loss=0.2035, simple_loss=0.2914, pruned_loss=0.05783, over 16229.00 frames. ], tot_loss[loss=0.2265, simple_loss=0.3073, pruned_loss=0.07284, over 3061502.19 frames. ], batch size: 165, lr: 6.75e-03, grad_scale: 8.0 2023-04-29 07:15:04,018 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99468.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 07:15:23,112 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99481.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:15:54,070 INFO [train.py:904] (4/8) Epoch 10, batch 8150, loss[loss=0.2411, simple_loss=0.306, pruned_loss=0.08813, over 11526.00 frames. ], tot_loss[loss=0.2233, simple_loss=0.3042, pruned_loss=0.07122, over 3075119.59 frames. ], batch size: 247, lr: 6.74e-03, grad_scale: 8.0 2023-04-29 07:16:35,885 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99529.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 07:16:49,708 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.193e+02 2.958e+02 3.724e+02 4.411e+02 7.985e+02, threshold=7.447e+02, percent-clipped=3.0 2023-04-29 07:17:01,404 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99546.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:17:06,089 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-29 07:17:08,967 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99551.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:17:09,684 INFO [train.py:904] (4/8) Epoch 10, batch 8200, loss[loss=0.2267, simple_loss=0.3138, pruned_loss=0.06973, over 16287.00 frames. ], tot_loss[loss=0.2215, simple_loss=0.3019, pruned_loss=0.07059, over 3079358.41 frames. ], batch size: 165, lr: 6.74e-03, grad_scale: 4.0 2023-04-29 07:17:36,119 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99568.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:18:04,664 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99585.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:18:11,137 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99589.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:18:20,129 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=99594.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:18:28,353 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=99599.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:18:32,973 INFO [train.py:904] (4/8) Epoch 10, batch 8250, loss[loss=0.2059, simple_loss=0.2964, pruned_loss=0.05768, over 16857.00 frames. ], tot_loss[loss=0.2186, simple_loss=0.3011, pruned_loss=0.06807, over 3082632.56 frames. ], batch size: 116, lr: 6.74e-03, grad_scale: 4.0 2023-04-29 07:19:10,190 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99624.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:19:16,593 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99627.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:19:19,860 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99629.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:19:36,138 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.952e+02 2.844e+02 3.459e+02 4.111e+02 7.773e+02, threshold=6.919e+02, percent-clipped=2.0 2023-04-29 07:19:55,548 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99650.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:19:57,953 INFO [train.py:904] (4/8) Epoch 10, batch 8300, loss[loss=0.2105, simple_loss=0.2934, pruned_loss=0.06384, over 16627.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.2977, pruned_loss=0.06448, over 3080817.88 frames. ], batch size: 57, lr: 6.74e-03, grad_scale: 4.0 2023-04-29 07:20:37,233 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=99675.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:20:44,554 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99680.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:20:44,665 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99680.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:21:21,641 INFO [train.py:904] (4/8) Epoch 10, batch 8350, loss[loss=0.2086, simple_loss=0.3019, pruned_loss=0.05762, over 15383.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.2975, pruned_loss=0.06286, over 3080401.75 frames. ], batch size: 191, lr: 6.74e-03, grad_scale: 4.0 2023-04-29 07:21:30,529 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99707.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:22:21,849 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.480e+02 2.332e+02 2.884e+02 3.614e+02 8.033e+02, threshold=5.769e+02, percent-clipped=2.0 2023-04-29 07:22:26,064 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99741.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:22:43,903 INFO [train.py:904] (4/8) Epoch 10, batch 8400, loss[loss=0.1856, simple_loss=0.2677, pruned_loss=0.0517, over 12530.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2947, pruned_loss=0.06075, over 3064560.20 frames. ], batch size: 249, lr: 6.74e-03, grad_scale: 8.0 2023-04-29 07:22:49,992 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=99755.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:23:12,214 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6158, 4.7713, 4.9790, 4.8106, 4.7667, 5.3738, 4.9104, 4.6407], device='cuda:4'), covar=tensor([0.0936, 0.1667, 0.1734, 0.1682, 0.2378, 0.0855, 0.1179, 0.2280], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0456, 0.0492, 0.0398, 0.0524, 0.0521, 0.0398, 0.0543], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 07:23:23,287 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8119, 5.1268, 4.8687, 4.8977, 4.6456, 4.5815, 4.5358, 5.1777], device='cuda:4'), covar=tensor([0.0883, 0.0854, 0.0957, 0.0574, 0.0732, 0.0900, 0.1022, 0.0871], device='cuda:4'), in_proj_covar=tensor([0.0501, 0.0628, 0.0523, 0.0434, 0.0391, 0.0417, 0.0526, 0.0478], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 07:23:25,432 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99776.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:24:06,248 INFO [train.py:904] (4/8) Epoch 10, batch 8450, loss[loss=0.1882, simple_loss=0.2797, pruned_loss=0.04832, over 16940.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.292, pruned_loss=0.05864, over 3066470.85 frames. ], batch size: 109, lr: 6.73e-03, grad_scale: 8.0 2023-04-29 07:24:28,570 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8154, 3.6478, 3.9369, 3.6997, 3.8757, 4.2715, 3.9426, 3.6705], device='cuda:4'), covar=tensor([0.1764, 0.2297, 0.1815, 0.2389, 0.2563, 0.1429, 0.1485, 0.2599], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0453, 0.0488, 0.0396, 0.0521, 0.0517, 0.0398, 0.0539], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 07:24:42,099 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99824.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 07:25:06,073 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.564e+02 2.551e+02 3.005e+02 3.693e+02 6.066e+02, threshold=6.011e+02, percent-clipped=2.0 2023-04-29 07:25:25,941 INFO [train.py:904] (4/8) Epoch 10, batch 8500, loss[loss=0.1811, simple_loss=0.2717, pruned_loss=0.0452, over 15268.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2879, pruned_loss=0.05621, over 3068521.50 frames. ], batch size: 190, lr: 6.73e-03, grad_scale: 8.0 2023-04-29 07:26:21,542 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99885.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:26:25,431 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-29 07:26:46,584 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2192, 2.0249, 2.1730, 3.8022, 2.0009, 2.4336, 2.1629, 2.1401], device='cuda:4'), covar=tensor([0.0832, 0.3464, 0.2096, 0.0372, 0.3845, 0.2198, 0.2957, 0.3339], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0371, 0.0308, 0.0304, 0.0394, 0.0416, 0.0331, 0.0429], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 07:26:50,314 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-29 07:26:50,575 INFO [train.py:904] (4/8) Epoch 10, batch 8550, loss[loss=0.1715, simple_loss=0.2752, pruned_loss=0.03389, over 16893.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2853, pruned_loss=0.05477, over 3064326.73 frames. ], batch size: 96, lr: 6.73e-03, grad_scale: 8.0 2023-04-29 07:27:34,859 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99924.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:27:34,901 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99924.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:27:50,547 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=99933.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:28:02,240 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.588e+02 2.540e+02 3.329e+02 3.854e+02 7.309e+02, threshold=6.657e+02, percent-clipped=4.0 2023-04-29 07:28:16,658 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99945.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:28:29,552 INFO [train.py:904] (4/8) Epoch 10, batch 8600, loss[loss=0.181, simple_loss=0.2786, pruned_loss=0.04171, over 15382.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.2858, pruned_loss=0.05401, over 3052100.64 frames. ], batch size: 191, lr: 6.73e-03, grad_scale: 8.0 2023-04-29 07:29:10,991 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=99972.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:29:26,477 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99980.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:30:11,015 INFO [train.py:904] (4/8) Epoch 10, batch 8650, loss[loss=0.1906, simple_loss=0.2705, pruned_loss=0.05534, over 12171.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.283, pruned_loss=0.0521, over 3047757.00 frames. ], batch size: 250, lr: 6.73e-03, grad_scale: 8.0 2023-04-29 07:30:24,829 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=100007.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 07:31:10,429 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=100028.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:31:21,183 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-29 07:31:27,926 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=100036.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:31:31,868 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.647e+02 2.482e+02 2.900e+02 4.130e+02 1.018e+03, threshold=5.800e+02, percent-clipped=4.0 2023-04-29 07:31:56,447 INFO [train.py:904] (4/8) Epoch 10, batch 8700, loss[loss=0.169, simple_loss=0.2606, pruned_loss=0.0387, over 16184.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2798, pruned_loss=0.05066, over 3042270.27 frames. ], batch size: 165, lr: 6.73e-03, grad_scale: 8.0 2023-04-29 07:32:28,550 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=100068.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 07:32:43,274 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100076.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:33:35,354 INFO [train.py:904] (4/8) Epoch 10, batch 8750, loss[loss=0.1914, simple_loss=0.2726, pruned_loss=0.05511, over 12392.00 frames. ], tot_loss[loss=0.189, simple_loss=0.279, pruned_loss=0.04953, over 3043717.59 frames. ], batch size: 248, lr: 6.72e-03, grad_scale: 8.0 2023-04-29 07:34:32,195 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=100124.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:34:32,251 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100124.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 07:34:54,766 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-29 07:35:02,703 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.689e+02 2.530e+02 3.072e+02 4.161e+02 6.852e+02, threshold=6.144e+02, percent-clipped=8.0 2023-04-29 07:35:29,749 INFO [train.py:904] (4/8) Epoch 10, batch 8800, loss[loss=0.1733, simple_loss=0.2727, pruned_loss=0.03696, over 16870.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2771, pruned_loss=0.04831, over 3055135.48 frames. ], batch size: 96, lr: 6.72e-03, grad_scale: 8.0 2023-04-29 07:36:12,283 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=100172.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 07:37:06,154 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0602, 5.0465, 4.8217, 4.4238, 4.8542, 1.7870, 4.6654, 4.7621], device='cuda:4'), covar=tensor([0.0041, 0.0043, 0.0109, 0.0213, 0.0057, 0.2085, 0.0080, 0.0118], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0105, 0.0149, 0.0142, 0.0123, 0.0170, 0.0137, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 07:37:15,875 INFO [train.py:904] (4/8) Epoch 10, batch 8850, loss[loss=0.173, simple_loss=0.2606, pruned_loss=0.04269, over 12656.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2793, pruned_loss=0.04808, over 3040910.19 frames. ], batch size: 247, lr: 6.72e-03, grad_scale: 8.0 2023-04-29 07:38:03,389 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100224.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:38:34,749 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=100238.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:38:36,932 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.257e+02 2.435e+02 3.067e+02 3.598e+02 6.365e+02, threshold=6.134e+02, percent-clipped=1.0 2023-04-29 07:38:48,316 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100245.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:39:01,668 INFO [train.py:904] (4/8) Epoch 10, batch 8900, loss[loss=0.191, simple_loss=0.2848, pruned_loss=0.04862, over 16900.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2792, pruned_loss=0.04718, over 3042765.06 frames. ], batch size: 96, lr: 6.72e-03, grad_scale: 4.0 2023-04-29 07:39:43,457 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=100272.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:40:45,777 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=100293.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:41:01,106 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=100299.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:41:05,909 INFO [train.py:904] (4/8) Epoch 10, batch 8950, loss[loss=0.1632, simple_loss=0.2606, pruned_loss=0.03292, over 16722.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2791, pruned_loss=0.04745, over 3053509.54 frames. ], batch size: 76, lr: 6.72e-03, grad_scale: 4.0 2023-04-29 07:42:08,199 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-29 07:42:22,264 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100336.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:42:29,137 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.659e+02 2.569e+02 2.987e+02 3.833e+02 6.749e+02, threshold=5.974e+02, percent-clipped=2.0 2023-04-29 07:42:55,809 INFO [train.py:904] (4/8) Epoch 10, batch 9000, loss[loss=0.1718, simple_loss=0.2685, pruned_loss=0.03752, over 15338.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2765, pruned_loss=0.04586, over 3072033.34 frames. ], batch size: 191, lr: 6.71e-03, grad_scale: 4.0 2023-04-29 07:42:55,809 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 07:43:05,469 INFO [train.py:938] (4/8) Epoch 10, validation: loss=0.1565, simple_loss=0.2604, pruned_loss=0.02634, over 944034.00 frames. 2023-04-29 07:43:05,470 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 07:43:30,279 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=100363.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 07:44:12,991 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=100384.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:44:50,483 INFO [train.py:904] (4/8) Epoch 10, batch 9050, loss[loss=0.1673, simple_loss=0.2536, pruned_loss=0.04055, over 16109.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2769, pruned_loss=0.04625, over 3085860.31 frames. ], batch size: 165, lr: 6.71e-03, grad_scale: 4.0 2023-04-29 07:44:52,156 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=100402.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:45:24,132 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3592, 3.4155, 1.8237, 3.6383, 2.4247, 3.6161, 2.0814, 2.7886], device='cuda:4'), covar=tensor([0.0190, 0.0280, 0.1565, 0.0149, 0.0810, 0.0435, 0.1388, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0151, 0.0178, 0.0113, 0.0157, 0.0188, 0.0186, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-29 07:46:08,563 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.813e+02 2.533e+02 2.923e+02 3.957e+02 1.423e+03, threshold=5.846e+02, percent-clipped=5.0 2023-04-29 07:46:20,883 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2103, 4.0098, 4.2338, 4.4047, 4.4844, 4.0848, 4.5284, 4.5502], device='cuda:4'), covar=tensor([0.1318, 0.0973, 0.1258, 0.0557, 0.0536, 0.1009, 0.0504, 0.0591], device='cuda:4'), in_proj_covar=tensor([0.0471, 0.0589, 0.0705, 0.0607, 0.0460, 0.0459, 0.0478, 0.0540], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 07:46:37,386 INFO [train.py:904] (4/8) Epoch 10, batch 9100, loss[loss=0.2029, simple_loss=0.2979, pruned_loss=0.05392, over 15246.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2765, pruned_loss=0.04675, over 3089902.94 frames. ], batch size: 190, lr: 6.71e-03, grad_scale: 4.0 2023-04-29 07:46:58,867 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=100463.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 07:47:55,558 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2124, 3.6505, 3.8591, 2.0125, 3.2362, 2.4687, 3.5574, 3.7887], device='cuda:4'), covar=tensor([0.0275, 0.0735, 0.0447, 0.1786, 0.0671, 0.0870, 0.0694, 0.0883], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0133, 0.0152, 0.0138, 0.0130, 0.0122, 0.0132, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 07:48:00,924 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-29 07:48:34,507 INFO [train.py:904] (4/8) Epoch 10, batch 9150, loss[loss=0.1604, simple_loss=0.2547, pruned_loss=0.033, over 16692.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.2768, pruned_loss=0.04641, over 3084902.53 frames. ], batch size: 57, lr: 6.71e-03, grad_scale: 4.0 2023-04-29 07:49:54,593 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.562e+02 2.281e+02 2.663e+02 3.521e+02 5.779e+02, threshold=5.326e+02, percent-clipped=0.0 2023-04-29 07:50:13,980 INFO [train.py:904] (4/8) Epoch 10, batch 9200, loss[loss=0.1974, simple_loss=0.295, pruned_loss=0.04992, over 15236.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2725, pruned_loss=0.04552, over 3073672.11 frames. ], batch size: 190, lr: 6.71e-03, grad_scale: 8.0 2023-04-29 07:51:31,221 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.34 vs. limit=5.0 2023-04-29 07:51:35,108 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=100594.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 07:51:50,957 INFO [train.py:904] (4/8) Epoch 10, batch 9250, loss[loss=0.1835, simple_loss=0.2661, pruned_loss=0.05048, over 12535.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.272, pruned_loss=0.04533, over 3079247.61 frames. ], batch size: 248, lr: 6.71e-03, grad_scale: 8.0 2023-04-29 07:53:14,008 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.494e+02 2.424e+02 2.858e+02 3.550e+02 5.723e+02, threshold=5.716e+02, percent-clipped=3.0 2023-04-29 07:53:42,536 INFO [train.py:904] (4/8) Epoch 10, batch 9300, loss[loss=0.1555, simple_loss=0.2432, pruned_loss=0.03388, over 16597.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2704, pruned_loss=0.04504, over 3044640.78 frames. ], batch size: 62, lr: 6.70e-03, grad_scale: 8.0 2023-04-29 07:54:09,130 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100663.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 07:55:29,731 INFO [train.py:904] (4/8) Epoch 10, batch 9350, loss[loss=0.2058, simple_loss=0.2949, pruned_loss=0.05838, over 15446.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2702, pruned_loss=0.045, over 3052192.46 frames. ], batch size: 191, lr: 6.70e-03, grad_scale: 8.0 2023-04-29 07:55:50,102 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=100711.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 07:56:10,100 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.29 vs. limit=5.0 2023-04-29 07:56:11,547 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-29 07:56:29,110 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=100730.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 07:56:42,755 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6601, 2.0530, 1.8022, 1.9009, 2.4444, 2.1029, 2.4312, 2.6142], device='cuda:4'), covar=tensor([0.0092, 0.0333, 0.0399, 0.0343, 0.0200, 0.0279, 0.0151, 0.0176], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0190, 0.0188, 0.0185, 0.0187, 0.0191, 0.0184, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 07:56:48,376 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.720e+02 2.395e+02 2.806e+02 3.199e+02 7.223e+02, threshold=5.612e+02, percent-clipped=2.0 2023-04-29 07:57:12,447 INFO [train.py:904] (4/8) Epoch 10, batch 9400, loss[loss=0.1905, simple_loss=0.2873, pruned_loss=0.0468, over 15518.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2713, pruned_loss=0.04505, over 3064278.34 frames. ], batch size: 191, lr: 6.70e-03, grad_scale: 8.0 2023-04-29 07:57:25,167 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=100758.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 07:58:33,052 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=100791.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 07:58:33,317 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 07:58:52,903 INFO [train.py:904] (4/8) Epoch 10, batch 9450, loss[loss=0.1902, simple_loss=0.2711, pruned_loss=0.05466, over 12460.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2727, pruned_loss=0.04515, over 3048954.31 frames. ], batch size: 247, lr: 6.70e-03, grad_scale: 8.0 2023-04-29 08:00:10,871 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 2.411e+02 3.092e+02 3.645e+02 7.379e+02, threshold=6.183e+02, percent-clipped=4.0 2023-04-29 08:00:34,635 INFO [train.py:904] (4/8) Epoch 10, batch 9500, loss[loss=0.1785, simple_loss=0.2705, pruned_loss=0.04324, over 15279.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.272, pruned_loss=0.04474, over 3059957.49 frames. ], batch size: 191, lr: 6.70e-03, grad_scale: 8.0 2023-04-29 08:00:41,477 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2588, 5.8464, 6.0652, 5.8207, 5.8991, 6.3733, 5.9116, 5.6607], device='cuda:4'), covar=tensor([0.0612, 0.1426, 0.1432, 0.1873, 0.1959, 0.0772, 0.1114, 0.2082], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0439, 0.0468, 0.0382, 0.0501, 0.0502, 0.0381, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:01:20,198 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7376, 1.7720, 2.0616, 2.6938, 2.5487, 2.8836, 1.8160, 2.8602], device='cuda:4'), covar=tensor([0.0124, 0.0343, 0.0253, 0.0187, 0.0204, 0.0127, 0.0342, 0.0110], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0166, 0.0148, 0.0150, 0.0162, 0.0114, 0.0166, 0.0106], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 08:01:56,297 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-29 08:02:03,483 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100894.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:02:20,731 INFO [train.py:904] (4/8) Epoch 10, batch 9550, loss[loss=0.1844, simple_loss=0.2669, pruned_loss=0.05092, over 12543.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2724, pruned_loss=0.0453, over 3069546.87 frames. ], batch size: 248, lr: 6.70e-03, grad_scale: 8.0 2023-04-29 08:03:40,249 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.820e+02 2.447e+02 2.728e+02 3.779e+02 7.191e+02, threshold=5.455e+02, percent-clipped=3.0 2023-04-29 08:03:45,437 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=100942.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:04:02,626 INFO [train.py:904] (4/8) Epoch 10, batch 9600, loss[loss=0.2128, simple_loss=0.3132, pruned_loss=0.05618, over 16705.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.2747, pruned_loss=0.04681, over 3052564.03 frames. ], batch size: 134, lr: 6.70e-03, grad_scale: 8.0 2023-04-29 08:04:57,883 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4783, 4.2870, 4.5296, 4.6938, 4.8409, 4.3061, 4.8435, 4.8067], device='cuda:4'), covar=tensor([0.1690, 0.1160, 0.1581, 0.0669, 0.0547, 0.0831, 0.0467, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0471, 0.0592, 0.0711, 0.0608, 0.0459, 0.0462, 0.0477, 0.0537], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:05:25,233 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5447, 4.3051, 4.5368, 4.7416, 4.9025, 4.3444, 4.9143, 4.8460], device='cuda:4'), covar=tensor([0.1441, 0.1188, 0.1642, 0.0654, 0.0495, 0.0853, 0.0401, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0470, 0.0592, 0.0711, 0.0608, 0.0458, 0.0461, 0.0476, 0.0537], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:05:52,878 INFO [train.py:904] (4/8) Epoch 10, batch 9650, loss[loss=0.1882, simple_loss=0.2691, pruned_loss=0.05361, over 12225.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.2758, pruned_loss=0.04651, over 3059113.25 frames. ], batch size: 250, lr: 6.69e-03, grad_scale: 8.0 2023-04-29 08:06:19,642 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1164, 1.4485, 1.7581, 2.0667, 2.1249, 2.2174, 1.6658, 2.0971], device='cuda:4'), covar=tensor([0.0162, 0.0334, 0.0185, 0.0188, 0.0188, 0.0159, 0.0335, 0.0088], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0164, 0.0147, 0.0148, 0.0159, 0.0113, 0.0165, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 08:06:36,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7697, 3.7058, 3.8599, 3.9852, 4.0541, 3.5927, 4.0184, 4.0710], device='cuda:4'), covar=tensor([0.1564, 0.0963, 0.1355, 0.0611, 0.0546, 0.1742, 0.0621, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0471, 0.0592, 0.0711, 0.0607, 0.0458, 0.0461, 0.0475, 0.0537], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:06:43,735 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-29 08:06:57,402 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5127, 4.0387, 4.0693, 2.8615, 3.6886, 4.0787, 3.8366, 2.1839], device='cuda:4'), covar=tensor([0.0387, 0.0022, 0.0026, 0.0304, 0.0064, 0.0046, 0.0044, 0.0390], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0062, 0.0065, 0.0120, 0.0074, 0.0082, 0.0071, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:07:15,510 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.746e+02 2.568e+02 3.331e+02 4.260e+02 1.013e+03, threshold=6.663e+02, percent-clipped=7.0 2023-04-29 08:07:41,460 INFO [train.py:904] (4/8) Epoch 10, batch 9700, loss[loss=0.1804, simple_loss=0.2634, pruned_loss=0.04867, over 12649.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2744, pruned_loss=0.04638, over 3046939.02 frames. ], batch size: 248, lr: 6.69e-03, grad_scale: 8.0 2023-04-29 08:07:52,537 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=101058.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:08:54,359 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=101086.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 08:09:25,195 INFO [train.py:904] (4/8) Epoch 10, batch 9750, loss[loss=0.2197, simple_loss=0.313, pruned_loss=0.06319, over 16972.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2738, pruned_loss=0.04675, over 3041015.84 frames. ], batch size: 109, lr: 6.69e-03, grad_scale: 8.0 2023-04-29 08:09:32,958 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=101106.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:09:34,805 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=101107.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:10:41,636 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-29 08:10:45,034 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.560e+02 2.458e+02 3.043e+02 3.868e+02 6.520e+02, threshold=6.087e+02, percent-clipped=0.0 2023-04-29 08:11:02,833 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-29 08:11:05,029 INFO [train.py:904] (4/8) Epoch 10, batch 9800, loss[loss=0.1795, simple_loss=0.2791, pruned_loss=0.04, over 16811.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2737, pruned_loss=0.04547, over 3060631.04 frames. ], batch size: 76, lr: 6.69e-03, grad_scale: 4.0 2023-04-29 08:11:36,086 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=101168.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:11:48,032 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5210, 3.5366, 3.4418, 2.9544, 3.4796, 1.9728, 3.2455, 2.9375], device='cuda:4'), covar=tensor([0.0096, 0.0082, 0.0114, 0.0165, 0.0067, 0.1922, 0.0098, 0.0173], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0106, 0.0148, 0.0137, 0.0122, 0.0170, 0.0136, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-29 08:12:49,194 INFO [train.py:904] (4/8) Epoch 10, batch 9850, loss[loss=0.1905, simple_loss=0.2797, pruned_loss=0.05062, over 15390.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2752, pruned_loss=0.04496, over 3065298.61 frames. ], batch size: 191, lr: 6.69e-03, grad_scale: 4.0 2023-04-29 08:13:54,817 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6982, 1.9896, 1.4849, 1.6533, 2.2964, 2.0076, 2.5978, 2.6219], device='cuda:4'), covar=tensor([0.0100, 0.0379, 0.0494, 0.0455, 0.0241, 0.0364, 0.0155, 0.0193], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0195, 0.0194, 0.0192, 0.0193, 0.0195, 0.0188, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:14:17,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.688e+02 2.461e+02 2.965e+02 3.704e+02 9.099e+02, threshold=5.931e+02, percent-clipped=1.0 2023-04-29 08:14:41,618 INFO [train.py:904] (4/8) Epoch 10, batch 9900, loss[loss=0.184, simple_loss=0.2807, pruned_loss=0.0437, over 16625.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2758, pruned_loss=0.04536, over 3056541.16 frames. ], batch size: 76, lr: 6.69e-03, grad_scale: 4.0 2023-04-29 08:15:39,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0960, 1.3811, 1.7452, 2.0727, 2.0924, 2.2670, 1.7095, 2.2084], device='cuda:4'), covar=tensor([0.0146, 0.0342, 0.0210, 0.0216, 0.0213, 0.0148, 0.0318, 0.0093], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0165, 0.0148, 0.0150, 0.0162, 0.0114, 0.0167, 0.0105], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 08:16:40,041 INFO [train.py:904] (4/8) Epoch 10, batch 9950, loss[loss=0.2218, simple_loss=0.3021, pruned_loss=0.07074, over 12550.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2779, pruned_loss=0.04571, over 3054569.11 frames. ], batch size: 248, lr: 6.68e-03, grad_scale: 4.0 2023-04-29 08:17:43,529 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1952, 3.6045, 3.5331, 2.4926, 3.3223, 3.5958, 3.4940, 2.0470], device='cuda:4'), covar=tensor([0.0394, 0.0024, 0.0031, 0.0270, 0.0062, 0.0050, 0.0040, 0.0362], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0062, 0.0065, 0.0120, 0.0074, 0.0081, 0.0072, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:18:13,666 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.784e+02 2.620e+02 2.929e+02 3.358e+02 7.407e+02, threshold=5.858e+02, percent-clipped=2.0 2023-04-29 08:18:17,739 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4049, 5.4268, 5.1106, 4.8263, 5.1796, 1.7669, 4.9778, 5.0953], device='cuda:4'), covar=tensor([0.0047, 0.0044, 0.0123, 0.0198, 0.0063, 0.2292, 0.0077, 0.0135], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0108, 0.0151, 0.0138, 0.0124, 0.0173, 0.0138, 0.0140], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:18:41,458 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4756, 3.5346, 3.8417, 1.6917, 4.0366, 4.1258, 3.0545, 3.0672], device='cuda:4'), covar=tensor([0.0724, 0.0190, 0.0166, 0.1266, 0.0043, 0.0079, 0.0320, 0.0381], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0096, 0.0081, 0.0136, 0.0065, 0.0093, 0.0115, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0003], device='cuda:4') 2023-04-29 08:18:42,237 INFO [train.py:904] (4/8) Epoch 10, batch 10000, loss[loss=0.1683, simple_loss=0.2722, pruned_loss=0.03218, over 16882.00 frames. ], tot_loss[loss=0.182, simple_loss=0.2753, pruned_loss=0.0444, over 3082792.90 frames. ], batch size: 102, lr: 6.68e-03, grad_scale: 8.0 2023-04-29 08:19:19,424 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7020, 1.7610, 2.0259, 2.6401, 2.4370, 2.8095, 1.8927, 2.7916], device='cuda:4'), covar=tensor([0.0143, 0.0342, 0.0261, 0.0188, 0.0225, 0.0144, 0.0353, 0.0090], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0164, 0.0148, 0.0149, 0.0162, 0.0113, 0.0167, 0.0104], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 08:19:54,045 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=101386.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 08:20:23,602 INFO [train.py:904] (4/8) Epoch 10, batch 10050, loss[loss=0.1981, simple_loss=0.2878, pruned_loss=0.05416, over 16524.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2755, pruned_loss=0.04444, over 3077478.75 frames. ], batch size: 62, lr: 6.68e-03, grad_scale: 8.0 2023-04-29 08:21:25,328 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=101434.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 08:21:36,391 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.448e+02 2.853e+02 3.607e+02 8.482e+02, threshold=5.706e+02, percent-clipped=4.0 2023-04-29 08:21:56,134 INFO [train.py:904] (4/8) Epoch 10, batch 10100, loss[loss=0.1924, simple_loss=0.2851, pruned_loss=0.04981, over 16440.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2756, pruned_loss=0.04453, over 3092688.38 frames. ], batch size: 35, lr: 6.68e-03, grad_scale: 8.0 2023-04-29 08:22:16,005 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=101463.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:23:38,433 INFO [train.py:904] (4/8) Epoch 11, batch 0, loss[loss=0.2, simple_loss=0.2848, pruned_loss=0.05755, over 17252.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2848, pruned_loss=0.05755, over 17252.00 frames. ], batch size: 52, lr: 6.37e-03, grad_scale: 8.0 2023-04-29 08:23:38,433 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 08:23:45,826 INFO [train.py:938] (4/8) Epoch 11, validation: loss=0.1557, simple_loss=0.2595, pruned_loss=0.02591, over 944034.00 frames. 2023-04-29 08:23:45,827 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 08:24:43,125 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.588e+02 2.573e+02 2.991e+02 3.932e+02 8.479e+02, threshold=5.981e+02, percent-clipped=4.0 2023-04-29 08:24:55,123 INFO [train.py:904] (4/8) Epoch 11, batch 50, loss[loss=0.2002, simple_loss=0.2782, pruned_loss=0.0611, over 16442.00 frames. ], tot_loss[loss=0.2084, simple_loss=0.2886, pruned_loss=0.06411, over 758387.62 frames. ], batch size: 75, lr: 6.37e-03, grad_scale: 1.0 2023-04-29 08:25:43,972 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 08:25:54,042 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.92 vs. limit=5.0 2023-04-29 08:26:05,597 INFO [train.py:904] (4/8) Epoch 11, batch 100, loss[loss=0.2044, simple_loss=0.2931, pruned_loss=0.05787, over 17021.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2823, pruned_loss=0.05998, over 1330502.50 frames. ], batch size: 55, lr: 6.37e-03, grad_scale: 1.0 2023-04-29 08:26:12,026 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5526, 3.5422, 3.7855, 2.6701, 3.6001, 3.7847, 3.6681, 2.0379], device='cuda:4'), covar=tensor([0.0399, 0.0170, 0.0051, 0.0315, 0.0081, 0.0097, 0.0076, 0.0464], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0066, 0.0067, 0.0124, 0.0076, 0.0084, 0.0074, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:26:45,512 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.02 vs. limit=2.0 2023-04-29 08:27:03,339 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.860e+02 2.523e+02 3.127e+02 3.796e+02 9.987e+02, threshold=6.254e+02, percent-clipped=2.0 2023-04-29 08:27:12,713 INFO [train.py:904] (4/8) Epoch 11, batch 150, loss[loss=0.1868, simple_loss=0.278, pruned_loss=0.04778, over 17113.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.2785, pruned_loss=0.0577, over 1782613.85 frames. ], batch size: 49, lr: 6.37e-03, grad_scale: 1.0 2023-04-29 08:27:52,230 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7832, 2.9534, 2.6206, 4.4815, 3.7193, 4.2948, 1.6119, 3.1139], device='cuda:4'), covar=tensor([0.1356, 0.0640, 0.1050, 0.0180, 0.0260, 0.0406, 0.1469, 0.0761], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0155, 0.0179, 0.0134, 0.0185, 0.0203, 0.0179, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 08:27:58,585 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=101684.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:28:23,285 INFO [train.py:904] (4/8) Epoch 11, batch 200, loss[loss=0.1898, simple_loss=0.2817, pruned_loss=0.04895, over 16664.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2789, pruned_loss=0.05816, over 2114520.25 frames. ], batch size: 62, lr: 6.37e-03, grad_scale: 1.0 2023-04-29 08:29:21,756 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.429e+02 2.923e+02 3.406e+02 5.474e+02, threshold=5.846e+02, percent-clipped=1.0 2023-04-29 08:29:23,273 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=101745.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:29:31,793 INFO [train.py:904] (4/8) Epoch 11, batch 250, loss[loss=0.2073, simple_loss=0.2765, pruned_loss=0.06909, over 16784.00 frames. ], tot_loss[loss=0.1954, simple_loss=0.2761, pruned_loss=0.05739, over 2386813.99 frames. ], batch size: 124, lr: 6.37e-03, grad_scale: 1.0 2023-04-29 08:29:46,113 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=101763.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:30:37,961 INFO [train.py:904] (4/8) Epoch 11, batch 300, loss[loss=0.1963, simple_loss=0.2781, pruned_loss=0.0572, over 15256.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.273, pruned_loss=0.05587, over 2588718.00 frames. ], batch size: 190, lr: 6.37e-03, grad_scale: 1.0 2023-04-29 08:30:51,057 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=101811.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:30:58,055 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0544, 5.7480, 5.9718, 5.6682, 5.7277, 6.2420, 5.8491, 5.5531], device='cuda:4'), covar=tensor([0.0987, 0.1751, 0.1616, 0.2060, 0.2858, 0.1018, 0.1471, 0.2341], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0481, 0.0516, 0.0416, 0.0554, 0.0546, 0.0413, 0.0561], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:31:18,094 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3855, 2.8531, 2.6033, 2.2610, 2.2627, 2.2293, 2.9691, 2.7764], device='cuda:4'), covar=tensor([0.2197, 0.0699, 0.1333, 0.1859, 0.2013, 0.1750, 0.0480, 0.1119], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0252, 0.0280, 0.0271, 0.0268, 0.0218, 0.0263, 0.0289], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:31:35,541 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.560e+02 2.476e+02 3.003e+02 3.668e+02 8.355e+02, threshold=6.006e+02, percent-clipped=2.0 2023-04-29 08:31:47,602 INFO [train.py:904] (4/8) Epoch 11, batch 350, loss[loss=0.1847, simple_loss=0.2755, pruned_loss=0.04694, over 16695.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2711, pruned_loss=0.05493, over 2751086.67 frames. ], batch size: 62, lr: 6.36e-03, grad_scale: 1.0 2023-04-29 08:32:01,971 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0029, 1.8119, 2.4728, 2.8138, 2.6643, 3.1959, 2.2297, 3.1274], device='cuda:4'), covar=tensor([0.0149, 0.0347, 0.0221, 0.0195, 0.0213, 0.0136, 0.0299, 0.0096], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0170, 0.0154, 0.0156, 0.0168, 0.0120, 0.0171, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 08:32:42,832 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1601, 4.1200, 4.5176, 4.5353, 4.6109, 4.2406, 4.2622, 4.1271], device='cuda:4'), covar=tensor([0.0332, 0.0637, 0.0459, 0.0442, 0.0431, 0.0392, 0.0828, 0.0510], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0328, 0.0332, 0.0311, 0.0373, 0.0346, 0.0444, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 08:32:56,331 INFO [train.py:904] (4/8) Epoch 11, batch 400, loss[loss=0.2031, simple_loss=0.2748, pruned_loss=0.06574, over 16867.00 frames. ], tot_loss[loss=0.1892, simple_loss=0.2697, pruned_loss=0.05432, over 2884745.83 frames. ], batch size: 96, lr: 6.36e-03, grad_scale: 2.0 2023-04-29 08:32:56,807 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7919, 2.4889, 1.9011, 2.2845, 2.8952, 2.6920, 3.0543, 3.0263], device='cuda:4'), covar=tensor([0.0151, 0.0266, 0.0375, 0.0320, 0.0147, 0.0235, 0.0176, 0.0156], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0201, 0.0199, 0.0197, 0.0200, 0.0202, 0.0201, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:33:22,047 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=101920.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 08:33:54,606 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.501e+02 2.298e+02 2.716e+02 3.213e+02 6.833e+02, threshold=5.433e+02, percent-clipped=1.0 2023-04-29 08:34:06,160 INFO [train.py:904] (4/8) Epoch 11, batch 450, loss[loss=0.1699, simple_loss=0.2624, pruned_loss=0.03869, over 17136.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.268, pruned_loss=0.0534, over 2986081.83 frames. ], batch size: 48, lr: 6.36e-03, grad_scale: 2.0 2023-04-29 08:34:46,985 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4030, 3.4714, 1.7428, 3.5718, 2.5498, 3.6800, 1.9009, 2.7561], device='cuda:4'), covar=tensor([0.0237, 0.0372, 0.1653, 0.0291, 0.0775, 0.0560, 0.1473, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0163, 0.0188, 0.0127, 0.0168, 0.0204, 0.0197, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 08:34:47,002 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=101981.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 08:34:54,621 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=101987.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:35:19,214 INFO [train.py:904] (4/8) Epoch 11, batch 500, loss[loss=0.171, simple_loss=0.2541, pruned_loss=0.04393, over 16492.00 frames. ], tot_loss[loss=0.1862, simple_loss=0.2672, pruned_loss=0.05256, over 3061478.78 frames. ], batch size: 68, lr: 6.36e-03, grad_scale: 2.0 2023-04-29 08:35:36,851 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2259, 3.0862, 3.3959, 2.4486, 3.1939, 3.4920, 3.3053, 1.9467], device='cuda:4'), covar=tensor([0.0386, 0.0112, 0.0044, 0.0285, 0.0074, 0.0074, 0.0069, 0.0370], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0068, 0.0069, 0.0125, 0.0077, 0.0086, 0.0076, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:36:13,474 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=102040.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:36:19,093 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.707e+02 2.413e+02 2.842e+02 3.807e+02 1.018e+03, threshold=5.684e+02, percent-clipped=5.0 2023-04-29 08:36:24,817 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=102048.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:36:29,127 INFO [train.py:904] (4/8) Epoch 11, batch 550, loss[loss=0.1979, simple_loss=0.2684, pruned_loss=0.06363, over 12283.00 frames. ], tot_loss[loss=0.186, simple_loss=0.267, pruned_loss=0.05247, over 3117671.04 frames. ], batch size: 246, lr: 6.36e-03, grad_scale: 2.0 2023-04-29 08:37:40,168 INFO [train.py:904] (4/8) Epoch 11, batch 600, loss[loss=0.1533, simple_loss=0.2362, pruned_loss=0.03523, over 17067.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.2654, pruned_loss=0.05155, over 3159830.14 frames. ], batch size: 41, lr: 6.36e-03, grad_scale: 2.0 2023-04-29 08:37:46,065 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0293, 1.7888, 2.4691, 2.9517, 2.7699, 3.2489, 2.0450, 3.1497], device='cuda:4'), covar=tensor([0.0163, 0.0358, 0.0225, 0.0226, 0.0210, 0.0130, 0.0336, 0.0149], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0170, 0.0154, 0.0157, 0.0168, 0.0121, 0.0171, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 08:38:38,909 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.551e+02 2.665e+02 3.250e+02 4.151e+02 1.142e+03, threshold=6.501e+02, percent-clipped=8.0 2023-04-29 08:38:49,712 INFO [train.py:904] (4/8) Epoch 11, batch 650, loss[loss=0.154, simple_loss=0.2289, pruned_loss=0.03953, over 16490.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2643, pruned_loss=0.05146, over 3194339.80 frames. ], batch size: 75, lr: 6.35e-03, grad_scale: 2.0 2023-04-29 08:39:16,550 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8738, 4.3970, 4.2455, 3.1295, 3.9220, 4.3164, 3.9763, 2.2579], device='cuda:4'), covar=tensor([0.0409, 0.0056, 0.0052, 0.0320, 0.0086, 0.0098, 0.0100, 0.0506], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0069, 0.0070, 0.0126, 0.0078, 0.0087, 0.0077, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:39:45,190 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5832, 5.9629, 5.7567, 5.7860, 5.2965, 5.4120, 5.4307, 6.1134], device='cuda:4'), covar=tensor([0.1194, 0.1031, 0.0967, 0.0767, 0.0886, 0.0640, 0.1016, 0.0854], device='cuda:4'), in_proj_covar=tensor([0.0546, 0.0679, 0.0563, 0.0474, 0.0431, 0.0444, 0.0573, 0.0519], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:39:58,899 INFO [train.py:904] (4/8) Epoch 11, batch 700, loss[loss=0.1786, simple_loss=0.2549, pruned_loss=0.05113, over 16870.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.264, pruned_loss=0.05086, over 3228741.21 frames. ], batch size: 96, lr: 6.35e-03, grad_scale: 2.0 2023-04-29 08:40:16,180 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-29 08:40:57,196 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.569e+02 2.406e+02 3.042e+02 3.688e+02 8.109e+02, threshold=6.084e+02, percent-clipped=2.0 2023-04-29 08:41:09,233 INFO [train.py:904] (4/8) Epoch 11, batch 750, loss[loss=0.1936, simple_loss=0.2607, pruned_loss=0.06321, over 16850.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2641, pruned_loss=0.05033, over 3246837.09 frames. ], batch size: 116, lr: 6.35e-03, grad_scale: 2.0 2023-04-29 08:41:42,359 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=102276.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 08:42:18,047 INFO [train.py:904] (4/8) Epoch 11, batch 800, loss[loss=0.1852, simple_loss=0.2523, pruned_loss=0.05906, over 16915.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2643, pruned_loss=0.05026, over 3271218.84 frames. ], batch size: 109, lr: 6.35e-03, grad_scale: 4.0 2023-04-29 08:42:26,375 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.81 vs. limit=5.0 2023-04-29 08:43:11,882 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=102340.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:43:15,336 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=102343.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:43:16,115 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.552e+02 2.524e+02 3.019e+02 3.517e+02 6.392e+02, threshold=6.037e+02, percent-clipped=1.0 2023-04-29 08:43:27,548 INFO [train.py:904] (4/8) Epoch 11, batch 850, loss[loss=0.1733, simple_loss=0.2603, pruned_loss=0.04313, over 16828.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2631, pruned_loss=0.04973, over 3282332.77 frames. ], batch size: 42, lr: 6.35e-03, grad_scale: 4.0 2023-04-29 08:44:17,573 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=102388.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:44:31,807 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8383, 4.1283, 4.3087, 3.0942, 3.7238, 4.2285, 3.8633, 2.4944], device='cuda:4'), covar=tensor([0.0368, 0.0059, 0.0033, 0.0306, 0.0081, 0.0080, 0.0068, 0.0363], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0071, 0.0071, 0.0128, 0.0079, 0.0089, 0.0078, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:44:37,481 INFO [train.py:904] (4/8) Epoch 11, batch 900, loss[loss=0.2306, simple_loss=0.2978, pruned_loss=0.08174, over 16481.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2634, pruned_loss=0.04952, over 3296900.51 frames. ], batch size: 146, lr: 6.35e-03, grad_scale: 4.0 2023-04-29 08:45:35,217 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.680e+02 2.314e+02 2.859e+02 3.467e+02 6.346e+02, threshold=5.718e+02, percent-clipped=1.0 2023-04-29 08:45:45,385 INFO [train.py:904] (4/8) Epoch 11, batch 950, loss[loss=0.1746, simple_loss=0.26, pruned_loss=0.04462, over 17227.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2636, pruned_loss=0.04978, over 3286954.93 frames. ], batch size: 44, lr: 6.34e-03, grad_scale: 4.0 2023-04-29 08:46:54,282 INFO [train.py:904] (4/8) Epoch 11, batch 1000, loss[loss=0.1501, simple_loss=0.2372, pruned_loss=0.03146, over 16847.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.263, pruned_loss=0.04978, over 3302909.79 frames. ], batch size: 42, lr: 6.34e-03, grad_scale: 4.0 2023-04-29 08:47:00,259 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7044, 2.7735, 2.3774, 3.9471, 3.0764, 3.9670, 1.5076, 2.8004], device='cuda:4'), covar=tensor([0.1439, 0.0620, 0.1153, 0.0163, 0.0200, 0.0375, 0.1510, 0.0838], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0157, 0.0180, 0.0139, 0.0193, 0.0209, 0.0180, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 08:47:52,029 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.689e+02 2.460e+02 2.841e+02 3.268e+02 6.263e+02, threshold=5.683e+02, percent-clipped=1.0 2023-04-29 08:48:02,250 INFO [train.py:904] (4/8) Epoch 11, batch 1050, loss[loss=0.1965, simple_loss=0.2621, pruned_loss=0.06548, over 12274.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2628, pruned_loss=0.05049, over 3298589.07 frames. ], batch size: 246, lr: 6.34e-03, grad_scale: 4.0 2023-04-29 08:48:36,344 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=102576.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 08:49:12,605 INFO [train.py:904] (4/8) Epoch 11, batch 1100, loss[loss=0.1733, simple_loss=0.2525, pruned_loss=0.04707, over 15575.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2619, pruned_loss=0.04978, over 3308568.60 frames. ], batch size: 190, lr: 6.34e-03, grad_scale: 4.0 2023-04-29 08:49:22,688 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4013, 3.3570, 3.7112, 2.5117, 3.3594, 3.6772, 3.4789, 2.0533], device='cuda:4'), covar=tensor([0.0390, 0.0136, 0.0040, 0.0297, 0.0080, 0.0088, 0.0065, 0.0368], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0070, 0.0070, 0.0126, 0.0078, 0.0088, 0.0077, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:49:26,749 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.30 vs. limit=5.0 2023-04-29 08:49:43,774 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=102624.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 08:49:52,278 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.56 vs. limit=2.0 2023-04-29 08:50:07,512 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=102643.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:50:09,057 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.567e+02 2.199e+02 2.657e+02 3.494e+02 5.570e+02, threshold=5.313e+02, percent-clipped=0.0 2023-04-29 08:50:20,438 INFO [train.py:904] (4/8) Epoch 11, batch 1150, loss[loss=0.1721, simple_loss=0.2629, pruned_loss=0.04067, over 17188.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2622, pruned_loss=0.04903, over 3314820.56 frames. ], batch size: 44, lr: 6.34e-03, grad_scale: 4.0 2023-04-29 08:50:49,926 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1914, 5.2706, 5.0056, 4.7120, 4.1357, 5.2376, 5.2783, 4.6971], device='cuda:4'), covar=tensor([0.0832, 0.0413, 0.0453, 0.0354, 0.2019, 0.0457, 0.0273, 0.0679], device='cuda:4'), in_proj_covar=tensor([0.0256, 0.0318, 0.0299, 0.0278, 0.0322, 0.0318, 0.0204, 0.0349], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:51:13,665 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-29 08:51:14,265 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=102691.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:51:27,908 INFO [train.py:904] (4/8) Epoch 11, batch 1200, loss[loss=0.1737, simple_loss=0.2645, pruned_loss=0.04147, over 17228.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2606, pruned_loss=0.0487, over 3305231.03 frames. ], batch size: 46, lr: 6.34e-03, grad_scale: 8.0 2023-04-29 08:51:30,813 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8890, 4.3926, 3.2572, 2.3671, 3.0977, 2.6076, 4.7659, 3.9334], device='cuda:4'), covar=tensor([0.2474, 0.0579, 0.1392, 0.2202, 0.2287, 0.1658, 0.0325, 0.0967], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0257, 0.0283, 0.0275, 0.0278, 0.0222, 0.0268, 0.0296], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:51:50,177 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0451, 4.5544, 3.5201, 2.4526, 3.0401, 2.7642, 4.8557, 3.9301], device='cuda:4'), covar=tensor([0.2361, 0.0536, 0.1300, 0.2152, 0.2442, 0.1619, 0.0362, 0.1072], device='cuda:4'), in_proj_covar=tensor([0.0304, 0.0257, 0.0282, 0.0275, 0.0278, 0.0222, 0.0268, 0.0296], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:52:27,632 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.474e+02 2.248e+02 2.693e+02 3.149e+02 5.178e+02, threshold=5.386e+02, percent-clipped=0.0 2023-04-29 08:52:39,081 INFO [train.py:904] (4/8) Epoch 11, batch 1250, loss[loss=0.1513, simple_loss=0.2393, pruned_loss=0.03164, over 16993.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2606, pruned_loss=0.04898, over 3304314.71 frames. ], batch size: 41, lr: 6.34e-03, grad_scale: 8.0 2023-04-29 08:52:41,163 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1083, 3.9922, 4.1733, 4.3539, 4.4660, 4.0479, 4.2185, 4.4047], device='cuda:4'), covar=tensor([0.1560, 0.0973, 0.1417, 0.0625, 0.0517, 0.1130, 0.1837, 0.0636], device='cuda:4'), in_proj_covar=tensor([0.0539, 0.0670, 0.0828, 0.0693, 0.0521, 0.0521, 0.0535, 0.0611], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:53:49,412 INFO [train.py:904] (4/8) Epoch 11, batch 1300, loss[loss=0.185, simple_loss=0.2561, pruned_loss=0.05696, over 16747.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2611, pruned_loss=0.04913, over 3307741.16 frames. ], batch size: 83, lr: 6.33e-03, grad_scale: 8.0 2023-04-29 08:53:55,156 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.03 vs. limit=5.0 2023-04-29 08:54:46,490 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.885e+02 2.481e+02 2.988e+02 3.715e+02 8.832e+02, threshold=5.975e+02, percent-clipped=5.0 2023-04-29 08:54:51,784 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4633, 2.2762, 1.8289, 2.0423, 2.6309, 2.4271, 2.7571, 2.7930], device='cuda:4'), covar=tensor([0.0130, 0.0256, 0.0373, 0.0324, 0.0139, 0.0225, 0.0142, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0206, 0.0202, 0.0202, 0.0206, 0.0205, 0.0211, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 08:54:57,227 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=102851.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:54:58,763 INFO [train.py:904] (4/8) Epoch 11, batch 1350, loss[loss=0.1763, simple_loss=0.2532, pruned_loss=0.04975, over 16757.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.261, pruned_loss=0.04868, over 3321337.40 frames. ], batch size: 83, lr: 6.33e-03, grad_scale: 8.0 2023-04-29 08:55:45,165 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=102887.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:56:05,623 INFO [train.py:904] (4/8) Epoch 11, batch 1400, loss[loss=0.1741, simple_loss=0.2571, pruned_loss=0.04555, over 17259.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2602, pruned_loss=0.04825, over 3328829.12 frames. ], batch size: 45, lr: 6.33e-03, grad_scale: 8.0 2023-04-29 08:56:20,052 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=102912.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:56:22,912 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.30 vs. limit=5.0 2023-04-29 08:56:39,756 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-29 08:57:02,702 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1049, 1.9961, 2.6335, 3.0743, 2.8681, 3.3841, 2.3501, 3.4317], device='cuda:4'), covar=tensor([0.0158, 0.0355, 0.0204, 0.0191, 0.0205, 0.0133, 0.0325, 0.0098], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0173, 0.0157, 0.0158, 0.0168, 0.0123, 0.0171, 0.0114], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 08:57:05,104 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.762e+02 2.362e+02 2.982e+02 3.517e+02 5.508e+02, threshold=5.963e+02, percent-clipped=1.0 2023-04-29 08:57:10,169 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=102948.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 08:57:15,710 INFO [train.py:904] (4/8) Epoch 11, batch 1450, loss[loss=0.1808, simple_loss=0.2704, pruned_loss=0.04557, over 16662.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2588, pruned_loss=0.04825, over 3321609.65 frames. ], batch size: 57, lr: 6.33e-03, grad_scale: 8.0 2023-04-29 08:58:25,002 INFO [train.py:904] (4/8) Epoch 11, batch 1500, loss[loss=0.2166, simple_loss=0.2799, pruned_loss=0.07666, over 16832.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.259, pruned_loss=0.04844, over 3321979.59 frames. ], batch size: 116, lr: 6.33e-03, grad_scale: 8.0 2023-04-29 08:58:59,941 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2802, 5.1370, 5.0482, 4.6500, 4.6555, 5.1268, 5.0779, 4.7191], device='cuda:4'), covar=tensor([0.0555, 0.0430, 0.0280, 0.0275, 0.1062, 0.0361, 0.0316, 0.0662], device='cuda:4'), in_proj_covar=tensor([0.0259, 0.0323, 0.0302, 0.0281, 0.0324, 0.0320, 0.0208, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:59:24,563 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.863e+02 2.416e+02 2.850e+02 3.321e+02 6.158e+02, threshold=5.699e+02, percent-clipped=1.0 2023-04-29 08:59:31,278 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4924, 4.3324, 4.4219, 3.0621, 3.7467, 4.3552, 3.8391, 2.5929], device='cuda:4'), covar=tensor([0.0395, 0.0034, 0.0028, 0.0272, 0.0075, 0.0058, 0.0062, 0.0334], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0069, 0.0070, 0.0124, 0.0077, 0.0088, 0.0077, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 08:59:35,066 INFO [train.py:904] (4/8) Epoch 11, batch 1550, loss[loss=0.1533, simple_loss=0.2373, pruned_loss=0.03465, over 15744.00 frames. ], tot_loss[loss=0.1794, simple_loss=0.26, pruned_loss=0.04944, over 3323695.13 frames. ], batch size: 35, lr: 6.33e-03, grad_scale: 8.0 2023-04-29 09:00:39,937 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103098.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:00:45,000 INFO [train.py:904] (4/8) Epoch 11, batch 1600, loss[loss=0.1929, simple_loss=0.2909, pruned_loss=0.04744, over 17118.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2628, pruned_loss=0.05078, over 3323775.01 frames. ], batch size: 48, lr: 6.32e-03, grad_scale: 8.0 2023-04-29 09:00:47,875 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0365, 4.0962, 2.6104, 4.5654, 3.0264, 4.5084, 2.6108, 3.2297], device='cuda:4'), covar=tensor([0.0224, 0.0269, 0.1296, 0.0193, 0.0702, 0.0485, 0.1342, 0.0641], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0167, 0.0188, 0.0133, 0.0170, 0.0210, 0.0196, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 09:01:25,964 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103132.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:01:42,625 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.757e+02 2.417e+02 3.013e+02 3.629e+02 9.117e+02, threshold=6.027e+02, percent-clipped=4.0 2023-04-29 09:01:53,520 INFO [train.py:904] (4/8) Epoch 11, batch 1650, loss[loss=0.1795, simple_loss=0.2752, pruned_loss=0.04192, over 17054.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2641, pruned_loss=0.05083, over 3322434.48 frames. ], batch size: 55, lr: 6.32e-03, grad_scale: 8.0 2023-04-29 09:02:03,343 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103159.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:02:50,049 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103193.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:02:58,455 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0755, 4.4964, 3.3787, 2.3622, 3.0458, 2.6612, 4.8450, 4.0735], device='cuda:4'), covar=tensor([0.2242, 0.0528, 0.1382, 0.2328, 0.2472, 0.1708, 0.0322, 0.0953], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0254, 0.0281, 0.0273, 0.0278, 0.0220, 0.0265, 0.0295], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:03:02,609 INFO [train.py:904] (4/8) Epoch 11, batch 1700, loss[loss=0.2219, simple_loss=0.3017, pruned_loss=0.07103, over 16397.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2664, pruned_loss=0.05173, over 3323905.21 frames. ], batch size: 146, lr: 6.32e-03, grad_scale: 8.0 2023-04-29 09:03:10,658 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103207.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:04:01,646 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103243.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:04:02,559 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.865e+02 2.478e+02 3.058e+02 3.640e+02 7.869e+02, threshold=6.116e+02, percent-clipped=2.0 2023-04-29 09:04:13,825 INFO [train.py:904] (4/8) Epoch 11, batch 1750, loss[loss=0.2522, simple_loss=0.3177, pruned_loss=0.09337, over 15268.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2667, pruned_loss=0.05153, over 3329245.31 frames. ], batch size: 190, lr: 6.32e-03, grad_scale: 4.0 2023-04-29 09:04:14,369 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5812, 2.4955, 2.1275, 2.3359, 2.8781, 2.5660, 3.4086, 3.1317], device='cuda:4'), covar=tensor([0.0079, 0.0298, 0.0401, 0.0390, 0.0207, 0.0314, 0.0169, 0.0202], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0206, 0.0202, 0.0203, 0.0206, 0.0206, 0.0213, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:04:55,954 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4557, 5.8205, 5.6192, 5.6910, 5.2358, 5.0019, 5.2745, 5.9701], device='cuda:4'), covar=tensor([0.1114, 0.0807, 0.0819, 0.0658, 0.0694, 0.0600, 0.0971, 0.0683], device='cuda:4'), in_proj_covar=tensor([0.0543, 0.0681, 0.0563, 0.0475, 0.0425, 0.0441, 0.0571, 0.0519], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:05:22,483 INFO [train.py:904] (4/8) Epoch 11, batch 1800, loss[loss=0.1859, simple_loss=0.2633, pruned_loss=0.05428, over 16321.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2689, pruned_loss=0.05233, over 3322191.56 frames. ], batch size: 165, lr: 6.32e-03, grad_scale: 4.0 2023-04-29 09:06:21,365 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.567e+02 2.500e+02 2.964e+02 3.680e+02 1.127e+03, threshold=5.929e+02, percent-clipped=5.0 2023-04-29 09:06:31,956 INFO [train.py:904] (4/8) Epoch 11, batch 1850, loss[loss=0.1543, simple_loss=0.2396, pruned_loss=0.03447, over 17014.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2699, pruned_loss=0.05276, over 3319698.34 frames. ], batch size: 41, lr: 6.32e-03, grad_scale: 4.0 2023-04-29 09:06:51,464 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103366.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 09:07:32,318 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9583, 1.6674, 2.4533, 2.8325, 2.6821, 3.0366, 2.0424, 3.0799], device='cuda:4'), covar=tensor([0.0131, 0.0365, 0.0216, 0.0204, 0.0225, 0.0162, 0.0343, 0.0106], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0174, 0.0159, 0.0160, 0.0170, 0.0124, 0.0171, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 09:07:33,297 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8923, 4.6823, 4.9273, 5.1898, 5.3643, 4.7295, 5.3354, 5.2784], device='cuda:4'), covar=tensor([0.1488, 0.1146, 0.1659, 0.0615, 0.0490, 0.0832, 0.0453, 0.0523], device='cuda:4'), in_proj_covar=tensor([0.0546, 0.0680, 0.0837, 0.0699, 0.0529, 0.0527, 0.0540, 0.0618], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:07:39,137 INFO [train.py:904] (4/8) Epoch 11, batch 1900, loss[loss=0.1941, simple_loss=0.268, pruned_loss=0.06012, over 16161.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2684, pruned_loss=0.05147, over 3324819.84 frames. ], batch size: 165, lr: 6.32e-03, grad_scale: 4.0 2023-04-29 09:08:16,107 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103427.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 09:08:40,923 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.233e+02 2.712e+02 3.149e+02 7.257e+02, threshold=5.425e+02, percent-clipped=1.0 2023-04-29 09:08:51,894 INFO [train.py:904] (4/8) Epoch 11, batch 1950, loss[loss=0.2041, simple_loss=0.2766, pruned_loss=0.06576, over 16808.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2678, pruned_loss=0.051, over 3322179.50 frames. ], batch size: 124, lr: 6.31e-03, grad_scale: 4.0 2023-04-29 09:08:55,092 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103454.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:09:04,327 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3371, 5.2510, 5.1735, 4.7770, 4.7551, 5.2132, 5.1850, 4.8184], device='cuda:4'), covar=tensor([0.0564, 0.0372, 0.0241, 0.0262, 0.0995, 0.0334, 0.0238, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0262, 0.0327, 0.0305, 0.0285, 0.0328, 0.0323, 0.0208, 0.0358], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 09:09:41,327 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103488.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:10:00,139 INFO [train.py:904] (4/8) Epoch 11, batch 2000, loss[loss=0.1888, simple_loss=0.2557, pruned_loss=0.06098, over 16738.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2674, pruned_loss=0.05131, over 3322476.33 frames. ], batch size: 102, lr: 6.31e-03, grad_scale: 8.0 2023-04-29 09:10:07,795 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=103507.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:10:42,084 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-29 09:10:58,905 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=103543.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:11:00,878 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.679e+02 2.374e+02 2.796e+02 3.607e+02 7.066e+02, threshold=5.592e+02, percent-clipped=4.0 2023-04-29 09:11:11,321 INFO [train.py:904] (4/8) Epoch 11, batch 2050, loss[loss=0.1946, simple_loss=0.2665, pruned_loss=0.06133, over 16856.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.2669, pruned_loss=0.05135, over 3314275.58 frames. ], batch size: 116, lr: 6.31e-03, grad_scale: 4.0 2023-04-29 09:11:16,433 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=103555.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:11:19,685 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.42 vs. limit=5.0 2023-04-29 09:11:40,946 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7967, 4.1089, 4.2827, 3.0906, 3.7156, 4.2761, 3.8904, 2.3021], device='cuda:4'), covar=tensor([0.0371, 0.0055, 0.0031, 0.0271, 0.0075, 0.0066, 0.0060, 0.0374], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0071, 0.0071, 0.0126, 0.0079, 0.0089, 0.0077, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 09:12:05,786 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=103591.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:12:21,596 INFO [train.py:904] (4/8) Epoch 11, batch 2100, loss[loss=0.2158, simple_loss=0.2929, pruned_loss=0.06936, over 16694.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2685, pruned_loss=0.05177, over 3321499.81 frames. ], batch size: 124, lr: 6.31e-03, grad_scale: 4.0 2023-04-29 09:13:22,846 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.732e+02 2.751e+02 3.128e+02 3.657e+02 9.004e+02, threshold=6.256e+02, percent-clipped=1.0 2023-04-29 09:13:31,355 INFO [train.py:904] (4/8) Epoch 11, batch 2150, loss[loss=0.1644, simple_loss=0.2623, pruned_loss=0.03324, over 17103.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2699, pruned_loss=0.05231, over 3324793.98 frames. ], batch size: 47, lr: 6.31e-03, grad_scale: 4.0 2023-04-29 09:14:42,107 INFO [train.py:904] (4/8) Epoch 11, batch 2200, loss[loss=0.2012, simple_loss=0.2783, pruned_loss=0.062, over 16443.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2704, pruned_loss=0.05254, over 3319976.58 frames. ], batch size: 146, lr: 6.31e-03, grad_scale: 4.0 2023-04-29 09:14:51,837 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-29 09:15:06,767 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.06 vs. limit=2.0 2023-04-29 09:15:10,359 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103722.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 09:15:35,668 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103740.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:15:44,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.506e+02 2.339e+02 2.759e+02 3.363e+02 7.654e+02, threshold=5.518e+02, percent-clipped=1.0 2023-04-29 09:15:51,103 INFO [train.py:904] (4/8) Epoch 11, batch 2250, loss[loss=0.1657, simple_loss=0.2501, pruned_loss=0.0406, over 16030.00 frames. ], tot_loss[loss=0.188, simple_loss=0.2706, pruned_loss=0.05267, over 3309957.92 frames. ], batch size: 35, lr: 6.31e-03, grad_scale: 4.0 2023-04-29 09:15:54,458 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=103754.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:16:12,332 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103765.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:16:39,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1578, 5.0612, 4.9821, 4.6097, 4.6470, 5.0534, 5.0694, 4.6645], device='cuda:4'), covar=tensor([0.0530, 0.0394, 0.0248, 0.0272, 0.0882, 0.0361, 0.0298, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0327, 0.0307, 0.0285, 0.0326, 0.0325, 0.0211, 0.0357], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 09:16:44,194 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=103788.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:16:55,564 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-29 09:17:01,640 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103801.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:17:02,364 INFO [train.py:904] (4/8) Epoch 11, batch 2300, loss[loss=0.1693, simple_loss=0.2621, pruned_loss=0.03823, over 16899.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2708, pruned_loss=0.05285, over 3322813.74 frames. ], batch size: 42, lr: 6.30e-03, grad_scale: 4.0 2023-04-29 09:17:02,658 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=103802.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:17:35,774 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103826.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:17:48,570 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=103836.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:18:02,532 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.826e+02 2.457e+02 2.837e+02 3.610e+02 7.865e+02, threshold=5.674e+02, percent-clipped=5.0 2023-04-29 09:18:11,650 INFO [train.py:904] (4/8) Epoch 11, batch 2350, loss[loss=0.2059, simple_loss=0.2755, pruned_loss=0.0682, over 16889.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2714, pruned_loss=0.05308, over 3328169.38 frames. ], batch size: 109, lr: 6.30e-03, grad_scale: 4.0 2023-04-29 09:18:45,300 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103876.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:19:19,669 INFO [train.py:904] (4/8) Epoch 11, batch 2400, loss[loss=0.2079, simple_loss=0.2894, pruned_loss=0.06319, over 16501.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2728, pruned_loss=0.05358, over 3312912.72 frames. ], batch size: 75, lr: 6.30e-03, grad_scale: 8.0 2023-04-29 09:20:09,844 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103937.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 09:20:20,789 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.671e+02 2.318e+02 2.704e+02 3.589e+02 9.063e+02, threshold=5.409e+02, percent-clipped=3.0 2023-04-29 09:20:28,624 INFO [train.py:904] (4/8) Epoch 11, batch 2450, loss[loss=0.1984, simple_loss=0.2733, pruned_loss=0.06176, over 16824.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2736, pruned_loss=0.05335, over 3308545.38 frames. ], batch size: 102, lr: 6.30e-03, grad_scale: 8.0 2023-04-29 09:20:33,875 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1153, 4.5405, 4.6400, 3.4378, 3.8232, 4.4952, 4.2099, 2.9457], device='cuda:4'), covar=tensor([0.0318, 0.0035, 0.0023, 0.0214, 0.0072, 0.0059, 0.0056, 0.0300], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0070, 0.0070, 0.0124, 0.0078, 0.0089, 0.0077, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 09:21:22,187 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103990.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:21:31,622 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103997.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:21:42,813 INFO [train.py:904] (4/8) Epoch 11, batch 2500, loss[loss=0.1978, simple_loss=0.2891, pruned_loss=0.05326, over 17114.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2737, pruned_loss=0.05304, over 3301587.85 frames. ], batch size: 49, lr: 6.30e-03, grad_scale: 8.0 2023-04-29 09:21:49,051 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.8172, 6.1015, 5.8285, 5.9465, 5.5388, 5.2756, 5.5579, 6.2327], device='cuda:4'), covar=tensor([0.1028, 0.0815, 0.0874, 0.0663, 0.0750, 0.0633, 0.0887, 0.0780], device='cuda:4'), in_proj_covar=tensor([0.0548, 0.0685, 0.0568, 0.0480, 0.0429, 0.0441, 0.0575, 0.0521], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:22:11,534 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104022.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 09:22:43,744 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.523e+02 2.442e+02 2.893e+02 3.416e+02 6.492e+02, threshold=5.787e+02, percent-clipped=3.0 2023-04-29 09:22:47,567 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=104049.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:22:49,903 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=104051.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:22:50,606 INFO [train.py:904] (4/8) Epoch 11, batch 2550, loss[loss=0.2029, simple_loss=0.2699, pruned_loss=0.06796, over 16931.00 frames. ], tot_loss[loss=0.1897, simple_loss=0.2736, pruned_loss=0.05297, over 3309877.44 frames. ], batch size: 109, lr: 6.30e-03, grad_scale: 8.0 2023-04-29 09:23:00,530 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=104058.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:23:13,743 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3473, 3.3921, 1.9353, 3.5897, 2.5504, 3.5521, 1.9889, 2.7714], device='cuda:4'), covar=tensor([0.0211, 0.0369, 0.1451, 0.0229, 0.0724, 0.0678, 0.1444, 0.0605], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0168, 0.0187, 0.0134, 0.0167, 0.0212, 0.0194, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 09:23:15,996 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=104070.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 09:23:17,494 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-29 09:23:51,812 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104096.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:23:59,570 INFO [train.py:904] (4/8) Epoch 11, batch 2600, loss[loss=0.2009, simple_loss=0.2903, pruned_loss=0.0558, over 17031.00 frames. ], tot_loss[loss=0.1892, simple_loss=0.2728, pruned_loss=0.05275, over 3310181.35 frames. ], batch size: 55, lr: 6.29e-03, grad_scale: 8.0 2023-04-29 09:24:11,995 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=104110.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:24:16,875 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8002, 1.6443, 2.3099, 2.7631, 2.6156, 3.3322, 1.9735, 3.1884], device='cuda:4'), covar=tensor([0.0196, 0.0398, 0.0267, 0.0254, 0.0233, 0.0133, 0.0381, 0.0123], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0174, 0.0159, 0.0161, 0.0170, 0.0125, 0.0172, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 09:24:26,682 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104121.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:25:01,151 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.592e+02 2.655e+02 3.049e+02 3.595e+02 6.269e+02, threshold=6.098e+02, percent-clipped=2.0 2023-04-29 09:25:10,376 INFO [train.py:904] (4/8) Epoch 11, batch 2650, loss[loss=0.1892, simple_loss=0.2808, pruned_loss=0.0488, over 16590.00 frames. ], tot_loss[loss=0.188, simple_loss=0.2725, pruned_loss=0.05177, over 3321850.63 frames. ], batch size: 75, lr: 6.29e-03, grad_scale: 8.0 2023-04-29 09:26:18,840 INFO [train.py:904] (4/8) Epoch 11, batch 2700, loss[loss=0.1838, simple_loss=0.2728, pruned_loss=0.04737, over 16681.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2725, pruned_loss=0.05172, over 3325306.85 frames. ], batch size: 62, lr: 6.29e-03, grad_scale: 8.0 2023-04-29 09:27:00,691 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104232.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 09:27:13,587 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0447, 5.0344, 5.5497, 5.5139, 5.5375, 5.1198, 5.1025, 4.8489], device='cuda:4'), covar=tensor([0.0293, 0.0459, 0.0328, 0.0445, 0.0452, 0.0331, 0.0868, 0.0399], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0347, 0.0350, 0.0328, 0.0396, 0.0366, 0.0466, 0.0293], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 09:27:19,004 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.677e+02 2.334e+02 2.747e+02 3.753e+02 9.915e+02, threshold=5.495e+02, percent-clipped=4.0 2023-04-29 09:27:27,257 INFO [train.py:904] (4/8) Epoch 11, batch 2750, loss[loss=0.1747, simple_loss=0.2698, pruned_loss=0.03976, over 17027.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2728, pruned_loss=0.05102, over 3323897.03 frames. ], batch size: 50, lr: 6.29e-03, grad_scale: 8.0 2023-04-29 09:28:10,914 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-29 09:28:32,582 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4650, 3.9573, 3.9904, 2.2067, 3.2906, 2.5202, 3.9747, 4.0035], device='cuda:4'), covar=tensor([0.0208, 0.0617, 0.0410, 0.1514, 0.0641, 0.0846, 0.0550, 0.0824], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0147, 0.0158, 0.0142, 0.0136, 0.0124, 0.0136, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 09:28:36,571 INFO [train.py:904] (4/8) Epoch 11, batch 2800, loss[loss=0.1987, simple_loss=0.2774, pruned_loss=0.05998, over 16798.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2728, pruned_loss=0.0506, over 3328232.38 frames. ], batch size: 102, lr: 6.29e-03, grad_scale: 8.0 2023-04-29 09:29:24,070 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-29 09:29:37,339 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.621e+02 2.406e+02 2.849e+02 3.526e+02 7.229e+02, threshold=5.698e+02, percent-clipped=5.0 2023-04-29 09:29:37,600 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104346.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:29:44,353 INFO [train.py:904] (4/8) Epoch 11, batch 2850, loss[loss=0.2097, simple_loss=0.279, pruned_loss=0.07018, over 16277.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2726, pruned_loss=0.05066, over 3329216.31 frames. ], batch size: 165, lr: 6.29e-03, grad_scale: 8.0 2023-04-29 09:29:45,797 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104353.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:30:28,864 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-29 09:30:44,750 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104396.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:30:52,869 INFO [train.py:904] (4/8) Epoch 11, batch 2900, loss[loss=0.1497, simple_loss=0.2295, pruned_loss=0.0349, over 17011.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2702, pruned_loss=0.05011, over 3333035.68 frames. ], batch size: 41, lr: 6.29e-03, grad_scale: 8.0 2023-04-29 09:30:57,319 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104405.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:31:20,046 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104421.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:31:43,223 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-29 09:31:48,911 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0593, 1.9842, 2.2880, 3.5487, 2.0603, 2.3145, 2.1741, 2.1320], device='cuda:4'), covar=tensor([0.1039, 0.3062, 0.2118, 0.0542, 0.3422, 0.2159, 0.2768, 0.3050], device='cuda:4'), in_proj_covar=tensor([0.0367, 0.0392, 0.0330, 0.0325, 0.0412, 0.0450, 0.0354, 0.0464], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:31:52,195 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=104444.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:31:55,136 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.712e+02 2.403e+02 2.933e+02 3.429e+02 6.056e+02, threshold=5.866e+02, percent-clipped=1.0 2023-04-29 09:32:04,091 INFO [train.py:904] (4/8) Epoch 11, batch 2950, loss[loss=0.2028, simple_loss=0.2714, pruned_loss=0.0671, over 16774.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2704, pruned_loss=0.05115, over 3317363.88 frames. ], batch size: 102, lr: 6.28e-03, grad_scale: 8.0 2023-04-29 09:32:18,937 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-29 09:32:27,832 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=104469.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:32:29,896 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7871, 3.8231, 2.8953, 2.3372, 2.7524, 2.3065, 3.9058, 3.7005], device='cuda:4'), covar=tensor([0.2112, 0.0567, 0.1440, 0.2155, 0.2054, 0.1670, 0.0480, 0.0954], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0258, 0.0282, 0.0276, 0.0284, 0.0222, 0.0269, 0.0301], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:33:12,380 INFO [train.py:904] (4/8) Epoch 11, batch 3000, loss[loss=0.2028, simple_loss=0.2744, pruned_loss=0.06555, over 16275.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2703, pruned_loss=0.05175, over 3326215.14 frames. ], batch size: 165, lr: 6.28e-03, grad_scale: 8.0 2023-04-29 09:33:12,380 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 09:33:22,061 INFO [train.py:938] (4/8) Epoch 11, validation: loss=0.1413, simple_loss=0.2475, pruned_loss=0.01754, over 944034.00 frames. 2023-04-29 09:33:22,062 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 09:34:04,228 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104532.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 09:34:20,848 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.514e+02 2.544e+02 3.030e+02 3.628e+02 6.112e+02, threshold=6.060e+02, percent-clipped=1.0 2023-04-29 09:34:30,327 INFO [train.py:904] (4/8) Epoch 11, batch 3050, loss[loss=0.1982, simple_loss=0.2735, pruned_loss=0.06144, over 16470.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2706, pruned_loss=0.05186, over 3333722.71 frames. ], batch size: 75, lr: 6.28e-03, grad_scale: 8.0 2023-04-29 09:35:07,352 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=104580.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:35:37,110 INFO [train.py:904] (4/8) Epoch 11, batch 3100, loss[loss=0.2174, simple_loss=0.2911, pruned_loss=0.07187, over 16227.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2706, pruned_loss=0.05289, over 3322694.66 frames. ], batch size: 165, lr: 6.28e-03, grad_scale: 8.0 2023-04-29 09:35:51,673 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6688, 3.7163, 3.8552, 1.9470, 3.9608, 4.0413, 3.1807, 3.0113], device='cuda:4'), covar=tensor([0.0690, 0.0157, 0.0165, 0.1102, 0.0068, 0.0102, 0.0317, 0.0379], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0100, 0.0091, 0.0139, 0.0072, 0.0105, 0.0122, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 09:36:03,283 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7548, 2.6003, 2.3576, 3.4592, 2.8443, 3.6317, 1.5223, 2.7601], device='cuda:4'), covar=tensor([0.1226, 0.0588, 0.1055, 0.0162, 0.0210, 0.0360, 0.1368, 0.0707], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0157, 0.0178, 0.0146, 0.0198, 0.0209, 0.0176, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 09:36:39,246 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.673e+02 2.465e+02 3.037e+02 3.411e+02 8.178e+02, threshold=6.075e+02, percent-clipped=4.0 2023-04-29 09:36:39,607 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104646.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:36:47,518 INFO [train.py:904] (4/8) Epoch 11, batch 3150, loss[loss=0.1893, simple_loss=0.2827, pruned_loss=0.04791, over 17053.00 frames. ], tot_loss[loss=0.1878, simple_loss=0.2696, pruned_loss=0.05296, over 3315455.66 frames. ], batch size: 50, lr: 6.28e-03, grad_scale: 8.0 2023-04-29 09:36:49,180 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104653.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:37:41,177 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6940, 4.0729, 2.9071, 2.1604, 3.0219, 2.3411, 4.3612, 3.7687], device='cuda:4'), covar=tensor([0.2704, 0.0674, 0.1776, 0.2432, 0.2429, 0.1929, 0.0468, 0.1048], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0260, 0.0285, 0.0278, 0.0288, 0.0225, 0.0271, 0.0305], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 09:37:46,268 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=104694.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:37:56,591 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=104701.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:37:57,453 INFO [train.py:904] (4/8) Epoch 11, batch 3200, loss[loss=0.1936, simple_loss=0.2821, pruned_loss=0.05258, over 16710.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2694, pruned_loss=0.05242, over 3316177.90 frames. ], batch size: 62, lr: 6.28e-03, grad_scale: 8.0 2023-04-29 09:38:01,261 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104705.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:38:57,881 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.739e+02 2.472e+02 2.965e+02 3.891e+02 6.936e+02, threshold=5.930e+02, percent-clipped=3.0 2023-04-29 09:39:06,599 INFO [train.py:904] (4/8) Epoch 11, batch 3250, loss[loss=0.2008, simple_loss=0.2882, pruned_loss=0.05666, over 16522.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.27, pruned_loss=0.05333, over 3305612.34 frames. ], batch size: 75, lr: 6.28e-03, grad_scale: 8.0 2023-04-29 09:39:08,055 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=104753.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:39:38,065 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0670, 2.7693, 2.7670, 2.0621, 2.6025, 2.1757, 2.7710, 2.8971], device='cuda:4'), covar=tensor([0.0262, 0.0595, 0.0460, 0.1397, 0.0661, 0.0803, 0.0494, 0.0538], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0148, 0.0158, 0.0143, 0.0136, 0.0125, 0.0137, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 09:40:15,729 INFO [train.py:904] (4/8) Epoch 11, batch 3300, loss[loss=0.2172, simple_loss=0.2952, pruned_loss=0.0696, over 15478.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2715, pruned_loss=0.05371, over 3310344.82 frames. ], batch size: 191, lr: 6.27e-03, grad_scale: 8.0 2023-04-29 09:40:56,917 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-29 09:41:16,320 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.576e+02 2.335e+02 2.858e+02 3.500e+02 6.085e+02, threshold=5.716e+02, percent-clipped=1.0 2023-04-29 09:41:24,648 INFO [train.py:904] (4/8) Epoch 11, batch 3350, loss[loss=0.1712, simple_loss=0.2567, pruned_loss=0.0428, over 16808.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2717, pruned_loss=0.05302, over 3314563.14 frames. ], batch size: 42, lr: 6.27e-03, grad_scale: 8.0 2023-04-29 09:41:26,233 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1437, 1.8533, 2.5136, 2.9495, 2.7562, 3.5222, 2.3725, 3.4590], device='cuda:4'), covar=tensor([0.0136, 0.0337, 0.0220, 0.0177, 0.0206, 0.0110, 0.0286, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0173, 0.0158, 0.0161, 0.0170, 0.0126, 0.0171, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 09:41:58,152 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0550, 2.3048, 2.5205, 4.8101, 2.3416, 2.8744, 2.4890, 2.6141], device='cuda:4'), covar=tensor([0.0760, 0.3440, 0.2176, 0.0310, 0.3677, 0.2342, 0.2813, 0.3290], device='cuda:4'), in_proj_covar=tensor([0.0368, 0.0393, 0.0331, 0.0325, 0.0413, 0.0453, 0.0356, 0.0464], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:42:33,961 INFO [train.py:904] (4/8) Epoch 11, batch 3400, loss[loss=0.2567, simple_loss=0.3262, pruned_loss=0.09363, over 12395.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2718, pruned_loss=0.05269, over 3307678.98 frames. ], batch size: 248, lr: 6.27e-03, grad_scale: 8.0 2023-04-29 09:43:23,021 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5927, 3.6527, 4.1320, 2.7630, 3.6429, 4.0981, 3.8068, 2.3967], device='cuda:4'), covar=tensor([0.0405, 0.0228, 0.0039, 0.0303, 0.0066, 0.0072, 0.0060, 0.0367], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0071, 0.0071, 0.0126, 0.0079, 0.0091, 0.0077, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 09:43:33,847 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.534e+02 2.318e+02 2.811e+02 3.324e+02 7.497e+02, threshold=5.622e+02, percent-clipped=1.0 2023-04-29 09:43:41,665 INFO [train.py:904] (4/8) Epoch 11, batch 3450, loss[loss=0.1623, simple_loss=0.2616, pruned_loss=0.03154, over 17034.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2705, pruned_loss=0.05204, over 3321207.25 frames. ], batch size: 50, lr: 6.27e-03, grad_scale: 8.0 2023-04-29 09:44:08,623 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7576, 2.8138, 2.4323, 4.0251, 3.3844, 4.0745, 1.3907, 2.8482], device='cuda:4'), covar=tensor([0.1318, 0.0575, 0.1018, 0.0189, 0.0194, 0.0369, 0.1423, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0158, 0.0178, 0.0148, 0.0200, 0.0211, 0.0177, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 09:44:13,511 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7771, 4.2326, 4.2059, 3.0402, 3.6388, 4.1570, 3.8432, 2.4184], device='cuda:4'), covar=tensor([0.0395, 0.0048, 0.0047, 0.0294, 0.0086, 0.0098, 0.0073, 0.0384], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0071, 0.0071, 0.0125, 0.0079, 0.0090, 0.0077, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 09:44:52,762 INFO [train.py:904] (4/8) Epoch 11, batch 3500, loss[loss=0.1655, simple_loss=0.2524, pruned_loss=0.03928, over 15867.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2687, pruned_loss=0.05153, over 3308435.98 frames. ], batch size: 35, lr: 6.27e-03, grad_scale: 8.0 2023-04-29 09:45:09,586 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-29 09:45:55,147 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.324e+02 2.297e+02 2.659e+02 3.238e+02 6.959e+02, threshold=5.317e+02, percent-clipped=2.0 2023-04-29 09:46:03,260 INFO [train.py:904] (4/8) Epoch 11, batch 3550, loss[loss=0.1481, simple_loss=0.228, pruned_loss=0.03416, over 16980.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2664, pruned_loss=0.05016, over 3310846.28 frames. ], batch size: 41, lr: 6.27e-03, grad_scale: 4.0 2023-04-29 09:47:12,591 INFO [train.py:904] (4/8) Epoch 11, batch 3600, loss[loss=0.2033, simple_loss=0.2764, pruned_loss=0.06508, over 15393.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2641, pruned_loss=0.04929, over 3304515.07 frames. ], batch size: 190, lr: 6.26e-03, grad_scale: 8.0 2023-04-29 09:47:29,892 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3343, 5.3052, 5.1808, 4.6910, 4.8140, 5.2374, 5.2266, 4.8417], device='cuda:4'), covar=tensor([0.0527, 0.0462, 0.0246, 0.0304, 0.1030, 0.0421, 0.0267, 0.0718], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0337, 0.0313, 0.0293, 0.0338, 0.0334, 0.0215, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 09:48:17,988 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.592e+02 2.381e+02 2.999e+02 3.559e+02 5.389e+02, threshold=5.997e+02, percent-clipped=2.0 2023-04-29 09:48:24,356 INFO [train.py:904] (4/8) Epoch 11, batch 3650, loss[loss=0.1726, simple_loss=0.2416, pruned_loss=0.05185, over 16827.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2634, pruned_loss=0.04992, over 3299781.51 frames. ], batch size: 96, lr: 6.26e-03, grad_scale: 8.0 2023-04-29 09:49:37,392 INFO [train.py:904] (4/8) Epoch 11, batch 3700, loss[loss=0.1809, simple_loss=0.2583, pruned_loss=0.05172, over 16814.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2628, pruned_loss=0.05204, over 3291348.05 frames. ], batch size: 102, lr: 6.26e-03, grad_scale: 8.0 2023-04-29 09:50:41,013 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.627e+02 2.380e+02 2.864e+02 3.354e+02 5.031e+02, threshold=5.728e+02, percent-clipped=0.0 2023-04-29 09:50:48,706 INFO [train.py:904] (4/8) Epoch 11, batch 3750, loss[loss=0.1739, simple_loss=0.2483, pruned_loss=0.0497, over 16845.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2639, pruned_loss=0.05401, over 3282756.80 frames. ], batch size: 96, lr: 6.26e-03, grad_scale: 8.0 2023-04-29 09:50:59,119 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-29 09:51:07,420 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-29 09:51:48,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7296, 2.8390, 2.2858, 3.9890, 3.2041, 3.9808, 1.5111, 2.6098], device='cuda:4'), covar=tensor([0.1318, 0.0574, 0.1163, 0.0157, 0.0162, 0.0355, 0.1398, 0.0862], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0157, 0.0179, 0.0148, 0.0199, 0.0210, 0.0177, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 09:51:57,078 INFO [train.py:904] (4/8) Epoch 11, batch 3800, loss[loss=0.2127, simple_loss=0.2898, pruned_loss=0.06787, over 15445.00 frames. ], tot_loss[loss=0.188, simple_loss=0.265, pruned_loss=0.0555, over 3285213.72 frames. ], batch size: 190, lr: 6.26e-03, grad_scale: 8.0 2023-04-29 09:52:16,747 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3841, 3.0210, 2.6409, 2.2099, 2.2152, 2.1733, 3.0545, 2.8606], device='cuda:4'), covar=tensor([0.2135, 0.0618, 0.1318, 0.1892, 0.2163, 0.1706, 0.0514, 0.0939], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0258, 0.0283, 0.0277, 0.0290, 0.0222, 0.0269, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 09:52:43,155 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7394, 3.7911, 4.0883, 2.0661, 4.1649, 4.1876, 3.2615, 3.0680], device='cuda:4'), covar=tensor([0.0683, 0.0173, 0.0113, 0.1131, 0.0049, 0.0114, 0.0310, 0.0378], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0100, 0.0090, 0.0140, 0.0071, 0.0106, 0.0122, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 09:53:02,333 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.313e+02 2.608e+02 3.308e+02 8.159e+02, threshold=5.217e+02, percent-clipped=3.0 2023-04-29 09:53:08,977 INFO [train.py:904] (4/8) Epoch 11, batch 3850, loss[loss=0.1783, simple_loss=0.2557, pruned_loss=0.05045, over 15611.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2654, pruned_loss=0.05614, over 3264406.12 frames. ], batch size: 190, lr: 6.26e-03, grad_scale: 8.0 2023-04-29 09:54:00,694 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105388.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:54:19,877 INFO [train.py:904] (4/8) Epoch 11, batch 3900, loss[loss=0.1811, simple_loss=0.2538, pruned_loss=0.05423, over 16454.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2649, pruned_loss=0.05645, over 3267392.45 frames. ], batch size: 68, lr: 6.26e-03, grad_scale: 8.0 2023-04-29 09:54:54,490 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-29 09:55:25,283 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.603e+02 2.279e+02 2.712e+02 3.395e+02 5.874e+02, threshold=5.424e+02, percent-clipped=4.0 2023-04-29 09:55:28,966 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105449.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 09:55:31,850 INFO [train.py:904] (4/8) Epoch 11, batch 3950, loss[loss=0.1688, simple_loss=0.2533, pruned_loss=0.0421, over 17059.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2649, pruned_loss=0.05702, over 3268817.57 frames. ], batch size: 50, lr: 6.25e-03, grad_scale: 8.0 2023-04-29 09:55:54,857 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6943, 2.7700, 2.5105, 4.0870, 3.4270, 4.0465, 1.5277, 2.8399], device='cuda:4'), covar=tensor([0.1330, 0.0589, 0.1001, 0.0155, 0.0142, 0.0345, 0.1409, 0.0731], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0159, 0.0181, 0.0148, 0.0201, 0.0212, 0.0179, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 09:56:21,736 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105486.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:56:44,038 INFO [train.py:904] (4/8) Epoch 11, batch 4000, loss[loss=0.196, simple_loss=0.2752, pruned_loss=0.05838, over 16530.00 frames. ], tot_loss[loss=0.189, simple_loss=0.2642, pruned_loss=0.0569, over 3277860.63 frames. ], batch size: 68, lr: 6.25e-03, grad_scale: 8.0 2023-04-29 09:56:44,519 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105502.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:56:48,604 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5185, 3.4885, 3.8228, 1.7823, 3.8960, 3.9231, 3.0626, 2.9411], device='cuda:4'), covar=tensor([0.0698, 0.0228, 0.0139, 0.1126, 0.0065, 0.0140, 0.0363, 0.0393], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0101, 0.0090, 0.0140, 0.0071, 0.0106, 0.0122, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 09:57:48,231 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.586e+02 2.126e+02 2.510e+02 3.113e+02 5.391e+02, threshold=5.020e+02, percent-clipped=0.0 2023-04-29 09:57:48,863 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105547.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:57:55,348 INFO [train.py:904] (4/8) Epoch 11, batch 4050, loss[loss=0.175, simple_loss=0.2535, pruned_loss=0.04826, over 17216.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2643, pruned_loss=0.0554, over 3280893.93 frames. ], batch size: 45, lr: 6.25e-03, grad_scale: 8.0 2023-04-29 09:58:11,089 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105563.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 09:59:08,731 INFO [train.py:904] (4/8) Epoch 11, batch 4100, loss[loss=0.2223, simple_loss=0.2915, pruned_loss=0.07655, over 12363.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2654, pruned_loss=0.05478, over 3274167.16 frames. ], batch size: 247, lr: 6.25e-03, grad_scale: 8.0 2023-04-29 09:59:52,319 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-29 10:00:18,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.211e+02 2.197e+02 2.575e+02 3.230e+02 7.189e+02, threshold=5.150e+02, percent-clipped=4.0 2023-04-29 10:00:24,121 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8416, 5.0450, 5.3286, 5.1188, 5.1459, 5.6952, 5.3041, 5.0063], device='cuda:4'), covar=tensor([0.0771, 0.1479, 0.1463, 0.1578, 0.2089, 0.0721, 0.1071, 0.2212], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0490, 0.0528, 0.0423, 0.0558, 0.0551, 0.0419, 0.0575], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 10:00:26,727 INFO [train.py:904] (4/8) Epoch 11, batch 4150, loss[loss=0.2848, simple_loss=0.3419, pruned_loss=0.1139, over 11467.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2734, pruned_loss=0.05836, over 3229888.01 frames. ], batch size: 246, lr: 6.25e-03, grad_scale: 8.0 2023-04-29 10:00:56,211 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5933, 3.6353, 2.7679, 2.1540, 2.4149, 2.2280, 3.8205, 3.3171], device='cuda:4'), covar=tensor([0.2398, 0.0569, 0.1497, 0.2142, 0.2380, 0.1774, 0.0424, 0.0912], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0256, 0.0284, 0.0278, 0.0289, 0.0223, 0.0268, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:01:44,700 INFO [train.py:904] (4/8) Epoch 11, batch 4200, loss[loss=0.2361, simple_loss=0.3174, pruned_loss=0.07741, over 16892.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2801, pruned_loss=0.05953, over 3215316.88 frames. ], batch size: 116, lr: 6.25e-03, grad_scale: 8.0 2023-04-29 10:02:27,640 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7745, 1.3083, 1.7036, 1.7031, 1.7629, 1.9249, 1.5584, 1.7594], device='cuda:4'), covar=tensor([0.0159, 0.0266, 0.0129, 0.0182, 0.0161, 0.0110, 0.0244, 0.0074], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0172, 0.0156, 0.0160, 0.0168, 0.0123, 0.0168, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 10:02:49,761 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=105744.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 10:02:53,623 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.565e+02 2.308e+02 2.787e+02 3.444e+02 5.705e+02, threshold=5.573e+02, percent-clipped=4.0 2023-04-29 10:02:59,809 INFO [train.py:904] (4/8) Epoch 11, batch 4250, loss[loss=0.1953, simple_loss=0.2919, pruned_loss=0.04939, over 17211.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2846, pruned_loss=0.06025, over 3190423.24 frames. ], batch size: 45, lr: 6.25e-03, grad_scale: 8.0 2023-04-29 10:03:55,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3253, 2.3969, 1.8555, 2.2156, 2.8578, 2.5391, 3.0649, 3.1458], device='cuda:4'), covar=tensor([0.0071, 0.0275, 0.0403, 0.0329, 0.0155, 0.0268, 0.0153, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0202, 0.0198, 0.0198, 0.0202, 0.0201, 0.0207, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:03:59,178 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105792.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:04:08,520 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105799.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:04:12,691 INFO [train.py:904] (4/8) Epoch 11, batch 4300, loss[loss=0.2154, simple_loss=0.3066, pruned_loss=0.06214, over 16641.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2855, pruned_loss=0.05901, over 3184669.23 frames. ], batch size: 62, lr: 6.24e-03, grad_scale: 8.0 2023-04-29 10:04:15,476 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-29 10:04:52,209 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3634, 5.6359, 5.3637, 5.4134, 5.1023, 4.9379, 5.0854, 5.7284], device='cuda:4'), covar=tensor([0.0802, 0.0583, 0.0801, 0.0568, 0.0689, 0.0622, 0.0842, 0.0714], device='cuda:4'), in_proj_covar=tensor([0.0539, 0.0666, 0.0555, 0.0469, 0.0424, 0.0435, 0.0556, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:05:06,311 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6127, 2.6556, 1.7733, 2.7898, 2.1530, 2.8123, 2.0741, 2.3711], device='cuda:4'), covar=tensor([0.0227, 0.0310, 0.1206, 0.0166, 0.0652, 0.0356, 0.1041, 0.0553], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0164, 0.0185, 0.0128, 0.0165, 0.0207, 0.0191, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-29 10:05:11,563 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=105842.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:05:17,304 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.532e+02 2.245e+02 2.727e+02 3.248e+02 5.954e+02, threshold=5.453e+02, percent-clipped=2.0 2023-04-29 10:05:25,481 INFO [train.py:904] (4/8) Epoch 11, batch 4350, loss[loss=0.2033, simple_loss=0.292, pruned_loss=0.05733, over 16789.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2894, pruned_loss=0.06007, over 3204420.43 frames. ], batch size: 83, lr: 6.24e-03, grad_scale: 8.0 2023-04-29 10:05:27,886 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105853.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:05:34,703 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=105858.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:05:37,425 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105860.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:06:38,335 INFO [train.py:904] (4/8) Epoch 11, batch 4400, loss[loss=0.2312, simple_loss=0.3008, pruned_loss=0.08084, over 11368.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2915, pruned_loss=0.06156, over 3188185.08 frames. ], batch size: 248, lr: 6.24e-03, grad_scale: 8.0 2023-04-29 10:07:32,128 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 10:07:40,706 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.666e+02 2.548e+02 2.924e+02 3.370e+02 5.572e+02, threshold=5.847e+02, percent-clipped=1.0 2023-04-29 10:07:48,892 INFO [train.py:904] (4/8) Epoch 11, batch 4450, loss[loss=0.2229, simple_loss=0.3132, pruned_loss=0.0663, over 16357.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2947, pruned_loss=0.06206, over 3195488.38 frames. ], batch size: 165, lr: 6.24e-03, grad_scale: 8.0 2023-04-29 10:09:04,901 INFO [train.py:904] (4/8) Epoch 11, batch 4500, loss[loss=0.1859, simple_loss=0.2668, pruned_loss=0.05255, over 16494.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2945, pruned_loss=0.06216, over 3200010.83 frames. ], batch size: 35, lr: 6.24e-03, grad_scale: 8.0 2023-04-29 10:09:41,036 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5618, 5.4945, 5.2828, 4.6441, 5.3952, 1.8195, 5.1121, 5.2081], device='cuda:4'), covar=tensor([0.0032, 0.0032, 0.0089, 0.0265, 0.0042, 0.2165, 0.0070, 0.0095], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0117, 0.0164, 0.0157, 0.0134, 0.0176, 0.0151, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:09:51,595 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-29 10:09:58,378 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0003, 4.1027, 2.1396, 4.9284, 3.0830, 4.7360, 2.5549, 3.0922], device='cuda:4'), covar=tensor([0.0177, 0.0246, 0.1610, 0.0067, 0.0683, 0.0288, 0.1262, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0164, 0.0186, 0.0128, 0.0166, 0.0208, 0.0193, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 10:10:05,602 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106044.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:10:09,392 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 1.945e+02 2.292e+02 2.586e+02 4.524e+02, threshold=4.584e+02, percent-clipped=0.0 2023-04-29 10:10:17,384 INFO [train.py:904] (4/8) Epoch 11, batch 4550, loss[loss=0.2158, simple_loss=0.2939, pruned_loss=0.06888, over 17223.00 frames. ], tot_loss[loss=0.2102, simple_loss=0.2948, pruned_loss=0.06279, over 3210639.19 frames. ], batch size: 45, lr: 6.24e-03, grad_scale: 8.0 2023-04-29 10:10:53,591 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=106077.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:11:15,421 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=106092.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:11:15,623 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0250, 3.9964, 1.9448, 4.8325, 3.0553, 4.6504, 2.2386, 3.0640], device='cuda:4'), covar=tensor([0.0152, 0.0220, 0.1717, 0.0053, 0.0654, 0.0203, 0.1531, 0.0610], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0163, 0.0185, 0.0127, 0.0165, 0.0206, 0.0192, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-29 10:11:25,349 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9966, 5.0758, 4.8284, 4.5165, 4.4410, 4.9474, 4.7775, 4.5600], device='cuda:4'), covar=tensor([0.0434, 0.0249, 0.0200, 0.0218, 0.0770, 0.0242, 0.0274, 0.0469], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0296, 0.0280, 0.0260, 0.0301, 0.0295, 0.0192, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:11:29,147 INFO [train.py:904] (4/8) Epoch 11, batch 4600, loss[loss=0.2055, simple_loss=0.2969, pruned_loss=0.05707, over 16839.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2953, pruned_loss=0.0626, over 3210181.54 frames. ], batch size: 83, lr: 6.24e-03, grad_scale: 8.0 2023-04-29 10:12:15,324 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-29 10:12:22,302 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=106138.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:12:28,024 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106142.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:12:31,856 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7228, 2.6049, 2.0870, 2.5957, 3.0976, 2.6944, 3.4058, 3.3289], device='cuda:4'), covar=tensor([0.0043, 0.0283, 0.0387, 0.0292, 0.0156, 0.0276, 0.0131, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0204, 0.0199, 0.0198, 0.0202, 0.0203, 0.0208, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:12:35,555 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.261e+02 1.926e+02 2.251e+02 2.695e+02 4.536e+02, threshold=4.502e+02, percent-clipped=0.0 2023-04-29 10:12:37,647 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106148.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:12:38,756 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3830, 5.6584, 5.3875, 5.4635, 5.0818, 4.9484, 5.0891, 5.8206], device='cuda:4'), covar=tensor([0.1005, 0.0667, 0.0907, 0.0629, 0.0757, 0.0672, 0.0869, 0.0685], device='cuda:4'), in_proj_covar=tensor([0.0526, 0.0655, 0.0541, 0.0454, 0.0416, 0.0422, 0.0544, 0.0500], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:12:41,942 INFO [train.py:904] (4/8) Epoch 11, batch 4650, loss[loss=0.1916, simple_loss=0.2766, pruned_loss=0.05337, over 16489.00 frames. ], tot_loss[loss=0.209, simple_loss=0.2937, pruned_loss=0.06212, over 3229115.18 frames. ], batch size: 68, lr: 6.23e-03, grad_scale: 8.0 2023-04-29 10:12:47,343 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106155.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:12:51,748 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106158.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:13:37,561 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=106190.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:13:55,745 INFO [train.py:904] (4/8) Epoch 11, batch 4700, loss[loss=0.1934, simple_loss=0.273, pruned_loss=0.05689, over 16596.00 frames. ], tot_loss[loss=0.207, simple_loss=0.2915, pruned_loss=0.06125, over 3225049.69 frames. ], batch size: 62, lr: 6.23e-03, grad_scale: 8.0 2023-04-29 10:14:01,835 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=106206.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:14:49,634 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1874, 4.1501, 4.1655, 3.0981, 4.0975, 1.4340, 3.8688, 3.7917], device='cuda:4'), covar=tensor([0.0134, 0.0121, 0.0145, 0.0577, 0.0113, 0.2804, 0.0159, 0.0296], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0116, 0.0163, 0.0157, 0.0133, 0.0176, 0.0150, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:15:01,675 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.393e+02 2.142e+02 2.451e+02 2.986e+02 5.737e+02, threshold=4.902e+02, percent-clipped=3.0 2023-04-29 10:15:03,995 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2250, 4.0995, 4.3154, 4.5167, 4.6203, 4.2291, 4.5725, 4.6228], device='cuda:4'), covar=tensor([0.1251, 0.1018, 0.1251, 0.0537, 0.0468, 0.0889, 0.0505, 0.0482], device='cuda:4'), in_proj_covar=tensor([0.0509, 0.0642, 0.0782, 0.0655, 0.0496, 0.0504, 0.0507, 0.0582], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:15:09,055 INFO [train.py:904] (4/8) Epoch 11, batch 4750, loss[loss=0.19, simple_loss=0.2764, pruned_loss=0.0518, over 16268.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.288, pruned_loss=0.05973, over 3208370.03 frames. ], batch size: 165, lr: 6.23e-03, grad_scale: 8.0 2023-04-29 10:15:32,882 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=106268.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:15:45,239 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6543, 4.0459, 4.3959, 1.7564, 4.3814, 4.6926, 3.1052, 3.6166], device='cuda:4'), covar=tensor([0.0805, 0.0178, 0.0179, 0.1387, 0.0179, 0.0062, 0.0396, 0.0362], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0100, 0.0090, 0.0138, 0.0070, 0.0102, 0.0121, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 10:16:07,484 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5082, 4.7820, 4.5413, 4.5988, 4.3202, 4.2978, 4.2402, 4.8469], device='cuda:4'), covar=tensor([0.1049, 0.0814, 0.0974, 0.0686, 0.0683, 0.1064, 0.0964, 0.0892], device='cuda:4'), in_proj_covar=tensor([0.0528, 0.0655, 0.0543, 0.0453, 0.0415, 0.0423, 0.0542, 0.0501], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:16:14,979 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2927, 3.9911, 3.9499, 2.3871, 3.3931, 3.8933, 3.6008, 1.9433], device='cuda:4'), covar=tensor([0.0428, 0.0028, 0.0029, 0.0345, 0.0075, 0.0080, 0.0072, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0069, 0.0070, 0.0125, 0.0078, 0.0089, 0.0076, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 10:16:22,066 INFO [train.py:904] (4/8) Epoch 11, batch 4800, loss[loss=0.2069, simple_loss=0.2972, pruned_loss=0.0583, over 16675.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2847, pruned_loss=0.05749, over 3215290.91 frames. ], batch size: 134, lr: 6.23e-03, grad_scale: 8.0 2023-04-29 10:16:45,126 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7682, 1.2767, 1.6147, 1.6677, 1.7887, 1.8791, 1.5589, 1.8298], device='cuda:4'), covar=tensor([0.0165, 0.0272, 0.0148, 0.0211, 0.0187, 0.0126, 0.0252, 0.0080], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0155, 0.0161, 0.0166, 0.0123, 0.0169, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 10:17:02,221 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=106329.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:17:12,310 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=106336.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:17:27,377 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.453e+02 2.167e+02 2.529e+02 2.965e+02 6.623e+02, threshold=5.058e+02, percent-clipped=3.0 2023-04-29 10:17:35,185 INFO [train.py:904] (4/8) Epoch 11, batch 4850, loss[loss=0.1881, simple_loss=0.2803, pruned_loss=0.04794, over 16797.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2854, pruned_loss=0.05725, over 3180918.24 frames. ], batch size: 83, lr: 6.23e-03, grad_scale: 8.0 2023-04-29 10:17:46,568 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9079, 4.6561, 4.9276, 5.1449, 5.3198, 4.7379, 5.3282, 5.2735], device='cuda:4'), covar=tensor([0.1260, 0.1119, 0.1410, 0.0586, 0.0408, 0.0655, 0.0383, 0.0563], device='cuda:4'), in_proj_covar=tensor([0.0503, 0.0635, 0.0775, 0.0651, 0.0494, 0.0500, 0.0504, 0.0578], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:17:49,037 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5533, 4.6313, 4.4070, 4.1650, 3.9935, 4.5027, 4.2777, 4.1443], device='cuda:4'), covar=tensor([0.0531, 0.0310, 0.0241, 0.0237, 0.0909, 0.0343, 0.0506, 0.0586], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0299, 0.0279, 0.0260, 0.0302, 0.0297, 0.0192, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:18:16,855 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-29 10:18:38,643 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=106397.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:18:45,470 INFO [train.py:904] (4/8) Epoch 11, batch 4900, loss[loss=0.1989, simple_loss=0.2799, pruned_loss=0.05889, over 16764.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2847, pruned_loss=0.05604, over 3181089.05 frames. ], batch size: 83, lr: 6.23e-03, grad_scale: 8.0 2023-04-29 10:18:49,710 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-29 10:19:16,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9011, 4.1910, 3.9535, 4.0408, 3.6664, 3.8256, 3.8014, 4.1418], device='cuda:4'), covar=tensor([0.1116, 0.0843, 0.0957, 0.0629, 0.0736, 0.1358, 0.0962, 0.1009], device='cuda:4'), in_proj_covar=tensor([0.0524, 0.0649, 0.0538, 0.0449, 0.0412, 0.0420, 0.0538, 0.0499], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:19:18,349 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-04-29 10:19:29,912 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106433.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:19:50,194 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.535e+02 2.157e+02 2.578e+02 2.904e+02 4.407e+02, threshold=5.156e+02, percent-clipped=0.0 2023-04-29 10:19:51,813 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106448.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:19:56,920 INFO [train.py:904] (4/8) Epoch 11, batch 4950, loss[loss=0.194, simple_loss=0.2909, pruned_loss=0.04851, over 16900.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.2841, pruned_loss=0.05527, over 3175460.84 frames. ], batch size: 102, lr: 6.22e-03, grad_scale: 8.0 2023-04-29 10:20:01,949 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=106454.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:20:02,977 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106455.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:20:04,249 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2106, 4.2893, 2.6243, 5.0298, 3.1951, 4.8222, 2.6191, 3.3036], device='cuda:4'), covar=tensor([0.0177, 0.0230, 0.1369, 0.0052, 0.0659, 0.0255, 0.1362, 0.0639], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0163, 0.0186, 0.0126, 0.0166, 0.0205, 0.0192, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-29 10:21:00,649 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=106496.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:21:08,412 INFO [train.py:904] (4/8) Epoch 11, batch 5000, loss[loss=0.1898, simple_loss=0.285, pruned_loss=0.04728, over 15380.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.2856, pruned_loss=0.05549, over 3180774.45 frames. ], batch size: 190, lr: 6.22e-03, grad_scale: 8.0 2023-04-29 10:21:10,910 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=106503.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:21:28,665 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=106515.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:22:14,238 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.886e+02 2.415e+02 2.751e+02 3.160e+02 5.367e+02, threshold=5.502e+02, percent-clipped=2.0 2023-04-29 10:22:21,433 INFO [train.py:904] (4/8) Epoch 11, batch 5050, loss[loss=0.1823, simple_loss=0.272, pruned_loss=0.04632, over 16514.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2852, pruned_loss=0.05479, over 3198241.65 frames. ], batch size: 68, lr: 6.22e-03, grad_scale: 8.0 2023-04-29 10:23:32,362 INFO [train.py:904] (4/8) Epoch 11, batch 5100, loss[loss=0.2306, simple_loss=0.3052, pruned_loss=0.07803, over 11954.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2842, pruned_loss=0.05463, over 3199749.03 frames. ], batch size: 247, lr: 6.22e-03, grad_scale: 8.0 2023-04-29 10:23:55,859 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4783, 4.5070, 4.7247, 4.5089, 4.5472, 5.0977, 4.6666, 4.3568], device='cuda:4'), covar=tensor([0.1002, 0.1620, 0.1488, 0.1662, 0.2413, 0.0906, 0.1261, 0.2281], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0467, 0.0505, 0.0406, 0.0537, 0.0536, 0.0403, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 10:24:03,609 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106624.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:24:38,775 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.440e+02 2.185e+02 2.500e+02 2.940e+02 7.724e+02, threshold=5.000e+02, percent-clipped=1.0 2023-04-29 10:24:44,356 INFO [train.py:904] (4/8) Epoch 11, batch 5150, loss[loss=0.2004, simple_loss=0.2986, pruned_loss=0.05112, over 15319.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.2846, pruned_loss=0.05407, over 3187276.65 frames. ], batch size: 191, lr: 6.22e-03, grad_scale: 4.0 2023-04-29 10:25:12,800 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6125, 2.2137, 1.8593, 2.0363, 2.5755, 2.3561, 2.6144, 2.7821], device='cuda:4'), covar=tensor([0.0122, 0.0298, 0.0380, 0.0358, 0.0150, 0.0264, 0.0153, 0.0196], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0199, 0.0195, 0.0193, 0.0197, 0.0200, 0.0201, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:25:13,970 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2924, 3.3577, 1.8370, 3.5656, 2.4651, 3.6152, 1.9747, 2.7370], device='cuda:4'), covar=tensor([0.0201, 0.0297, 0.1591, 0.0163, 0.0833, 0.0451, 0.1537, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0162, 0.0185, 0.0124, 0.0165, 0.0203, 0.0191, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-29 10:25:43,744 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106692.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:25:56,050 INFO [train.py:904] (4/8) Epoch 11, batch 5200, loss[loss=0.18, simple_loss=0.2653, pruned_loss=0.04735, over 16925.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2839, pruned_loss=0.05388, over 3188261.72 frames. ], batch size: 116, lr: 6.22e-03, grad_scale: 8.0 2023-04-29 10:26:42,777 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106733.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:27:04,335 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.660e+02 2.187e+02 2.709e+02 3.054e+02 5.980e+02, threshold=5.418e+02, percent-clipped=1.0 2023-04-29 10:27:11,356 INFO [train.py:904] (4/8) Epoch 11, batch 5250, loss[loss=0.2341, simple_loss=0.298, pruned_loss=0.08512, over 12749.00 frames. ], tot_loss[loss=0.194, simple_loss=0.281, pruned_loss=0.05354, over 3182070.28 frames. ], batch size: 246, lr: 6.22e-03, grad_scale: 8.0 2023-04-29 10:27:35,021 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-29 10:27:52,546 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=106781.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:28:22,266 INFO [train.py:904] (4/8) Epoch 11, batch 5300, loss[loss=0.147, simple_loss=0.2325, pruned_loss=0.03075, over 16810.00 frames. ], tot_loss[loss=0.1919, simple_loss=0.2782, pruned_loss=0.05282, over 3178559.92 frames. ], batch size: 42, lr: 6.21e-03, grad_scale: 8.0 2023-04-29 10:28:29,612 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3964, 4.2700, 4.4323, 4.6410, 4.7710, 4.3025, 4.7314, 4.7492], device='cuda:4'), covar=tensor([0.1301, 0.0911, 0.1328, 0.0530, 0.0406, 0.0920, 0.0416, 0.0459], device='cuda:4'), in_proj_covar=tensor([0.0512, 0.0644, 0.0788, 0.0659, 0.0500, 0.0506, 0.0509, 0.0584], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:28:32,955 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106810.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:28:47,194 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5282, 4.5934, 4.8347, 4.6052, 4.6699, 5.2180, 4.7487, 4.3771], device='cuda:4'), covar=tensor([0.1099, 0.1566, 0.1291, 0.1613, 0.2158, 0.0798, 0.1186, 0.2411], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0472, 0.0513, 0.0414, 0.0548, 0.0542, 0.0410, 0.0565], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 10:29:27,199 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.387e+02 2.160e+02 2.565e+02 2.956e+02 5.223e+02, threshold=5.130e+02, percent-clipped=0.0 2023-04-29 10:29:33,906 INFO [train.py:904] (4/8) Epoch 11, batch 5350, loss[loss=0.2031, simple_loss=0.2893, pruned_loss=0.05848, over 16685.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2756, pruned_loss=0.05133, over 3191108.82 frames. ], batch size: 62, lr: 6.21e-03, grad_scale: 8.0 2023-04-29 10:30:45,878 INFO [train.py:904] (4/8) Epoch 11, batch 5400, loss[loss=0.2337, simple_loss=0.3072, pruned_loss=0.08004, over 12159.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2779, pruned_loss=0.05209, over 3183626.00 frames. ], batch size: 246, lr: 6.21e-03, grad_scale: 8.0 2023-04-29 10:30:51,519 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2713, 4.0928, 4.2781, 4.4359, 4.5773, 4.1620, 4.5218, 4.5693], device='cuda:4'), covar=tensor([0.1315, 0.0960, 0.1381, 0.0628, 0.0472, 0.1065, 0.0515, 0.0524], device='cuda:4'), in_proj_covar=tensor([0.0518, 0.0650, 0.0793, 0.0666, 0.0507, 0.0513, 0.0512, 0.0591], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:31:18,225 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106924.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:31:54,541 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.468e+02 2.177e+02 2.658e+02 3.216e+02 5.881e+02, threshold=5.316e+02, percent-clipped=3.0 2023-04-29 10:32:02,030 INFO [train.py:904] (4/8) Epoch 11, batch 5450, loss[loss=0.2336, simple_loss=0.3085, pruned_loss=0.07932, over 12041.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2809, pruned_loss=0.05366, over 3166781.26 frames. ], batch size: 250, lr: 6.21e-03, grad_scale: 8.0 2023-04-29 10:32:11,887 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7680, 1.2604, 1.5995, 1.6586, 1.7163, 1.8690, 1.4810, 1.7848], device='cuda:4'), covar=tensor([0.0194, 0.0263, 0.0142, 0.0192, 0.0177, 0.0111, 0.0255, 0.0069], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0173, 0.0155, 0.0162, 0.0168, 0.0124, 0.0170, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 10:32:34,433 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=106972.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:33:03,912 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106992.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:33:18,592 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1747, 4.1585, 4.0880, 3.4641, 4.0818, 1.5827, 3.9061, 3.8187], device='cuda:4'), covar=tensor([0.0093, 0.0087, 0.0132, 0.0309, 0.0082, 0.2436, 0.0112, 0.0192], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0116, 0.0164, 0.0159, 0.0135, 0.0177, 0.0150, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:33:19,270 INFO [train.py:904] (4/8) Epoch 11, batch 5500, loss[loss=0.2361, simple_loss=0.3132, pruned_loss=0.07956, over 16268.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.2892, pruned_loss=0.05915, over 3142976.13 frames. ], batch size: 165, lr: 6.21e-03, grad_scale: 8.0 2023-04-29 10:34:18,908 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=107040.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:34:31,622 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.975e+02 3.119e+02 3.959e+02 4.735e+02 9.206e+02, threshold=7.917e+02, percent-clipped=14.0 2023-04-29 10:34:37,901 INFO [train.py:904] (4/8) Epoch 11, batch 5550, loss[loss=0.3231, simple_loss=0.3768, pruned_loss=0.1347, over 11182.00 frames. ], tot_loss[loss=0.2136, simple_loss=0.297, pruned_loss=0.06505, over 3117727.71 frames. ], batch size: 247, lr: 6.21e-03, grad_scale: 8.0 2023-04-29 10:35:57,901 INFO [train.py:904] (4/8) Epoch 11, batch 5600, loss[loss=0.3326, simple_loss=0.3818, pruned_loss=0.1417, over 11435.00 frames. ], tot_loss[loss=0.2207, simple_loss=0.3024, pruned_loss=0.0695, over 3090293.02 frames. ], batch size: 247, lr: 6.21e-03, grad_scale: 8.0 2023-04-29 10:36:12,122 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=107110.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:37:02,618 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107140.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 10:37:17,301 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.615e+02 3.503e+02 4.339e+02 5.380e+02 1.040e+03, threshold=8.679e+02, percent-clipped=3.0 2023-04-29 10:37:21,711 INFO [train.py:904] (4/8) Epoch 11, batch 5650, loss[loss=0.1938, simple_loss=0.2874, pruned_loss=0.05011, over 16734.00 frames. ], tot_loss[loss=0.2278, simple_loss=0.3078, pruned_loss=0.07396, over 3078014.46 frames. ], batch size: 89, lr: 6.20e-03, grad_scale: 4.0 2023-04-29 10:37:22,228 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107152.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:37:32,841 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=107158.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:38:29,829 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3484, 3.3842, 1.8638, 3.6984, 2.4413, 3.7163, 1.9906, 2.6639], device='cuda:4'), covar=tensor([0.0218, 0.0341, 0.1566, 0.0153, 0.0811, 0.0451, 0.1459, 0.0692], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0163, 0.0186, 0.0125, 0.0166, 0.0203, 0.0192, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 10:38:42,728 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107201.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 10:38:43,535 INFO [train.py:904] (4/8) Epoch 11, batch 5700, loss[loss=0.2691, simple_loss=0.3458, pruned_loss=0.09619, over 15223.00 frames. ], tot_loss[loss=0.2293, simple_loss=0.3086, pruned_loss=0.07505, over 3084763.00 frames. ], batch size: 190, lr: 6.20e-03, grad_scale: 4.0 2023-04-29 10:38:47,181 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4835, 3.4777, 2.7096, 2.1829, 2.6201, 2.2643, 3.8334, 3.3654], device='cuda:4'), covar=tensor([0.2614, 0.0769, 0.1643, 0.2090, 0.2242, 0.1777, 0.0415, 0.1040], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0255, 0.0281, 0.0275, 0.0280, 0.0219, 0.0265, 0.0291], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:39:02,874 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107213.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 10:39:34,266 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 10:39:59,317 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.558e+02 3.709e+02 4.582e+02 6.080e+02 1.249e+03, threshold=9.164e+02, percent-clipped=1.0 2023-04-29 10:40:04,436 INFO [train.py:904] (4/8) Epoch 11, batch 5750, loss[loss=0.2267, simple_loss=0.3085, pruned_loss=0.07245, over 16768.00 frames. ], tot_loss[loss=0.2326, simple_loss=0.3122, pruned_loss=0.07647, over 3090725.16 frames. ], batch size: 83, lr: 6.20e-03, grad_scale: 4.0 2023-04-29 10:40:35,485 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6767, 2.6203, 1.8911, 2.6950, 2.2334, 2.7783, 2.1103, 2.4491], device='cuda:4'), covar=tensor([0.0265, 0.0405, 0.1191, 0.0214, 0.0639, 0.0511, 0.1170, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0164, 0.0187, 0.0126, 0.0168, 0.0205, 0.0195, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 10:41:25,722 INFO [train.py:904] (4/8) Epoch 11, batch 5800, loss[loss=0.2174, simple_loss=0.3078, pruned_loss=0.06344, over 16630.00 frames. ], tot_loss[loss=0.2309, simple_loss=0.311, pruned_loss=0.07535, over 3065848.32 frames. ], batch size: 57, lr: 6.20e-03, grad_scale: 4.0 2023-04-29 10:41:59,226 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-29 10:42:39,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.001e+02 2.956e+02 3.560e+02 4.660e+02 9.709e+02, threshold=7.120e+02, percent-clipped=1.0 2023-04-29 10:42:43,688 INFO [train.py:904] (4/8) Epoch 11, batch 5850, loss[loss=0.1969, simple_loss=0.2828, pruned_loss=0.05554, over 16534.00 frames. ], tot_loss[loss=0.2278, simple_loss=0.3087, pruned_loss=0.07348, over 3069592.04 frames. ], batch size: 68, lr: 6.20e-03, grad_scale: 4.0 2023-04-29 10:42:52,965 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5825, 5.9069, 5.6192, 5.6960, 5.3040, 5.1849, 5.4097, 6.0406], device='cuda:4'), covar=tensor([0.0999, 0.0746, 0.0852, 0.0676, 0.0761, 0.0574, 0.0873, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0539, 0.0664, 0.0552, 0.0457, 0.0421, 0.0432, 0.0553, 0.0507], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:43:19,697 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.84 vs. limit=5.0 2023-04-29 10:44:05,220 INFO [train.py:904] (4/8) Epoch 11, batch 5900, loss[loss=0.2786, simple_loss=0.3334, pruned_loss=0.1119, over 11350.00 frames. ], tot_loss[loss=0.2276, simple_loss=0.3088, pruned_loss=0.07321, over 3065415.70 frames. ], batch size: 247, lr: 6.20e-03, grad_scale: 4.0 2023-04-29 10:45:21,995 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.026e+02 2.914e+02 3.759e+02 4.619e+02 1.179e+03, threshold=7.519e+02, percent-clipped=2.0 2023-04-29 10:45:26,030 INFO [train.py:904] (4/8) Epoch 11, batch 5950, loss[loss=0.1973, simple_loss=0.2931, pruned_loss=0.05081, over 16866.00 frames. ], tot_loss[loss=0.2262, simple_loss=0.3089, pruned_loss=0.07177, over 3088206.57 frames. ], batch size: 96, lr: 6.20e-03, grad_scale: 4.0 2023-04-29 10:45:39,020 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5860, 3.7520, 4.1006, 2.0074, 4.2696, 4.2905, 3.0636, 3.1820], device='cuda:4'), covar=tensor([0.0659, 0.0158, 0.0108, 0.1029, 0.0045, 0.0082, 0.0293, 0.0357], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0098, 0.0087, 0.0137, 0.0069, 0.0101, 0.0118, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 10:46:40,585 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=107496.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 10:46:48,947 INFO [train.py:904] (4/8) Epoch 11, batch 6000, loss[loss=0.2019, simple_loss=0.2883, pruned_loss=0.0577, over 16872.00 frames. ], tot_loss[loss=0.225, simple_loss=0.3076, pruned_loss=0.07121, over 3094010.89 frames. ], batch size: 96, lr: 6.19e-03, grad_scale: 8.0 2023-04-29 10:46:48,947 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 10:46:59,887 INFO [train.py:938] (4/8) Epoch 11, validation: loss=0.163, simple_loss=0.2761, pruned_loss=0.02492, over 944034.00 frames. 2023-04-29 10:46:59,888 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 10:47:06,077 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9409, 4.2876, 3.3622, 2.4507, 3.1593, 2.6190, 4.7008, 4.1048], device='cuda:4'), covar=tensor([0.2479, 0.0642, 0.1423, 0.1875, 0.2217, 0.1583, 0.0335, 0.0781], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0256, 0.0281, 0.0276, 0.0281, 0.0219, 0.0266, 0.0291], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:47:10,234 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=107508.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 10:47:13,526 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107510.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:47:13,688 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1253, 1.9748, 2.1432, 3.6892, 1.9838, 2.4041, 2.1419, 2.1611], device='cuda:4'), covar=tensor([0.0990, 0.3122, 0.2164, 0.0435, 0.3627, 0.2148, 0.2834, 0.3013], device='cuda:4'), in_proj_covar=tensor([0.0358, 0.0385, 0.0323, 0.0315, 0.0406, 0.0443, 0.0350, 0.0452], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:48:07,130 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-04-29 10:48:12,719 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.272e+02 3.327e+02 3.927e+02 4.902e+02 9.680e+02, threshold=7.854e+02, percent-clipped=6.0 2023-04-29 10:48:18,415 INFO [train.py:904] (4/8) Epoch 11, batch 6050, loss[loss=0.2106, simple_loss=0.3062, pruned_loss=0.05746, over 17029.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.3066, pruned_loss=0.07091, over 3091434.73 frames. ], batch size: 55, lr: 6.19e-03, grad_scale: 8.0 2023-04-29 10:48:48,369 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107571.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:49:28,893 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-29 10:49:34,932 INFO [train.py:904] (4/8) Epoch 11, batch 6100, loss[loss=0.3188, simple_loss=0.3686, pruned_loss=0.1346, over 11135.00 frames. ], tot_loss[loss=0.2232, simple_loss=0.3063, pruned_loss=0.07008, over 3095596.95 frames. ], batch size: 246, lr: 6.19e-03, grad_scale: 8.0 2023-04-29 10:50:51,420 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.904e+02 2.980e+02 3.646e+02 4.413e+02 7.935e+02, threshold=7.292e+02, percent-clipped=3.0 2023-04-29 10:50:56,527 INFO [train.py:904] (4/8) Epoch 11, batch 6150, loss[loss=0.2098, simple_loss=0.2971, pruned_loss=0.0613, over 16743.00 frames. ], tot_loss[loss=0.2213, simple_loss=0.3039, pruned_loss=0.06934, over 3087747.22 frames. ], batch size: 83, lr: 6.19e-03, grad_scale: 8.0 2023-04-29 10:51:07,615 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107659.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:51:11,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1717, 3.3470, 3.5382, 3.5244, 3.5289, 3.3400, 3.3531, 3.4042], device='cuda:4'), covar=tensor([0.0355, 0.0614, 0.0405, 0.0423, 0.0491, 0.0452, 0.0879, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0341, 0.0343, 0.0325, 0.0388, 0.0361, 0.0463, 0.0294], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 10:51:36,454 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9922, 2.7597, 2.7848, 2.1274, 2.6373, 2.1993, 2.7583, 2.9200], device='cuda:4'), covar=tensor([0.0242, 0.0585, 0.0437, 0.1447, 0.0639, 0.0802, 0.0503, 0.0665], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0142, 0.0156, 0.0143, 0.0135, 0.0124, 0.0136, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 10:52:14,168 INFO [train.py:904] (4/8) Epoch 11, batch 6200, loss[loss=0.2199, simple_loss=0.3059, pruned_loss=0.067, over 16518.00 frames. ], tot_loss[loss=0.22, simple_loss=0.3022, pruned_loss=0.06897, over 3091912.46 frames. ], batch size: 62, lr: 6.19e-03, grad_scale: 4.0 2023-04-29 10:52:17,841 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3817, 3.6194, 3.6581, 2.0939, 3.1095, 2.5129, 3.6881, 3.8190], device='cuda:4'), covar=tensor([0.0233, 0.0671, 0.0491, 0.1793, 0.0742, 0.0856, 0.0613, 0.0882], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0142, 0.0156, 0.0143, 0.0135, 0.0124, 0.0136, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 10:52:20,365 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.36 vs. limit=5.0 2023-04-29 10:52:39,383 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8399, 4.1353, 3.9161, 3.9698, 3.6208, 3.7334, 3.8349, 4.1141], device='cuda:4'), covar=tensor([0.1082, 0.0823, 0.0997, 0.0704, 0.0812, 0.1385, 0.0908, 0.0970], device='cuda:4'), in_proj_covar=tensor([0.0547, 0.0674, 0.0564, 0.0466, 0.0426, 0.0438, 0.0562, 0.0514], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:52:42,553 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107720.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:52:57,579 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0922, 3.7530, 3.7889, 2.4820, 3.4175, 3.7955, 3.5941, 2.1189], device='cuda:4'), covar=tensor([0.0455, 0.0034, 0.0035, 0.0327, 0.0070, 0.0075, 0.0060, 0.0371], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0068, 0.0069, 0.0125, 0.0077, 0.0090, 0.0076, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 10:53:07,928 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107737.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:53:27,428 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.232e+02 3.321e+02 3.959e+02 5.175e+02 1.024e+03, threshold=7.919e+02, percent-clipped=5.0 2023-04-29 10:53:29,995 INFO [train.py:904] (4/8) Epoch 11, batch 6250, loss[loss=0.2149, simple_loss=0.3017, pruned_loss=0.06401, over 15415.00 frames. ], tot_loss[loss=0.2196, simple_loss=0.3016, pruned_loss=0.06878, over 3092353.01 frames. ], batch size: 191, lr: 6.19e-03, grad_scale: 4.0 2023-04-29 10:53:55,913 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0016, 2.3523, 2.3784, 2.9724, 2.0961, 3.2567, 1.7559, 2.6961], device='cuda:4'), covar=tensor([0.1092, 0.0503, 0.0923, 0.0141, 0.0130, 0.0389, 0.1245, 0.0651], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0157, 0.0180, 0.0143, 0.0197, 0.0207, 0.0179, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 10:54:36,283 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=107796.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 10:54:37,540 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4946, 3.5728, 3.2848, 3.0661, 3.1037, 3.4280, 3.2658, 3.1941], device='cuda:4'), covar=tensor([0.0575, 0.0515, 0.0253, 0.0239, 0.0540, 0.0404, 0.1461, 0.0457], device='cuda:4'), in_proj_covar=tensor([0.0237, 0.0303, 0.0279, 0.0258, 0.0301, 0.0298, 0.0193, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:54:39,013 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107798.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 10:54:42,618 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1878, 1.9414, 1.6397, 1.7543, 2.2087, 1.9332, 1.9838, 2.3542], device='cuda:4'), covar=tensor([0.0110, 0.0256, 0.0337, 0.0320, 0.0160, 0.0251, 0.0148, 0.0159], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0200, 0.0197, 0.0197, 0.0200, 0.0201, 0.0205, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:54:45,102 INFO [train.py:904] (4/8) Epoch 11, batch 6300, loss[loss=0.2189, simple_loss=0.2997, pruned_loss=0.06904, over 16720.00 frames. ], tot_loss[loss=0.2189, simple_loss=0.3017, pruned_loss=0.06808, over 3101747.26 frames. ], batch size: 134, lr: 6.19e-03, grad_scale: 4.0 2023-04-29 10:54:54,829 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=107808.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:55:52,343 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=107844.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 10:56:00,513 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.985e+02 3.066e+02 3.848e+02 4.977e+02 9.936e+02, threshold=7.696e+02, percent-clipped=3.0 2023-04-29 10:56:03,040 INFO [train.py:904] (4/8) Epoch 11, batch 6350, loss[loss=0.2412, simple_loss=0.3138, pruned_loss=0.08427, over 15458.00 frames. ], tot_loss[loss=0.2205, simple_loss=0.3024, pruned_loss=0.06927, over 3092810.09 frames. ], batch size: 191, lr: 6.18e-03, grad_scale: 4.0 2023-04-29 10:56:10,569 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=107856.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:56:25,828 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=107866.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 10:57:20,955 INFO [train.py:904] (4/8) Epoch 11, batch 6400, loss[loss=0.2041, simple_loss=0.2794, pruned_loss=0.06443, over 16305.00 frames. ], tot_loss[loss=0.2224, simple_loss=0.3033, pruned_loss=0.07072, over 3088012.70 frames. ], batch size: 35, lr: 6.18e-03, grad_scale: 8.0 2023-04-29 10:57:26,284 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-29 10:57:32,748 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0077, 3.7424, 3.7406, 4.1889, 4.2586, 3.9866, 4.1497, 4.2640], device='cuda:4'), covar=tensor([0.1473, 0.1257, 0.2337, 0.1004, 0.0884, 0.1376, 0.1220, 0.1050], device='cuda:4'), in_proj_covar=tensor([0.0510, 0.0636, 0.0767, 0.0645, 0.0495, 0.0498, 0.0513, 0.0577], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:58:23,906 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-29 10:58:35,864 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.893e+02 3.293e+02 4.168e+02 5.156e+02 1.369e+03, threshold=8.335e+02, percent-clipped=6.0 2023-04-29 10:58:35,879 INFO [train.py:904] (4/8) Epoch 11, batch 6450, loss[loss=0.2311, simple_loss=0.314, pruned_loss=0.07408, over 16801.00 frames. ], tot_loss[loss=0.221, simple_loss=0.303, pruned_loss=0.06953, over 3093068.29 frames. ], batch size: 83, lr: 6.18e-03, grad_scale: 2.0 2023-04-29 10:58:52,462 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2308, 1.9702, 1.6540, 1.7993, 2.2325, 1.9260, 2.0644, 2.3742], device='cuda:4'), covar=tensor([0.0138, 0.0253, 0.0382, 0.0335, 0.0176, 0.0270, 0.0177, 0.0204], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0198, 0.0196, 0.0196, 0.0198, 0.0200, 0.0203, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 10:59:54,943 INFO [train.py:904] (4/8) Epoch 11, batch 6500, loss[loss=0.1813, simple_loss=0.2666, pruned_loss=0.04801, over 16690.00 frames. ], tot_loss[loss=0.2192, simple_loss=0.3005, pruned_loss=0.06893, over 3082988.87 frames. ], batch size: 89, lr: 6.18e-03, grad_scale: 2.0 2023-04-29 11:00:14,833 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108015.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:00:48,139 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0531, 2.3106, 2.4307, 2.8370, 2.0758, 3.1973, 1.7345, 2.7157], device='cuda:4'), covar=tensor([0.1018, 0.0500, 0.0922, 0.0128, 0.0139, 0.0317, 0.1251, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0158, 0.0182, 0.0143, 0.0199, 0.0208, 0.0180, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 11:01:12,918 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.008e+02 3.175e+02 3.633e+02 4.590e+02 1.124e+03, threshold=7.266e+02, percent-clipped=3.0 2023-04-29 11:01:12,933 INFO [train.py:904] (4/8) Epoch 11, batch 6550, loss[loss=0.2169, simple_loss=0.3114, pruned_loss=0.06118, over 16658.00 frames. ], tot_loss[loss=0.2213, simple_loss=0.3031, pruned_loss=0.06978, over 3076799.13 frames. ], batch size: 134, lr: 6.18e-03, grad_scale: 2.0 2023-04-29 11:01:26,260 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1116, 1.4661, 1.7952, 2.0496, 2.1966, 2.2817, 1.6341, 2.3006], device='cuda:4'), covar=tensor([0.0170, 0.0351, 0.0188, 0.0235, 0.0210, 0.0156, 0.0319, 0.0093], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0171, 0.0150, 0.0158, 0.0168, 0.0124, 0.0168, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 11:02:13,589 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108093.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 11:02:25,858 INFO [train.py:904] (4/8) Epoch 11, batch 6600, loss[loss=0.2126, simple_loss=0.3061, pruned_loss=0.05961, over 16482.00 frames. ], tot_loss[loss=0.2224, simple_loss=0.3053, pruned_loss=0.06978, over 3108190.27 frames. ], batch size: 68, lr: 6.18e-03, grad_scale: 2.0 2023-04-29 11:02:41,079 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1861, 3.3333, 3.5893, 3.5285, 3.5417, 3.3313, 3.3835, 3.4223], device='cuda:4'), covar=tensor([0.0370, 0.0577, 0.0387, 0.0420, 0.0487, 0.0508, 0.0813, 0.0473], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0341, 0.0343, 0.0322, 0.0389, 0.0360, 0.0461, 0.0291], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 11:03:41,606 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.097e+02 3.251e+02 4.038e+02 4.757e+02 1.371e+03, threshold=8.077e+02, percent-clipped=5.0 2023-04-29 11:03:41,625 INFO [train.py:904] (4/8) Epoch 11, batch 6650, loss[loss=0.2825, simple_loss=0.3448, pruned_loss=0.1101, over 11675.00 frames. ], tot_loss[loss=0.2237, simple_loss=0.306, pruned_loss=0.0707, over 3097694.84 frames. ], batch size: 248, lr: 6.18e-03, grad_scale: 2.0 2023-04-29 11:03:51,284 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-29 11:04:03,296 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=108166.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:04:56,954 INFO [train.py:904] (4/8) Epoch 11, batch 6700, loss[loss=0.2524, simple_loss=0.3222, pruned_loss=0.09131, over 16962.00 frames. ], tot_loss[loss=0.2224, simple_loss=0.3039, pruned_loss=0.07043, over 3088566.70 frames. ], batch size: 109, lr: 6.17e-03, grad_scale: 2.0 2023-04-29 11:05:14,876 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=108214.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:05:28,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9886, 3.9672, 4.3870, 4.3616, 4.3673, 4.0773, 4.0991, 3.9949], device='cuda:4'), covar=tensor([0.0327, 0.0562, 0.0406, 0.0450, 0.0487, 0.0365, 0.0806, 0.0474], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0341, 0.0346, 0.0323, 0.0392, 0.0360, 0.0462, 0.0291], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 11:06:13,521 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.884e+02 3.397e+02 4.413e+02 5.907e+02 2.679e+03, threshold=8.826e+02, percent-clipped=7.0 2023-04-29 11:06:13,536 INFO [train.py:904] (4/8) Epoch 11, batch 6750, loss[loss=0.197, simple_loss=0.2832, pruned_loss=0.05538, over 16737.00 frames. ], tot_loss[loss=0.2214, simple_loss=0.3026, pruned_loss=0.07007, over 3095053.86 frames. ], batch size: 124, lr: 6.17e-03, grad_scale: 2.0 2023-04-29 11:07:28,508 INFO [train.py:904] (4/8) Epoch 11, batch 6800, loss[loss=0.2762, simple_loss=0.3309, pruned_loss=0.1107, over 11387.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.303, pruned_loss=0.07034, over 3076736.06 frames. ], batch size: 246, lr: 6.17e-03, grad_scale: 4.0 2023-04-29 11:07:49,236 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=108315.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:08:05,602 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=108325.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:08:21,326 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9154, 4.1368, 3.9613, 3.9822, 3.6398, 3.7418, 3.8784, 4.1165], device='cuda:4'), covar=tensor([0.1002, 0.0824, 0.0936, 0.0670, 0.0749, 0.1585, 0.0842, 0.0959], device='cuda:4'), in_proj_covar=tensor([0.0541, 0.0660, 0.0551, 0.0457, 0.0419, 0.0432, 0.0554, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:08:21,447 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1418, 2.5842, 2.6301, 1.8453, 2.7860, 2.8206, 2.4044, 2.4341], device='cuda:4'), covar=tensor([0.0601, 0.0193, 0.0214, 0.0948, 0.0096, 0.0190, 0.0403, 0.0379], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0099, 0.0088, 0.0138, 0.0069, 0.0102, 0.0120, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 11:08:45,531 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.125e+02 2.977e+02 3.381e+02 4.053e+02 7.016e+02, threshold=6.762e+02, percent-clipped=0.0 2023-04-29 11:08:45,546 INFO [train.py:904] (4/8) Epoch 11, batch 6850, loss[loss=0.216, simple_loss=0.316, pruned_loss=0.05803, over 17282.00 frames. ], tot_loss[loss=0.2227, simple_loss=0.3044, pruned_loss=0.0705, over 3088637.87 frames. ], batch size: 52, lr: 6.17e-03, grad_scale: 4.0 2023-04-29 11:09:01,708 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=108363.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:09:07,962 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7805, 3.7930, 4.0878, 1.8380, 4.3087, 4.3669, 3.2079, 3.1007], device='cuda:4'), covar=tensor([0.0652, 0.0171, 0.0120, 0.1224, 0.0043, 0.0082, 0.0300, 0.0430], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0099, 0.0088, 0.0139, 0.0069, 0.0102, 0.0120, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 11:09:35,681 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=108386.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:09:46,464 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=108393.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:09:59,715 INFO [train.py:904] (4/8) Epoch 11, batch 6900, loss[loss=0.2551, simple_loss=0.3337, pruned_loss=0.0882, over 16499.00 frames. ], tot_loss[loss=0.224, simple_loss=0.3074, pruned_loss=0.07025, over 3097600.90 frames. ], batch size: 146, lr: 6.17e-03, grad_scale: 2.0 2023-04-29 11:10:12,496 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6155, 2.8035, 2.3361, 4.0643, 2.8616, 3.9666, 1.3582, 2.7373], device='cuda:4'), covar=tensor([0.1417, 0.0659, 0.1228, 0.0144, 0.0243, 0.0384, 0.1649, 0.0843], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0159, 0.0182, 0.0144, 0.0201, 0.0209, 0.0182, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 11:10:45,872 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=108432.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:10:59,697 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=108441.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:11:15,788 INFO [train.py:904] (4/8) Epoch 11, batch 6950, loss[loss=0.2211, simple_loss=0.2998, pruned_loss=0.07117, over 16550.00 frames. ], tot_loss[loss=0.2281, simple_loss=0.3099, pruned_loss=0.07315, over 3067144.83 frames. ], batch size: 62, lr: 6.17e-03, grad_scale: 2.0 2023-04-29 11:11:17,883 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.069e+02 2.933e+02 3.744e+02 4.621e+02 9.342e+02, threshold=7.489e+02, percent-clipped=9.0 2023-04-29 11:11:31,128 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4659, 3.4509, 3.4122, 2.8593, 3.3390, 2.1598, 3.1397, 2.8414], device='cuda:4'), covar=tensor([0.0116, 0.0097, 0.0143, 0.0226, 0.0088, 0.1744, 0.0113, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0113, 0.0160, 0.0151, 0.0131, 0.0175, 0.0146, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:12:20,836 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=108493.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 11:12:33,547 INFO [train.py:904] (4/8) Epoch 11, batch 7000, loss[loss=0.2139, simple_loss=0.3071, pruned_loss=0.06031, over 17073.00 frames. ], tot_loss[loss=0.2256, simple_loss=0.3088, pruned_loss=0.07114, over 3100888.31 frames. ], batch size: 53, lr: 6.17e-03, grad_scale: 2.0 2023-04-29 11:12:56,340 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.72 vs. limit=5.0 2023-04-29 11:13:19,954 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-29 11:13:52,274 INFO [train.py:904] (4/8) Epoch 11, batch 7050, loss[loss=0.222, simple_loss=0.3086, pruned_loss=0.06771, over 16599.00 frames. ], tot_loss[loss=0.2258, simple_loss=0.3095, pruned_loss=0.07102, over 3103170.47 frames. ], batch size: 57, lr: 6.16e-03, grad_scale: 2.0 2023-04-29 11:13:53,482 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.056e+02 3.044e+02 3.927e+02 4.822e+02 1.171e+03, threshold=7.854e+02, percent-clipped=4.0 2023-04-29 11:13:59,957 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-29 11:14:20,341 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8837, 4.1492, 3.8318, 3.6757, 3.2039, 4.0492, 3.7423, 3.6558], device='cuda:4'), covar=tensor([0.0893, 0.0467, 0.0460, 0.0382, 0.1565, 0.0460, 0.1040, 0.0664], device='cuda:4'), in_proj_covar=tensor([0.0231, 0.0298, 0.0274, 0.0251, 0.0294, 0.0292, 0.0191, 0.0318], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:14:33,471 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=108577.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:14:39,224 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3950, 4.3469, 4.2601, 3.6258, 4.2554, 1.6634, 4.0423, 4.0983], device='cuda:4'), covar=tensor([0.0087, 0.0078, 0.0143, 0.0308, 0.0080, 0.2313, 0.0117, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0113, 0.0161, 0.0152, 0.0132, 0.0177, 0.0148, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:15:11,199 INFO [train.py:904] (4/8) Epoch 11, batch 7100, loss[loss=0.2176, simple_loss=0.2991, pruned_loss=0.06806, over 15428.00 frames. ], tot_loss[loss=0.225, simple_loss=0.3077, pruned_loss=0.07115, over 3073230.66 frames. ], batch size: 190, lr: 6.16e-03, grad_scale: 2.0 2023-04-29 11:16:02,719 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8835, 2.7016, 2.7383, 1.9428, 2.5263, 2.6828, 2.6564, 1.9089], device='cuda:4'), covar=tensor([0.0333, 0.0046, 0.0047, 0.0291, 0.0097, 0.0091, 0.0066, 0.0305], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0068, 0.0068, 0.0125, 0.0077, 0.0090, 0.0077, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 11:16:07,299 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=108638.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:16:27,131 INFO [train.py:904] (4/8) Epoch 11, batch 7150, loss[loss=0.2334, simple_loss=0.3097, pruned_loss=0.07858, over 16317.00 frames. ], tot_loss[loss=0.2236, simple_loss=0.3056, pruned_loss=0.07084, over 3089359.26 frames. ], batch size: 146, lr: 6.16e-03, grad_scale: 2.0 2023-04-29 11:16:28,930 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.138e+02 3.054e+02 3.516e+02 4.486e+02 8.068e+02, threshold=7.031e+02, percent-clipped=2.0 2023-04-29 11:17:10,532 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108681.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:17:15,808 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.09 vs. limit=5.0 2023-04-29 11:17:29,122 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-29 11:17:41,815 INFO [train.py:904] (4/8) Epoch 11, batch 7200, loss[loss=0.1886, simple_loss=0.2814, pruned_loss=0.04791, over 16673.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.3029, pruned_loss=0.0686, over 3086667.12 frames. ], batch size: 134, lr: 6.16e-03, grad_scale: 4.0 2023-04-29 11:19:02,058 INFO [train.py:904] (4/8) Epoch 11, batch 7250, loss[loss=0.2082, simple_loss=0.2905, pruned_loss=0.06288, over 15404.00 frames. ], tot_loss[loss=0.2173, simple_loss=0.3005, pruned_loss=0.06702, over 3099982.70 frames. ], batch size: 191, lr: 6.16e-03, grad_scale: 4.0 2023-04-29 11:19:03,145 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.903e+02 3.497e+02 4.380e+02 7.660e+02, threshold=6.994e+02, percent-clipped=1.0 2023-04-29 11:19:34,212 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6430, 4.1364, 4.3729, 2.3023, 3.4943, 2.9971, 4.0370, 4.2027], device='cuda:4'), covar=tensor([0.0224, 0.0523, 0.0398, 0.1654, 0.0663, 0.0720, 0.0513, 0.0768], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0145, 0.0159, 0.0146, 0.0139, 0.0127, 0.0138, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 11:19:55,675 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108788.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 11:20:01,010 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=108791.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:20:16,098 INFO [train.py:904] (4/8) Epoch 11, batch 7300, loss[loss=0.2367, simple_loss=0.3231, pruned_loss=0.07514, over 15389.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.3007, pruned_loss=0.06761, over 3096949.95 frames. ], batch size: 190, lr: 6.16e-03, grad_scale: 4.0 2023-04-29 11:21:34,169 INFO [train.py:904] (4/8) Epoch 11, batch 7350, loss[loss=0.2587, simple_loss=0.3194, pruned_loss=0.09903, over 11122.00 frames. ], tot_loss[loss=0.2197, simple_loss=0.3018, pruned_loss=0.06886, over 3060474.93 frames. ], batch size: 248, lr: 6.16e-03, grad_scale: 4.0 2023-04-29 11:21:34,692 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=108852.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:21:35,287 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.782e+02 3.104e+02 3.844e+02 4.749e+02 1.066e+03, threshold=7.688e+02, percent-clipped=6.0 2023-04-29 11:21:38,321 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2422, 2.0622, 2.0773, 4.0047, 2.0960, 2.4322, 2.2208, 2.2600], device='cuda:4'), covar=tensor([0.0981, 0.3151, 0.2405, 0.0424, 0.3740, 0.2214, 0.2940, 0.3129], device='cuda:4'), in_proj_covar=tensor([0.0360, 0.0389, 0.0326, 0.0317, 0.0413, 0.0448, 0.0354, 0.0458], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:22:54,323 INFO [train.py:904] (4/8) Epoch 11, batch 7400, loss[loss=0.2236, simple_loss=0.3104, pruned_loss=0.06841, over 15402.00 frames. ], tot_loss[loss=0.221, simple_loss=0.3028, pruned_loss=0.06966, over 3064311.79 frames. ], batch size: 191, lr: 6.15e-03, grad_scale: 4.0 2023-04-29 11:23:42,501 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108933.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:24:11,388 INFO [train.py:904] (4/8) Epoch 11, batch 7450, loss[loss=0.2173, simple_loss=0.3, pruned_loss=0.06728, over 16824.00 frames. ], tot_loss[loss=0.2232, simple_loss=0.3047, pruned_loss=0.07092, over 3072820.22 frames. ], batch size: 116, lr: 6.15e-03, grad_scale: 4.0 2023-04-29 11:24:13,635 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.939e+02 3.048e+02 3.705e+02 5.111e+02 8.676e+02, threshold=7.410e+02, percent-clipped=2.0 2023-04-29 11:24:56,127 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2599, 3.6191, 3.6744, 1.8787, 2.9917, 2.3476, 3.5778, 3.7216], device='cuda:4'), covar=tensor([0.0262, 0.0690, 0.0548, 0.2080, 0.0827, 0.0989, 0.0657, 0.0935], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0145, 0.0159, 0.0145, 0.0138, 0.0127, 0.0138, 0.0156], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 11:24:59,253 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=108981.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:25:00,417 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1362, 4.1810, 4.0243, 3.8293, 3.6747, 4.1099, 3.8656, 3.8810], device='cuda:4'), covar=tensor([0.0591, 0.0516, 0.0260, 0.0261, 0.0818, 0.0392, 0.0773, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0233, 0.0301, 0.0275, 0.0253, 0.0295, 0.0294, 0.0191, 0.0321], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:25:31,627 INFO [train.py:904] (4/8) Epoch 11, batch 7500, loss[loss=0.1834, simple_loss=0.2637, pruned_loss=0.05161, over 16459.00 frames. ], tot_loss[loss=0.2227, simple_loss=0.3046, pruned_loss=0.07042, over 3058386.87 frames. ], batch size: 68, lr: 6.15e-03, grad_scale: 4.0 2023-04-29 11:26:16,354 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=109029.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:26:28,525 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109037.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 11:26:51,082 INFO [train.py:904] (4/8) Epoch 11, batch 7550, loss[loss=0.2143, simple_loss=0.2945, pruned_loss=0.0671, over 16224.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.3035, pruned_loss=0.07013, over 3072514.77 frames. ], batch size: 165, lr: 6.15e-03, grad_scale: 4.0 2023-04-29 11:26:52,322 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.931e+02 3.123e+02 3.909e+02 4.676e+02 1.371e+03, threshold=7.817e+02, percent-clipped=2.0 2023-04-29 11:27:30,741 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3960, 4.6638, 4.4480, 4.4498, 4.2213, 4.1153, 4.2609, 4.6988], device='cuda:4'), covar=tensor([0.0934, 0.0777, 0.0914, 0.0702, 0.0696, 0.1310, 0.0955, 0.0826], device='cuda:4'), in_proj_covar=tensor([0.0535, 0.0658, 0.0548, 0.0454, 0.0416, 0.0432, 0.0550, 0.0501], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:27:39,836 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9777, 4.0177, 3.8262, 3.6605, 3.5459, 3.9558, 3.6849, 3.6736], device='cuda:4'), covar=tensor([0.0581, 0.0420, 0.0252, 0.0253, 0.0815, 0.0396, 0.0870, 0.0605], device='cuda:4'), in_proj_covar=tensor([0.0234, 0.0304, 0.0278, 0.0255, 0.0297, 0.0296, 0.0193, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:27:46,114 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109088.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:28:02,025 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109098.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 11:28:06,702 INFO [train.py:904] (4/8) Epoch 11, batch 7600, loss[loss=0.2395, simple_loss=0.3066, pruned_loss=0.08624, over 11508.00 frames. ], tot_loss[loss=0.2211, simple_loss=0.3022, pruned_loss=0.07001, over 3074079.17 frames. ], batch size: 248, lr: 6.15e-03, grad_scale: 8.0 2023-04-29 11:28:41,392 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-29 11:28:53,757 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1129, 1.4675, 1.8576, 2.0116, 2.2024, 2.3172, 1.6501, 2.2359], device='cuda:4'), covar=tensor([0.0156, 0.0335, 0.0175, 0.0228, 0.0178, 0.0134, 0.0325, 0.0084], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0167, 0.0148, 0.0155, 0.0163, 0.0121, 0.0166, 0.0111], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 11:28:56,879 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=109136.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:29:08,118 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-29 11:29:12,236 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109147.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:29:20,692 INFO [train.py:904] (4/8) Epoch 11, batch 7650, loss[loss=0.2617, simple_loss=0.3219, pruned_loss=0.1008, over 11400.00 frames. ], tot_loss[loss=0.2224, simple_loss=0.3034, pruned_loss=0.0707, over 3070793.02 frames. ], batch size: 246, lr: 6.15e-03, grad_scale: 4.0 2023-04-29 11:29:23,629 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.975e+02 2.787e+02 3.621e+02 4.390e+02 9.045e+02, threshold=7.242e+02, percent-clipped=3.0 2023-04-29 11:30:25,799 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109194.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:30:36,525 INFO [train.py:904] (4/8) Epoch 11, batch 7700, loss[loss=0.1888, simple_loss=0.2737, pruned_loss=0.05193, over 16439.00 frames. ], tot_loss[loss=0.223, simple_loss=0.3031, pruned_loss=0.07149, over 3059421.82 frames. ], batch size: 68, lr: 6.15e-03, grad_scale: 4.0 2023-04-29 11:31:25,276 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109233.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:31:53,770 INFO [train.py:904] (4/8) Epoch 11, batch 7750, loss[loss=0.2051, simple_loss=0.2908, pruned_loss=0.05971, over 16917.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.303, pruned_loss=0.07113, over 3079522.01 frames. ], batch size: 96, lr: 6.14e-03, grad_scale: 4.0 2023-04-29 11:31:56,716 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.029e+02 3.061e+02 3.488e+02 4.422e+02 1.299e+03, threshold=6.977e+02, percent-clipped=6.0 2023-04-29 11:31:59,732 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109255.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:32:09,868 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-29 11:32:19,733 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6363, 3.7487, 4.1104, 4.0692, 4.0530, 3.7915, 3.8145, 3.7752], device='cuda:4'), covar=tensor([0.0381, 0.0635, 0.0385, 0.0416, 0.0509, 0.0448, 0.0962, 0.0546], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0335, 0.0340, 0.0317, 0.0386, 0.0354, 0.0460, 0.0289], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 11:32:39,672 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=109281.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:32:58,584 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109294.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:33:10,716 INFO [train.py:904] (4/8) Epoch 11, batch 7800, loss[loss=0.2385, simple_loss=0.3233, pruned_loss=0.07685, over 16861.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.3046, pruned_loss=0.0719, over 3078387.49 frames. ], batch size: 116, lr: 6.14e-03, grad_scale: 4.0 2023-04-29 11:33:35,165 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109318.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:33:40,287 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-29 11:34:25,569 INFO [train.py:904] (4/8) Epoch 11, batch 7850, loss[loss=0.2033, simple_loss=0.3001, pruned_loss=0.05327, over 16594.00 frames. ], tot_loss[loss=0.2251, simple_loss=0.3057, pruned_loss=0.07221, over 3062482.75 frames. ], batch size: 62, lr: 6.14e-03, grad_scale: 2.0 2023-04-29 11:34:30,495 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.254e+02 2.996e+02 3.807e+02 4.830e+02 8.310e+02, threshold=7.614e+02, percent-clipped=7.0 2023-04-29 11:34:30,986 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109355.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:34:59,301 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-29 11:35:07,062 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109379.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 11:35:17,728 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3853, 3.0673, 2.9233, 1.9336, 2.7157, 2.1271, 2.9242, 3.2502], device='cuda:4'), covar=tensor([0.0337, 0.0569, 0.0614, 0.1849, 0.0811, 0.0921, 0.0779, 0.0823], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0143, 0.0156, 0.0143, 0.0136, 0.0125, 0.0136, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 11:35:27,276 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109393.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 11:35:41,205 INFO [train.py:904] (4/8) Epoch 11, batch 7900, loss[loss=0.2153, simple_loss=0.3032, pruned_loss=0.06377, over 16439.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.3049, pruned_loss=0.07178, over 3067095.51 frames. ], batch size: 146, lr: 6.14e-03, grad_scale: 2.0 2023-04-29 11:35:57,680 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4294, 2.5447, 2.0277, 2.3139, 2.9052, 2.5447, 3.1857, 3.1177], device='cuda:4'), covar=tensor([0.0062, 0.0248, 0.0371, 0.0309, 0.0164, 0.0258, 0.0151, 0.0151], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0198, 0.0195, 0.0193, 0.0197, 0.0198, 0.0201, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:36:52,137 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109447.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:37:00,092 INFO [train.py:904] (4/8) Epoch 11, batch 7950, loss[loss=0.2183, simple_loss=0.2955, pruned_loss=0.07053, over 15366.00 frames. ], tot_loss[loss=0.2239, simple_loss=0.3047, pruned_loss=0.07149, over 3071748.05 frames. ], batch size: 190, lr: 6.14e-03, grad_scale: 2.0 2023-04-29 11:37:04,710 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.170e+02 3.029e+02 3.447e+02 4.404e+02 6.785e+02, threshold=6.894e+02, percent-clipped=0.0 2023-04-29 11:37:43,963 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4543, 4.2749, 4.5354, 4.6944, 4.8391, 4.4152, 4.7799, 4.8030], device='cuda:4'), covar=tensor([0.1564, 0.1159, 0.1427, 0.0621, 0.0524, 0.0842, 0.0552, 0.0625], device='cuda:4'), in_proj_covar=tensor([0.0517, 0.0643, 0.0772, 0.0652, 0.0506, 0.0500, 0.0517, 0.0586], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:38:04,154 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=109495.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:38:14,152 INFO [train.py:904] (4/8) Epoch 11, batch 8000, loss[loss=0.2571, simple_loss=0.3195, pruned_loss=0.09737, over 11280.00 frames. ], tot_loss[loss=0.2258, simple_loss=0.3058, pruned_loss=0.07292, over 3036428.87 frames. ], batch size: 247, lr: 6.14e-03, grad_scale: 4.0 2023-04-29 11:39:24,916 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109550.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:39:27,474 INFO [train.py:904] (4/8) Epoch 11, batch 8050, loss[loss=0.2375, simple_loss=0.322, pruned_loss=0.07651, over 16913.00 frames. ], tot_loss[loss=0.2243, simple_loss=0.3047, pruned_loss=0.07196, over 3047945.01 frames. ], batch size: 116, lr: 6.14e-03, grad_scale: 4.0 2023-04-29 11:39:31,004 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.160e+02 2.984e+02 3.790e+02 4.579e+02 1.063e+03, threshold=7.580e+02, percent-clipped=3.0 2023-04-29 11:39:40,818 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5786, 4.4492, 4.4996, 2.7583, 3.8375, 4.3924, 3.9654, 2.6005], device='cuda:4'), covar=tensor([0.0408, 0.0029, 0.0026, 0.0330, 0.0066, 0.0078, 0.0053, 0.0329], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0068, 0.0069, 0.0125, 0.0078, 0.0090, 0.0077, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 11:40:06,475 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109579.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:40:40,895 INFO [train.py:904] (4/8) Epoch 11, batch 8100, loss[loss=0.2231, simple_loss=0.3059, pruned_loss=0.07015, over 16314.00 frames. ], tot_loss[loss=0.2222, simple_loss=0.3035, pruned_loss=0.07047, over 3071081.61 frames. ], batch size: 146, lr: 6.13e-03, grad_scale: 4.0 2023-04-29 11:41:39,232 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109640.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:41:54,529 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109650.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:41:55,947 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8873, 3.0532, 3.1373, 1.9083, 2.9091, 3.0857, 2.9960, 1.9455], device='cuda:4'), covar=tensor([0.0455, 0.0046, 0.0043, 0.0378, 0.0085, 0.0094, 0.0072, 0.0351], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0067, 0.0069, 0.0125, 0.0078, 0.0090, 0.0077, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 11:41:56,612 INFO [train.py:904] (4/8) Epoch 11, batch 8150, loss[loss=0.2074, simple_loss=0.2853, pruned_loss=0.0648, over 15059.00 frames. ], tot_loss[loss=0.2199, simple_loss=0.3011, pruned_loss=0.06935, over 3075544.92 frames. ], batch size: 190, lr: 6.13e-03, grad_scale: 4.0 2023-04-29 11:42:01,361 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.997e+02 3.110e+02 3.865e+02 4.805e+02 7.636e+02, threshold=7.730e+02, percent-clipped=1.0 2023-04-29 11:42:08,194 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109659.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:42:29,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5539, 4.5549, 4.4261, 4.1588, 4.0772, 4.5031, 4.3196, 4.1733], device='cuda:4'), covar=tensor([0.0596, 0.0473, 0.0262, 0.0280, 0.0903, 0.0399, 0.0430, 0.0689], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0306, 0.0278, 0.0257, 0.0296, 0.0297, 0.0193, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:42:30,922 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109674.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 11:42:59,313 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109693.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 11:43:12,475 INFO [train.py:904] (4/8) Epoch 11, batch 8200, loss[loss=0.2663, simple_loss=0.3186, pruned_loss=0.107, over 11303.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.2992, pruned_loss=0.06888, over 3090472.95 frames. ], batch size: 247, lr: 6.13e-03, grad_scale: 4.0 2023-04-29 11:43:41,763 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109720.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:44:16,310 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=109741.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 11:44:33,540 INFO [train.py:904] (4/8) Epoch 11, batch 8250, loss[loss=0.1875, simple_loss=0.2861, pruned_loss=0.04448, over 16845.00 frames. ], tot_loss[loss=0.2158, simple_loss=0.2985, pruned_loss=0.06656, over 3074785.13 frames. ], batch size: 102, lr: 6.13e-03, grad_scale: 4.0 2023-04-29 11:44:38,005 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.186e+02 3.150e+02 3.966e+02 4.988e+02 1.179e+03, threshold=7.932e+02, percent-clipped=8.0 2023-04-29 11:44:50,527 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7781, 3.7881, 4.1076, 4.0859, 4.0874, 3.8820, 3.8474, 3.8380], device='cuda:4'), covar=tensor([0.0301, 0.0560, 0.0416, 0.0430, 0.0425, 0.0381, 0.0950, 0.0473], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0335, 0.0339, 0.0316, 0.0387, 0.0353, 0.0461, 0.0288], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 11:45:52,507 INFO [train.py:904] (4/8) Epoch 11, batch 8300, loss[loss=0.207, simple_loss=0.3001, pruned_loss=0.05696, over 16925.00 frames. ], tot_loss[loss=0.2111, simple_loss=0.2952, pruned_loss=0.06347, over 3057007.77 frames. ], batch size: 109, lr: 6.13e-03, grad_scale: 4.0 2023-04-29 11:47:09,109 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109850.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:47:11,568 INFO [train.py:904] (4/8) Epoch 11, batch 8350, loss[loss=0.2212, simple_loss=0.2994, pruned_loss=0.07146, over 12037.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2946, pruned_loss=0.06154, over 3046003.35 frames. ], batch size: 248, lr: 6.13e-03, grad_scale: 4.0 2023-04-29 11:47:16,940 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.870e+02 2.495e+02 2.844e+02 3.376e+02 6.294e+02, threshold=5.687e+02, percent-clipped=0.0 2023-04-29 11:47:46,042 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.61 vs. limit=2.0 2023-04-29 11:48:25,292 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=109898.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:48:30,268 INFO [train.py:904] (4/8) Epoch 11, batch 8400, loss[loss=0.186, simple_loss=0.2762, pruned_loss=0.04792, over 16837.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.2919, pruned_loss=0.05923, over 3040871.30 frames. ], batch size: 116, lr: 6.13e-03, grad_scale: 8.0 2023-04-29 11:48:42,950 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7283, 4.0110, 3.1228, 2.1441, 2.6267, 2.3271, 4.1844, 3.5628], device='cuda:4'), covar=tensor([0.2508, 0.0518, 0.1330, 0.2284, 0.2343, 0.1844, 0.0340, 0.0922], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0250, 0.0276, 0.0269, 0.0273, 0.0218, 0.0259, 0.0286], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:49:21,771 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109935.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:49:45,311 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109950.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:49:47,555 INFO [train.py:904] (4/8) Epoch 11, batch 8450, loss[loss=0.1804, simple_loss=0.2779, pruned_loss=0.04149, over 16465.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2898, pruned_loss=0.05713, over 3049833.26 frames. ], batch size: 146, lr: 6.13e-03, grad_scale: 8.0 2023-04-29 11:49:52,353 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.575e+02 2.195e+02 2.664e+02 3.370e+02 8.341e+02, threshold=5.327e+02, percent-clipped=3.0 2023-04-29 11:49:54,843 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6778, 3.9838, 4.1220, 2.0432, 4.2821, 4.4102, 3.3429, 3.1663], device='cuda:4'), covar=tensor([0.0809, 0.0125, 0.0139, 0.1114, 0.0052, 0.0073, 0.0284, 0.0462], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0098, 0.0085, 0.0135, 0.0067, 0.0099, 0.0118, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 11:50:23,530 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109974.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:51:00,422 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=109998.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:51:09,669 INFO [train.py:904] (4/8) Epoch 11, batch 8500, loss[loss=0.1869, simple_loss=0.2716, pruned_loss=0.05113, over 16435.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2855, pruned_loss=0.0547, over 3046565.31 frames. ], batch size: 146, lr: 6.12e-03, grad_scale: 8.0 2023-04-29 11:51:16,010 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-29 11:51:27,696 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3907, 3.0384, 2.7127, 2.2294, 2.2187, 2.2678, 2.9630, 2.8714], device='cuda:4'), covar=tensor([0.2151, 0.0679, 0.1266, 0.2124, 0.2113, 0.1758, 0.0374, 0.1056], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0247, 0.0271, 0.0266, 0.0267, 0.0215, 0.0254, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:51:31,229 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110015.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:51:33,946 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110017.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:51:41,405 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=110022.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:52:17,065 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110044.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:52:31,122 INFO [train.py:904] (4/8) Epoch 11, batch 8550, loss[loss=0.1897, simple_loss=0.2875, pruned_loss=0.04592, over 16851.00 frames. ], tot_loss[loss=0.1949, simple_loss=0.2832, pruned_loss=0.05328, over 3035486.62 frames. ], batch size: 102, lr: 6.12e-03, grad_scale: 8.0 2023-04-29 11:52:37,011 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.639e+02 2.340e+02 2.855e+02 3.457e+02 5.505e+02, threshold=5.709e+02, percent-clipped=2.0 2023-04-29 11:53:19,586 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110078.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:54:04,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8494, 3.5461, 2.7505, 5.0472, 3.8940, 4.6614, 1.6967, 3.4300], device='cuda:4'), covar=tensor([0.1314, 0.0484, 0.0996, 0.0099, 0.0127, 0.0248, 0.1463, 0.0587], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0157, 0.0178, 0.0141, 0.0195, 0.0206, 0.0180, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 11:54:07,122 INFO [train.py:904] (4/8) Epoch 11, batch 8600, loss[loss=0.175, simple_loss=0.277, pruned_loss=0.03646, over 16901.00 frames. ], tot_loss[loss=0.1935, simple_loss=0.2829, pruned_loss=0.052, over 3029979.77 frames. ], batch size: 102, lr: 6.12e-03, grad_scale: 8.0 2023-04-29 11:54:14,606 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110105.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:54:23,065 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110110.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:54:35,055 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110115.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:55:43,444 INFO [train.py:904] (4/8) Epoch 11, batch 8650, loss[loss=0.1712, simple_loss=0.2731, pruned_loss=0.03459, over 16860.00 frames. ], tot_loss[loss=0.1908, simple_loss=0.2811, pruned_loss=0.05026, over 3042832.23 frames. ], batch size: 102, lr: 6.12e-03, grad_scale: 4.0 2023-04-29 11:55:53,865 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.622e+02 2.527e+02 3.196e+02 4.321e+02 7.577e+02, threshold=6.393e+02, percent-clipped=5.0 2023-04-29 11:56:16,579 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6573, 4.0016, 3.5420, 3.9210, 3.4431, 3.5810, 3.5232, 3.9667], device='cuda:4'), covar=tensor([0.2521, 0.1819, 0.2952, 0.1273, 0.1911, 0.3111, 0.2641, 0.2012], device='cuda:4'), in_proj_covar=tensor([0.0521, 0.0642, 0.0533, 0.0447, 0.0401, 0.0424, 0.0537, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:56:28,133 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110171.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:56:40,268 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110176.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:57:08,858 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-29 11:57:30,807 INFO [train.py:904] (4/8) Epoch 11, batch 8700, loss[loss=0.1769, simple_loss=0.273, pruned_loss=0.04041, over 16962.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2781, pruned_loss=0.04891, over 3055099.32 frames. ], batch size: 109, lr: 6.12e-03, grad_scale: 4.0 2023-04-29 11:57:44,804 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4502, 4.5520, 4.3489, 4.0781, 4.0656, 4.4112, 4.2681, 4.1251], device='cuda:4'), covar=tensor([0.0538, 0.0502, 0.0248, 0.0254, 0.0737, 0.0462, 0.0365, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0233, 0.0301, 0.0274, 0.0255, 0.0290, 0.0294, 0.0191, 0.0319], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 11:57:58,314 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6630, 2.8376, 2.4394, 3.9974, 2.6700, 4.0001, 1.2794, 3.0055], device='cuda:4'), covar=tensor([0.1421, 0.0623, 0.1158, 0.0133, 0.0137, 0.0390, 0.1721, 0.0667], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0156, 0.0178, 0.0140, 0.0193, 0.0205, 0.0179, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 11:58:34,008 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110235.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 11:59:05,959 INFO [train.py:904] (4/8) Epoch 11, batch 8750, loss[loss=0.1936, simple_loss=0.2918, pruned_loss=0.04776, over 16429.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2778, pruned_loss=0.04817, over 3068691.87 frames. ], batch size: 147, lr: 6.12e-03, grad_scale: 4.0 2023-04-29 11:59:15,717 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.697e+02 2.315e+02 2.719e+02 3.353e+02 7.427e+02, threshold=5.437e+02, percent-clipped=1.0 2023-04-29 12:00:22,507 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=110283.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:00:22,681 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110283.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:01:00,283 INFO [train.py:904] (4/8) Epoch 11, batch 8800, loss[loss=0.18, simple_loss=0.2762, pruned_loss=0.04192, over 16790.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2762, pruned_loss=0.04717, over 3057360.06 frames. ], batch size: 124, lr: 6.12e-03, grad_scale: 8.0 2023-04-29 12:01:03,772 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110303.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:01:27,800 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110315.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:01:34,225 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110318.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:02:29,289 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110344.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:02:45,834 INFO [train.py:904] (4/8) Epoch 11, batch 8850, loss[loss=0.1679, simple_loss=0.2533, pruned_loss=0.04125, over 12048.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2775, pruned_loss=0.0463, over 3043420.06 frames. ], batch size: 248, lr: 6.11e-03, grad_scale: 8.0 2023-04-29 12:02:52,480 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.350e+02 2.510e+02 2.922e+02 3.907e+02 6.547e+02, threshold=5.844e+02, percent-clipped=7.0 2023-04-29 12:03:09,148 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=110363.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:03:11,331 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110364.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:03:29,776 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110373.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:03:42,468 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110379.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:04:27,923 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110400.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:04:30,862 INFO [train.py:904] (4/8) Epoch 11, batch 8900, loss[loss=0.1971, simple_loss=0.284, pruned_loss=0.05509, over 16947.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2779, pruned_loss=0.04591, over 3031140.56 frames. ], batch size: 116, lr: 6.11e-03, grad_scale: 8.0 2023-04-29 12:05:18,282 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3337, 3.1218, 2.5577, 2.1905, 2.2905, 2.0974, 3.0619, 2.8992], device='cuda:4'), covar=tensor([0.2560, 0.0802, 0.1693, 0.2283, 0.2284, 0.2127, 0.0707, 0.1202], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0250, 0.0276, 0.0269, 0.0265, 0.0217, 0.0259, 0.0283], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:06:21,525 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.65 vs. limit=2.0 2023-04-29 12:06:36,591 INFO [train.py:904] (4/8) Epoch 11, batch 8950, loss[loss=0.1761, simple_loss=0.2652, pruned_loss=0.04354, over 16434.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2777, pruned_loss=0.04649, over 3043365.41 frames. ], batch size: 68, lr: 6.11e-03, grad_scale: 8.0 2023-04-29 12:06:45,507 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.580e+02 2.155e+02 2.674e+02 3.353e+02 7.252e+02, threshold=5.349e+02, percent-clipped=1.0 2023-04-29 12:06:46,795 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-29 12:07:08,189 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110466.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:07:17,913 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110471.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:07:30,241 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2358, 2.0001, 2.1937, 3.8335, 1.9644, 2.3720, 2.1006, 2.1733], device='cuda:4'), covar=tensor([0.0952, 0.3418, 0.2338, 0.0403, 0.4026, 0.2343, 0.3355, 0.3360], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0379, 0.0321, 0.0307, 0.0401, 0.0430, 0.0342, 0.0441], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:08:26,527 INFO [train.py:904] (4/8) Epoch 11, batch 9000, loss[loss=0.1765, simple_loss=0.2637, pruned_loss=0.04463, over 16379.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2743, pruned_loss=0.04492, over 3036635.92 frames. ], batch size: 146, lr: 6.11e-03, grad_scale: 8.0 2023-04-29 12:08:26,527 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 12:08:36,935 INFO [train.py:938] (4/8) Epoch 11, validation: loss=0.1545, simple_loss=0.2586, pruned_loss=0.02523, over 944034.00 frames. 2023-04-29 12:08:36,936 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 12:10:21,906 INFO [train.py:904] (4/8) Epoch 11, batch 9050, loss[loss=0.1916, simple_loss=0.2777, pruned_loss=0.05275, over 12755.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2748, pruned_loss=0.04543, over 3029274.04 frames. ], batch size: 248, lr: 6.11e-03, grad_scale: 8.0 2023-04-29 12:10:28,904 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.635e+02 2.375e+02 3.006e+02 3.857e+02 1.104e+03, threshold=6.012e+02, percent-clipped=5.0 2023-04-29 12:10:53,867 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-29 12:12:06,397 INFO [train.py:904] (4/8) Epoch 11, batch 9100, loss[loss=0.1895, simple_loss=0.2833, pruned_loss=0.0478, over 16432.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2744, pruned_loss=0.04585, over 3040862.22 frames. ], batch size: 146, lr: 6.11e-03, grad_scale: 8.0 2023-04-29 12:12:25,145 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110611.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:13:36,696 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110639.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:14:06,378 INFO [train.py:904] (4/8) Epoch 11, batch 9150, loss[loss=0.1873, simple_loss=0.2765, pruned_loss=0.04905, over 16405.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2756, pruned_loss=0.04571, over 3056665.86 frames. ], batch size: 146, lr: 6.11e-03, grad_scale: 8.0 2023-04-29 12:14:15,979 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.599e+02 2.395e+02 2.852e+02 3.560e+02 7.952e+02, threshold=5.704e+02, percent-clipped=1.0 2023-04-29 12:14:23,371 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110659.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:14:52,265 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110672.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:14:53,800 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110673.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:14:54,195 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-29 12:14:56,548 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110674.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:15:46,085 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110700.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:15:48,928 INFO [train.py:904] (4/8) Epoch 11, batch 9200, loss[loss=0.183, simple_loss=0.2766, pruned_loss=0.04471, over 16722.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2716, pruned_loss=0.04483, over 3047622.23 frames. ], batch size: 134, lr: 6.10e-03, grad_scale: 8.0 2023-04-29 12:16:10,386 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.38 vs. limit=5.0 2023-04-29 12:16:24,195 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=110721.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:16:52,966 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110736.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:17:18,046 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=110748.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:17:24,346 INFO [train.py:904] (4/8) Epoch 11, batch 9250, loss[loss=0.1725, simple_loss=0.2668, pruned_loss=0.03908, over 16337.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2714, pruned_loss=0.04438, over 3061067.64 frames. ], batch size: 146, lr: 6.10e-03, grad_scale: 8.0 2023-04-29 12:17:32,962 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.609e+02 2.583e+02 3.099e+02 3.621e+02 8.759e+02, threshold=6.198e+02, percent-clipped=4.0 2023-04-29 12:17:53,590 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110766.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:18:05,080 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110771.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:18:31,865 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6861, 2.2108, 2.3229, 4.4026, 2.1012, 2.7541, 2.3239, 2.4063], device='cuda:4'), covar=tensor([0.0749, 0.3266, 0.2223, 0.0303, 0.3851, 0.2063, 0.2912, 0.3239], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0375, 0.0320, 0.0306, 0.0399, 0.0425, 0.0339, 0.0438], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:19:06,569 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110797.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 12:19:16,916 INFO [train.py:904] (4/8) Epoch 11, batch 9300, loss[loss=0.165, simple_loss=0.2593, pruned_loss=0.03535, over 16720.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2693, pruned_loss=0.04339, over 3075129.51 frames. ], batch size: 134, lr: 6.10e-03, grad_scale: 8.0 2023-04-29 12:19:45,570 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=110814.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:19:59,650 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=110819.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:20:01,665 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5382, 3.5233, 3.4636, 2.9223, 3.3334, 2.0022, 3.1742, 2.9150], device='cuda:4'), covar=tensor([0.0114, 0.0115, 0.0150, 0.0198, 0.0094, 0.2058, 0.0127, 0.0198], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0109, 0.0156, 0.0142, 0.0128, 0.0175, 0.0143, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:20:49,145 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-29 12:21:01,158 INFO [train.py:904] (4/8) Epoch 11, batch 9350, loss[loss=0.2071, simple_loss=0.2935, pruned_loss=0.06033, over 16733.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2694, pruned_loss=0.04352, over 3066020.02 frames. ], batch size: 124, lr: 6.10e-03, grad_scale: 8.0 2023-04-29 12:21:10,062 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.415e+02 2.201e+02 2.572e+02 3.084e+02 6.751e+02, threshold=5.144e+02, percent-clipped=1.0 2023-04-29 12:21:35,929 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110868.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:21:58,563 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-29 12:22:41,698 INFO [train.py:904] (4/8) Epoch 11, batch 9400, loss[loss=0.2083, simple_loss=0.2983, pruned_loss=0.05909, over 16901.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2697, pruned_loss=0.04349, over 3062640.00 frames. ], batch size: 116, lr: 6.10e-03, grad_scale: 8.0 2023-04-29 12:23:10,366 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-29 12:23:36,518 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110929.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:23:57,896 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110939.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:24:21,277 INFO [train.py:904] (4/8) Epoch 11, batch 9450, loss[loss=0.1906, simple_loss=0.284, pruned_loss=0.04864, over 16401.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2719, pruned_loss=0.04409, over 3066108.03 frames. ], batch size: 146, lr: 6.10e-03, grad_scale: 8.0 2023-04-29 12:24:27,346 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.534e+02 2.422e+02 3.077e+02 3.855e+02 7.743e+02, threshold=6.155e+02, percent-clipped=6.0 2023-04-29 12:24:33,085 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110958.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:24:34,533 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110959.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:24:50,261 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110967.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:25:06,688 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110974.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:25:33,499 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=110987.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:26:00,932 INFO [train.py:904] (4/8) Epoch 11, batch 9500, loss[loss=0.1977, simple_loss=0.2844, pruned_loss=0.05544, over 16317.00 frames. ], tot_loss[loss=0.1794, simple_loss=0.2712, pruned_loss=0.04375, over 3069312.71 frames. ], batch size: 146, lr: 6.10e-03, grad_scale: 8.0 2023-04-29 12:26:07,781 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0215, 1.8125, 1.5879, 1.4927, 1.9156, 1.6196, 1.7143, 2.0224], device='cuda:4'), covar=tensor([0.0096, 0.0226, 0.0301, 0.0292, 0.0168, 0.0208, 0.0123, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0199, 0.0194, 0.0193, 0.0197, 0.0196, 0.0194, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:26:07,841 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4317, 3.0498, 2.6981, 2.2422, 2.1320, 2.1294, 3.0506, 2.9258], device='cuda:4'), covar=tensor([0.2293, 0.0665, 0.1363, 0.2192, 0.2363, 0.1937, 0.0452, 0.1141], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0250, 0.0276, 0.0271, 0.0259, 0.0217, 0.0259, 0.0282], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:26:15,566 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=111007.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:26:22,084 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0098, 4.9270, 4.7230, 4.3598, 4.7439, 1.9732, 4.5374, 4.7380], device='cuda:4'), covar=tensor([0.0065, 0.0066, 0.0145, 0.0256, 0.0079, 0.2152, 0.0109, 0.0144], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0110, 0.0157, 0.0143, 0.0128, 0.0177, 0.0145, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:26:38,111 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=111019.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:26:43,960 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=111022.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:27:46,818 INFO [train.py:904] (4/8) Epoch 11, batch 9550, loss[loss=0.2179, simple_loss=0.3064, pruned_loss=0.06467, over 16390.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2711, pruned_loss=0.044, over 3070244.58 frames. ], batch size: 166, lr: 6.09e-03, grad_scale: 8.0 2023-04-29 12:27:55,298 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.451e+02 2.456e+02 2.878e+02 3.283e+02 5.727e+02, threshold=5.755e+02, percent-clipped=0.0 2023-04-29 12:29:10,460 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111092.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 12:29:25,192 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-29 12:29:26,768 INFO [train.py:904] (4/8) Epoch 11, batch 9600, loss[loss=0.1932, simple_loss=0.2968, pruned_loss=0.04478, over 15334.00 frames. ], tot_loss[loss=0.1805, simple_loss=0.2721, pruned_loss=0.04446, over 3067464.44 frames. ], batch size: 191, lr: 6.09e-03, grad_scale: 8.0 2023-04-29 12:30:46,152 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1753, 3.4030, 3.5928, 3.5864, 3.6075, 3.4101, 3.4456, 3.4464], device='cuda:4'), covar=tensor([0.0384, 0.0565, 0.0517, 0.0517, 0.0505, 0.0462, 0.0801, 0.0422], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0313, 0.0317, 0.0297, 0.0358, 0.0334, 0.0425, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002], device='cuda:4') 2023-04-29 12:31:14,961 INFO [train.py:904] (4/8) Epoch 11, batch 9650, loss[loss=0.1812, simple_loss=0.2718, pruned_loss=0.04534, over 16240.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2735, pruned_loss=0.04438, over 3068838.29 frames. ], batch size: 165, lr: 6.09e-03, grad_scale: 8.0 2023-04-29 12:31:19,752 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-29 12:31:24,134 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.791e+02 2.367e+02 2.756e+02 3.328e+02 5.495e+02, threshold=5.512e+02, percent-clipped=0.0 2023-04-29 12:32:59,101 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0088, 2.4122, 2.3226, 3.0017, 2.0989, 3.3604, 1.6773, 2.8410], device='cuda:4'), covar=tensor([0.1107, 0.0516, 0.0885, 0.0134, 0.0098, 0.0356, 0.1246, 0.0537], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0154, 0.0176, 0.0138, 0.0184, 0.0202, 0.0179, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 12:33:03,901 INFO [train.py:904] (4/8) Epoch 11, batch 9700, loss[loss=0.1683, simple_loss=0.2655, pruned_loss=0.03551, over 16681.00 frames. ], tot_loss[loss=0.1805, simple_loss=0.2726, pruned_loss=0.04425, over 3063088.48 frames. ], batch size: 83, lr: 6.09e-03, grad_scale: 8.0 2023-04-29 12:33:07,378 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-29 12:33:49,082 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111224.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:33:52,460 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=111225.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:34:05,964 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-29 12:34:35,555 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-29 12:34:46,293 INFO [train.py:904] (4/8) Epoch 11, batch 9750, loss[loss=0.1747, simple_loss=0.2563, pruned_loss=0.04658, over 12137.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2711, pruned_loss=0.04431, over 3063012.58 frames. ], batch size: 246, lr: 6.09e-03, grad_scale: 8.0 2023-04-29 12:34:53,729 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.704e+02 2.278e+02 2.798e+02 3.645e+02 8.933e+02, threshold=5.595e+02, percent-clipped=9.0 2023-04-29 12:34:54,720 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.02 vs. limit=5.0 2023-04-29 12:35:17,236 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111267.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:35:36,610 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3837, 3.0152, 2.6066, 2.2350, 2.1374, 2.1979, 2.9371, 2.8025], device='cuda:4'), covar=tensor([0.2306, 0.0777, 0.1487, 0.1999, 0.2275, 0.1863, 0.0488, 0.1099], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0249, 0.0274, 0.0268, 0.0255, 0.0216, 0.0257, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:35:46,664 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5422, 4.6422, 4.8873, 4.6603, 4.7356, 5.2412, 4.7689, 4.4736], device='cuda:4'), covar=tensor([0.1077, 0.1877, 0.1772, 0.1961, 0.2438, 0.1079, 0.1557, 0.2553], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0459, 0.0501, 0.0394, 0.0521, 0.0532, 0.0398, 0.0533], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:35:58,544 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=111286.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:36:25,538 INFO [train.py:904] (4/8) Epoch 11, batch 9800, loss[loss=0.1771, simple_loss=0.2703, pruned_loss=0.04193, over 15567.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2713, pruned_loss=0.04346, over 3069061.04 frames. ], batch size: 191, lr: 6.09e-03, grad_scale: 8.0 2023-04-29 12:36:49,118 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111314.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:36:51,185 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=111315.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:37:32,728 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7132, 1.2746, 1.4642, 1.6889, 1.8429, 1.8284, 1.5927, 1.7755], device='cuda:4'), covar=tensor([0.0217, 0.0286, 0.0163, 0.0226, 0.0199, 0.0144, 0.0286, 0.0095], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0167, 0.0149, 0.0152, 0.0161, 0.0119, 0.0166, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 12:38:09,307 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6177, 3.7569, 2.8367, 2.2119, 2.3657, 2.3115, 3.9774, 3.2661], device='cuda:4'), covar=tensor([0.2576, 0.0665, 0.1614, 0.2327, 0.2353, 0.1823, 0.0372, 0.1146], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0248, 0.0274, 0.0269, 0.0255, 0.0216, 0.0256, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:38:11,445 INFO [train.py:904] (4/8) Epoch 11, batch 9850, loss[loss=0.2069, simple_loss=0.2904, pruned_loss=0.06168, over 15355.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2725, pruned_loss=0.04355, over 3056588.38 frames. ], batch size: 191, lr: 6.09e-03, grad_scale: 8.0 2023-04-29 12:38:20,188 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.514e+02 2.271e+02 2.860e+02 3.428e+02 8.615e+02, threshold=5.720e+02, percent-clipped=1.0 2023-04-29 12:39:36,487 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1826, 3.6076, 3.8332, 2.0796, 3.0096, 2.3710, 3.6177, 3.6528], device='cuda:4'), covar=tensor([0.0283, 0.0599, 0.0418, 0.1768, 0.0752, 0.0916, 0.0661, 0.0899], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0133, 0.0152, 0.0139, 0.0133, 0.0122, 0.0132, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 12:39:42,356 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111392.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 12:39:56,323 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6800, 4.5271, 4.6821, 4.9021, 5.0615, 4.5215, 5.0569, 5.0706], device='cuda:4'), covar=tensor([0.1505, 0.0994, 0.1584, 0.0643, 0.0473, 0.0753, 0.0547, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0486, 0.0607, 0.0729, 0.0617, 0.0469, 0.0477, 0.0492, 0.0555], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:40:02,790 INFO [train.py:904] (4/8) Epoch 11, batch 9900, loss[loss=0.1786, simple_loss=0.2762, pruned_loss=0.04057, over 16915.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.272, pruned_loss=0.04323, over 3046098.51 frames. ], batch size: 116, lr: 6.09e-03, grad_scale: 8.0 2023-04-29 12:41:12,038 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6815, 1.6491, 1.8899, 2.6302, 2.3801, 2.9435, 1.8074, 2.9118], device='cuda:4'), covar=tensor([0.0156, 0.0366, 0.0293, 0.0205, 0.0232, 0.0118, 0.0383, 0.0093], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0167, 0.0149, 0.0153, 0.0162, 0.0118, 0.0166, 0.0109], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 12:41:18,693 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8566, 3.8620, 3.9858, 3.7293, 3.8221, 4.2751, 3.9987, 3.7095], device='cuda:4'), covar=tensor([0.1747, 0.2038, 0.1953, 0.2415, 0.2849, 0.1660, 0.1492, 0.2819], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0453, 0.0496, 0.0390, 0.0516, 0.0529, 0.0393, 0.0528], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:41:34,258 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=111440.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:42:01,104 INFO [train.py:904] (4/8) Epoch 11, batch 9950, loss[loss=0.1934, simple_loss=0.288, pruned_loss=0.04946, over 16222.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2742, pruned_loss=0.0436, over 3046548.04 frames. ], batch size: 165, lr: 6.08e-03, grad_scale: 8.0 2023-04-29 12:42:10,737 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-29 12:42:11,465 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.497e+02 2.371e+02 2.780e+02 3.480e+02 6.168e+02, threshold=5.560e+02, percent-clipped=1.0 2023-04-29 12:42:40,498 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=111468.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:44:02,036 INFO [train.py:904] (4/8) Epoch 11, batch 10000, loss[loss=0.1882, simple_loss=0.2857, pruned_loss=0.04532, over 15455.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2725, pruned_loss=0.04277, over 3070196.26 frames. ], batch size: 191, lr: 6.08e-03, grad_scale: 8.0 2023-04-29 12:44:23,719 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9839, 3.3863, 2.6410, 4.9778, 3.8962, 4.5334, 1.5949, 3.3975], device='cuda:4'), covar=tensor([0.1190, 0.0523, 0.1091, 0.0077, 0.0140, 0.0290, 0.1474, 0.0556], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0154, 0.0176, 0.0137, 0.0181, 0.0202, 0.0179, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 12:44:45,183 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111524.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:44:51,337 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8656, 2.2253, 1.8114, 1.9366, 2.5696, 2.2805, 2.6365, 2.7270], device='cuda:4'), covar=tensor([0.0105, 0.0333, 0.0427, 0.0402, 0.0237, 0.0298, 0.0183, 0.0213], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0203, 0.0196, 0.0196, 0.0199, 0.0198, 0.0195, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:44:55,617 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=111529.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:44:58,700 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-29 12:45:17,508 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8506, 1.8583, 2.1775, 3.1800, 1.9877, 2.0465, 2.1098, 1.9380], device='cuda:4'), covar=tensor([0.1066, 0.3823, 0.2192, 0.0619, 0.4590, 0.2841, 0.3250, 0.4022], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0375, 0.0321, 0.0305, 0.0400, 0.0424, 0.0340, 0.0436], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:45:42,148 INFO [train.py:904] (4/8) Epoch 11, batch 10050, loss[loss=0.1839, simple_loss=0.2764, pruned_loss=0.04567, over 15350.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2722, pruned_loss=0.04283, over 3053189.61 frames. ], batch size: 191, lr: 6.08e-03, grad_scale: 8.0 2023-04-29 12:45:50,236 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.518e+02 2.160e+02 2.693e+02 3.599e+02 6.171e+02, threshold=5.385e+02, percent-clipped=1.0 2023-04-29 12:46:22,294 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=111572.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:46:38,797 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111581.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:47:14,825 INFO [train.py:904] (4/8) Epoch 11, batch 10100, loss[loss=0.1944, simple_loss=0.2844, pruned_loss=0.05224, over 16872.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2728, pruned_loss=0.04323, over 3058477.99 frames. ], batch size: 116, lr: 6.08e-03, grad_scale: 8.0 2023-04-29 12:47:35,954 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111614.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:48:58,024 INFO [train.py:904] (4/8) Epoch 12, batch 0, loss[loss=0.1964, simple_loss=0.2766, pruned_loss=0.05809, over 15983.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.2766, pruned_loss=0.05809, over 15983.00 frames. ], batch size: 35, lr: 5.82e-03, grad_scale: 8.0 2023-04-29 12:48:58,025 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 12:49:05,311 INFO [train.py:938] (4/8) Epoch 12, validation: loss=0.1543, simple_loss=0.2577, pruned_loss=0.0254, over 944034.00 frames. 2023-04-29 12:49:05,311 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 12:49:12,535 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.743e+02 2.560e+02 3.107e+02 3.982e+02 7.820e+02, threshold=6.214e+02, percent-clipped=3.0 2023-04-29 12:49:21,579 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=111662.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:49:22,807 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2088, 5.7311, 5.8789, 5.6511, 5.7315, 6.2107, 5.7456, 5.5339], device='cuda:4'), covar=tensor([0.0729, 0.1657, 0.1978, 0.1921, 0.2164, 0.0878, 0.1419, 0.2261], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0454, 0.0497, 0.0394, 0.0519, 0.0532, 0.0400, 0.0527], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:49:54,648 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2016, 4.1862, 4.5501, 2.1989, 4.7387, 4.7375, 3.4059, 3.6868], device='cuda:4'), covar=tensor([0.0655, 0.0172, 0.0151, 0.1095, 0.0046, 0.0094, 0.0346, 0.0307], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0096, 0.0082, 0.0136, 0.0066, 0.0098, 0.0116, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 12:50:15,964 INFO [train.py:904] (4/8) Epoch 12, batch 50, loss[loss=0.1875, simple_loss=0.2805, pruned_loss=0.04723, over 16762.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2858, pruned_loss=0.0601, over 753826.98 frames. ], batch size: 57, lr: 5.82e-03, grad_scale: 2.0 2023-04-29 12:51:25,699 INFO [train.py:904] (4/8) Epoch 12, batch 100, loss[loss=0.1607, simple_loss=0.2473, pruned_loss=0.03709, over 16996.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2814, pruned_loss=0.05796, over 1319814.51 frames. ], batch size: 41, lr: 5.82e-03, grad_scale: 2.0 2023-04-29 12:51:34,343 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.874e+02 2.444e+02 2.855e+02 3.643e+02 7.519e+02, threshold=5.710e+02, percent-clipped=2.0 2023-04-29 12:52:18,881 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=111791.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 12:52:31,943 INFO [train.py:904] (4/8) Epoch 12, batch 150, loss[loss=0.1758, simple_loss=0.2709, pruned_loss=0.04037, over 17136.00 frames. ], tot_loss[loss=0.1962, simple_loss=0.2784, pruned_loss=0.05696, over 1764591.06 frames. ], batch size: 48, lr: 5.82e-03, grad_scale: 2.0 2023-04-29 12:52:38,838 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-29 12:53:03,165 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111824.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:53:13,690 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1700, 5.1262, 4.9600, 4.5031, 4.8322, 1.9030, 4.6748, 4.9550], device='cuda:4'), covar=tensor([0.0071, 0.0067, 0.0162, 0.0302, 0.0088, 0.2292, 0.0126, 0.0173], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0112, 0.0162, 0.0147, 0.0132, 0.0181, 0.0149, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 12:53:40,725 INFO [train.py:904] (4/8) Epoch 12, batch 200, loss[loss=0.1863, simple_loss=0.2806, pruned_loss=0.04601, over 16816.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2779, pruned_loss=0.05684, over 2112589.57 frames. ], batch size: 42, lr: 5.82e-03, grad_scale: 2.0 2023-04-29 12:53:41,259 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=111852.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 12:53:50,233 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.787e+02 2.566e+02 3.072e+02 3.744e+02 9.632e+02, threshold=6.144e+02, percent-clipped=5.0 2023-04-29 12:53:52,989 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-29 12:54:21,378 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111881.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:54:49,915 INFO [train.py:904] (4/8) Epoch 12, batch 250, loss[loss=0.1819, simple_loss=0.2768, pruned_loss=0.04353, over 17235.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2761, pruned_loss=0.05691, over 2381748.56 frames. ], batch size: 45, lr: 5.82e-03, grad_scale: 2.0 2023-04-29 12:55:27,223 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=111929.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:55:32,716 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9157, 1.9627, 2.2169, 2.9433, 2.6070, 3.4075, 2.2231, 3.2536], device='cuda:4'), covar=tensor([0.0167, 0.0346, 0.0256, 0.0198, 0.0242, 0.0122, 0.0316, 0.0110], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0171, 0.0153, 0.0159, 0.0168, 0.0123, 0.0170, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 12:55:57,751 INFO [train.py:904] (4/8) Epoch 12, batch 300, loss[loss=0.1786, simple_loss=0.2709, pruned_loss=0.04312, over 17066.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2728, pruned_loss=0.05487, over 2592645.79 frames. ], batch size: 50, lr: 5.82e-03, grad_scale: 1.0 2023-04-29 12:56:09,470 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.668e+02 2.376e+02 2.773e+02 3.192e+02 6.381e+02, threshold=5.545e+02, percent-clipped=1.0 2023-04-29 12:57:10,667 INFO [train.py:904] (4/8) Epoch 12, batch 350, loss[loss=0.1555, simple_loss=0.2334, pruned_loss=0.0388, over 16069.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2705, pruned_loss=0.05338, over 2761026.84 frames. ], batch size: 35, lr: 5.81e-03, grad_scale: 1.0 2023-04-29 12:58:17,741 INFO [train.py:904] (4/8) Epoch 12, batch 400, loss[loss=0.2018, simple_loss=0.2744, pruned_loss=0.06459, over 16900.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2692, pruned_loss=0.05246, over 2892414.52 frames. ], batch size: 109, lr: 5.81e-03, grad_scale: 2.0 2023-04-29 12:58:27,674 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.698e+02 2.508e+02 2.981e+02 3.635e+02 6.653e+02, threshold=5.962e+02, percent-clipped=1.0 2023-04-29 12:58:37,003 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112066.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 12:58:53,586 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.52 vs. limit=5.0 2023-04-29 12:59:26,047 INFO [train.py:904] (4/8) Epoch 12, batch 450, loss[loss=0.1713, simple_loss=0.2616, pruned_loss=0.04046, over 17036.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2671, pruned_loss=0.05198, over 2990800.32 frames. ], batch size: 55, lr: 5.81e-03, grad_scale: 2.0 2023-04-29 12:59:46,996 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0385, 5.5908, 5.7165, 5.4117, 5.4197, 6.0819, 5.5887, 5.3857], device='cuda:4'), covar=tensor([0.0833, 0.1840, 0.2356, 0.2204, 0.3254, 0.1094, 0.1451, 0.2441], device='cuda:4'), in_proj_covar=tensor([0.0351, 0.0501, 0.0550, 0.0435, 0.0581, 0.0579, 0.0436, 0.0583], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 12:59:55,934 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=112124.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:00:00,727 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112127.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:00:27,067 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112147.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 13:00:33,812 INFO [train.py:904] (4/8) Epoch 12, batch 500, loss[loss=0.1664, simple_loss=0.2572, pruned_loss=0.03776, over 16645.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2657, pruned_loss=0.05123, over 3065105.71 frames. ], batch size: 62, lr: 5.81e-03, grad_scale: 2.0 2023-04-29 13:00:45,214 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.367e+02 2.270e+02 2.759e+02 3.532e+02 6.724e+02, threshold=5.519e+02, percent-clipped=2.0 2023-04-29 13:01:02,760 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=112172.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:01:19,685 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-29 13:01:29,291 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 13:01:32,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6799, 3.7740, 4.1125, 4.0899, 4.0936, 3.8364, 3.8638, 3.8308], device='cuda:4'), covar=tensor([0.0386, 0.0646, 0.0423, 0.0428, 0.0454, 0.0433, 0.0801, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0346, 0.0348, 0.0328, 0.0389, 0.0367, 0.0468, 0.0298], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 13:01:44,736 INFO [train.py:904] (4/8) Epoch 12, batch 550, loss[loss=0.1663, simple_loss=0.2524, pruned_loss=0.04011, over 16616.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.264, pruned_loss=0.05005, over 3118626.53 frames. ], batch size: 68, lr: 5.81e-03, grad_scale: 2.0 2023-04-29 13:02:55,015 INFO [train.py:904] (4/8) Epoch 12, batch 600, loss[loss=0.1904, simple_loss=0.2559, pruned_loss=0.06241, over 16700.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2636, pruned_loss=0.05123, over 3151258.46 frames. ], batch size: 134, lr: 5.81e-03, grad_scale: 2.0 2023-04-29 13:03:06,853 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.658e+02 2.365e+02 2.773e+02 3.420e+02 1.272e+03, threshold=5.547e+02, percent-clipped=1.0 2023-04-29 13:04:02,126 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-29 13:04:04,843 INFO [train.py:904] (4/8) Epoch 12, batch 650, loss[loss=0.1619, simple_loss=0.2513, pruned_loss=0.03623, over 17225.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2612, pruned_loss=0.05038, over 3192910.37 frames. ], batch size: 44, lr: 5.81e-03, grad_scale: 2.0 2023-04-29 13:05:14,225 INFO [train.py:904] (4/8) Epoch 12, batch 700, loss[loss=0.1819, simple_loss=0.2675, pruned_loss=0.04816, over 17116.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2616, pruned_loss=0.0503, over 3221437.53 frames. ], batch size: 47, lr: 5.81e-03, grad_scale: 2.0 2023-04-29 13:05:14,650 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8055, 4.4109, 4.5089, 4.9972, 5.1433, 4.5761, 5.1954, 5.1199], device='cuda:4'), covar=tensor([0.1625, 0.1606, 0.3010, 0.1101, 0.0848, 0.0911, 0.0771, 0.0896], device='cuda:4'), in_proj_covar=tensor([0.0543, 0.0676, 0.0820, 0.0683, 0.0521, 0.0527, 0.0542, 0.0618], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:05:26,019 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.678e+02 2.509e+02 2.916e+02 3.539e+02 5.225e+02, threshold=5.832e+02, percent-clipped=0.0 2023-04-29 13:06:07,737 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112389.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:06:13,256 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6878, 1.6843, 1.5881, 1.5673, 1.9005, 1.5569, 1.6713, 1.9429], device='cuda:4'), covar=tensor([0.0157, 0.0254, 0.0343, 0.0319, 0.0176, 0.0251, 0.0166, 0.0176], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0210, 0.0202, 0.0202, 0.0209, 0.0208, 0.0212, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:06:14,378 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0584, 4.5401, 4.6046, 3.3647, 3.8415, 4.5570, 4.0753, 2.5787], device='cuda:4'), covar=tensor([0.0347, 0.0041, 0.0027, 0.0252, 0.0072, 0.0062, 0.0057, 0.0356], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0071, 0.0070, 0.0127, 0.0080, 0.0090, 0.0079, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:06:24,418 INFO [train.py:904] (4/8) Epoch 12, batch 750, loss[loss=0.1565, simple_loss=0.2444, pruned_loss=0.03429, over 16866.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2633, pruned_loss=0.0511, over 3234014.79 frames. ], batch size: 42, lr: 5.80e-03, grad_scale: 2.0 2023-04-29 13:06:24,850 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7001, 4.1323, 4.2942, 3.0241, 3.5918, 4.1927, 3.8113, 2.2174], device='cuda:4'), covar=tensor([0.0389, 0.0068, 0.0032, 0.0274, 0.0081, 0.0073, 0.0062, 0.0379], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0070, 0.0070, 0.0127, 0.0080, 0.0090, 0.0079, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:06:28,647 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8892, 1.3205, 1.5815, 1.7831, 1.8588, 1.9710, 1.5754, 1.7932], device='cuda:4'), covar=tensor([0.0176, 0.0298, 0.0146, 0.0213, 0.0183, 0.0138, 0.0304, 0.0090], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0174, 0.0154, 0.0161, 0.0169, 0.0126, 0.0173, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 13:06:51,814 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112422.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:07:19,612 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5119, 2.1612, 2.4541, 4.2792, 2.2159, 2.6659, 2.2669, 2.3919], device='cuda:4'), covar=tensor([0.0984, 0.3252, 0.2094, 0.0454, 0.3502, 0.2121, 0.3099, 0.2781], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0394, 0.0333, 0.0322, 0.0415, 0.0449, 0.0357, 0.0461], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:07:29,136 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=112447.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 13:07:32,700 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112450.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:07:34,635 INFO [train.py:904] (4/8) Epoch 12, batch 800, loss[loss=0.1802, simple_loss=0.2574, pruned_loss=0.05153, over 16782.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2631, pruned_loss=0.05062, over 3250751.30 frames. ], batch size: 102, lr: 5.80e-03, grad_scale: 4.0 2023-04-29 13:07:45,051 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.661e+02 2.488e+02 3.052e+02 3.720e+02 1.064e+03, threshold=6.105e+02, percent-clipped=2.0 2023-04-29 13:08:33,906 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=112495.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 13:08:42,798 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-29 13:08:42,875 INFO [train.py:904] (4/8) Epoch 12, batch 850, loss[loss=0.1927, simple_loss=0.2853, pruned_loss=0.05005, over 17048.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.262, pruned_loss=0.04924, over 3278983.31 frames. ], batch size: 55, lr: 5.80e-03, grad_scale: 4.0 2023-04-29 13:08:58,608 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6440, 4.4196, 4.6426, 4.8788, 4.9842, 4.4817, 4.9038, 4.9441], device='cuda:4'), covar=tensor([0.1351, 0.1064, 0.1369, 0.0562, 0.0503, 0.0920, 0.0862, 0.0580], device='cuda:4'), in_proj_covar=tensor([0.0560, 0.0693, 0.0841, 0.0703, 0.0534, 0.0540, 0.0553, 0.0635], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:09:26,327 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.77 vs. limit=5.0 2023-04-29 13:09:51,999 INFO [train.py:904] (4/8) Epoch 12, batch 900, loss[loss=0.174, simple_loss=0.2676, pruned_loss=0.0402, over 17121.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2608, pruned_loss=0.0479, over 3297641.74 frames. ], batch size: 47, lr: 5.80e-03, grad_scale: 4.0 2023-04-29 13:10:02,362 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.529e+02 2.271e+02 2.835e+02 3.426e+02 5.348e+02, threshold=5.671e+02, percent-clipped=0.0 2023-04-29 13:10:05,729 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112562.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:10:26,083 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112577.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:10:35,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8966, 4.4893, 3.2729, 2.3178, 2.7988, 2.5636, 4.6849, 3.7868], device='cuda:4'), covar=tensor([0.2502, 0.0540, 0.1488, 0.2356, 0.2573, 0.1727, 0.0337, 0.1042], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0259, 0.0284, 0.0277, 0.0275, 0.0223, 0.0266, 0.0298], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:10:59,465 INFO [train.py:904] (4/8) Epoch 12, batch 950, loss[loss=0.1569, simple_loss=0.2506, pruned_loss=0.03158, over 17206.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2606, pruned_loss=0.04733, over 3310587.27 frames. ], batch size: 45, lr: 5.80e-03, grad_scale: 4.0 2023-04-29 13:11:21,294 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112617.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:11:28,367 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112623.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:11:50,141 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112638.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 13:12:07,158 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9301, 4.1599, 2.4095, 4.6705, 3.0456, 4.5724, 2.4989, 3.2771], device='cuda:4'), covar=tensor([0.0237, 0.0284, 0.1494, 0.0167, 0.0692, 0.0522, 0.1484, 0.0649], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0166, 0.0189, 0.0134, 0.0167, 0.0208, 0.0198, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 13:12:07,851 INFO [train.py:904] (4/8) Epoch 12, batch 1000, loss[loss=0.1919, simple_loss=0.278, pruned_loss=0.05293, over 17233.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2608, pruned_loss=0.04801, over 3307384.67 frames. ], batch size: 45, lr: 5.80e-03, grad_scale: 4.0 2023-04-29 13:12:18,367 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.478e+02 2.238e+02 2.624e+02 3.056e+02 7.407e+02, threshold=5.248e+02, percent-clipped=1.0 2023-04-29 13:12:43,260 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112678.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 13:12:49,693 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4676, 2.9784, 2.6687, 2.2540, 2.2133, 2.2381, 2.9701, 2.8582], device='cuda:4'), covar=tensor([0.2227, 0.0777, 0.1425, 0.1975, 0.2251, 0.1815, 0.0493, 0.1039], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0258, 0.0283, 0.0276, 0.0275, 0.0223, 0.0266, 0.0298], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:13:15,633 INFO [train.py:904] (4/8) Epoch 12, batch 1050, loss[loss=0.1849, simple_loss=0.2674, pruned_loss=0.05118, over 16513.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2602, pruned_loss=0.04804, over 3300158.43 frames. ], batch size: 68, lr: 5.80e-03, grad_scale: 4.0 2023-04-29 13:13:32,593 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2053, 3.3285, 3.6364, 2.5730, 3.2960, 3.6560, 3.4093, 2.1556], device='cuda:4'), covar=tensor([0.0410, 0.0145, 0.0040, 0.0279, 0.0081, 0.0073, 0.0070, 0.0346], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0070, 0.0069, 0.0125, 0.0079, 0.0089, 0.0078, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:13:35,473 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-29 13:13:36,676 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7056, 3.7358, 3.9894, 2.0405, 4.1286, 4.1269, 3.2383, 3.2053], device='cuda:4'), covar=tensor([0.0693, 0.0176, 0.0185, 0.1090, 0.0057, 0.0126, 0.0332, 0.0314], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0101, 0.0089, 0.0141, 0.0070, 0.0106, 0.0121, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 13:13:42,904 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=112722.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:14:14,517 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112745.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:14:22,687 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0071, 4.5471, 4.6032, 3.5699, 3.9040, 4.4928, 4.0375, 2.7037], device='cuda:4'), covar=tensor([0.0371, 0.0036, 0.0022, 0.0225, 0.0068, 0.0059, 0.0058, 0.0333], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0070, 0.0069, 0.0125, 0.0079, 0.0089, 0.0078, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:14:23,335 INFO [train.py:904] (4/8) Epoch 12, batch 1100, loss[loss=0.1597, simple_loss=0.2454, pruned_loss=0.03699, over 17004.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2599, pruned_loss=0.04771, over 3305819.65 frames. ], batch size: 41, lr: 5.79e-03, grad_scale: 4.0 2023-04-29 13:14:34,071 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.585e+02 2.374e+02 2.893e+02 3.790e+02 9.509e+02, threshold=5.785e+02, percent-clipped=6.0 2023-04-29 13:14:48,742 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=112770.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:15:33,615 INFO [train.py:904] (4/8) Epoch 12, batch 1150, loss[loss=0.1579, simple_loss=0.2339, pruned_loss=0.04094, over 16868.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2586, pruned_loss=0.04747, over 3296880.79 frames. ], batch size: 96, lr: 5.79e-03, grad_scale: 4.0 2023-04-29 13:15:34,032 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0143, 4.3501, 4.4317, 3.5507, 3.7577, 4.4215, 3.9759, 2.7200], device='cuda:4'), covar=tensor([0.0319, 0.0032, 0.0026, 0.0196, 0.0068, 0.0060, 0.0059, 0.0321], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0071, 0.0070, 0.0127, 0.0080, 0.0089, 0.0078, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:15:50,997 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-29 13:15:57,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8286, 3.9893, 2.3402, 4.5393, 2.9096, 4.4762, 2.2548, 3.1584], device='cuda:4'), covar=tensor([0.0222, 0.0327, 0.1518, 0.0210, 0.0798, 0.0421, 0.1653, 0.0684], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0168, 0.0191, 0.0136, 0.0168, 0.0210, 0.0200, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 13:16:42,797 INFO [train.py:904] (4/8) Epoch 12, batch 1200, loss[loss=0.1736, simple_loss=0.2586, pruned_loss=0.04428, over 16749.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2581, pruned_loss=0.04696, over 3300310.58 frames. ], batch size: 57, lr: 5.79e-03, grad_scale: 8.0 2023-04-29 13:16:52,675 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.635e+02 2.211e+02 2.743e+02 3.250e+02 7.821e+02, threshold=5.486e+02, percent-clipped=2.0 2023-04-29 13:17:49,331 INFO [train.py:904] (4/8) Epoch 12, batch 1250, loss[loss=0.1601, simple_loss=0.2403, pruned_loss=0.03996, over 16000.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2575, pruned_loss=0.04703, over 3308166.57 frames. ], batch size: 35, lr: 5.79e-03, grad_scale: 8.0 2023-04-29 13:18:12,308 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112918.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:18:16,586 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112921.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:18:32,350 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112933.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 13:18:41,600 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-29 13:18:57,927 INFO [train.py:904] (4/8) Epoch 12, batch 1300, loss[loss=0.1964, simple_loss=0.2901, pruned_loss=0.05131, over 17060.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2568, pruned_loss=0.04686, over 3314526.38 frames. ], batch size: 53, lr: 5.79e-03, grad_scale: 8.0 2023-04-29 13:19:09,586 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.880e+02 2.412e+02 2.831e+02 3.329e+02 6.613e+02, threshold=5.661e+02, percent-clipped=2.0 2023-04-29 13:19:27,111 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112973.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 13:19:40,973 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112982.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:20:08,546 INFO [train.py:904] (4/8) Epoch 12, batch 1350, loss[loss=0.1634, simple_loss=0.2509, pruned_loss=0.03797, over 17200.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2578, pruned_loss=0.04664, over 3317192.79 frames. ], batch size: 44, lr: 5.79e-03, grad_scale: 8.0 2023-04-29 13:21:07,363 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113045.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:21:16,368 INFO [train.py:904] (4/8) Epoch 12, batch 1400, loss[loss=0.1618, simple_loss=0.2346, pruned_loss=0.04445, over 16289.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2578, pruned_loss=0.04672, over 3318268.53 frames. ], batch size: 36, lr: 5.79e-03, grad_scale: 8.0 2023-04-29 13:21:26,436 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.826e+02 2.440e+02 3.018e+02 3.818e+02 8.239e+02, threshold=6.035e+02, percent-clipped=8.0 2023-04-29 13:21:46,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8924, 2.8980, 2.4241, 2.9143, 3.1504, 3.0632, 3.6313, 3.3766], device='cuda:4'), covar=tensor([0.0090, 0.0263, 0.0358, 0.0270, 0.0200, 0.0236, 0.0186, 0.0187], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0214, 0.0206, 0.0205, 0.0213, 0.0213, 0.0218, 0.0203], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:22:13,293 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=113093.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:22:24,188 INFO [train.py:904] (4/8) Epoch 12, batch 1450, loss[loss=0.1618, simple_loss=0.2472, pruned_loss=0.03818, over 17221.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2575, pruned_loss=0.04603, over 3330416.61 frames. ], batch size: 44, lr: 5.79e-03, grad_scale: 8.0 2023-04-29 13:22:32,969 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-29 13:23:06,578 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.07 vs. limit=5.0 2023-04-29 13:23:35,102 INFO [train.py:904] (4/8) Epoch 12, batch 1500, loss[loss=0.1642, simple_loss=0.2553, pruned_loss=0.03653, over 17031.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2579, pruned_loss=0.0465, over 3323988.57 frames. ], batch size: 50, lr: 5.78e-03, grad_scale: 8.0 2023-04-29 13:23:45,776 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.735e+02 2.499e+02 2.892e+02 3.478e+02 9.992e+02, threshold=5.783e+02, percent-clipped=3.0 2023-04-29 13:23:50,354 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7696, 3.9401, 2.1475, 4.4171, 3.0795, 4.4127, 2.2621, 3.0080], device='cuda:4'), covar=tensor([0.0248, 0.0327, 0.1664, 0.0252, 0.0675, 0.0474, 0.1596, 0.0755], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0169, 0.0191, 0.0137, 0.0168, 0.0212, 0.0199, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 13:24:02,314 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-29 13:24:10,284 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6863, 4.5588, 4.5922, 4.3342, 4.2499, 4.6250, 4.4595, 4.4034], device='cuda:4'), covar=tensor([0.0672, 0.0695, 0.0294, 0.0281, 0.0930, 0.0468, 0.0488, 0.0684], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0342, 0.0312, 0.0289, 0.0332, 0.0337, 0.0213, 0.0362], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:24:43,442 INFO [train.py:904] (4/8) Epoch 12, batch 1550, loss[loss=0.2003, simple_loss=0.287, pruned_loss=0.05681, over 17061.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2595, pruned_loss=0.04737, over 3320599.94 frames. ], batch size: 53, lr: 5.78e-03, grad_scale: 8.0 2023-04-29 13:25:06,568 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113218.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:25:27,409 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113233.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:25:28,571 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=113234.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:25:54,128 INFO [train.py:904] (4/8) Epoch 12, batch 1600, loss[loss=0.1996, simple_loss=0.2887, pruned_loss=0.05528, over 16571.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2612, pruned_loss=0.04802, over 3322527.72 frames. ], batch size: 62, lr: 5.78e-03, grad_scale: 8.0 2023-04-29 13:26:04,702 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.814e+02 2.404e+02 2.815e+02 3.461e+02 5.184e+02, threshold=5.631e+02, percent-clipped=0.0 2023-04-29 13:26:12,638 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=113266.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:26:23,394 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113273.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 13:26:28,587 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=113277.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:26:33,950 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=113281.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:26:53,638 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=113295.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:27:02,727 INFO [train.py:904] (4/8) Epoch 12, batch 1650, loss[loss=0.1839, simple_loss=0.2615, pruned_loss=0.05315, over 16763.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2624, pruned_loss=0.0489, over 3321620.25 frames. ], batch size: 83, lr: 5.78e-03, grad_scale: 8.0 2023-04-29 13:27:29,756 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=113321.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:27:45,243 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4406, 3.3070, 2.6756, 2.1425, 2.2775, 2.1782, 3.3118, 3.1019], device='cuda:4'), covar=tensor([0.2534, 0.0766, 0.1550, 0.2285, 0.2285, 0.1883, 0.0525, 0.1200], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0258, 0.0282, 0.0278, 0.0278, 0.0224, 0.0266, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:27:57,963 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5597, 2.9117, 2.6179, 4.8222, 3.9805, 4.3699, 1.5213, 3.1140], device='cuda:4'), covar=tensor([0.1412, 0.0682, 0.1150, 0.0177, 0.0270, 0.0387, 0.1530, 0.0778], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0158, 0.0179, 0.0148, 0.0193, 0.0209, 0.0180, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 13:28:12,393 INFO [train.py:904] (4/8) Epoch 12, batch 1700, loss[loss=0.1999, simple_loss=0.2811, pruned_loss=0.05931, over 16816.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2645, pruned_loss=0.04995, over 3323823.60 frames. ], batch size: 83, lr: 5.78e-03, grad_scale: 8.0 2023-04-29 13:28:23,616 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.528e+02 2.537e+02 3.070e+02 3.736e+02 6.116e+02, threshold=6.140e+02, percent-clipped=1.0 2023-04-29 13:29:08,151 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3280, 2.0935, 2.3165, 3.9526, 2.1664, 2.4952, 2.2063, 2.3116], device='cuda:4'), covar=tensor([0.1066, 0.3305, 0.2267, 0.0476, 0.3349, 0.2248, 0.3171, 0.2796], device='cuda:4'), in_proj_covar=tensor([0.0370, 0.0398, 0.0336, 0.0325, 0.0416, 0.0456, 0.0361, 0.0467], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:29:22,273 INFO [train.py:904] (4/8) Epoch 12, batch 1750, loss[loss=0.1832, simple_loss=0.2712, pruned_loss=0.04761, over 15847.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2659, pruned_loss=0.05051, over 3321542.08 frames. ], batch size: 35, lr: 5.78e-03, grad_scale: 8.0 2023-04-29 13:29:53,648 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8247, 4.5656, 4.8465, 5.0347, 5.2268, 4.6282, 5.1795, 5.1832], device='cuda:4'), covar=tensor([0.1434, 0.1094, 0.1543, 0.0672, 0.0445, 0.0893, 0.0513, 0.0500], device='cuda:4'), in_proj_covar=tensor([0.0564, 0.0703, 0.0851, 0.0717, 0.0539, 0.0550, 0.0557, 0.0645], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:30:32,313 INFO [train.py:904] (4/8) Epoch 12, batch 1800, loss[loss=0.1577, simple_loss=0.2378, pruned_loss=0.03882, over 16954.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2665, pruned_loss=0.04993, over 3319883.44 frames. ], batch size: 41, lr: 5.78e-03, grad_scale: 8.0 2023-04-29 13:30:43,443 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.551e+02 2.362e+02 2.904e+02 3.594e+02 5.616e+02, threshold=5.809e+02, percent-clipped=0.0 2023-04-29 13:31:03,062 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6863, 4.6389, 4.5586, 4.0377, 4.5760, 1.8155, 4.3436, 4.3085], device='cuda:4'), covar=tensor([0.0098, 0.0082, 0.0141, 0.0304, 0.0091, 0.2321, 0.0121, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0124, 0.0173, 0.0162, 0.0144, 0.0188, 0.0161, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:31:42,917 INFO [train.py:904] (4/8) Epoch 12, batch 1850, loss[loss=0.2037, simple_loss=0.2882, pruned_loss=0.05956, over 16235.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2678, pruned_loss=0.0504, over 3320472.38 frames. ], batch size: 165, lr: 5.78e-03, grad_scale: 8.0 2023-04-29 13:32:53,282 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-29 13:32:53,584 INFO [train.py:904] (4/8) Epoch 12, batch 1900, loss[loss=0.1664, simple_loss=0.2574, pruned_loss=0.03772, over 16437.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2676, pruned_loss=0.04969, over 3325065.38 frames. ], batch size: 68, lr: 5.77e-03, grad_scale: 8.0 2023-04-29 13:33:04,793 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.279e+02 2.197e+02 2.521e+02 3.135e+02 9.394e+02, threshold=5.041e+02, percent-clipped=1.0 2023-04-29 13:33:29,736 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113577.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:33:48,964 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=113590.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:34:05,628 INFO [train.py:904] (4/8) Epoch 12, batch 1950, loss[loss=0.1694, simple_loss=0.2525, pruned_loss=0.04311, over 17210.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2682, pruned_loss=0.04979, over 3317851.12 frames. ], batch size: 44, lr: 5.77e-03, grad_scale: 8.0 2023-04-29 13:34:27,378 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=113617.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:34:37,513 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7283, 2.6597, 2.3712, 2.5019, 2.9151, 2.8085, 3.4259, 3.2449], device='cuda:4'), covar=tensor([0.0081, 0.0293, 0.0353, 0.0342, 0.0197, 0.0277, 0.0195, 0.0197], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0213, 0.0206, 0.0205, 0.0214, 0.0212, 0.0221, 0.0203], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:34:39,273 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=113625.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:35:08,128 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=113646.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:35:16,104 INFO [train.py:904] (4/8) Epoch 12, batch 2000, loss[loss=0.1809, simple_loss=0.258, pruned_loss=0.05191, over 16848.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.2674, pruned_loss=0.04964, over 3318188.62 frames. ], batch size: 96, lr: 5.77e-03, grad_scale: 8.0 2023-04-29 13:35:27,901 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.513e+02 2.257e+02 2.755e+02 3.569e+02 6.259e+02, threshold=5.509e+02, percent-clipped=3.0 2023-04-29 13:35:52,975 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=113678.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:36:25,541 INFO [train.py:904] (4/8) Epoch 12, batch 2050, loss[loss=0.2199, simple_loss=0.2918, pruned_loss=0.07402, over 16487.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2668, pruned_loss=0.05031, over 3317504.53 frames. ], batch size: 146, lr: 5.77e-03, grad_scale: 8.0 2023-04-29 13:36:32,485 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=113707.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:37:33,719 INFO [train.py:904] (4/8) Epoch 12, batch 2100, loss[loss=0.1654, simple_loss=0.2507, pruned_loss=0.04008, over 17015.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.2673, pruned_loss=0.05077, over 3315811.53 frames. ], batch size: 41, lr: 5.77e-03, grad_scale: 8.0 2023-04-29 13:37:45,425 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.512e+02 2.602e+02 3.022e+02 3.669e+02 5.748e+02, threshold=6.044e+02, percent-clipped=1.0 2023-04-29 13:38:30,011 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3052, 5.1222, 5.1473, 4.7607, 4.7544, 5.1505, 5.1491, 4.7646], device='cuda:4'), covar=tensor([0.0530, 0.0395, 0.0284, 0.0273, 0.1063, 0.0357, 0.0347, 0.0774], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0348, 0.0316, 0.0294, 0.0337, 0.0340, 0.0215, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:38:39,673 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8458, 1.8450, 2.2812, 2.7534, 2.7965, 2.6922, 1.8670, 2.9435], device='cuda:4'), covar=tensor([0.0133, 0.0347, 0.0249, 0.0179, 0.0176, 0.0183, 0.0342, 0.0085], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0177, 0.0158, 0.0163, 0.0173, 0.0130, 0.0176, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 13:38:44,795 INFO [train.py:904] (4/8) Epoch 12, batch 2150, loss[loss=0.2167, simple_loss=0.3105, pruned_loss=0.06148, over 16985.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.268, pruned_loss=0.05079, over 3321462.24 frames. ], batch size: 55, lr: 5.77e-03, grad_scale: 8.0 2023-04-29 13:39:54,115 INFO [train.py:904] (4/8) Epoch 12, batch 2200, loss[loss=0.184, simple_loss=0.2793, pruned_loss=0.04436, over 17059.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2684, pruned_loss=0.05085, over 3321619.02 frames. ], batch size: 50, lr: 5.77e-03, grad_scale: 8.0 2023-04-29 13:40:02,383 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0886, 4.3069, 4.3762, 3.3348, 3.7070, 4.1833, 3.8906, 2.5964], device='cuda:4'), covar=tensor([0.0332, 0.0059, 0.0027, 0.0238, 0.0087, 0.0080, 0.0069, 0.0343], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0072, 0.0070, 0.0125, 0.0080, 0.0090, 0.0079, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:40:05,118 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 2.389e+02 2.779e+02 3.462e+02 6.586e+02, threshold=5.558e+02, percent-clipped=1.0 2023-04-29 13:40:05,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0017, 2.8618, 3.1517, 2.2308, 2.9275, 3.1729, 2.9535, 1.9414], device='cuda:4'), covar=tensor([0.0409, 0.0117, 0.0048, 0.0293, 0.0099, 0.0087, 0.0086, 0.0333], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0072, 0.0071, 0.0125, 0.0080, 0.0090, 0.0079, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:40:48,886 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113890.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:41:03,698 INFO [train.py:904] (4/8) Epoch 12, batch 2250, loss[loss=0.1634, simple_loss=0.2521, pruned_loss=0.03739, over 17112.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2682, pruned_loss=0.05076, over 3326596.23 frames. ], batch size: 47, lr: 5.77e-03, grad_scale: 8.0 2023-04-29 13:41:17,198 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6637, 2.5436, 2.1468, 2.4679, 2.8071, 2.6498, 3.4200, 3.1229], device='cuda:4'), covar=tensor([0.0083, 0.0291, 0.0367, 0.0320, 0.0212, 0.0273, 0.0163, 0.0190], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0211, 0.0203, 0.0203, 0.0211, 0.0209, 0.0218, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:41:54,643 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=113938.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:42:14,191 INFO [train.py:904] (4/8) Epoch 12, batch 2300, loss[loss=0.1837, simple_loss=0.2663, pruned_loss=0.05054, over 16726.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2686, pruned_loss=0.05039, over 3334040.78 frames. ], batch size: 134, lr: 5.76e-03, grad_scale: 16.0 2023-04-29 13:42:24,216 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.424e+02 2.413e+02 2.819e+02 3.398e+02 7.599e+02, threshold=5.637e+02, percent-clipped=2.0 2023-04-29 13:42:43,120 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=113973.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:43:25,573 INFO [train.py:904] (4/8) Epoch 12, batch 2350, loss[loss=0.2459, simple_loss=0.3105, pruned_loss=0.0907, over 16912.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2694, pruned_loss=0.05137, over 3335813.71 frames. ], batch size: 109, lr: 5.76e-03, grad_scale: 16.0 2023-04-29 13:43:25,835 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=114002.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:43:51,728 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-29 13:44:06,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6703, 4.5759, 4.5865, 4.3123, 4.2766, 4.6296, 4.4364, 4.3513], device='cuda:4'), covar=tensor([0.0675, 0.0780, 0.0275, 0.0323, 0.0792, 0.0422, 0.0531, 0.0710], device='cuda:4'), in_proj_covar=tensor([0.0265, 0.0346, 0.0313, 0.0293, 0.0332, 0.0338, 0.0214, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:44:35,991 INFO [train.py:904] (4/8) Epoch 12, batch 2400, loss[loss=0.1912, simple_loss=0.2764, pruned_loss=0.05299, over 16508.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2699, pruned_loss=0.05183, over 3340510.19 frames. ], batch size: 68, lr: 5.76e-03, grad_scale: 16.0 2023-04-29 13:44:48,487 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.062e+02 2.572e+02 3.152e+02 4.032e+02 9.919e+02, threshold=6.305e+02, percent-clipped=6.0 2023-04-29 13:45:49,044 INFO [train.py:904] (4/8) Epoch 12, batch 2450, loss[loss=0.1751, simple_loss=0.2654, pruned_loss=0.04241, over 17123.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2703, pruned_loss=0.05123, over 3336847.78 frames. ], batch size: 48, lr: 5.76e-03, grad_scale: 16.0 2023-04-29 13:46:23,307 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=114127.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:46:57,384 INFO [train.py:904] (4/8) Epoch 12, batch 2500, loss[loss=0.1818, simple_loss=0.2592, pruned_loss=0.05222, over 16905.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2696, pruned_loss=0.05046, over 3332200.87 frames. ], batch size: 116, lr: 5.76e-03, grad_scale: 16.0 2023-04-29 13:47:04,216 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4911, 3.4573, 3.7151, 2.5982, 3.3540, 3.7599, 3.5526, 2.1776], device='cuda:4'), covar=tensor([0.0372, 0.0098, 0.0035, 0.0284, 0.0080, 0.0066, 0.0055, 0.0338], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0070, 0.0070, 0.0124, 0.0080, 0.0089, 0.0079, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:47:09,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.591e+02 2.290e+02 2.668e+02 3.158e+02 5.614e+02, threshold=5.335e+02, percent-clipped=0.0 2023-04-29 13:47:21,231 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-29 13:47:42,653 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.47 vs. limit=5.0 2023-04-29 13:47:42,850 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-29 13:47:48,618 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=114188.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:48:06,982 INFO [train.py:904] (4/8) Epoch 12, batch 2550, loss[loss=0.2152, simple_loss=0.2855, pruned_loss=0.07248, over 16659.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.271, pruned_loss=0.05128, over 3328760.02 frames. ], batch size: 134, lr: 5.76e-03, grad_scale: 16.0 2023-04-29 13:48:17,236 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-29 13:48:31,875 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9028, 1.8681, 2.2831, 2.7908, 2.7343, 3.3473, 2.1347, 3.1670], device='cuda:4'), covar=tensor([0.0173, 0.0391, 0.0281, 0.0253, 0.0227, 0.0133, 0.0356, 0.0120], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0179, 0.0159, 0.0166, 0.0174, 0.0131, 0.0176, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:48:52,658 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9349, 3.2163, 2.9026, 5.0678, 4.0676, 4.5256, 1.8027, 3.2861], device='cuda:4'), covar=tensor([0.1233, 0.0658, 0.1035, 0.0192, 0.0320, 0.0391, 0.1395, 0.0697], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0160, 0.0182, 0.0152, 0.0197, 0.0213, 0.0182, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 13:49:05,337 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3830, 5.3685, 5.1418, 4.4682, 5.1770, 2.1166, 4.9600, 5.1764], device='cuda:4'), covar=tensor([0.0086, 0.0070, 0.0167, 0.0384, 0.0090, 0.2254, 0.0123, 0.0160], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0125, 0.0176, 0.0163, 0.0146, 0.0187, 0.0163, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:49:15,388 INFO [train.py:904] (4/8) Epoch 12, batch 2600, loss[loss=0.1814, simple_loss=0.2615, pruned_loss=0.0507, over 16838.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2703, pruned_loss=0.05071, over 3331274.40 frames. ], batch size: 96, lr: 5.76e-03, grad_scale: 16.0 2023-04-29 13:49:25,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.778e+02 2.446e+02 2.880e+02 3.491e+02 5.288e+02, threshold=5.760e+02, percent-clipped=0.0 2023-04-29 13:49:45,801 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=114273.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:50:24,177 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-29 13:50:24,314 INFO [train.py:904] (4/8) Epoch 12, batch 2650, loss[loss=0.1715, simple_loss=0.2534, pruned_loss=0.04479, over 16838.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2713, pruned_loss=0.0509, over 3337289.23 frames. ], batch size: 39, lr: 5.76e-03, grad_scale: 16.0 2023-04-29 13:50:24,612 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=114302.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:50:41,599 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5484, 2.2464, 1.7343, 2.0852, 2.6859, 2.4790, 2.7714, 2.8225], device='cuda:4'), covar=tensor([0.0150, 0.0310, 0.0454, 0.0355, 0.0175, 0.0241, 0.0180, 0.0197], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0214, 0.0205, 0.0206, 0.0213, 0.0213, 0.0222, 0.0205], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:50:45,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8635, 1.7380, 2.3012, 2.8475, 2.7346, 3.3281, 2.0436, 3.2610], device='cuda:4'), covar=tensor([0.0171, 0.0398, 0.0251, 0.0207, 0.0219, 0.0126, 0.0364, 0.0095], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0177, 0.0157, 0.0164, 0.0172, 0.0130, 0.0174, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 13:50:51,415 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=114321.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:50:51,935 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.17 vs. limit=5.0 2023-04-29 13:51:32,414 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=114350.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:51:35,147 INFO [train.py:904] (4/8) Epoch 12, batch 2700, loss[loss=0.1907, simple_loss=0.2704, pruned_loss=0.05555, over 16743.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2717, pruned_loss=0.05079, over 3335497.11 frames. ], batch size: 124, lr: 5.75e-03, grad_scale: 16.0 2023-04-29 13:51:45,484 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.582e+02 2.425e+02 2.890e+02 3.523e+02 1.000e+03, threshold=5.781e+02, percent-clipped=5.0 2023-04-29 13:52:22,019 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1550, 1.9843, 1.5545, 1.7824, 2.2800, 2.0914, 2.2108, 2.4605], device='cuda:4'), covar=tensor([0.0197, 0.0297, 0.0385, 0.0323, 0.0171, 0.0230, 0.0182, 0.0191], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0215, 0.0206, 0.0206, 0.0214, 0.0213, 0.0223, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 13:52:44,856 INFO [train.py:904] (4/8) Epoch 12, batch 2750, loss[loss=0.1812, simple_loss=0.257, pruned_loss=0.0527, over 16941.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2722, pruned_loss=0.05053, over 3330581.98 frames. ], batch size: 96, lr: 5.75e-03, grad_scale: 16.0 2023-04-29 13:53:23,528 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9055, 4.0111, 2.6127, 4.6218, 3.0260, 4.5777, 2.4425, 3.2118], device='cuda:4'), covar=tensor([0.0217, 0.0312, 0.1214, 0.0165, 0.0758, 0.0395, 0.1423, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0170, 0.0190, 0.0141, 0.0171, 0.0215, 0.0199, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 13:53:54,633 INFO [train.py:904] (4/8) Epoch 12, batch 2800, loss[loss=0.1572, simple_loss=0.2399, pruned_loss=0.03722, over 17199.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2716, pruned_loss=0.04972, over 3329690.63 frames. ], batch size: 44, lr: 5.75e-03, grad_scale: 16.0 2023-04-29 13:54:06,075 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.709e+02 2.386e+02 2.760e+02 3.486e+02 6.287e+02, threshold=5.520e+02, percent-clipped=1.0 2023-04-29 13:54:38,612 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=114483.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 13:55:04,045 INFO [train.py:904] (4/8) Epoch 12, batch 2850, loss[loss=0.1957, simple_loss=0.29, pruned_loss=0.0507, over 16610.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2708, pruned_loss=0.04959, over 3326378.24 frames. ], batch size: 57, lr: 5.75e-03, grad_scale: 16.0 2023-04-29 13:55:27,389 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4917, 3.5958, 3.7972, 2.6212, 3.4633, 3.8228, 3.6681, 2.2976], device='cuda:4'), covar=tensor([0.0412, 0.0124, 0.0048, 0.0314, 0.0085, 0.0091, 0.0066, 0.0370], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0072, 0.0071, 0.0126, 0.0082, 0.0091, 0.0081, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 13:56:13,202 INFO [train.py:904] (4/8) Epoch 12, batch 2900, loss[loss=0.2095, simple_loss=0.2785, pruned_loss=0.07029, over 16728.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2694, pruned_loss=0.04988, over 3326404.87 frames. ], batch size: 134, lr: 5.75e-03, grad_scale: 8.0 2023-04-29 13:56:24,537 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.629e+02 2.459e+02 2.961e+02 3.423e+02 5.764e+02, threshold=5.923e+02, percent-clipped=1.0 2023-04-29 13:57:20,462 INFO [train.py:904] (4/8) Epoch 12, batch 2950, loss[loss=0.2014, simple_loss=0.2811, pruned_loss=0.06091, over 15600.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2692, pruned_loss=0.05047, over 3327500.17 frames. ], batch size: 191, lr: 5.75e-03, grad_scale: 8.0 2023-04-29 13:58:28,575 INFO [train.py:904] (4/8) Epoch 12, batch 3000, loss[loss=0.1853, simple_loss=0.2636, pruned_loss=0.05345, over 16400.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2689, pruned_loss=0.0505, over 3328240.54 frames. ], batch size: 146, lr: 5.75e-03, grad_scale: 8.0 2023-04-29 13:58:28,575 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 13:58:38,475 INFO [train.py:938] (4/8) Epoch 12, validation: loss=0.14, simple_loss=0.2459, pruned_loss=0.01708, over 944034.00 frames. 2023-04-29 13:58:38,476 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 13:58:50,228 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.524e+02 2.323e+02 2.752e+02 3.480e+02 1.028e+03, threshold=5.504e+02, percent-clipped=2.0 2023-04-29 13:59:48,674 INFO [train.py:904] (4/8) Epoch 12, batch 3050, loss[loss=0.1779, simple_loss=0.2676, pruned_loss=0.04405, over 17125.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2688, pruned_loss=0.05025, over 3325167.94 frames. ], batch size: 48, lr: 5.75e-03, grad_scale: 8.0 2023-04-29 13:59:54,095 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3371, 4.3408, 4.2536, 4.0831, 3.9501, 4.3129, 4.1329, 4.0424], device='cuda:4'), covar=tensor([0.0739, 0.0604, 0.0346, 0.0287, 0.0914, 0.0480, 0.0611, 0.0679], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0348, 0.0320, 0.0296, 0.0338, 0.0341, 0.0216, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 14:00:08,926 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=114716.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:00:52,873 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9764, 2.0041, 2.3330, 2.9959, 2.7149, 3.3886, 2.0372, 3.3132], device='cuda:4'), covar=tensor([0.0181, 0.0396, 0.0261, 0.0220, 0.0240, 0.0132, 0.0401, 0.0120], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0177, 0.0160, 0.0166, 0.0175, 0.0131, 0.0176, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 14:00:56,623 INFO [train.py:904] (4/8) Epoch 12, batch 3100, loss[loss=0.1748, simple_loss=0.2504, pruned_loss=0.04957, over 16840.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2673, pruned_loss=0.04961, over 3326785.01 frames. ], batch size: 90, lr: 5.74e-03, grad_scale: 8.0 2023-04-29 14:00:56,996 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1903, 3.2397, 3.4897, 2.3610, 3.1991, 3.5460, 3.3522, 2.0392], device='cuda:4'), covar=tensor([0.0352, 0.0087, 0.0042, 0.0281, 0.0082, 0.0073, 0.0068, 0.0330], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0071, 0.0071, 0.0124, 0.0081, 0.0089, 0.0079, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 14:01:10,357 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.272e+02 2.804e+02 3.413e+02 6.523e+02, threshold=5.608e+02, percent-clipped=1.0 2023-04-29 14:01:31,835 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=114777.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:01:40,366 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=114783.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:01:56,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0719, 3.2814, 3.2465, 2.0832, 2.7596, 2.2926, 3.5410, 3.5384], device='cuda:4'), covar=tensor([0.0230, 0.0737, 0.0547, 0.1626, 0.0783, 0.0897, 0.0506, 0.0823], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0148, 0.0159, 0.0144, 0.0136, 0.0124, 0.0138, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 14:02:05,282 INFO [train.py:904] (4/8) Epoch 12, batch 3150, loss[loss=0.1801, simple_loss=0.269, pruned_loss=0.04556, over 16582.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2673, pruned_loss=0.04982, over 3319797.69 frames. ], batch size: 68, lr: 5.74e-03, grad_scale: 8.0 2023-04-29 14:02:45,192 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=114831.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:03:14,011 INFO [train.py:904] (4/8) Epoch 12, batch 3200, loss[loss=0.174, simple_loss=0.2666, pruned_loss=0.0407, over 17241.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2665, pruned_loss=0.0495, over 3312918.27 frames. ], batch size: 45, lr: 5.74e-03, grad_scale: 8.0 2023-04-29 14:03:26,055 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.649e+02 2.483e+02 2.978e+02 3.563e+02 6.234e+02, threshold=5.956e+02, percent-clipped=4.0 2023-04-29 14:03:40,745 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=114871.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:04:22,406 INFO [train.py:904] (4/8) Epoch 12, batch 3250, loss[loss=0.1891, simple_loss=0.2848, pruned_loss=0.04672, over 17099.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2676, pruned_loss=0.05023, over 3314559.79 frames. ], batch size: 47, lr: 5.74e-03, grad_scale: 4.0 2023-04-29 14:05:05,264 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=114932.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:05:32,344 INFO [train.py:904] (4/8) Epoch 12, batch 3300, loss[loss=0.2017, simple_loss=0.2823, pruned_loss=0.06054, over 16880.00 frames. ], tot_loss[loss=0.1845, simple_loss=0.2686, pruned_loss=0.05022, over 3320078.70 frames. ], batch size: 116, lr: 5.74e-03, grad_scale: 4.0 2023-04-29 14:05:45,364 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.497e+02 2.351e+02 3.077e+02 3.881e+02 7.792e+02, threshold=6.153e+02, percent-clipped=7.0 2023-04-29 14:06:24,930 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7290, 5.0714, 4.8519, 4.8488, 4.5848, 4.5251, 4.5290, 5.1338], device='cuda:4'), covar=tensor([0.1103, 0.0834, 0.0953, 0.0667, 0.0713, 0.0960, 0.1072, 0.0843], device='cuda:4'), in_proj_covar=tensor([0.0588, 0.0729, 0.0597, 0.0509, 0.0456, 0.0464, 0.0606, 0.0563], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:06:42,021 INFO [train.py:904] (4/8) Epoch 12, batch 3350, loss[loss=0.1744, simple_loss=0.2677, pruned_loss=0.04057, over 17119.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.2685, pruned_loss=0.04992, over 3325511.39 frames. ], batch size: 48, lr: 5.74e-03, grad_scale: 4.0 2023-04-29 14:07:13,200 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-29 14:07:50,787 INFO [train.py:904] (4/8) Epoch 12, batch 3400, loss[loss=0.1542, simple_loss=0.2358, pruned_loss=0.03627, over 17224.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2682, pruned_loss=0.04959, over 3325575.97 frames. ], batch size: 45, lr: 5.74e-03, grad_scale: 4.0 2023-04-29 14:08:04,045 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.344e+02 2.703e+02 3.237e+02 1.034e+03, threshold=5.406e+02, percent-clipped=1.0 2023-04-29 14:08:08,703 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9310, 3.3558, 2.9584, 5.0799, 4.2619, 4.5542, 1.7307, 3.2902], device='cuda:4'), covar=tensor([0.1230, 0.0606, 0.0961, 0.0136, 0.0176, 0.0354, 0.1419, 0.0696], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0161, 0.0181, 0.0154, 0.0200, 0.0214, 0.0182, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 14:08:18,301 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115072.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:08:37,794 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=115086.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:09:00,024 INFO [train.py:904] (4/8) Epoch 12, batch 3450, loss[loss=0.1584, simple_loss=0.2499, pruned_loss=0.03338, over 17238.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2669, pruned_loss=0.04922, over 3322446.71 frames. ], batch size: 45, lr: 5.74e-03, grad_scale: 4.0 2023-04-29 14:09:15,890 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.40 vs. limit=5.0 2023-04-29 14:09:29,940 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6300, 3.6022, 2.8279, 2.1516, 2.4131, 2.2023, 3.5642, 3.2166], device='cuda:4'), covar=tensor([0.2431, 0.0631, 0.1481, 0.2506, 0.2457, 0.1886, 0.0516, 0.1299], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0259, 0.0285, 0.0281, 0.0284, 0.0224, 0.0269, 0.0304], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:10:02,233 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=115147.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:10:08,328 INFO [train.py:904] (4/8) Epoch 12, batch 3500, loss[loss=0.1623, simple_loss=0.2548, pruned_loss=0.03483, over 17157.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2658, pruned_loss=0.04927, over 3323904.07 frames. ], batch size: 48, lr: 5.73e-03, grad_scale: 4.0 2023-04-29 14:10:23,310 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.509e+02 2.367e+02 2.772e+02 3.300e+02 1.193e+03, threshold=5.543e+02, percent-clipped=2.0 2023-04-29 14:11:19,454 INFO [train.py:904] (4/8) Epoch 12, batch 3550, loss[loss=0.1522, simple_loss=0.2373, pruned_loss=0.03352, over 16236.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2643, pruned_loss=0.04851, over 3318728.90 frames. ], batch size: 36, lr: 5.73e-03, grad_scale: 4.0 2023-04-29 14:11:53,680 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2748, 4.6966, 4.0963, 4.5619, 4.3562, 4.2569, 4.2549, 4.7597], device='cuda:4'), covar=tensor([0.2288, 0.1686, 0.2974, 0.1377, 0.1326, 0.2236, 0.2185, 0.1804], device='cuda:4'), in_proj_covar=tensor([0.0584, 0.0729, 0.0595, 0.0510, 0.0456, 0.0462, 0.0604, 0.0565], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:11:53,688 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115227.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:12:28,666 INFO [train.py:904] (4/8) Epoch 12, batch 3600, loss[loss=0.1702, simple_loss=0.2424, pruned_loss=0.04898, over 16946.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2642, pruned_loss=0.04875, over 3311912.56 frames. ], batch size: 109, lr: 5.73e-03, grad_scale: 8.0 2023-04-29 14:12:28,982 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7428, 4.8003, 5.0263, 4.7989, 4.7391, 5.4501, 5.0700, 4.6997], device='cuda:4'), covar=tensor([0.1284, 0.2220, 0.1952, 0.2172, 0.3253, 0.1073, 0.1538, 0.2684], device='cuda:4'), in_proj_covar=tensor([0.0364, 0.0518, 0.0561, 0.0440, 0.0598, 0.0583, 0.0444, 0.0596], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 14:12:43,638 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.95 vs. limit=5.0 2023-04-29 14:12:43,850 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.756e+02 2.259e+02 2.631e+02 3.384e+02 1.021e+03, threshold=5.262e+02, percent-clipped=2.0 2023-04-29 14:12:57,383 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1441, 2.9541, 3.1699, 1.6631, 3.3215, 3.1984, 2.6660, 2.6065], device='cuda:4'), covar=tensor([0.0754, 0.0219, 0.0197, 0.1149, 0.0082, 0.0206, 0.0418, 0.0396], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0101, 0.0088, 0.0138, 0.0070, 0.0109, 0.0121, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 14:13:40,322 INFO [train.py:904] (4/8) Epoch 12, batch 3650, loss[loss=0.1591, simple_loss=0.2474, pruned_loss=0.0354, over 17192.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.263, pruned_loss=0.04927, over 3300232.00 frames. ], batch size: 44, lr: 5.73e-03, grad_scale: 8.0 2023-04-29 14:14:55,146 INFO [train.py:904] (4/8) Epoch 12, batch 3700, loss[loss=0.1801, simple_loss=0.2515, pruned_loss=0.05433, over 16505.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2616, pruned_loss=0.05098, over 3296941.45 frames. ], batch size: 68, lr: 5.73e-03, grad_scale: 8.0 2023-04-29 14:15:09,328 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.748e+02 2.309e+02 2.720e+02 3.193e+02 6.265e+02, threshold=5.440e+02, percent-clipped=2.0 2023-04-29 14:15:22,937 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=115370.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:15:24,736 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=115371.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:15:25,798 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=115372.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:15:36,175 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5409, 1.6767, 2.1872, 2.4255, 2.5665, 2.3787, 1.6836, 2.6034], device='cuda:4'), covar=tensor([0.0138, 0.0364, 0.0208, 0.0193, 0.0186, 0.0214, 0.0376, 0.0097], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0177, 0.0160, 0.0164, 0.0177, 0.0132, 0.0177, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:16:10,080 INFO [train.py:904] (4/8) Epoch 12, batch 3750, loss[loss=0.2087, simple_loss=0.2899, pruned_loss=0.06375, over 11506.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.2623, pruned_loss=0.05214, over 3274174.18 frames. ], batch size: 249, lr: 5.73e-03, grad_scale: 8.0 2023-04-29 14:16:36,698 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=115420.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:16:39,968 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-29 14:16:54,179 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=115431.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:16:55,339 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=115432.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:17:09,418 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115442.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:17:18,702 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5898, 1.6165, 2.1936, 2.5678, 2.6708, 2.4311, 1.6833, 2.6462], device='cuda:4'), covar=tensor([0.0148, 0.0402, 0.0250, 0.0208, 0.0188, 0.0224, 0.0391, 0.0107], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0179, 0.0161, 0.0165, 0.0178, 0.0133, 0.0178, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:17:23,444 INFO [train.py:904] (4/8) Epoch 12, batch 3800, loss[loss=0.1958, simple_loss=0.2728, pruned_loss=0.05936, over 16408.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2638, pruned_loss=0.05367, over 3286742.98 frames. ], batch size: 146, lr: 5.73e-03, grad_scale: 8.0 2023-04-29 14:17:38,965 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.548e+02 2.466e+02 2.817e+02 3.424e+02 6.065e+02, threshold=5.633e+02, percent-clipped=2.0 2023-04-29 14:18:28,649 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6708, 3.9317, 4.1986, 2.8988, 3.6959, 4.1842, 3.8950, 2.4181], device='cuda:4'), covar=tensor([0.0386, 0.0127, 0.0037, 0.0297, 0.0076, 0.0070, 0.0054, 0.0352], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0071, 0.0071, 0.0125, 0.0080, 0.0090, 0.0080, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 14:18:37,589 INFO [train.py:904] (4/8) Epoch 12, batch 3850, loss[loss=0.1746, simple_loss=0.2605, pruned_loss=0.04438, over 17269.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2641, pruned_loss=0.05422, over 3284693.02 frames. ], batch size: 52, lr: 5.73e-03, grad_scale: 8.0 2023-04-29 14:19:16,878 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=115527.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:19:52,697 INFO [train.py:904] (4/8) Epoch 12, batch 3900, loss[loss=0.1758, simple_loss=0.2488, pruned_loss=0.05142, over 16445.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2631, pruned_loss=0.05449, over 3288946.17 frames. ], batch size: 146, lr: 5.72e-03, grad_scale: 8.0 2023-04-29 14:20:06,432 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-29 14:20:07,961 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.661e+02 2.153e+02 2.652e+02 3.211e+02 6.333e+02, threshold=5.304e+02, percent-clipped=2.0 2023-04-29 14:20:29,517 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=115575.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:20:49,638 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9300, 4.0938, 3.0832, 2.4394, 2.8120, 2.5278, 4.1185, 3.7404], device='cuda:4'), covar=tensor([0.2095, 0.0556, 0.1359, 0.2181, 0.2284, 0.1629, 0.0484, 0.1011], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0258, 0.0285, 0.0280, 0.0287, 0.0224, 0.0269, 0.0303], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:20:56,471 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0543, 1.9538, 2.5875, 2.9596, 2.9025, 3.4573, 1.8713, 3.1357], device='cuda:4'), covar=tensor([0.0152, 0.0417, 0.0238, 0.0244, 0.0221, 0.0083, 0.0426, 0.0087], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0176, 0.0159, 0.0163, 0.0175, 0.0131, 0.0174, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:21:08,905 INFO [train.py:904] (4/8) Epoch 12, batch 3950, loss[loss=0.2047, simple_loss=0.2813, pruned_loss=0.06409, over 16505.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2623, pruned_loss=0.05466, over 3283920.87 frames. ], batch size: 68, lr: 5.72e-03, grad_scale: 8.0 2023-04-29 14:22:18,855 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4012, 5.4574, 5.3030, 5.0168, 4.9030, 5.3549, 5.1988, 5.0266], device='cuda:4'), covar=tensor([0.0524, 0.0219, 0.0224, 0.0226, 0.0939, 0.0270, 0.0267, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0262, 0.0340, 0.0311, 0.0288, 0.0330, 0.0333, 0.0209, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:22:21,472 INFO [train.py:904] (4/8) Epoch 12, batch 4000, loss[loss=0.1796, simple_loss=0.2569, pruned_loss=0.05109, over 16406.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2629, pruned_loss=0.05534, over 3274317.57 frames. ], batch size: 35, lr: 5.72e-03, grad_scale: 8.0 2023-04-29 14:22:34,737 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.680e+02 2.376e+02 2.718e+02 3.245e+02 5.190e+02, threshold=5.435e+02, percent-clipped=0.0 2023-04-29 14:23:35,776 INFO [train.py:904] (4/8) Epoch 12, batch 4050, loss[loss=0.1853, simple_loss=0.2651, pruned_loss=0.05278, over 16900.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2623, pruned_loss=0.05381, over 3276063.11 frames. ], batch size: 109, lr: 5.72e-03, grad_scale: 8.0 2023-04-29 14:23:36,394 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-29 14:23:41,266 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3900, 3.5525, 2.0566, 3.8495, 2.7316, 3.9766, 2.2102, 2.7136], device='cuda:4'), covar=tensor([0.0268, 0.0363, 0.1605, 0.0102, 0.0808, 0.0255, 0.1570, 0.0783], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0166, 0.0188, 0.0137, 0.0168, 0.0210, 0.0194, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 14:24:12,118 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115726.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:24:13,876 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115727.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:24:35,680 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=115742.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:24:50,500 INFO [train.py:904] (4/8) Epoch 12, batch 4100, loss[loss=0.1951, simple_loss=0.2851, pruned_loss=0.05255, over 16656.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2642, pruned_loss=0.05337, over 3271486.82 frames. ], batch size: 134, lr: 5.72e-03, grad_scale: 8.0 2023-04-29 14:25:05,533 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.445e+02 1.888e+02 2.192e+02 2.609e+02 4.553e+02, threshold=4.384e+02, percent-clipped=0.0 2023-04-29 14:25:25,509 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-29 14:25:48,321 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=115790.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:26:05,810 INFO [train.py:904] (4/8) Epoch 12, batch 4150, loss[loss=0.2184, simple_loss=0.3097, pruned_loss=0.06357, over 16230.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2718, pruned_loss=0.05624, over 3227807.07 frames. ], batch size: 165, lr: 5.72e-03, grad_scale: 8.0 2023-04-29 14:27:02,024 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5753, 2.0689, 1.6644, 1.9004, 2.4620, 2.1753, 2.4932, 2.7059], device='cuda:4'), covar=tensor([0.0120, 0.0320, 0.0397, 0.0362, 0.0180, 0.0278, 0.0133, 0.0179], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0208, 0.0201, 0.0202, 0.0208, 0.0206, 0.0214, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:27:23,037 INFO [train.py:904] (4/8) Epoch 12, batch 4200, loss[loss=0.2297, simple_loss=0.3018, pruned_loss=0.07886, over 11063.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2788, pruned_loss=0.0581, over 3190565.85 frames. ], batch size: 248, lr: 5.72e-03, grad_scale: 8.0 2023-04-29 14:27:37,143 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.566e+02 2.546e+02 2.889e+02 3.538e+02 7.743e+02, threshold=5.778e+02, percent-clipped=11.0 2023-04-29 14:28:22,427 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=115892.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 14:28:36,619 INFO [train.py:904] (4/8) Epoch 12, batch 4250, loss[loss=0.179, simple_loss=0.2777, pruned_loss=0.04013, over 16712.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2819, pruned_loss=0.05751, over 3194086.92 frames. ], batch size: 89, lr: 5.72e-03, grad_scale: 8.0 2023-04-29 14:28:56,086 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6628, 4.9441, 4.7145, 4.7393, 4.4504, 4.3843, 4.3962, 4.9872], device='cuda:4'), covar=tensor([0.0988, 0.0792, 0.0909, 0.0672, 0.0739, 0.0957, 0.0886, 0.0888], device='cuda:4'), in_proj_covar=tensor([0.0559, 0.0697, 0.0573, 0.0491, 0.0438, 0.0449, 0.0576, 0.0539], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:29:01,146 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4703, 2.9591, 2.9390, 1.9199, 2.6049, 2.1353, 2.9634, 3.1080], device='cuda:4'), covar=tensor([0.0299, 0.0688, 0.0608, 0.1671, 0.0821, 0.0912, 0.0692, 0.0708], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0149, 0.0158, 0.0144, 0.0136, 0.0124, 0.0138, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 14:29:49,145 INFO [train.py:904] (4/8) Epoch 12, batch 4300, loss[loss=0.2137, simple_loss=0.3031, pruned_loss=0.06208, over 16478.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2835, pruned_loss=0.05646, over 3204920.39 frames. ], batch size: 75, lr: 5.71e-03, grad_scale: 8.0 2023-04-29 14:29:50,864 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=115953.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 14:30:04,741 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.680e+02 2.520e+02 2.885e+02 3.320e+02 6.529e+02, threshold=5.769e+02, percent-clipped=1.0 2023-04-29 14:30:18,033 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5954, 3.8372, 2.8913, 2.2115, 2.7321, 2.3929, 4.0983, 3.4558], device='cuda:4'), covar=tensor([0.2725, 0.0661, 0.1554, 0.2231, 0.2395, 0.1754, 0.0432, 0.0953], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0260, 0.0286, 0.0282, 0.0287, 0.0225, 0.0271, 0.0301], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:30:52,374 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-29 14:31:07,411 INFO [train.py:904] (4/8) Epoch 12, batch 4350, loss[loss=0.2331, simple_loss=0.3133, pruned_loss=0.07642, over 16858.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2876, pruned_loss=0.05788, over 3192667.70 frames. ], batch size: 116, lr: 5.71e-03, grad_scale: 8.0 2023-04-29 14:31:27,396 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116014.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:31:46,122 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116026.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:31:47,883 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116027.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:32:22,077 INFO [train.py:904] (4/8) Epoch 12, batch 4400, loss[loss=0.1791, simple_loss=0.269, pruned_loss=0.04458, over 16854.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2904, pruned_loss=0.05987, over 3179441.77 frames. ], batch size: 42, lr: 5.71e-03, grad_scale: 8.0 2023-04-29 14:32:37,554 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.612e+02 2.512e+02 3.045e+02 3.645e+02 6.621e+02, threshold=6.090e+02, percent-clipped=4.0 2023-04-29 14:32:54,756 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=116074.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:32:57,359 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=116075.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:32:57,495 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116075.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 14:33:35,427 INFO [train.py:904] (4/8) Epoch 12, batch 4450, loss[loss=0.2171, simple_loss=0.3052, pruned_loss=0.06451, over 16545.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2936, pruned_loss=0.06101, over 3189890.43 frames. ], batch size: 68, lr: 5.71e-03, grad_scale: 8.0 2023-04-29 14:33:40,122 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116105.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:33:57,004 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116116.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:34:18,678 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7689, 2.0966, 1.6399, 1.9123, 2.4578, 2.1903, 2.5839, 2.7548], device='cuda:4'), covar=tensor([0.0106, 0.0325, 0.0477, 0.0378, 0.0198, 0.0301, 0.0171, 0.0176], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0206, 0.0201, 0.0200, 0.0206, 0.0206, 0.0212, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:34:49,326 INFO [train.py:904] (4/8) Epoch 12, batch 4500, loss[loss=0.194, simple_loss=0.2805, pruned_loss=0.05374, over 16668.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2936, pruned_loss=0.06133, over 3200240.86 frames. ], batch size: 57, lr: 5.71e-03, grad_scale: 8.0 2023-04-29 14:35:03,473 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.385e+02 2.115e+02 2.423e+02 2.863e+02 5.089e+02, threshold=4.846e+02, percent-clipped=0.0 2023-04-29 14:35:10,875 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116166.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:35:25,710 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116177.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:36:02,108 INFO [train.py:904] (4/8) Epoch 12, batch 4550, loss[loss=0.2, simple_loss=0.2947, pruned_loss=0.05264, over 16694.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2932, pruned_loss=0.06142, over 3207331.59 frames. ], batch size: 76, lr: 5.71e-03, grad_scale: 8.0 2023-04-29 14:37:04,519 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.4610, 2.4448, 2.4263, 3.2219, 2.6621, 3.5950, 1.2442, 2.9268], device='cuda:4'), covar=tensor([0.1624, 0.0825, 0.1243, 0.0137, 0.0321, 0.0352, 0.1945, 0.0744], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0160, 0.0181, 0.0151, 0.0200, 0.0209, 0.0182, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 14:37:08,829 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116248.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 14:37:14,079 INFO [train.py:904] (4/8) Epoch 12, batch 4600, loss[loss=0.2178, simple_loss=0.3, pruned_loss=0.06781, over 16864.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.294, pruned_loss=0.06158, over 3216308.77 frames. ], batch size: 116, lr: 5.71e-03, grad_scale: 8.0 2023-04-29 14:37:29,432 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.395e+02 1.985e+02 2.258e+02 2.658e+02 3.690e+02, threshold=4.517e+02, percent-clipped=0.0 2023-04-29 14:38:26,069 INFO [train.py:904] (4/8) Epoch 12, batch 4650, loss[loss=0.1632, simple_loss=0.2532, pruned_loss=0.03663, over 16317.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.2925, pruned_loss=0.06141, over 3217258.17 frames. ], batch size: 35, lr: 5.71e-03, grad_scale: 8.0 2023-04-29 14:38:55,829 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116323.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:39:01,240 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1878, 1.8924, 1.5794, 1.6040, 2.1701, 1.8333, 2.0113, 2.3211], device='cuda:4'), covar=tensor([0.0152, 0.0296, 0.0413, 0.0373, 0.0185, 0.0304, 0.0143, 0.0189], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0207, 0.0203, 0.0201, 0.0207, 0.0207, 0.0213, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:39:38,316 INFO [train.py:904] (4/8) Epoch 12, batch 4700, loss[loss=0.1628, simple_loss=0.2555, pruned_loss=0.03507, over 16828.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2891, pruned_loss=0.05973, over 3239626.73 frames. ], batch size: 102, lr: 5.70e-03, grad_scale: 8.0 2023-04-29 14:39:53,944 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.398e+02 2.029e+02 2.445e+02 3.086e+02 5.514e+02, threshold=4.890e+02, percent-clipped=4.0 2023-04-29 14:40:05,777 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116370.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 14:40:27,505 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116384.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:40:54,105 INFO [train.py:904] (4/8) Epoch 12, batch 4750, loss[loss=0.1665, simple_loss=0.2597, pruned_loss=0.03666, over 16833.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2849, pruned_loss=0.05746, over 3232936.24 frames. ], batch size: 102, lr: 5.70e-03, grad_scale: 8.0 2023-04-29 14:41:40,936 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5274, 2.3185, 2.3510, 4.3146, 2.1925, 2.7834, 2.3174, 2.5381], device='cuda:4'), covar=tensor([0.0992, 0.3109, 0.2278, 0.0356, 0.3594, 0.2035, 0.2960, 0.2841], device='cuda:4'), in_proj_covar=tensor([0.0371, 0.0400, 0.0333, 0.0321, 0.0418, 0.0463, 0.0364, 0.0466], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:41:58,960 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116446.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:42:04,473 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-29 14:42:07,122 INFO [train.py:904] (4/8) Epoch 12, batch 4800, loss[loss=0.1771, simple_loss=0.2685, pruned_loss=0.04286, over 16763.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2816, pruned_loss=0.0557, over 3223998.29 frames. ], batch size: 89, lr: 5.70e-03, grad_scale: 8.0 2023-04-29 14:42:17,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6391, 2.3162, 2.1862, 3.2713, 2.1495, 3.5630, 1.4446, 2.7299], device='cuda:4'), covar=tensor([0.1377, 0.0732, 0.1237, 0.0154, 0.0158, 0.0388, 0.1593, 0.0780], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0161, 0.0182, 0.0152, 0.0199, 0.0209, 0.0183, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 14:42:22,176 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116461.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:42:23,017 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.261e+02 1.880e+02 2.223e+02 2.728e+02 4.441e+02, threshold=4.446e+02, percent-clipped=0.0 2023-04-29 14:42:39,674 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116472.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:43:11,656 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6103, 2.6098, 1.7665, 2.7546, 2.1336, 2.7789, 2.0131, 2.4073], device='cuda:4'), covar=tensor([0.0253, 0.0333, 0.1221, 0.0143, 0.0712, 0.0447, 0.1171, 0.0567], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0164, 0.0188, 0.0132, 0.0165, 0.0206, 0.0195, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 14:43:23,406 INFO [train.py:904] (4/8) Epoch 12, batch 4850, loss[loss=0.1945, simple_loss=0.2833, pruned_loss=0.0529, over 16503.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.283, pruned_loss=0.05575, over 3183328.63 frames. ], batch size: 75, lr: 5.70e-03, grad_scale: 8.0 2023-04-29 14:43:28,304 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116505.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:43:30,815 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116507.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 14:43:59,241 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0056, 2.5580, 2.6333, 1.8282, 2.7330, 2.8185, 2.4704, 2.3498], device='cuda:4'), covar=tensor([0.0731, 0.0218, 0.0182, 0.0989, 0.0107, 0.0167, 0.0409, 0.0453], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0099, 0.0088, 0.0137, 0.0070, 0.0105, 0.0121, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 14:44:05,030 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-29 14:44:31,766 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116548.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 14:44:38,140 INFO [train.py:904] (4/8) Epoch 12, batch 4900, loss[loss=0.1853, simple_loss=0.2668, pruned_loss=0.05188, over 17106.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2819, pruned_loss=0.05454, over 3167825.68 frames. ], batch size: 49, lr: 5.70e-03, grad_scale: 8.0 2023-04-29 14:44:52,629 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.574e+02 2.192e+02 2.607e+02 3.082e+02 6.652e+02, threshold=5.215e+02, percent-clipped=3.0 2023-04-29 14:44:54,444 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8442, 4.8659, 4.6794, 4.4013, 4.3042, 4.7603, 4.6981, 4.4714], device='cuda:4'), covar=tensor([0.0545, 0.0386, 0.0263, 0.0251, 0.0966, 0.0391, 0.0272, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0318, 0.0289, 0.0269, 0.0309, 0.0309, 0.0197, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:44:58,520 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116566.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:45:39,973 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6488, 4.6617, 5.0324, 5.0084, 5.0155, 4.6969, 4.6534, 4.4378], device='cuda:4'), covar=tensor([0.0247, 0.0375, 0.0289, 0.0342, 0.0405, 0.0262, 0.0774, 0.0419], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0351, 0.0350, 0.0331, 0.0399, 0.0371, 0.0475, 0.0296], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 14:45:42,830 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=116596.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 14:45:52,019 INFO [train.py:904] (4/8) Epoch 12, batch 4950, loss[loss=0.2045, simple_loss=0.2951, pruned_loss=0.057, over 16903.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2815, pruned_loss=0.0538, over 3183500.48 frames. ], batch size: 90, lr: 5.70e-03, grad_scale: 8.0 2023-04-29 14:46:16,305 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116618.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:46:23,349 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9880, 4.0834, 3.8419, 3.6363, 3.5798, 3.9970, 3.7199, 3.7198], device='cuda:4'), covar=tensor([0.0569, 0.0448, 0.0303, 0.0260, 0.0866, 0.0399, 0.0964, 0.0565], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0320, 0.0290, 0.0270, 0.0311, 0.0311, 0.0198, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:47:04,274 INFO [train.py:904] (4/8) Epoch 12, batch 5000, loss[loss=0.1843, simple_loss=0.2669, pruned_loss=0.05086, over 17244.00 frames. ], tot_loss[loss=0.1956, simple_loss=0.2832, pruned_loss=0.05399, over 3204900.47 frames. ], batch size: 45, lr: 5.70e-03, grad_scale: 8.0 2023-04-29 14:47:09,850 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-29 14:47:17,030 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.587e+02 2.254e+02 2.645e+02 3.532e+02 7.072e+02, threshold=5.290e+02, percent-clipped=1.0 2023-04-29 14:47:30,344 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116670.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 14:47:43,240 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116679.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:47:43,365 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0068, 3.9500, 3.9600, 3.2403, 3.9235, 1.6602, 3.7341, 3.5010], device='cuda:4'), covar=tensor([0.0108, 0.0098, 0.0120, 0.0343, 0.0088, 0.2453, 0.0130, 0.0225], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0120, 0.0165, 0.0157, 0.0138, 0.0181, 0.0156, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:47:43,397 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116679.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:48:15,712 INFO [train.py:904] (4/8) Epoch 12, batch 5050, loss[loss=0.1813, simple_loss=0.2649, pruned_loss=0.04884, over 16613.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2841, pruned_loss=0.05401, over 3208176.59 frames. ], batch size: 62, lr: 5.70e-03, grad_scale: 8.0 2023-04-29 14:48:18,378 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116704.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:48:38,169 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=116718.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:49:22,696 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5639, 4.5317, 4.4777, 2.9948, 3.7280, 4.4364, 3.8588, 2.5247], device='cuda:4'), covar=tensor([0.0420, 0.0020, 0.0021, 0.0294, 0.0073, 0.0059, 0.0070, 0.0316], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0068, 0.0069, 0.0124, 0.0079, 0.0089, 0.0079, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 14:49:24,607 INFO [train.py:904] (4/8) Epoch 12, batch 5100, loss[loss=0.1731, simple_loss=0.2637, pruned_loss=0.04131, over 15415.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.2821, pruned_loss=0.05292, over 3209971.79 frames. ], batch size: 190, lr: 5.69e-03, grad_scale: 8.0 2023-04-29 14:49:27,620 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8600, 1.8674, 2.3924, 2.8948, 2.8171, 3.3257, 2.0054, 3.0813], device='cuda:4'), covar=tensor([0.0149, 0.0402, 0.0239, 0.0195, 0.0207, 0.0096, 0.0378, 0.0098], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0172, 0.0155, 0.0160, 0.0170, 0.0126, 0.0174, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 14:49:37,467 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116761.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:49:38,224 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.346e+02 2.103e+02 2.626e+02 3.146e+02 5.078e+02, threshold=5.251e+02, percent-clipped=1.0 2023-04-29 14:49:38,731 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8039, 1.3101, 1.6784, 1.7602, 1.8473, 1.9352, 1.5302, 1.8366], device='cuda:4'), covar=tensor([0.0191, 0.0314, 0.0156, 0.0207, 0.0187, 0.0121, 0.0314, 0.0093], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0172, 0.0155, 0.0160, 0.0170, 0.0126, 0.0174, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 14:49:43,702 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116765.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:49:52,532 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116772.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:50:20,531 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0150, 4.7819, 4.9970, 5.2800, 5.4052, 4.8426, 5.4149, 5.3569], device='cuda:4'), covar=tensor([0.1443, 0.1249, 0.1597, 0.0700, 0.0655, 0.0690, 0.0502, 0.0658], device='cuda:4'), in_proj_covar=tensor([0.0524, 0.0657, 0.0787, 0.0670, 0.0497, 0.0513, 0.0520, 0.0605], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:50:35,772 INFO [train.py:904] (4/8) Epoch 12, batch 5150, loss[loss=0.2016, simple_loss=0.2971, pruned_loss=0.05302, over 16899.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.2827, pruned_loss=0.05251, over 3204118.33 frames. ], batch size: 109, lr: 5.69e-03, grad_scale: 8.0 2023-04-29 14:50:36,730 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116802.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 14:50:36,798 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116802.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:50:47,799 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=116809.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:51:03,103 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=116820.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:51:15,149 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2221, 3.6252, 3.5821, 2.0758, 3.1384, 2.4250, 3.5742, 3.7664], device='cuda:4'), covar=tensor([0.0236, 0.0643, 0.0547, 0.1681, 0.0717, 0.0853, 0.0608, 0.0696], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0148, 0.0160, 0.0145, 0.0137, 0.0125, 0.0138, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 14:51:44,504 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-29 14:51:47,912 INFO [train.py:904] (4/8) Epoch 12, batch 5200, loss[loss=0.2066, simple_loss=0.2897, pruned_loss=0.06174, over 16381.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2815, pruned_loss=0.05223, over 3203256.35 frames. ], batch size: 146, lr: 5.69e-03, grad_scale: 8.0 2023-04-29 14:51:55,102 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-29 14:52:00,670 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116861.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:52:01,643 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.630e+02 2.168e+02 2.438e+02 3.019e+02 6.927e+02, threshold=4.876e+02, percent-clipped=1.0 2023-04-29 14:52:03,400 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116863.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:52:58,581 INFO [train.py:904] (4/8) Epoch 12, batch 5250, loss[loss=0.1917, simple_loss=0.2792, pruned_loss=0.05211, over 17238.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.279, pruned_loss=0.0516, over 3209801.21 frames. ], batch size: 45, lr: 5.69e-03, grad_scale: 16.0 2023-04-29 14:54:01,811 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 14:54:11,424 INFO [train.py:904] (4/8) Epoch 12, batch 5300, loss[loss=0.1609, simple_loss=0.2453, pruned_loss=0.03821, over 16848.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2754, pruned_loss=0.05042, over 3219416.52 frames. ], batch size: 116, lr: 5.69e-03, grad_scale: 8.0 2023-04-29 14:54:27,261 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.433e+02 2.208e+02 2.588e+02 3.045e+02 5.450e+02, threshold=5.175e+02, percent-clipped=1.0 2023-04-29 14:54:42,022 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116974.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:54:48,088 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8651, 1.3282, 1.6139, 1.7314, 1.8563, 1.9022, 1.5467, 1.7987], device='cuda:4'), covar=tensor([0.0176, 0.0296, 0.0147, 0.0208, 0.0185, 0.0133, 0.0291, 0.0087], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0172, 0.0154, 0.0160, 0.0169, 0.0125, 0.0172, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 14:54:49,790 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116979.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:55:11,182 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-29 14:55:21,984 INFO [train.py:904] (4/8) Epoch 12, batch 5350, loss[loss=0.1744, simple_loss=0.2641, pruned_loss=0.04232, over 17029.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2736, pruned_loss=0.04964, over 3213203.85 frames. ], batch size: 50, lr: 5.69e-03, grad_scale: 4.0 2023-04-29 14:55:44,045 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3369, 3.4204, 1.8687, 3.7776, 2.5004, 3.6801, 2.0689, 2.6817], device='cuda:4'), covar=tensor([0.0223, 0.0284, 0.1548, 0.0113, 0.0767, 0.0473, 0.1446, 0.0695], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0163, 0.0187, 0.0129, 0.0165, 0.0203, 0.0194, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-29 14:55:58,442 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=117027.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:56:31,809 INFO [train.py:904] (4/8) Epoch 12, batch 5400, loss[loss=0.2362, simple_loss=0.3147, pruned_loss=0.07885, over 12045.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2758, pruned_loss=0.05039, over 3199021.93 frames. ], batch size: 248, lr: 5.69e-03, grad_scale: 4.0 2023-04-29 14:56:43,919 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=117060.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:56:49,066 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.392e+02 2.253e+02 2.451e+02 2.837e+02 5.241e+02, threshold=4.902e+02, percent-clipped=1.0 2023-04-29 14:57:15,786 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.06 vs. limit=5.0 2023-04-29 14:57:32,178 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2721, 4.1043, 4.3092, 4.5018, 4.5876, 4.2055, 4.5678, 4.6284], device='cuda:4'), covar=tensor([0.1481, 0.1070, 0.1557, 0.0602, 0.0555, 0.1033, 0.0591, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0533, 0.0670, 0.0809, 0.0680, 0.0508, 0.0527, 0.0532, 0.0615], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 14:57:46,200 INFO [train.py:904] (4/8) Epoch 12, batch 5450, loss[loss=0.2006, simple_loss=0.2781, pruned_loss=0.0616, over 12105.00 frames. ], tot_loss[loss=0.1917, simple_loss=0.2793, pruned_loss=0.052, over 3207347.15 frames. ], batch size: 246, lr: 5.69e-03, grad_scale: 4.0 2023-04-29 14:57:46,808 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117102.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:58:55,229 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-29 14:58:57,077 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=117150.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:59:00,350 INFO [train.py:904] (4/8) Epoch 12, batch 5500, loss[loss=0.2492, simple_loss=0.3342, pruned_loss=0.08211, over 16790.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2866, pruned_loss=0.05646, over 3198640.16 frames. ], batch size: 39, lr: 5.69e-03, grad_scale: 4.0 2023-04-29 14:59:03,181 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0147, 3.5792, 3.4761, 1.8821, 3.0388, 2.2769, 3.3821, 3.7449], device='cuda:4'), covar=tensor([0.0277, 0.0576, 0.0540, 0.1851, 0.0696, 0.0903, 0.0650, 0.0760], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0148, 0.0160, 0.0144, 0.0138, 0.0125, 0.0139, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 14:59:09,336 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=117158.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:59:13,816 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117161.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 14:59:18,159 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.872e+02 2.773e+02 3.440e+02 4.409e+02 8.971e+02, threshold=6.880e+02, percent-clipped=17.0 2023-04-29 15:00:18,010 INFO [train.py:904] (4/8) Epoch 12, batch 5550, loss[loss=0.3068, simple_loss=0.358, pruned_loss=0.1279, over 11181.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2939, pruned_loss=0.06186, over 3174930.55 frames. ], batch size: 248, lr: 5.68e-03, grad_scale: 4.0 2023-04-29 15:00:30,332 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=117209.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:00:43,529 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-29 15:00:49,855 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=117221.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:01:39,144 INFO [train.py:904] (4/8) Epoch 12, batch 5600, loss[loss=0.3222, simple_loss=0.3644, pruned_loss=0.14, over 11040.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.2994, pruned_loss=0.06664, over 3129786.90 frames. ], batch size: 248, lr: 5.68e-03, grad_scale: 8.0 2023-04-29 15:01:58,901 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.404e+02 3.697e+02 4.299e+02 5.006e+02 8.998e+02, threshold=8.599e+02, percent-clipped=4.0 2023-04-29 15:02:16,764 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117274.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:02:30,383 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=117282.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:03:02,068 INFO [train.py:904] (4/8) Epoch 12, batch 5650, loss[loss=0.2188, simple_loss=0.3099, pruned_loss=0.06387, over 17007.00 frames. ], tot_loss[loss=0.2229, simple_loss=0.3044, pruned_loss=0.07071, over 3095756.42 frames. ], batch size: 41, lr: 5.68e-03, grad_scale: 4.0 2023-04-29 15:03:33,930 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=117322.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:03:44,981 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=117329.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:04:18,929 INFO [train.py:904] (4/8) Epoch 12, batch 5700, loss[loss=0.2248, simple_loss=0.3126, pruned_loss=0.06853, over 16362.00 frames. ], tot_loss[loss=0.2253, simple_loss=0.3059, pruned_loss=0.07236, over 3087182.54 frames. ], batch size: 146, lr: 5.68e-03, grad_scale: 2.0 2023-04-29 15:04:32,781 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117360.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:04:41,587 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.520e+02 3.684e+02 4.316e+02 5.371e+02 1.144e+03, threshold=8.631e+02, percent-clipped=1.0 2023-04-29 15:05:21,453 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=117390.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:05:33,423 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-29 15:05:39,185 INFO [train.py:904] (4/8) Epoch 12, batch 5750, loss[loss=0.2125, simple_loss=0.3043, pruned_loss=0.06035, over 16780.00 frames. ], tot_loss[loss=0.2279, simple_loss=0.3086, pruned_loss=0.07354, over 3083735.54 frames. ], batch size: 83, lr: 5.68e-03, grad_scale: 2.0 2023-04-29 15:05:48,655 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=117408.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:05:58,354 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8809, 2.1075, 2.4218, 3.1958, 2.1645, 2.4014, 2.2947, 2.2014], device='cuda:4'), covar=tensor([0.1066, 0.2948, 0.1653, 0.0519, 0.3344, 0.1909, 0.2518, 0.2694], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0395, 0.0331, 0.0316, 0.0412, 0.0454, 0.0360, 0.0461], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:07:00,135 INFO [train.py:904] (4/8) Epoch 12, batch 5800, loss[loss=0.2057, simple_loss=0.2956, pruned_loss=0.05792, over 16823.00 frames. ], tot_loss[loss=0.2275, simple_loss=0.3086, pruned_loss=0.07322, over 3061010.03 frames. ], batch size: 42, lr: 5.68e-03, grad_scale: 2.0 2023-04-29 15:07:09,875 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117458.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:07:21,329 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.077e+02 3.389e+02 3.943e+02 4.772e+02 8.236e+02, threshold=7.885e+02, percent-clipped=0.0 2023-04-29 15:07:49,194 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7816, 3.6908, 3.8402, 3.9915, 4.0340, 3.5774, 3.9845, 4.0533], device='cuda:4'), covar=tensor([0.1287, 0.0982, 0.1238, 0.0563, 0.0548, 0.2017, 0.0719, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0519, 0.0650, 0.0778, 0.0660, 0.0495, 0.0514, 0.0517, 0.0597], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:08:16,568 INFO [train.py:904] (4/8) Epoch 12, batch 5850, loss[loss=0.2009, simple_loss=0.2864, pruned_loss=0.05768, over 16540.00 frames. ], tot_loss[loss=0.2248, simple_loss=0.3063, pruned_loss=0.07161, over 3053512.66 frames. ], batch size: 62, lr: 5.68e-03, grad_scale: 2.0 2023-04-29 15:08:24,194 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=117506.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:09:37,155 INFO [train.py:904] (4/8) Epoch 12, batch 5900, loss[loss=0.2249, simple_loss=0.302, pruned_loss=0.07385, over 16742.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.3057, pruned_loss=0.0713, over 3058454.99 frames. ], batch size: 124, lr: 5.68e-03, grad_scale: 2.0 2023-04-29 15:10:01,560 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.797e+02 2.838e+02 3.606e+02 4.204e+02 7.799e+02, threshold=7.213e+02, percent-clipped=0.0 2023-04-29 15:10:02,224 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4173, 3.4738, 2.8192, 2.0234, 2.3521, 2.1651, 3.7172, 3.2436], device='cuda:4'), covar=tensor([0.2723, 0.0565, 0.1521, 0.2367, 0.2215, 0.1849, 0.0397, 0.1011], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0258, 0.0285, 0.0280, 0.0281, 0.0221, 0.0267, 0.0299], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:10:18,958 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=117577.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:10:49,776 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8056, 1.7069, 1.4502, 1.4699, 1.7682, 1.4325, 1.6151, 1.8838], device='cuda:4'), covar=tensor([0.0168, 0.0267, 0.0394, 0.0315, 0.0174, 0.0254, 0.0170, 0.0171], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0202, 0.0199, 0.0197, 0.0202, 0.0201, 0.0208, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:10:54,907 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8604, 4.9112, 5.3475, 5.2928, 5.3168, 4.9606, 4.9417, 4.5682], device='cuda:4'), covar=tensor([0.0282, 0.0473, 0.0314, 0.0419, 0.0410, 0.0360, 0.0885, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0355, 0.0357, 0.0335, 0.0406, 0.0378, 0.0479, 0.0303], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 15:10:55,831 INFO [train.py:904] (4/8) Epoch 12, batch 5950, loss[loss=0.2122, simple_loss=0.2921, pruned_loss=0.06617, over 17014.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.3061, pruned_loss=0.06952, over 3077826.58 frames. ], batch size: 55, lr: 5.67e-03, grad_scale: 2.0 2023-04-29 15:12:14,105 INFO [train.py:904] (4/8) Epoch 12, batch 6000, loss[loss=0.1912, simple_loss=0.2726, pruned_loss=0.05488, over 17202.00 frames. ], tot_loss[loss=0.2217, simple_loss=0.3052, pruned_loss=0.06912, over 3097333.95 frames. ], batch size: 44, lr: 5.67e-03, grad_scale: 4.0 2023-04-29 15:12:14,105 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 15:12:25,320 INFO [train.py:938] (4/8) Epoch 12, validation: loss=0.161, simple_loss=0.2739, pruned_loss=0.02405, over 944034.00 frames. 2023-04-29 15:12:25,320 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 15:12:46,504 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.036e+02 2.944e+02 3.518e+02 4.227e+02 7.444e+02, threshold=7.035e+02, percent-clipped=1.0 2023-04-29 15:13:18,108 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=117685.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:13:46,105 INFO [train.py:904] (4/8) Epoch 12, batch 6050, loss[loss=0.2236, simple_loss=0.3086, pruned_loss=0.06925, over 16303.00 frames. ], tot_loss[loss=0.2212, simple_loss=0.3044, pruned_loss=0.06905, over 3083159.88 frames. ], batch size: 146, lr: 5.67e-03, grad_scale: 4.0 2023-04-29 15:14:54,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8007, 2.5152, 2.2418, 3.9263, 2.5807, 3.8949, 1.5005, 2.6345], device='cuda:4'), covar=tensor([0.1281, 0.0786, 0.1346, 0.0178, 0.0252, 0.0411, 0.1573, 0.0938], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0163, 0.0182, 0.0151, 0.0199, 0.0208, 0.0183, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 15:15:02,175 INFO [train.py:904] (4/8) Epoch 12, batch 6100, loss[loss=0.2428, simple_loss=0.3026, pruned_loss=0.09146, over 11425.00 frames. ], tot_loss[loss=0.2205, simple_loss=0.304, pruned_loss=0.0685, over 3092389.21 frames. ], batch size: 247, lr: 5.67e-03, grad_scale: 4.0 2023-04-29 15:15:13,774 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4836, 3.5519, 3.3143, 3.0950, 3.1685, 3.4364, 3.2528, 3.2272], device='cuda:4'), covar=tensor([0.0570, 0.0503, 0.0257, 0.0262, 0.0522, 0.0397, 0.1274, 0.0463], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0323, 0.0290, 0.0271, 0.0312, 0.0312, 0.0200, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:15:24,819 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.907e+02 2.974e+02 3.801e+02 4.781e+02 1.516e+03, threshold=7.602e+02, percent-clipped=11.0 2023-04-29 15:16:08,282 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.41 vs. limit=5.0 2023-04-29 15:16:12,624 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4903, 4.6021, 4.7337, 4.5768, 4.6185, 5.1328, 4.7017, 4.5087], device='cuda:4'), covar=tensor([0.1220, 0.1679, 0.1783, 0.1950, 0.2492, 0.1101, 0.1486, 0.2322], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0490, 0.0540, 0.0422, 0.0579, 0.0564, 0.0429, 0.0581], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 15:16:19,682 INFO [train.py:904] (4/8) Epoch 12, batch 6150, loss[loss=0.2474, simple_loss=0.3094, pruned_loss=0.09271, over 11203.00 frames. ], tot_loss[loss=0.219, simple_loss=0.302, pruned_loss=0.068, over 3096401.48 frames. ], batch size: 246, lr: 5.67e-03, grad_scale: 4.0 2023-04-29 15:17:38,949 INFO [train.py:904] (4/8) Epoch 12, batch 6200, loss[loss=0.2105, simple_loss=0.296, pruned_loss=0.06252, over 16744.00 frames. ], tot_loss[loss=0.2176, simple_loss=0.3005, pruned_loss=0.06735, over 3103451.43 frames. ], batch size: 134, lr: 5.67e-03, grad_scale: 4.0 2023-04-29 15:18:00,667 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.030e+02 3.055e+02 3.626e+02 4.280e+02 7.203e+02, threshold=7.253e+02, percent-clipped=0.0 2023-04-29 15:18:18,041 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117877.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:18:25,717 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3177, 5.7196, 5.3366, 5.4191, 4.9958, 5.0084, 5.1175, 5.7745], device='cuda:4'), covar=tensor([0.1055, 0.0706, 0.1082, 0.0762, 0.0857, 0.0717, 0.1068, 0.0818], device='cuda:4'), in_proj_covar=tensor([0.0554, 0.0689, 0.0568, 0.0486, 0.0438, 0.0449, 0.0571, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:18:52,875 INFO [train.py:904] (4/8) Epoch 12, batch 6250, loss[loss=0.2152, simple_loss=0.3089, pruned_loss=0.06076, over 16843.00 frames. ], tot_loss[loss=0.2173, simple_loss=0.3004, pruned_loss=0.06713, over 3114801.24 frames. ], batch size: 116, lr: 5.67e-03, grad_scale: 4.0 2023-04-29 15:19:27,990 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=117925.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:19:46,782 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1415, 4.2434, 4.3976, 4.2469, 4.2953, 4.7867, 4.3485, 4.1539], device='cuda:4'), covar=tensor([0.1713, 0.1841, 0.1955, 0.1823, 0.2362, 0.1096, 0.1552, 0.2384], device='cuda:4'), in_proj_covar=tensor([0.0353, 0.0491, 0.0541, 0.0423, 0.0576, 0.0562, 0.0427, 0.0578], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 15:20:06,464 INFO [train.py:904] (4/8) Epoch 12, batch 6300, loss[loss=0.2094, simple_loss=0.2935, pruned_loss=0.06259, over 15240.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.2999, pruned_loss=0.06632, over 3111702.98 frames. ], batch size: 190, lr: 5.67e-03, grad_scale: 4.0 2023-04-29 15:20:22,804 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7827, 3.7501, 4.3720, 2.0541, 4.6004, 4.5438, 3.2158, 3.2148], device='cuda:4'), covar=tensor([0.0671, 0.0225, 0.0137, 0.1111, 0.0037, 0.0108, 0.0327, 0.0419], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0100, 0.0088, 0.0138, 0.0070, 0.0105, 0.0121, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 15:20:28,834 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.927e+02 3.047e+02 3.596e+02 4.345e+02 1.152e+03, threshold=7.193e+02, percent-clipped=4.0 2023-04-29 15:20:57,586 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117985.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:21:25,210 INFO [train.py:904] (4/8) Epoch 12, batch 6350, loss[loss=0.2155, simple_loss=0.302, pruned_loss=0.06455, over 16863.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.301, pruned_loss=0.06799, over 3091920.81 frames. ], batch size: 102, lr: 5.66e-03, grad_scale: 4.0 2023-04-29 15:22:04,031 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6252, 1.5164, 2.1971, 2.5092, 2.5467, 2.8075, 1.5114, 2.7044], device='cuda:4'), covar=tensor([0.0139, 0.0436, 0.0205, 0.0188, 0.0188, 0.0122, 0.0527, 0.0072], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0172, 0.0155, 0.0160, 0.0170, 0.0127, 0.0173, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 15:22:11,807 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=118033.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:22:21,858 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118040.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:22:37,545 INFO [train.py:904] (4/8) Epoch 12, batch 6400, loss[loss=0.1952, simple_loss=0.2819, pruned_loss=0.05426, over 16894.00 frames. ], tot_loss[loss=0.2203, simple_loss=0.3022, pruned_loss=0.06922, over 3091030.74 frames. ], batch size: 116, lr: 5.66e-03, grad_scale: 8.0 2023-04-29 15:22:57,664 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.235e+02 3.125e+02 4.036e+02 4.702e+02 8.359e+02, threshold=8.072e+02, percent-clipped=7.0 2023-04-29 15:23:32,923 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2999, 3.2336, 3.2825, 3.4668, 3.4600, 3.2363, 3.4137, 3.5106], device='cuda:4'), covar=tensor([0.1160, 0.1046, 0.1317, 0.0727, 0.0850, 0.2788, 0.1300, 0.0820], device='cuda:4'), in_proj_covar=tensor([0.0527, 0.0657, 0.0789, 0.0669, 0.0503, 0.0518, 0.0529, 0.0605], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:23:50,428 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118101.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:23:50,964 INFO [train.py:904] (4/8) Epoch 12, batch 6450, loss[loss=0.2382, simple_loss=0.2931, pruned_loss=0.09169, over 11814.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.3008, pruned_loss=0.06766, over 3107426.65 frames. ], batch size: 248, lr: 5.66e-03, grad_scale: 8.0 2023-04-29 15:23:57,548 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4745, 3.5645, 3.3084, 3.0488, 3.1541, 3.4068, 3.2722, 3.2293], device='cuda:4'), covar=tensor([0.0570, 0.0492, 0.0270, 0.0268, 0.0506, 0.0408, 0.1297, 0.0523], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0322, 0.0288, 0.0268, 0.0307, 0.0309, 0.0198, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:24:33,650 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-29 15:25:08,003 INFO [train.py:904] (4/8) Epoch 12, batch 6500, loss[loss=0.2, simple_loss=0.2889, pruned_loss=0.05554, over 16152.00 frames. ], tot_loss[loss=0.2169, simple_loss=0.299, pruned_loss=0.0674, over 3111140.94 frames. ], batch size: 165, lr: 5.66e-03, grad_scale: 8.0 2023-04-29 15:25:29,347 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.059e+02 2.874e+02 3.539e+02 4.309e+02 7.479e+02, threshold=7.078e+02, percent-clipped=0.0 2023-04-29 15:25:32,722 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118168.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:26:28,541 INFO [train.py:904] (4/8) Epoch 12, batch 6550, loss[loss=0.2289, simple_loss=0.3321, pruned_loss=0.06287, over 16699.00 frames. ], tot_loss[loss=0.2183, simple_loss=0.3014, pruned_loss=0.06758, over 3112889.36 frames. ], batch size: 134, lr: 5.66e-03, grad_scale: 8.0 2023-04-29 15:27:10,818 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118229.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 15:27:42,921 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-29 15:27:44,389 INFO [train.py:904] (4/8) Epoch 12, batch 6600, loss[loss=0.215, simple_loss=0.3074, pruned_loss=0.0613, over 16416.00 frames. ], tot_loss[loss=0.2204, simple_loss=0.3037, pruned_loss=0.06853, over 3096210.22 frames. ], batch size: 68, lr: 5.66e-03, grad_scale: 8.0 2023-04-29 15:28:05,475 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.686e+02 3.186e+02 3.899e+02 4.864e+02 8.341e+02, threshold=7.798e+02, percent-clipped=4.0 2023-04-29 15:28:10,226 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118269.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:28:30,024 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0536, 3.9655, 3.8753, 2.4914, 3.5713, 3.9046, 3.5252, 2.2201], device='cuda:4'), covar=tensor([0.0516, 0.0027, 0.0039, 0.0356, 0.0066, 0.0091, 0.0075, 0.0373], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0069, 0.0072, 0.0126, 0.0081, 0.0092, 0.0081, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 15:28:41,366 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118289.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:29:00,306 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1743, 3.3308, 3.5559, 3.5458, 3.5498, 3.3541, 3.3844, 3.4311], device='cuda:4'), covar=tensor([0.0374, 0.0664, 0.0434, 0.0452, 0.0526, 0.0520, 0.0830, 0.0504], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0355, 0.0357, 0.0340, 0.0406, 0.0379, 0.0480, 0.0304], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 15:29:01,033 INFO [train.py:904] (4/8) Epoch 12, batch 6650, loss[loss=0.2584, simple_loss=0.3168, pruned_loss=0.1, over 11524.00 frames. ], tot_loss[loss=0.221, simple_loss=0.3036, pruned_loss=0.06925, over 3090806.95 frames. ], batch size: 248, lr: 5.66e-03, grad_scale: 8.0 2023-04-29 15:29:43,708 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118329.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:29:45,508 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118330.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:30:16,386 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118350.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:30:18,842 INFO [train.py:904] (4/8) Epoch 12, batch 6700, loss[loss=0.2378, simple_loss=0.3032, pruned_loss=0.08624, over 11695.00 frames. ], tot_loss[loss=0.2204, simple_loss=0.3021, pruned_loss=0.06939, over 3083063.69 frames. ], batch size: 248, lr: 5.66e-03, grad_scale: 8.0 2023-04-29 15:30:39,918 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.931e+02 3.120e+02 3.560e+02 4.149e+02 6.786e+02, threshold=7.119e+02, percent-clipped=0.0 2023-04-29 15:31:17,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118390.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:31:27,185 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118396.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:31:35,210 INFO [train.py:904] (4/8) Epoch 12, batch 6750, loss[loss=0.1951, simple_loss=0.2846, pruned_loss=0.05278, over 16791.00 frames. ], tot_loss[loss=0.2188, simple_loss=0.3006, pruned_loss=0.06853, over 3105686.80 frames. ], batch size: 102, lr: 5.66e-03, grad_scale: 8.0 2023-04-29 15:32:13,623 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0029, 5.4187, 5.6153, 5.4019, 5.3751, 5.9530, 5.4609, 5.2832], device='cuda:4'), covar=tensor([0.0831, 0.1462, 0.1675, 0.1599, 0.2131, 0.0831, 0.1367, 0.2148], device='cuda:4'), in_proj_covar=tensor([0.0359, 0.0495, 0.0546, 0.0430, 0.0582, 0.0570, 0.0431, 0.0584], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 15:32:49,923 INFO [train.py:904] (4/8) Epoch 12, batch 6800, loss[loss=0.246, simple_loss=0.3162, pruned_loss=0.08789, over 11755.00 frames. ], tot_loss[loss=0.2192, simple_loss=0.3007, pruned_loss=0.06887, over 3084313.98 frames. ], batch size: 248, lr: 5.65e-03, grad_scale: 8.0 2023-04-29 15:33:11,647 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.917e+02 3.071e+02 3.777e+02 4.759e+02 7.416e+02, threshold=7.554e+02, percent-clipped=1.0 2023-04-29 15:34:04,798 INFO [train.py:904] (4/8) Epoch 12, batch 6850, loss[loss=0.2228, simple_loss=0.3317, pruned_loss=0.05695, over 16589.00 frames. ], tot_loss[loss=0.2197, simple_loss=0.3019, pruned_loss=0.06874, over 3106938.26 frames. ], batch size: 62, lr: 5.65e-03, grad_scale: 8.0 2023-04-29 15:34:36,243 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118524.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 15:34:36,801 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-29 15:35:15,879 INFO [train.py:904] (4/8) Epoch 12, batch 6900, loss[loss=0.2371, simple_loss=0.3172, pruned_loss=0.07847, over 16405.00 frames. ], tot_loss[loss=0.2208, simple_loss=0.304, pruned_loss=0.06879, over 3096143.88 frames. ], batch size: 146, lr: 5.65e-03, grad_scale: 8.0 2023-04-29 15:35:36,844 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.921e+02 3.007e+02 3.634e+02 4.757e+02 1.105e+03, threshold=7.268e+02, percent-clipped=1.0 2023-04-29 15:36:30,544 INFO [train.py:904] (4/8) Epoch 12, batch 6950, loss[loss=0.2856, simple_loss=0.3337, pruned_loss=0.1188, over 11028.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.3051, pruned_loss=0.06994, over 3108510.16 frames. ], batch size: 248, lr: 5.65e-03, grad_scale: 8.0 2023-04-29 15:37:04,588 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118625.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:37:14,682 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118631.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:37:35,444 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118645.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:37:44,800 INFO [train.py:904] (4/8) Epoch 12, batch 7000, loss[loss=0.221, simple_loss=0.3119, pruned_loss=0.06502, over 16325.00 frames. ], tot_loss[loss=0.2218, simple_loss=0.3054, pruned_loss=0.06908, over 3106189.10 frames. ], batch size: 165, lr: 5.65e-03, grad_scale: 8.0 2023-04-29 15:38:05,443 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.978e+02 2.982e+02 3.795e+02 4.664e+02 8.323e+02, threshold=7.591e+02, percent-clipped=2.0 2023-04-29 15:38:33,776 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118685.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:38:44,764 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118692.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:38:50,366 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118696.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:38:59,534 INFO [train.py:904] (4/8) Epoch 12, batch 7050, loss[loss=0.2353, simple_loss=0.3155, pruned_loss=0.07754, over 15449.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.3062, pruned_loss=0.0688, over 3111568.92 frames. ], batch size: 191, lr: 5.65e-03, grad_scale: 8.0 2023-04-29 15:39:08,728 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-29 15:39:10,252 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.88 vs. limit=5.0 2023-04-29 15:40:01,768 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=118744.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:40:14,608 INFO [train.py:904] (4/8) Epoch 12, batch 7100, loss[loss=0.2101, simple_loss=0.3037, pruned_loss=0.05828, over 16783.00 frames. ], tot_loss[loss=0.2212, simple_loss=0.3049, pruned_loss=0.06873, over 3109797.82 frames. ], batch size: 89, lr: 5.65e-03, grad_scale: 8.0 2023-04-29 15:40:36,860 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.757e+02 3.120e+02 3.747e+02 4.649e+02 1.761e+03, threshold=7.495e+02, percent-clipped=5.0 2023-04-29 15:41:29,294 INFO [train.py:904] (4/8) Epoch 12, batch 7150, loss[loss=0.21, simple_loss=0.2972, pruned_loss=0.06139, over 16460.00 frames. ], tot_loss[loss=0.22, simple_loss=0.3028, pruned_loss=0.06863, over 3087989.09 frames. ], batch size: 75, lr: 5.65e-03, grad_scale: 8.0 2023-04-29 15:41:32,291 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6818, 2.2975, 2.2096, 3.2025, 2.3354, 3.5756, 1.3499, 2.7024], device='cuda:4'), covar=tensor([0.1309, 0.0747, 0.1216, 0.0167, 0.0183, 0.0347, 0.1612, 0.0735], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0162, 0.0183, 0.0152, 0.0200, 0.0209, 0.0184, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 15:42:01,923 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118824.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:42:20,408 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-29 15:42:41,519 INFO [train.py:904] (4/8) Epoch 12, batch 7200, loss[loss=0.2057, simple_loss=0.299, pruned_loss=0.05623, over 16336.00 frames. ], tot_loss[loss=0.2168, simple_loss=0.3006, pruned_loss=0.06655, over 3086840.49 frames. ], batch size: 165, lr: 5.64e-03, grad_scale: 8.0 2023-04-29 15:43:03,912 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.822e+02 2.566e+02 3.185e+02 3.996e+02 6.189e+02, threshold=6.369e+02, percent-clipped=0.0 2023-04-29 15:43:12,503 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=118872.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:43:38,286 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 15:44:00,069 INFO [train.py:904] (4/8) Epoch 12, batch 7250, loss[loss=0.1872, simple_loss=0.2701, pruned_loss=0.0522, over 16906.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.2982, pruned_loss=0.06539, over 3086880.54 frames. ], batch size: 109, lr: 5.64e-03, grad_scale: 4.0 2023-04-29 15:44:35,490 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118925.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:45:04,493 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118945.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:45:15,111 INFO [train.py:904] (4/8) Epoch 12, batch 7300, loss[loss=0.247, simple_loss=0.3148, pruned_loss=0.08964, over 11431.00 frames. ], tot_loss[loss=0.2148, simple_loss=0.298, pruned_loss=0.06579, over 3074495.88 frames. ], batch size: 247, lr: 5.64e-03, grad_scale: 4.0 2023-04-29 15:45:26,546 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-29 15:45:31,096 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4640, 3.4528, 3.3804, 2.7704, 3.2865, 2.1460, 3.1347, 2.7761], device='cuda:4'), covar=tensor([0.0106, 0.0083, 0.0151, 0.0187, 0.0079, 0.1838, 0.0103, 0.0161], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0119, 0.0164, 0.0156, 0.0136, 0.0179, 0.0152, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:45:33,125 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0461, 3.3837, 3.2752, 2.1482, 3.0617, 3.3448, 3.1217, 1.7101], device='cuda:4'), covar=tensor([0.0451, 0.0034, 0.0044, 0.0359, 0.0089, 0.0079, 0.0080, 0.0423], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0068, 0.0070, 0.0124, 0.0079, 0.0090, 0.0079, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 15:45:36,388 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.980e+02 3.034e+02 3.599e+02 4.380e+02 7.583e+02, threshold=7.199e+02, percent-clipped=5.0 2023-04-29 15:45:45,784 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=118973.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:46:03,496 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118985.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:46:06,453 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118987.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:46:15,623 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=118993.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:46:28,512 INFO [train.py:904] (4/8) Epoch 12, batch 7350, loss[loss=0.1965, simple_loss=0.2908, pruned_loss=0.05106, over 16847.00 frames. ], tot_loss[loss=0.2169, simple_loss=0.2994, pruned_loss=0.06715, over 3040328.07 frames. ], batch size: 96, lr: 5.64e-03, grad_scale: 4.0 2023-04-29 15:46:57,031 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-29 15:47:14,407 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=119033.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:47:42,935 INFO [train.py:904] (4/8) Epoch 12, batch 7400, loss[loss=0.2326, simple_loss=0.3047, pruned_loss=0.08029, over 15387.00 frames. ], tot_loss[loss=0.2188, simple_loss=0.3013, pruned_loss=0.06819, over 3042871.35 frames. ], batch size: 190, lr: 5.64e-03, grad_scale: 4.0 2023-04-29 15:47:51,780 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-29 15:48:06,306 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.884e+02 3.304e+02 4.015e+02 4.796e+02 1.422e+03, threshold=8.030e+02, percent-clipped=7.0 2023-04-29 15:48:19,157 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=119075.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:48:27,642 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5625, 4.2744, 4.2117, 3.1392, 3.7393, 4.1968, 3.8687, 2.1654], device='cuda:4'), covar=tensor([0.0392, 0.0026, 0.0034, 0.0247, 0.0075, 0.0090, 0.0061, 0.0387], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0068, 0.0071, 0.0125, 0.0080, 0.0091, 0.0080, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 15:48:37,941 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-04-29 15:48:47,031 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=119093.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:49:01,208 INFO [train.py:904] (4/8) Epoch 12, batch 7450, loss[loss=0.2321, simple_loss=0.3135, pruned_loss=0.07533, over 16889.00 frames. ], tot_loss[loss=0.2194, simple_loss=0.3018, pruned_loss=0.06851, over 3070773.04 frames. ], batch size: 116, lr: 5.64e-03, grad_scale: 4.0 2023-04-29 15:49:42,622 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-29 15:49:55,907 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0296, 2.9571, 3.1276, 1.7203, 3.2936, 3.3525, 2.6472, 2.5657], device='cuda:4'), covar=tensor([0.0845, 0.0211, 0.0206, 0.1118, 0.0084, 0.0195, 0.0413, 0.0453], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0100, 0.0088, 0.0137, 0.0069, 0.0105, 0.0121, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 15:49:55,932 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=119136.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:50:20,479 INFO [train.py:904] (4/8) Epoch 12, batch 7500, loss[loss=0.1731, simple_loss=0.2567, pruned_loss=0.04479, over 17123.00 frames. ], tot_loss[loss=0.219, simple_loss=0.3024, pruned_loss=0.06777, over 3079160.14 frames. ], batch size: 47, lr: 5.64e-03, grad_scale: 4.0 2023-04-29 15:50:24,117 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=119154.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 15:50:42,266 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.862e+02 2.952e+02 3.435e+02 4.438e+02 7.679e+02, threshold=6.870e+02, percent-clipped=0.0 2023-04-29 15:51:08,996 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-29 15:51:30,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8013, 1.8166, 1.6345, 1.6026, 1.8894, 1.5747, 1.6785, 1.9142], device='cuda:4'), covar=tensor([0.0113, 0.0191, 0.0277, 0.0260, 0.0138, 0.0198, 0.0139, 0.0130], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0202, 0.0197, 0.0197, 0.0202, 0.0200, 0.0204, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:51:35,616 INFO [train.py:904] (4/8) Epoch 12, batch 7550, loss[loss=0.2097, simple_loss=0.2888, pruned_loss=0.06528, over 16703.00 frames. ], tot_loss[loss=0.2204, simple_loss=0.3023, pruned_loss=0.06922, over 3046144.30 frames. ], batch size: 134, lr: 5.64e-03, grad_scale: 4.0 2023-04-29 15:51:42,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1053, 4.8578, 5.1530, 5.3278, 5.4911, 4.8219, 5.4857, 5.4554], device='cuda:4'), covar=tensor([0.1627, 0.1171, 0.1517, 0.0574, 0.0477, 0.0834, 0.0475, 0.0476], device='cuda:4'), in_proj_covar=tensor([0.0521, 0.0653, 0.0786, 0.0665, 0.0503, 0.0517, 0.0534, 0.0604], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:52:03,736 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-29 15:52:50,119 INFO [train.py:904] (4/8) Epoch 12, batch 7600, loss[loss=0.2651, simple_loss=0.3174, pruned_loss=0.1064, over 11303.00 frames. ], tot_loss[loss=0.2204, simple_loss=0.3016, pruned_loss=0.06959, over 3033007.82 frames. ], batch size: 248, lr: 5.64e-03, grad_scale: 8.0 2023-04-29 15:53:12,404 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.101e+02 2.984e+02 3.563e+02 4.291e+02 1.017e+03, threshold=7.126e+02, percent-clipped=2.0 2023-04-29 15:53:34,500 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5195, 3.5205, 3.4516, 2.8046, 3.3351, 2.1000, 3.1165, 2.8457], device='cuda:4'), covar=tensor([0.0120, 0.0103, 0.0141, 0.0218, 0.0084, 0.2009, 0.0108, 0.0191], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0117, 0.0162, 0.0154, 0.0135, 0.0177, 0.0150, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:53:43,852 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=119287.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:54:05,863 INFO [train.py:904] (4/8) Epoch 12, batch 7650, loss[loss=0.2402, simple_loss=0.3225, pruned_loss=0.07896, over 15391.00 frames. ], tot_loss[loss=0.2209, simple_loss=0.3016, pruned_loss=0.07016, over 3022831.26 frames. ], batch size: 191, lr: 5.63e-03, grad_scale: 8.0 2023-04-29 15:54:54,778 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0130, 1.9922, 2.2404, 3.4975, 1.9577, 2.3706, 2.1396, 2.1315], device='cuda:4'), covar=tensor([0.1041, 0.3165, 0.2185, 0.0497, 0.3787, 0.2165, 0.2969, 0.2941], device='cuda:4'), in_proj_covar=tensor([0.0363, 0.0394, 0.0331, 0.0316, 0.0413, 0.0452, 0.0359, 0.0458], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:54:55,610 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=119335.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:55:20,164 INFO [train.py:904] (4/8) Epoch 12, batch 7700, loss[loss=0.2075, simple_loss=0.2953, pruned_loss=0.05984, over 16770.00 frames. ], tot_loss[loss=0.2216, simple_loss=0.3019, pruned_loss=0.07062, over 3032356.15 frames. ], batch size: 124, lr: 5.63e-03, grad_scale: 8.0 2023-04-29 15:55:42,614 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.234e+02 3.396e+02 4.418e+02 5.495e+02 1.012e+03, threshold=8.835e+02, percent-clipped=5.0 2023-04-29 15:56:36,113 INFO [train.py:904] (4/8) Epoch 12, batch 7750, loss[loss=0.2093, simple_loss=0.3006, pruned_loss=0.05903, over 16701.00 frames. ], tot_loss[loss=0.2206, simple_loss=0.3021, pruned_loss=0.06962, over 3058744.77 frames. ], batch size: 62, lr: 5.63e-03, grad_scale: 8.0 2023-04-29 15:57:18,795 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=119431.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 15:57:31,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6554, 4.5237, 4.7314, 4.8942, 5.0391, 4.5481, 5.0643, 5.0149], device='cuda:4'), covar=tensor([0.1738, 0.1118, 0.1539, 0.0600, 0.0562, 0.0824, 0.0567, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0517, 0.0650, 0.0780, 0.0664, 0.0501, 0.0510, 0.0531, 0.0600], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:57:39,516 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1574, 3.1482, 3.3114, 1.6691, 3.5017, 3.5437, 2.7251, 2.5886], device='cuda:4'), covar=tensor([0.0768, 0.0200, 0.0178, 0.1102, 0.0063, 0.0137, 0.0388, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0100, 0.0086, 0.0136, 0.0068, 0.0104, 0.0119, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 15:57:45,518 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=119449.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 15:57:47,931 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8190, 5.0651, 5.2974, 4.9967, 5.1384, 5.6769, 5.1637, 4.9022], device='cuda:4'), covar=tensor([0.0923, 0.1731, 0.2126, 0.1931, 0.2128, 0.0832, 0.1425, 0.2295], device='cuda:4'), in_proj_covar=tensor([0.0360, 0.0502, 0.0549, 0.0434, 0.0581, 0.0571, 0.0433, 0.0589], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 15:57:48,704 INFO [train.py:904] (4/8) Epoch 12, batch 7800, loss[loss=0.1901, simple_loss=0.2791, pruned_loss=0.05057, over 16794.00 frames. ], tot_loss[loss=0.2206, simple_loss=0.3023, pruned_loss=0.06942, over 3086121.87 frames. ], batch size: 83, lr: 5.63e-03, grad_scale: 8.0 2023-04-29 15:58:11,196 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.951e+02 3.141e+02 3.879e+02 4.510e+02 7.221e+02, threshold=7.757e+02, percent-clipped=0.0 2023-04-29 15:58:42,984 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1898, 3.3258, 3.5693, 3.5598, 3.5501, 3.3531, 3.3814, 3.4071], device='cuda:4'), covar=tensor([0.0404, 0.0702, 0.0442, 0.0439, 0.0522, 0.0521, 0.0825, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0360, 0.0357, 0.0339, 0.0408, 0.0381, 0.0477, 0.0305], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 15:58:51,587 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4344, 2.9398, 2.7023, 2.2071, 2.2606, 2.1793, 2.9205, 2.8503], device='cuda:4'), covar=tensor([0.2250, 0.0653, 0.1377, 0.2094, 0.2121, 0.1932, 0.0448, 0.1069], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0256, 0.0286, 0.0281, 0.0279, 0.0222, 0.0267, 0.0296], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 15:59:04,882 INFO [train.py:904] (4/8) Epoch 12, batch 7850, loss[loss=0.2069, simple_loss=0.3005, pruned_loss=0.05665, over 16870.00 frames. ], tot_loss[loss=0.2208, simple_loss=0.3033, pruned_loss=0.06915, over 3095627.11 frames. ], batch size: 96, lr: 5.63e-03, grad_scale: 4.0 2023-04-29 15:59:24,398 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=119514.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:00:05,605 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8751, 1.2856, 1.6897, 1.7646, 1.8675, 1.9559, 1.4497, 1.8750], device='cuda:4'), covar=tensor([0.0155, 0.0304, 0.0146, 0.0177, 0.0169, 0.0120, 0.0340, 0.0082], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0170, 0.0155, 0.0158, 0.0170, 0.0124, 0.0171, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 16:00:19,055 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5623, 3.6620, 4.0686, 1.7643, 4.2689, 4.2424, 2.9569, 3.0945], device='cuda:4'), covar=tensor([0.0740, 0.0187, 0.0164, 0.1183, 0.0044, 0.0112, 0.0390, 0.0395], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0100, 0.0087, 0.0137, 0.0069, 0.0105, 0.0120, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 16:00:21,535 INFO [train.py:904] (4/8) Epoch 12, batch 7900, loss[loss=0.2128, simple_loss=0.2953, pruned_loss=0.06516, over 16655.00 frames. ], tot_loss[loss=0.219, simple_loss=0.3019, pruned_loss=0.06803, over 3107414.08 frames. ], batch size: 62, lr: 5.63e-03, grad_scale: 4.0 2023-04-29 16:00:22,002 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1413, 5.1437, 4.9619, 4.7067, 4.6215, 4.9641, 4.9062, 4.6404], device='cuda:4'), covar=tensor([0.0516, 0.0379, 0.0228, 0.0233, 0.0821, 0.0453, 0.0293, 0.0644], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0324, 0.0286, 0.0264, 0.0304, 0.0309, 0.0198, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:00:45,733 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.963e+02 2.838e+02 3.537e+02 4.302e+02 9.445e+02, threshold=7.074e+02, percent-clipped=1.0 2023-04-29 16:00:56,362 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=119575.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:00:59,288 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8893, 2.3188, 2.3851, 2.8064, 2.1750, 3.1557, 1.6968, 2.7509], device='cuda:4'), covar=tensor([0.1276, 0.0632, 0.0999, 0.0175, 0.0158, 0.0349, 0.1423, 0.0692], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0162, 0.0182, 0.0151, 0.0199, 0.0208, 0.0183, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 16:01:16,969 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5998, 2.5243, 2.2753, 3.4335, 2.2574, 3.5987, 1.3981, 2.7579], device='cuda:4'), covar=tensor([0.1406, 0.0690, 0.1189, 0.0174, 0.0173, 0.0375, 0.1718, 0.0748], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0162, 0.0182, 0.0151, 0.0199, 0.0208, 0.0183, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 16:01:38,578 INFO [train.py:904] (4/8) Epoch 12, batch 7950, loss[loss=0.2034, simple_loss=0.2848, pruned_loss=0.06098, over 16700.00 frames. ], tot_loss[loss=0.2196, simple_loss=0.302, pruned_loss=0.06859, over 3100301.49 frames. ], batch size: 62, lr: 5.63e-03, grad_scale: 4.0 2023-04-29 16:02:53,353 INFO [train.py:904] (4/8) Epoch 12, batch 8000, loss[loss=0.2355, simple_loss=0.3205, pruned_loss=0.07526, over 16859.00 frames. ], tot_loss[loss=0.2216, simple_loss=0.3033, pruned_loss=0.06999, over 3080517.61 frames. ], batch size: 116, lr: 5.63e-03, grad_scale: 8.0 2023-04-29 16:03:12,810 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=119665.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:03:17,122 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.726e+02 2.898e+02 3.526e+02 5.291e+02 1.062e+03, threshold=7.052e+02, percent-clipped=9.0 2023-04-29 16:03:29,298 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.62 vs. limit=5.0 2023-04-29 16:04:07,955 INFO [train.py:904] (4/8) Epoch 12, batch 8050, loss[loss=0.2057, simple_loss=0.2945, pruned_loss=0.05843, over 16329.00 frames. ], tot_loss[loss=0.221, simple_loss=0.3027, pruned_loss=0.06962, over 3073232.08 frames. ], batch size: 146, lr: 5.62e-03, grad_scale: 4.0 2023-04-29 16:04:42,911 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=119726.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:04:49,746 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=119731.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:04:57,097 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0238, 3.1575, 3.1667, 2.1485, 2.9506, 3.1590, 3.0597, 1.8675], device='cuda:4'), covar=tensor([0.0426, 0.0048, 0.0042, 0.0323, 0.0077, 0.0091, 0.0068, 0.0368], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0069, 0.0071, 0.0126, 0.0081, 0.0091, 0.0080, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 16:05:17,354 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=119749.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 16:05:21,185 INFO [train.py:904] (4/8) Epoch 12, batch 8100, loss[loss=0.2021, simple_loss=0.2836, pruned_loss=0.06032, over 16807.00 frames. ], tot_loss[loss=0.2207, simple_loss=0.3027, pruned_loss=0.06933, over 3060053.18 frames. ], batch size: 39, lr: 5.62e-03, grad_scale: 4.0 2023-04-29 16:05:47,745 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.274e+02 3.055e+02 3.528e+02 4.290e+02 7.598e+02, threshold=7.056e+02, percent-clipped=1.0 2023-04-29 16:05:52,395 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6173, 2.7364, 2.2059, 2.4517, 3.0690, 2.8022, 3.4015, 3.3294], device='cuda:4'), covar=tensor([0.0076, 0.0283, 0.0395, 0.0312, 0.0179, 0.0276, 0.0156, 0.0172], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0204, 0.0200, 0.0200, 0.0205, 0.0202, 0.0208, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:06:00,175 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=119779.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:06:14,920 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0765, 2.4188, 1.9267, 2.1886, 2.7549, 2.4779, 2.9206, 2.9690], device='cuda:4'), covar=tensor([0.0106, 0.0323, 0.0418, 0.0338, 0.0181, 0.0274, 0.0168, 0.0203], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0203, 0.0199, 0.0199, 0.0204, 0.0201, 0.0206, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:06:28,056 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=119797.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:06:34,772 INFO [train.py:904] (4/8) Epoch 12, batch 8150, loss[loss=0.2152, simple_loss=0.2842, pruned_loss=0.07307, over 11220.00 frames. ], tot_loss[loss=0.2173, simple_loss=0.2992, pruned_loss=0.06771, over 3082378.92 frames. ], batch size: 247, lr: 5.62e-03, grad_scale: 2.0 2023-04-29 16:07:50,608 INFO [train.py:904] (4/8) Epoch 12, batch 8200, loss[loss=0.2058, simple_loss=0.2854, pruned_loss=0.06307, over 16838.00 frames. ], tot_loss[loss=0.215, simple_loss=0.2965, pruned_loss=0.06673, over 3095005.77 frames. ], batch size: 42, lr: 5.62e-03, grad_scale: 2.0 2023-04-29 16:08:18,226 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.831e+02 3.028e+02 3.797e+02 4.737e+02 9.915e+02, threshold=7.593e+02, percent-clipped=4.0 2023-04-29 16:08:19,312 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=119870.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:08:36,981 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-29 16:08:38,920 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-29 16:09:09,078 INFO [train.py:904] (4/8) Epoch 12, batch 8250, loss[loss=0.1925, simple_loss=0.2838, pruned_loss=0.05065, over 16817.00 frames. ], tot_loss[loss=0.2127, simple_loss=0.2957, pruned_loss=0.0648, over 3076645.24 frames. ], batch size: 83, lr: 5.62e-03, grad_scale: 2.0 2023-04-29 16:10:28,023 INFO [train.py:904] (4/8) Epoch 12, batch 8300, loss[loss=0.1943, simple_loss=0.2881, pruned_loss=0.05027, over 15311.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2931, pruned_loss=0.06155, over 3064136.45 frames. ], batch size: 190, lr: 5.62e-03, grad_scale: 2.0 2023-04-29 16:10:57,564 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.672e+02 2.352e+02 2.901e+02 3.420e+02 5.863e+02, threshold=5.801e+02, percent-clipped=0.0 2023-04-29 16:11:09,023 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5944, 3.5621, 3.5071, 2.8838, 3.4105, 2.0217, 3.2907, 2.9177], device='cuda:4'), covar=tensor([0.0100, 0.0093, 0.0129, 0.0188, 0.0080, 0.2072, 0.0092, 0.0200], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0117, 0.0162, 0.0153, 0.0134, 0.0178, 0.0150, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:11:52,965 INFO [train.py:904] (4/8) Epoch 12, batch 8350, loss[loss=0.1743, simple_loss=0.2687, pruned_loss=0.03994, over 17198.00 frames. ], tot_loss[loss=0.2049, simple_loss=0.2917, pruned_loss=0.05904, over 3065190.96 frames. ], batch size: 44, lr: 5.62e-03, grad_scale: 2.0 2023-04-29 16:12:09,927 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8901, 4.2174, 4.0280, 4.0369, 3.6613, 3.8109, 3.8072, 4.1823], device='cuda:4'), covar=tensor([0.0970, 0.0848, 0.0917, 0.0713, 0.0873, 0.1565, 0.0998, 0.1007], device='cuda:4'), in_proj_covar=tensor([0.0540, 0.0669, 0.0555, 0.0475, 0.0425, 0.0444, 0.0560, 0.0519], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:12:24,536 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=120021.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:12:43,045 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7956, 3.9721, 3.1169, 2.1586, 2.6342, 2.3624, 4.2615, 3.5382], device='cuda:4'), covar=tensor([0.2345, 0.0551, 0.1396, 0.2625, 0.2565, 0.1922, 0.0331, 0.1108], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0250, 0.0279, 0.0275, 0.0273, 0.0218, 0.0261, 0.0289], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:13:14,449 INFO [train.py:904] (4/8) Epoch 12, batch 8400, loss[loss=0.1827, simple_loss=0.2742, pruned_loss=0.04559, over 16839.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2888, pruned_loss=0.0567, over 3053754.44 frames. ], batch size: 116, lr: 5.62e-03, grad_scale: 4.0 2023-04-29 16:13:42,957 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.720e+02 2.479e+02 2.965e+02 3.428e+02 6.852e+02, threshold=5.930e+02, percent-clipped=2.0 2023-04-29 16:14:01,885 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9732, 2.3627, 2.3412, 2.9539, 2.0109, 3.3121, 1.7083, 2.8206], device='cuda:4'), covar=tensor([0.1183, 0.0557, 0.0963, 0.0135, 0.0112, 0.0375, 0.1414, 0.0610], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0160, 0.0181, 0.0148, 0.0196, 0.0206, 0.0182, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 16:14:31,500 INFO [train.py:904] (4/8) Epoch 12, batch 8450, loss[loss=0.1784, simple_loss=0.2741, pruned_loss=0.04134, over 16807.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2868, pruned_loss=0.05498, over 3054371.06 frames. ], batch size: 116, lr: 5.62e-03, grad_scale: 4.0 2023-04-29 16:15:47,450 INFO [train.py:904] (4/8) Epoch 12, batch 8500, loss[loss=0.1986, simple_loss=0.296, pruned_loss=0.05061, over 16745.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2831, pruned_loss=0.05259, over 3049301.21 frames. ], batch size: 124, lr: 5.61e-03, grad_scale: 4.0 2023-04-29 16:16:15,429 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.656e+02 2.373e+02 2.767e+02 3.491e+02 7.213e+02, threshold=5.534e+02, percent-clipped=3.0 2023-04-29 16:16:15,857 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=120170.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:16:43,114 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-29 16:17:07,725 INFO [train.py:904] (4/8) Epoch 12, batch 8550, loss[loss=0.1987, simple_loss=0.2738, pruned_loss=0.06174, over 11932.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2799, pruned_loss=0.05132, over 3027358.36 frames. ], batch size: 248, lr: 5.61e-03, grad_scale: 4.0 2023-04-29 16:17:37,325 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=120218.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:18:37,419 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7450, 3.7327, 4.1101, 4.1047, 4.1062, 3.8735, 3.8906, 3.8385], device='cuda:4'), covar=tensor([0.0324, 0.0601, 0.0400, 0.0399, 0.0406, 0.0396, 0.0816, 0.0473], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0352, 0.0346, 0.0330, 0.0399, 0.0372, 0.0465, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 16:18:41,889 INFO [train.py:904] (4/8) Epoch 12, batch 8600, loss[loss=0.1656, simple_loss=0.252, pruned_loss=0.03959, over 12575.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2796, pruned_loss=0.05011, over 3016948.38 frames. ], batch size: 248, lr: 5.61e-03, grad_scale: 4.0 2023-04-29 16:19:19,442 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.485e+02 2.557e+02 3.073e+02 4.030e+02 6.956e+02, threshold=6.147e+02, percent-clipped=5.0 2023-04-29 16:20:19,200 INFO [train.py:904] (4/8) Epoch 12, batch 8650, loss[loss=0.1734, simple_loss=0.2673, pruned_loss=0.0397, over 15351.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2779, pruned_loss=0.04857, over 3019892.74 frames. ], batch size: 190, lr: 5.61e-03, grad_scale: 4.0 2023-04-29 16:20:32,390 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-29 16:21:01,199 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=120321.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:21:25,228 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.99 vs. limit=5.0 2023-04-29 16:22:02,147 INFO [train.py:904] (4/8) Epoch 12, batch 8700, loss[loss=0.1621, simple_loss=0.2543, pruned_loss=0.03495, over 16707.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.2752, pruned_loss=0.0472, over 3035377.77 frames. ], batch size: 83, lr: 5.61e-03, grad_scale: 4.0 2023-04-29 16:22:33,633 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=120369.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:22:34,433 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.439e+02 2.303e+02 2.810e+02 3.260e+02 5.191e+02, threshold=5.620e+02, percent-clipped=0.0 2023-04-29 16:23:36,538 INFO [train.py:904] (4/8) Epoch 12, batch 8750, loss[loss=0.2206, simple_loss=0.3111, pruned_loss=0.06506, over 16252.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.275, pruned_loss=0.04685, over 3031002.05 frames. ], batch size: 166, lr: 5.61e-03, grad_scale: 4.0 2023-04-29 16:24:28,491 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7400, 2.5287, 2.4828, 3.4408, 2.2798, 3.6724, 1.4930, 2.8690], device='cuda:4'), covar=tensor([0.1357, 0.0707, 0.1127, 0.0143, 0.0111, 0.0383, 0.1624, 0.0734], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0159, 0.0180, 0.0146, 0.0192, 0.0204, 0.0182, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 16:25:30,725 INFO [train.py:904] (4/8) Epoch 12, batch 8800, loss[loss=0.1825, simple_loss=0.2767, pruned_loss=0.04415, over 15463.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2734, pruned_loss=0.04548, over 3058422.75 frames. ], batch size: 191, lr: 5.61e-03, grad_scale: 8.0 2023-04-29 16:26:08,426 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.720e+02 2.538e+02 3.135e+02 3.629e+02 6.620e+02, threshold=6.271e+02, percent-clipped=2.0 2023-04-29 16:26:45,753 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.88 vs. limit=2.0 2023-04-29 16:26:59,019 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-29 16:27:17,117 INFO [train.py:904] (4/8) Epoch 12, batch 8850, loss[loss=0.183, simple_loss=0.282, pruned_loss=0.04193, over 15488.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2766, pruned_loss=0.04533, over 3051850.94 frames. ], batch size: 191, lr: 5.61e-03, grad_scale: 8.0 2023-04-29 16:27:36,592 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-29 16:27:38,339 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4907, 3.8242, 2.7015, 1.9924, 2.5396, 2.1568, 3.8692, 3.3424], device='cuda:4'), covar=tensor([0.2946, 0.0604, 0.1746, 0.2571, 0.2588, 0.2031, 0.0497, 0.1053], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0246, 0.0277, 0.0271, 0.0264, 0.0216, 0.0258, 0.0284], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:27:42,797 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.01 vs. limit=5.0 2023-04-29 16:29:04,467 INFO [train.py:904] (4/8) Epoch 12, batch 8900, loss[loss=0.18, simple_loss=0.2754, pruned_loss=0.04228, over 16204.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2754, pruned_loss=0.04447, over 3038474.64 frames. ], batch size: 165, lr: 5.60e-03, grad_scale: 8.0 2023-04-29 16:29:10,215 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-29 16:29:39,305 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.788e+02 2.488e+02 2.995e+02 3.537e+02 1.098e+03, threshold=5.991e+02, percent-clipped=1.0 2023-04-29 16:31:00,340 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.11 vs. limit=5.0 2023-04-29 16:31:10,991 INFO [train.py:904] (4/8) Epoch 12, batch 8950, loss[loss=0.1658, simple_loss=0.2597, pruned_loss=0.0359, over 16253.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2752, pruned_loss=0.04449, over 3073613.51 frames. ], batch size: 165, lr: 5.60e-03, grad_scale: 8.0 2023-04-29 16:33:00,370 INFO [train.py:904] (4/8) Epoch 12, batch 9000, loss[loss=0.1686, simple_loss=0.2617, pruned_loss=0.03771, over 16147.00 frames. ], tot_loss[loss=0.1794, simple_loss=0.2724, pruned_loss=0.04321, over 3076105.51 frames. ], batch size: 165, lr: 5.60e-03, grad_scale: 8.0 2023-04-29 16:33:00,370 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 16:33:10,348 INFO [train.py:938] (4/8) Epoch 12, validation: loss=0.1532, simple_loss=0.2571, pruned_loss=0.02465, over 944034.00 frames. 2023-04-29 16:33:10,349 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 16:33:49,107 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.498e+02 2.337e+02 2.622e+02 3.241e+02 7.245e+02, threshold=5.244e+02, percent-clipped=4.0 2023-04-29 16:34:01,230 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=120676.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:34:54,268 INFO [train.py:904] (4/8) Epoch 12, batch 9050, loss[loss=0.1884, simple_loss=0.2734, pruned_loss=0.05164, over 16954.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2739, pruned_loss=0.04388, over 3094115.50 frames. ], batch size: 109, lr: 5.60e-03, grad_scale: 8.0 2023-04-29 16:35:27,440 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9193, 3.0828, 3.1743, 2.1558, 2.9296, 3.2125, 3.1161, 1.8402], device='cuda:4'), covar=tensor([0.0475, 0.0043, 0.0042, 0.0326, 0.0073, 0.0062, 0.0061, 0.0438], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0067, 0.0068, 0.0124, 0.0079, 0.0088, 0.0079, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 16:35:43,815 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1484, 5.1296, 4.9832, 4.5526, 4.6072, 5.0089, 5.0091, 4.6755], device='cuda:4'), covar=tensor([0.0543, 0.0457, 0.0273, 0.0303, 0.0981, 0.0386, 0.0270, 0.0646], device='cuda:4'), in_proj_covar=tensor([0.0232, 0.0312, 0.0275, 0.0256, 0.0291, 0.0296, 0.0191, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:36:04,408 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=120737.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:36:36,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6043, 6.0288, 5.6802, 5.7694, 5.3726, 5.3146, 5.3824, 6.0848], device='cuda:4'), covar=tensor([0.1127, 0.0843, 0.0966, 0.0691, 0.0703, 0.0577, 0.1003, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0528, 0.0654, 0.0539, 0.0465, 0.0419, 0.0433, 0.0546, 0.0510], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:36:39,406 INFO [train.py:904] (4/8) Epoch 12, batch 9100, loss[loss=0.1984, simple_loss=0.3002, pruned_loss=0.04829, over 16464.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.274, pruned_loss=0.04468, over 3088697.18 frames. ], batch size: 146, lr: 5.60e-03, grad_scale: 8.0 2023-04-29 16:37:15,444 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.896e+02 2.513e+02 2.910e+02 3.503e+02 5.095e+02, threshold=5.819e+02, percent-clipped=0.0 2023-04-29 16:37:27,109 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9825, 3.0992, 3.1336, 2.1764, 2.9455, 3.1871, 3.0870, 1.9193], device='cuda:4'), covar=tensor([0.0457, 0.0041, 0.0037, 0.0324, 0.0082, 0.0064, 0.0061, 0.0413], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0066, 0.0067, 0.0122, 0.0078, 0.0086, 0.0077, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 16:38:36,944 INFO [train.py:904] (4/8) Epoch 12, batch 9150, loss[loss=0.1912, simple_loss=0.2875, pruned_loss=0.0475, over 16380.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.274, pruned_loss=0.04418, over 3073915.42 frames. ], batch size: 147, lr: 5.60e-03, grad_scale: 8.0 2023-04-29 16:40:21,513 INFO [train.py:904] (4/8) Epoch 12, batch 9200, loss[loss=0.172, simple_loss=0.262, pruned_loss=0.04097, over 15190.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2709, pruned_loss=0.04369, over 3067389.69 frames. ], batch size: 190, lr: 5.60e-03, grad_scale: 8.0 2023-04-29 16:40:55,536 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.751e+02 2.400e+02 2.881e+02 3.588e+02 6.775e+02, threshold=5.761e+02, percent-clipped=4.0 2023-04-29 16:41:00,513 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-29 16:41:03,835 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6671, 2.5059, 2.1812, 3.6007, 2.0127, 3.6136, 1.3512, 2.7256], device='cuda:4'), covar=tensor([0.1499, 0.0768, 0.1286, 0.0165, 0.0130, 0.0393, 0.1854, 0.0800], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0158, 0.0179, 0.0143, 0.0187, 0.0202, 0.0181, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 16:41:14,951 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.75 vs. limit=2.0 2023-04-29 16:42:00,504 INFO [train.py:904] (4/8) Epoch 12, batch 9250, loss[loss=0.1603, simple_loss=0.2382, pruned_loss=0.04116, over 12312.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2699, pruned_loss=0.04386, over 3041767.21 frames. ], batch size: 250, lr: 5.60e-03, grad_scale: 8.0 2023-04-29 16:43:49,104 INFO [train.py:904] (4/8) Epoch 12, batch 9300, loss[loss=0.1569, simple_loss=0.2406, pruned_loss=0.03662, over 12469.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2685, pruned_loss=0.04304, over 3046829.10 frames. ], batch size: 248, lr: 5.60e-03, grad_scale: 4.0 2023-04-29 16:44:20,082 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6526, 4.7198, 4.5164, 4.1793, 4.2121, 4.6196, 4.4540, 4.3137], device='cuda:4'), covar=tensor([0.0550, 0.0453, 0.0248, 0.0256, 0.0772, 0.0430, 0.0351, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0234, 0.0313, 0.0276, 0.0256, 0.0290, 0.0297, 0.0191, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:44:31,798 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.540e+02 2.549e+02 3.003e+02 3.626e+02 5.994e+02, threshold=6.005e+02, percent-clipped=1.0 2023-04-29 16:45:22,336 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9458, 4.2317, 4.0475, 4.0722, 3.7189, 3.7955, 3.8375, 4.1952], device='cuda:4'), covar=tensor([0.0887, 0.0871, 0.0856, 0.0659, 0.0746, 0.1390, 0.0832, 0.0925], device='cuda:4'), in_proj_covar=tensor([0.0524, 0.0653, 0.0537, 0.0462, 0.0417, 0.0434, 0.0543, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:45:32,149 INFO [train.py:904] (4/8) Epoch 12, batch 9350, loss[loss=0.191, simple_loss=0.282, pruned_loss=0.05001, over 15558.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2679, pruned_loss=0.04265, over 3057842.97 frames. ], batch size: 191, lr: 5.59e-03, grad_scale: 4.0 2023-04-29 16:45:35,617 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121003.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:46:24,108 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1474, 4.2139, 4.5020, 2.1516, 4.6894, 4.7485, 3.4227, 3.6430], device='cuda:4'), covar=tensor([0.0571, 0.0149, 0.0135, 0.1027, 0.0035, 0.0057, 0.0313, 0.0309], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0095, 0.0082, 0.0132, 0.0065, 0.0099, 0.0116, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 16:46:33,080 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121032.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:47:13,259 INFO [train.py:904] (4/8) Epoch 12, batch 9400, loss[loss=0.1584, simple_loss=0.2425, pruned_loss=0.03716, over 12345.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2683, pruned_loss=0.0429, over 3046842.67 frames. ], batch size: 248, lr: 5.59e-03, grad_scale: 4.0 2023-04-29 16:47:34,786 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4175, 1.9471, 1.7406, 1.7039, 2.2746, 1.8797, 2.0304, 2.3728], device='cuda:4'), covar=tensor([0.0093, 0.0316, 0.0391, 0.0376, 0.0208, 0.0286, 0.0121, 0.0171], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0204, 0.0199, 0.0198, 0.0204, 0.0201, 0.0200, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:47:39,111 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=121064.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:47:50,521 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.435e+02 2.897e+02 3.662e+02 7.614e+02, threshold=5.793e+02, percent-clipped=2.0 2023-04-29 16:48:04,400 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2367, 4.2164, 4.1215, 3.6107, 4.1180, 1.6302, 3.9193, 3.7831], device='cuda:4'), covar=tensor([0.0071, 0.0068, 0.0121, 0.0194, 0.0074, 0.2357, 0.0097, 0.0195], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0115, 0.0157, 0.0146, 0.0132, 0.0178, 0.0147, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 16:48:55,032 INFO [train.py:904] (4/8) Epoch 12, batch 9450, loss[loss=0.1845, simple_loss=0.2753, pruned_loss=0.04684, over 15430.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2708, pruned_loss=0.04333, over 3055923.95 frames. ], batch size: 191, lr: 5.59e-03, grad_scale: 4.0 2023-04-29 16:49:39,711 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0452, 1.4228, 1.7903, 2.1084, 2.1845, 2.2648, 1.5997, 2.2319], device='cuda:4'), covar=tensor([0.0191, 0.0397, 0.0238, 0.0259, 0.0238, 0.0156, 0.0388, 0.0121], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0170, 0.0156, 0.0157, 0.0167, 0.0121, 0.0170, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 16:49:47,137 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0793, 3.4938, 3.7462, 1.9160, 3.1643, 2.4362, 3.6310, 3.4742], device='cuda:4'), covar=tensor([0.0253, 0.0766, 0.0475, 0.1911, 0.0695, 0.0880, 0.0641, 0.1003], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0138, 0.0155, 0.0141, 0.0133, 0.0123, 0.0133, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 16:50:34,751 INFO [train.py:904] (4/8) Epoch 12, batch 9500, loss[loss=0.1759, simple_loss=0.2708, pruned_loss=0.04047, over 16191.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2702, pruned_loss=0.04305, over 3042582.24 frames. ], batch size: 165, lr: 5.59e-03, grad_scale: 4.0 2023-04-29 16:51:01,834 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121164.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:51:13,616 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.587e+02 2.296e+02 2.757e+02 3.587e+02 6.751e+02, threshold=5.515e+02, percent-clipped=2.0 2023-04-29 16:52:19,995 INFO [train.py:904] (4/8) Epoch 12, batch 9550, loss[loss=0.1948, simple_loss=0.291, pruned_loss=0.04932, over 15246.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2697, pruned_loss=0.0429, over 3052616.53 frames. ], batch size: 190, lr: 5.59e-03, grad_scale: 4.0 2023-04-29 16:53:08,735 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=121225.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:53:59,925 INFO [train.py:904] (4/8) Epoch 12, batch 9600, loss[loss=0.1878, simple_loss=0.2696, pruned_loss=0.053, over 12456.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.271, pruned_loss=0.04356, over 3056532.90 frames. ], batch size: 246, lr: 5.59e-03, grad_scale: 8.0 2023-04-29 16:54:35,187 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.521e+02 2.441e+02 3.037e+02 3.482e+02 7.490e+02, threshold=6.075e+02, percent-clipped=2.0 2023-04-29 16:55:12,433 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4079, 4.4697, 4.5913, 4.3963, 4.4497, 5.0152, 4.5611, 4.2692], device='cuda:4'), covar=tensor([0.1158, 0.1909, 0.1711, 0.2227, 0.2629, 0.0946, 0.1420, 0.2358], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0470, 0.0524, 0.0406, 0.0541, 0.0545, 0.0410, 0.0544], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 16:55:45,648 INFO [train.py:904] (4/8) Epoch 12, batch 9650, loss[loss=0.1963, simple_loss=0.2848, pruned_loss=0.05384, over 16331.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2733, pruned_loss=0.04431, over 3047512.08 frames. ], batch size: 146, lr: 5.59e-03, grad_scale: 8.0 2023-04-29 16:56:09,932 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121311.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 16:56:51,958 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=121332.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:57:28,668 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1310, 2.6162, 2.7285, 1.8936, 2.8675, 2.9106, 2.5257, 2.4611], device='cuda:4'), covar=tensor([0.0646, 0.0197, 0.0160, 0.0954, 0.0075, 0.0181, 0.0433, 0.0408], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0098, 0.0084, 0.0135, 0.0067, 0.0102, 0.0119, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 16:57:30,748 INFO [train.py:904] (4/8) Epoch 12, batch 9700, loss[loss=0.174, simple_loss=0.2563, pruned_loss=0.04583, over 12462.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2716, pruned_loss=0.04393, over 3042523.24 frames. ], batch size: 248, lr: 5.59e-03, grad_scale: 8.0 2023-04-29 16:57:44,329 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121359.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:58:07,288 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.590e+02 2.472e+02 3.175e+02 3.928e+02 9.920e+02, threshold=6.351e+02, percent-clipped=4.0 2023-04-29 16:58:11,037 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=121372.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 16:58:31,088 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=121380.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:58:33,108 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-29 16:59:04,700 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9972, 4.0761, 3.8706, 3.5923, 3.6019, 3.9957, 3.6949, 3.7685], device='cuda:4'), covar=tensor([0.0599, 0.0526, 0.0295, 0.0274, 0.0798, 0.0403, 0.0919, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0232, 0.0304, 0.0273, 0.0252, 0.0287, 0.0293, 0.0188, 0.0317], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-29 16:59:14,333 INFO [train.py:904] (4/8) Epoch 12, batch 9750, loss[loss=0.1687, simple_loss=0.2537, pruned_loss=0.0419, over 12296.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2705, pruned_loss=0.04407, over 3040171.97 frames. ], batch size: 247, lr: 5.59e-03, grad_scale: 8.0 2023-04-29 16:59:23,933 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121407.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 16:59:52,977 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-29 16:59:58,335 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-29 17:00:39,139 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7645, 1.2310, 1.6635, 1.6887, 1.8296, 1.8810, 1.6298, 1.8422], device='cuda:4'), covar=tensor([0.0249, 0.0305, 0.0165, 0.0216, 0.0211, 0.0158, 0.0283, 0.0097], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0168, 0.0154, 0.0156, 0.0166, 0.0120, 0.0167, 0.0113], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001], device='cuda:4') 2023-04-29 17:00:53,746 INFO [train.py:904] (4/8) Epoch 12, batch 9800, loss[loss=0.178, simple_loss=0.2797, pruned_loss=0.03814, over 16401.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2702, pruned_loss=0.04279, over 3046678.15 frames. ], batch size: 146, lr: 5.58e-03, grad_scale: 8.0 2023-04-29 17:01:24,645 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=121468.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:01:29,898 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.684e+02 2.372e+02 2.708e+02 3.035e+02 5.819e+02, threshold=5.415e+02, percent-clipped=0.0 2023-04-29 17:02:38,495 INFO [train.py:904] (4/8) Epoch 12, batch 9850, loss[loss=0.1771, simple_loss=0.2744, pruned_loss=0.03994, over 16771.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2712, pruned_loss=0.0422, over 3059340.90 frames. ], batch size: 83, lr: 5.58e-03, grad_scale: 8.0 2023-04-29 17:03:17,654 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121520.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:03:23,765 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1637, 4.0996, 4.2924, 4.4091, 4.5434, 4.0655, 4.5081, 4.5651], device='cuda:4'), covar=tensor([0.1504, 0.1037, 0.1317, 0.0661, 0.0508, 0.1046, 0.0597, 0.0582], device='cuda:4'), in_proj_covar=tensor([0.0500, 0.0625, 0.0750, 0.0640, 0.0482, 0.0495, 0.0508, 0.0581], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:04:30,193 INFO [train.py:904] (4/8) Epoch 12, batch 9900, loss[loss=0.1752, simple_loss=0.2784, pruned_loss=0.03601, over 15350.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2717, pruned_loss=0.04211, over 3057967.78 frames. ], batch size: 191, lr: 5.58e-03, grad_scale: 8.0 2023-04-29 17:05:12,985 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.540e+02 2.385e+02 2.886e+02 3.421e+02 6.283e+02, threshold=5.771e+02, percent-clipped=3.0 2023-04-29 17:06:15,896 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-29 17:06:26,639 INFO [train.py:904] (4/8) Epoch 12, batch 9950, loss[loss=0.1861, simple_loss=0.283, pruned_loss=0.04461, over 17013.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2733, pruned_loss=0.04233, over 3058482.14 frames. ], batch size: 109, lr: 5.58e-03, grad_scale: 8.0 2023-04-29 17:08:27,932 INFO [train.py:904] (4/8) Epoch 12, batch 10000, loss[loss=0.1924, simple_loss=0.2857, pruned_loss=0.04959, over 12546.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2721, pruned_loss=0.04217, over 3072043.69 frames. ], batch size: 247, lr: 5.58e-03, grad_scale: 8.0 2023-04-29 17:08:44,252 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=121659.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:08:50,123 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4190, 3.3911, 3.4823, 3.5647, 3.6148, 3.2950, 3.5747, 3.6271], device='cuda:4'), covar=tensor([0.1114, 0.0865, 0.1042, 0.0587, 0.0546, 0.1985, 0.0730, 0.0690], device='cuda:4'), in_proj_covar=tensor([0.0497, 0.0625, 0.0748, 0.0638, 0.0480, 0.0491, 0.0504, 0.0578], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:08:56,058 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9312, 5.2241, 5.0088, 4.9780, 4.6774, 4.6249, 4.5775, 5.2520], device='cuda:4'), covar=tensor([0.0845, 0.0759, 0.0860, 0.0633, 0.0690, 0.0806, 0.0982, 0.0767], device='cuda:4'), in_proj_covar=tensor([0.0517, 0.0651, 0.0531, 0.0459, 0.0415, 0.0428, 0.0538, 0.0499], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:08:59,916 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121667.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 17:09:06,349 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.215e+02 2.322e+02 2.719e+02 3.378e+02 7.215e+02, threshold=5.438e+02, percent-clipped=2.0 2023-04-29 17:10:10,827 INFO [train.py:904] (4/8) Epoch 12, batch 10050, loss[loss=0.1762, simple_loss=0.2655, pruned_loss=0.04345, over 11795.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2718, pruned_loss=0.04178, over 3066727.83 frames. ], batch size: 247, lr: 5.58e-03, grad_scale: 8.0 2023-04-29 17:10:21,423 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=121707.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:11:46,088 INFO [train.py:904] (4/8) Epoch 12, batch 10100, loss[loss=0.1659, simple_loss=0.255, pruned_loss=0.03836, over 16882.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2722, pruned_loss=0.04221, over 3058738.28 frames. ], batch size: 116, lr: 5.58e-03, grad_scale: 8.0 2023-04-29 17:12:06,123 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121763.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:12:23,726 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.628e+02 2.328e+02 2.737e+02 3.349e+02 5.090e+02, threshold=5.474e+02, percent-clipped=0.0 2023-04-29 17:13:30,553 INFO [train.py:904] (4/8) Epoch 13, batch 0, loss[loss=0.2041, simple_loss=0.2915, pruned_loss=0.05837, over 17095.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2915, pruned_loss=0.05837, over 17095.00 frames. ], batch size: 48, lr: 5.36e-03, grad_scale: 8.0 2023-04-29 17:13:30,553 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 17:13:38,108 INFO [train.py:938] (4/8) Epoch 13, validation: loss=0.1523, simple_loss=0.2559, pruned_loss=0.02431, over 944034.00 frames. 2023-04-29 17:13:38,109 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 17:14:04,079 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=121820.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:14:30,460 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8189, 5.0244, 5.1991, 5.0571, 4.9560, 5.6131, 5.1964, 4.9238], device='cuda:4'), covar=tensor([0.1044, 0.1918, 0.2270, 0.2067, 0.2625, 0.1016, 0.1554, 0.2202], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0476, 0.0525, 0.0413, 0.0548, 0.0547, 0.0416, 0.0546], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 17:14:49,500 INFO [train.py:904] (4/8) Epoch 13, batch 50, loss[loss=0.1649, simple_loss=0.2506, pruned_loss=0.03957, over 15859.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2824, pruned_loss=0.05941, over 734114.43 frames. ], batch size: 35, lr: 5.36e-03, grad_scale: 2.0 2023-04-29 17:15:11,442 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=121868.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:15:18,817 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.864e+02 2.677e+02 3.187e+02 4.069e+02 9.250e+02, threshold=6.374e+02, percent-clipped=6.0 2023-04-29 17:15:58,269 INFO [train.py:904] (4/8) Epoch 13, batch 100, loss[loss=0.2084, simple_loss=0.2974, pruned_loss=0.05975, over 16655.00 frames. ], tot_loss[loss=0.1949, simple_loss=0.2775, pruned_loss=0.05613, over 1312645.57 frames. ], batch size: 62, lr: 5.36e-03, grad_scale: 2.0 2023-04-29 17:16:10,829 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7354, 4.0558, 4.3412, 3.1861, 3.7429, 4.2504, 3.9962, 2.7248], device='cuda:4'), covar=tensor([0.0380, 0.0052, 0.0025, 0.0260, 0.0068, 0.0055, 0.0046, 0.0301], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0070, 0.0070, 0.0127, 0.0080, 0.0088, 0.0079, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 17:16:27,130 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-29 17:16:46,533 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8044, 2.9703, 2.5603, 5.0000, 3.8291, 4.2233, 1.5793, 3.0573], device='cuda:4'), covar=tensor([0.1402, 0.0761, 0.1355, 0.0141, 0.0337, 0.0442, 0.1666, 0.0849], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0156, 0.0179, 0.0144, 0.0185, 0.0203, 0.0181, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 17:17:04,498 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121950.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:17:07,072 INFO [train.py:904] (4/8) Epoch 13, batch 150, loss[loss=0.1577, simple_loss=0.2464, pruned_loss=0.0345, over 17234.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2763, pruned_loss=0.05545, over 1760994.29 frames. ], batch size: 45, lr: 5.35e-03, grad_scale: 2.0 2023-04-29 17:17:27,826 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=121967.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 17:17:35,629 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.742e+02 2.538e+02 3.283e+02 4.116e+02 1.264e+03, threshold=6.566e+02, percent-clipped=3.0 2023-04-29 17:18:00,010 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5696, 4.9280, 4.7006, 4.6582, 4.4493, 4.4417, 4.4318, 4.9963], device='cuda:4'), covar=tensor([0.1132, 0.0863, 0.1042, 0.0769, 0.0786, 0.1080, 0.0998, 0.0880], device='cuda:4'), in_proj_covar=tensor([0.0544, 0.0684, 0.0558, 0.0482, 0.0437, 0.0446, 0.0570, 0.0522], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:18:18,197 INFO [train.py:904] (4/8) Epoch 13, batch 200, loss[loss=0.162, simple_loss=0.2569, pruned_loss=0.03353, over 17226.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2761, pruned_loss=0.05414, over 2108036.25 frames. ], batch size: 44, lr: 5.35e-03, grad_scale: 2.0 2023-04-29 17:18:30,514 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122011.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:18:35,253 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=122015.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 17:18:42,696 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.60 vs. limit=2.0 2023-04-29 17:19:26,246 INFO [train.py:904] (4/8) Epoch 13, batch 250, loss[loss=0.1839, simple_loss=0.2745, pruned_loss=0.04668, over 17101.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2736, pruned_loss=0.05363, over 2382370.87 frames. ], batch size: 53, lr: 5.35e-03, grad_scale: 2.0 2023-04-29 17:19:41,321 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122063.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:19:54,203 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.354e+02 2.915e+02 3.517e+02 1.210e+03, threshold=5.831e+02, percent-clipped=2.0 2023-04-29 17:20:34,740 INFO [train.py:904] (4/8) Epoch 13, batch 300, loss[loss=0.2018, simple_loss=0.2688, pruned_loss=0.0674, over 16586.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2712, pruned_loss=0.05321, over 2580373.29 frames. ], batch size: 75, lr: 5.35e-03, grad_scale: 2.0 2023-04-29 17:20:48,063 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=122111.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:20:55,779 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122117.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:21:42,719 INFO [train.py:904] (4/8) Epoch 13, batch 350, loss[loss=0.1792, simple_loss=0.245, pruned_loss=0.05663, over 16881.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2672, pruned_loss=0.05137, over 2749155.08 frames. ], batch size: 116, lr: 5.35e-03, grad_scale: 2.0 2023-04-29 17:22:13,942 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.554e+02 2.369e+02 2.759e+02 3.305e+02 7.145e+02, threshold=5.518e+02, percent-clipped=2.0 2023-04-29 17:22:20,576 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122178.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 17:22:44,488 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122195.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:22:53,135 INFO [train.py:904] (4/8) Epoch 13, batch 400, loss[loss=0.1891, simple_loss=0.2627, pruned_loss=0.05775, over 16907.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2657, pruned_loss=0.05125, over 2876789.50 frames. ], batch size: 90, lr: 5.35e-03, grad_scale: 4.0 2023-04-29 17:24:03,759 INFO [train.py:904] (4/8) Epoch 13, batch 450, loss[loss=0.1618, simple_loss=0.2542, pruned_loss=0.03467, over 17129.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2643, pruned_loss=0.04991, over 2977549.78 frames. ], batch size: 49, lr: 5.35e-03, grad_scale: 4.0 2023-04-29 17:24:09,623 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122256.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:24:29,386 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0248, 2.0493, 2.5129, 3.0777, 2.7669, 3.5374, 2.2040, 3.3825], device='cuda:4'), covar=tensor([0.0172, 0.0345, 0.0239, 0.0229, 0.0231, 0.0125, 0.0330, 0.0106], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0174, 0.0161, 0.0164, 0.0174, 0.0127, 0.0175, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 17:24:32,943 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.250e+02 2.134e+02 2.674e+02 3.325e+02 9.122e+02, threshold=5.349e+02, percent-clipped=2.0 2023-04-29 17:25:13,885 INFO [train.py:904] (4/8) Epoch 13, batch 500, loss[loss=0.1914, simple_loss=0.2629, pruned_loss=0.05993, over 12203.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2624, pruned_loss=0.04914, over 3049590.70 frames. ], batch size: 247, lr: 5.35e-03, grad_scale: 4.0 2023-04-29 17:25:19,448 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122306.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:25:36,855 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122319.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:25:59,324 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0757, 5.0162, 4.8484, 4.4009, 4.4354, 4.8924, 4.8599, 4.5733], device='cuda:4'), covar=tensor([0.0547, 0.0441, 0.0305, 0.0311, 0.1071, 0.0419, 0.0331, 0.0664], device='cuda:4'), in_proj_covar=tensor([0.0252, 0.0338, 0.0301, 0.0280, 0.0319, 0.0320, 0.0204, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:26:07,819 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5838, 3.6096, 3.9596, 1.9329, 4.0523, 4.0714, 3.0421, 3.0513], device='cuda:4'), covar=tensor([0.0777, 0.0190, 0.0143, 0.1123, 0.0061, 0.0156, 0.0407, 0.0443], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0102, 0.0089, 0.0139, 0.0069, 0.0108, 0.0122, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 17:26:21,438 INFO [train.py:904] (4/8) Epoch 13, batch 550, loss[loss=0.1701, simple_loss=0.2428, pruned_loss=0.04874, over 16852.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2611, pruned_loss=0.04816, over 3111549.05 frames. ], batch size: 102, lr: 5.35e-03, grad_scale: 4.0 2023-04-29 17:26:50,249 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 2.266e+02 2.671e+02 3.065e+02 5.763e+02, threshold=5.342e+02, percent-clipped=1.0 2023-04-29 17:27:00,786 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122380.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:27:22,539 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8900, 1.9188, 2.3686, 2.8663, 2.6403, 3.2005, 2.1651, 3.2508], device='cuda:4'), covar=tensor([0.0193, 0.0385, 0.0246, 0.0247, 0.0249, 0.0169, 0.0325, 0.0121], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0160, 0.0163, 0.0174, 0.0128, 0.0174, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 17:27:30,036 INFO [train.py:904] (4/8) Epoch 13, batch 600, loss[loss=0.1742, simple_loss=0.2473, pruned_loss=0.05053, over 16878.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2607, pruned_loss=0.04815, over 3160887.97 frames. ], batch size: 102, lr: 5.35e-03, grad_scale: 4.0 2023-04-29 17:27:39,208 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9293, 4.0057, 2.0786, 4.4984, 2.8695, 4.5001, 2.3011, 3.1662], device='cuda:4'), covar=tensor([0.0208, 0.0368, 0.1509, 0.0169, 0.0731, 0.0355, 0.1314, 0.0575], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0164, 0.0188, 0.0136, 0.0170, 0.0205, 0.0197, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 17:27:49,652 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9180, 3.1300, 2.6933, 4.5982, 3.8038, 4.3768, 1.8183, 3.0785], device='cuda:4'), covar=tensor([0.1202, 0.0584, 0.0972, 0.0151, 0.0182, 0.0344, 0.1301, 0.0673], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0157, 0.0179, 0.0147, 0.0189, 0.0204, 0.0180, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 17:28:17,225 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.58 vs. limit=5.0 2023-04-29 17:28:37,794 INFO [train.py:904] (4/8) Epoch 13, batch 650, loss[loss=0.176, simple_loss=0.251, pruned_loss=0.0505, over 16908.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2601, pruned_loss=0.0478, over 3202511.87 frames. ], batch size: 90, lr: 5.34e-03, grad_scale: 4.0 2023-04-29 17:29:01,231 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122468.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:29:07,722 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.495e+02 2.494e+02 2.897e+02 3.411e+02 6.772e+02, threshold=5.794e+02, percent-clipped=6.0 2023-04-29 17:29:08,039 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122473.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 17:29:47,551 INFO [train.py:904] (4/8) Epoch 13, batch 700, loss[loss=0.1818, simple_loss=0.2556, pruned_loss=0.05405, over 12347.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2597, pruned_loss=0.04748, over 3224429.28 frames. ], batch size: 246, lr: 5.34e-03, grad_scale: 4.0 2023-04-29 17:30:26,071 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122529.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:30:56,880 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122551.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:30:57,598 INFO [train.py:904] (4/8) Epoch 13, batch 750, loss[loss=0.1903, simple_loss=0.2621, pruned_loss=0.0593, over 16617.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2604, pruned_loss=0.04731, over 3254982.54 frames. ], batch size: 134, lr: 5.34e-03, grad_scale: 4.0 2023-04-29 17:31:00,518 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3915, 4.3935, 4.8357, 4.8222, 4.8658, 4.4963, 4.5418, 4.3530], device='cuda:4'), covar=tensor([0.0399, 0.0585, 0.0429, 0.0463, 0.0462, 0.0370, 0.0885, 0.0584], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0374, 0.0371, 0.0353, 0.0420, 0.0396, 0.0493, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 17:31:27,840 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.426e+02 2.324e+02 2.801e+02 3.357e+02 5.795e+02, threshold=5.603e+02, percent-clipped=1.0 2023-04-29 17:32:03,025 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5649, 5.5254, 5.3912, 4.9871, 5.0123, 5.4220, 5.4220, 5.0718], device='cuda:4'), covar=tensor([0.0516, 0.0373, 0.0232, 0.0225, 0.0889, 0.0298, 0.0205, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0257, 0.0345, 0.0307, 0.0287, 0.0327, 0.0328, 0.0208, 0.0358], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:32:09,553 INFO [train.py:904] (4/8) Epoch 13, batch 800, loss[loss=0.1963, simple_loss=0.2681, pruned_loss=0.0622, over 12231.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2604, pruned_loss=0.04755, over 3260194.95 frames. ], batch size: 246, lr: 5.34e-03, grad_scale: 8.0 2023-04-29 17:32:15,181 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122606.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:33:15,730 INFO [train.py:904] (4/8) Epoch 13, batch 850, loss[loss=0.1748, simple_loss=0.2764, pruned_loss=0.03666, over 17148.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2603, pruned_loss=0.04756, over 3272989.72 frames. ], batch size: 49, lr: 5.34e-03, grad_scale: 8.0 2023-04-29 17:33:18,234 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=122654.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:33:23,686 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9719, 4.2158, 2.6683, 4.7138, 3.3074, 4.6949, 2.6090, 3.2956], device='cuda:4'), covar=tensor([0.0258, 0.0301, 0.1308, 0.0227, 0.0658, 0.0459, 0.1328, 0.0660], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0164, 0.0188, 0.0136, 0.0168, 0.0206, 0.0196, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 17:33:44,224 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.660e+02 2.256e+02 2.755e+02 3.448e+02 5.006e+02, threshold=5.510e+02, percent-clipped=0.0 2023-04-29 17:33:46,988 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122675.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:34:23,500 INFO [train.py:904] (4/8) Epoch 13, batch 900, loss[loss=0.1644, simple_loss=0.2571, pruned_loss=0.03579, over 17110.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.259, pruned_loss=0.04631, over 3284459.97 frames. ], batch size: 49, lr: 5.34e-03, grad_scale: 8.0 2023-04-29 17:35:33,154 INFO [train.py:904] (4/8) Epoch 13, batch 950, loss[loss=0.1823, simple_loss=0.2751, pruned_loss=0.04476, over 16682.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2586, pruned_loss=0.04595, over 3288465.13 frames. ], batch size: 57, lr: 5.34e-03, grad_scale: 8.0 2023-04-29 17:36:02,666 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.570e+02 2.301e+02 2.627e+02 3.023e+02 6.278e+02, threshold=5.253e+02, percent-clipped=3.0 2023-04-29 17:36:03,707 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122773.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 17:36:40,537 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1961, 5.2581, 5.7495, 5.7169, 5.7220, 5.3703, 5.2917, 5.1657], device='cuda:4'), covar=tensor([0.0270, 0.0455, 0.0295, 0.0368, 0.0378, 0.0311, 0.0814, 0.0352], device='cuda:4'), in_proj_covar=tensor([0.0355, 0.0374, 0.0369, 0.0354, 0.0421, 0.0398, 0.0494, 0.0321], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 17:36:43,312 INFO [train.py:904] (4/8) Epoch 13, batch 1000, loss[loss=0.1874, simple_loss=0.2552, pruned_loss=0.05977, over 16467.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.258, pruned_loss=0.04612, over 3292276.73 frames. ], batch size: 146, lr: 5.34e-03, grad_scale: 8.0 2023-04-29 17:37:09,710 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=122821.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:37:14,624 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122824.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:37:46,848 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-29 17:37:51,960 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122851.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:37:52,802 INFO [train.py:904] (4/8) Epoch 13, batch 1050, loss[loss=0.1763, simple_loss=0.2604, pruned_loss=0.04611, over 16357.00 frames. ], tot_loss[loss=0.1757, simple_loss=0.2582, pruned_loss=0.04662, over 3308497.64 frames. ], batch size: 68, lr: 5.34e-03, grad_scale: 8.0 2023-04-29 17:38:20,513 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122872.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:38:21,177 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.558e+02 2.183e+02 2.662e+02 3.059e+02 5.432e+02, threshold=5.323e+02, percent-clipped=2.0 2023-04-29 17:38:32,643 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7423, 3.5974, 3.9876, 3.0198, 3.6365, 4.0089, 3.7590, 2.2590], device='cuda:4'), covar=tensor([0.0401, 0.0224, 0.0043, 0.0267, 0.0077, 0.0076, 0.0064, 0.0413], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0073, 0.0072, 0.0129, 0.0084, 0.0092, 0.0082, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 17:38:58,266 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=122899.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:39:01,457 INFO [train.py:904] (4/8) Epoch 13, batch 1100, loss[loss=0.1763, simple_loss=0.2563, pruned_loss=0.04817, over 16820.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2575, pruned_loss=0.04612, over 3310909.05 frames. ], batch size: 102, lr: 5.33e-03, grad_scale: 8.0 2023-04-29 17:39:45,833 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122933.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:40:11,764 INFO [train.py:904] (4/8) Epoch 13, batch 1150, loss[loss=0.1697, simple_loss=0.2561, pruned_loss=0.04166, over 17119.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2565, pruned_loss=0.04578, over 3309981.44 frames. ], batch size: 47, lr: 5.33e-03, grad_scale: 8.0 2023-04-29 17:40:39,511 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.467e+02 2.197e+02 2.560e+02 3.227e+02 9.975e+02, threshold=5.120e+02, percent-clipped=3.0 2023-04-29 17:40:43,377 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122975.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:41:20,385 INFO [train.py:904] (4/8) Epoch 13, batch 1200, loss[loss=0.1819, simple_loss=0.2519, pruned_loss=0.05597, over 16884.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2557, pruned_loss=0.04513, over 3322836.68 frames. ], batch size: 116, lr: 5.33e-03, grad_scale: 8.0 2023-04-29 17:41:47,870 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8084, 4.7827, 4.6761, 4.1714, 4.7282, 1.9465, 4.4857, 4.4718], device='cuda:4'), covar=tensor([0.0113, 0.0080, 0.0151, 0.0286, 0.0092, 0.2158, 0.0114, 0.0171], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0127, 0.0172, 0.0161, 0.0145, 0.0188, 0.0161, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:41:50,192 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=123023.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:42:17,211 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7684, 3.1703, 2.6428, 4.3657, 3.6147, 4.2305, 1.6621, 2.9473], device='cuda:4'), covar=tensor([0.1492, 0.0555, 0.1055, 0.0181, 0.0219, 0.0375, 0.1521, 0.0774], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0160, 0.0183, 0.0152, 0.0194, 0.0210, 0.0183, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 17:42:30,138 INFO [train.py:904] (4/8) Epoch 13, batch 1250, loss[loss=0.1728, simple_loss=0.2474, pruned_loss=0.04911, over 16749.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.256, pruned_loss=0.04623, over 3321835.17 frames. ], batch size: 83, lr: 5.33e-03, grad_scale: 8.0 2023-04-29 17:42:34,865 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7946, 2.7168, 2.3299, 2.5562, 3.0839, 2.8972, 3.5698, 3.3321], device='cuda:4'), covar=tensor([0.0099, 0.0321, 0.0390, 0.0368, 0.0214, 0.0308, 0.0172, 0.0219], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0219, 0.0211, 0.0210, 0.0218, 0.0215, 0.0222, 0.0205], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:42:59,814 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.819e+02 2.369e+02 2.845e+02 3.353e+02 6.021e+02, threshold=5.690e+02, percent-clipped=1.0 2023-04-29 17:43:02,470 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-29 17:43:40,470 INFO [train.py:904] (4/8) Epoch 13, batch 1300, loss[loss=0.1919, simple_loss=0.2825, pruned_loss=0.05066, over 16699.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.256, pruned_loss=0.04571, over 3320033.90 frames. ], batch size: 57, lr: 5.33e-03, grad_scale: 8.0 2023-04-29 17:44:12,382 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1713, 4.3208, 4.1078, 3.9370, 3.4324, 4.3883, 4.1245, 3.9527], device='cuda:4'), covar=tensor([0.1053, 0.0784, 0.0551, 0.0422, 0.1626, 0.0510, 0.0755, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0355, 0.0316, 0.0295, 0.0336, 0.0337, 0.0212, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:44:12,392 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=123124.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:44:25,309 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7300, 4.7059, 4.5836, 3.4609, 4.6529, 1.9615, 4.3327, 4.2805], device='cuda:4'), covar=tensor([0.0149, 0.0127, 0.0216, 0.0658, 0.0130, 0.2796, 0.0210, 0.0313], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0128, 0.0174, 0.0162, 0.0146, 0.0189, 0.0163, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:44:49,681 INFO [train.py:904] (4/8) Epoch 13, batch 1350, loss[loss=0.1518, simple_loss=0.2389, pruned_loss=0.03242, over 16834.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2565, pruned_loss=0.04593, over 3320805.93 frames. ], batch size: 42, lr: 5.33e-03, grad_scale: 8.0 2023-04-29 17:45:17,072 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=123172.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:45:17,975 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.686e+02 2.284e+02 2.773e+02 3.220e+02 7.351e+02, threshold=5.547e+02, percent-clipped=2.0 2023-04-29 17:45:58,577 INFO [train.py:904] (4/8) Epoch 13, batch 1400, loss[loss=0.1647, simple_loss=0.2546, pruned_loss=0.03742, over 17232.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2564, pruned_loss=0.04647, over 3309487.56 frames. ], batch size: 45, lr: 5.33e-03, grad_scale: 8.0 2023-04-29 17:46:35,859 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=123228.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:47:09,270 INFO [train.py:904] (4/8) Epoch 13, batch 1450, loss[loss=0.1728, simple_loss=0.2684, pruned_loss=0.03859, over 17055.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.256, pruned_loss=0.04579, over 3310877.86 frames. ], batch size: 50, lr: 5.33e-03, grad_scale: 8.0 2023-04-29 17:47:38,907 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.758e+02 2.301e+02 2.596e+02 3.249e+02 6.793e+02, threshold=5.192e+02, percent-clipped=2.0 2023-04-29 17:47:59,627 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1618, 4.8947, 5.1644, 5.3657, 5.5849, 4.7869, 5.5345, 5.5077], device='cuda:4'), covar=tensor([0.1534, 0.1184, 0.1615, 0.0658, 0.0439, 0.0641, 0.0391, 0.0528], device='cuda:4'), in_proj_covar=tensor([0.0571, 0.0716, 0.0865, 0.0723, 0.0546, 0.0567, 0.0578, 0.0666], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:48:02,074 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9287, 5.3142, 5.0348, 5.0624, 4.7835, 4.7275, 4.7700, 5.3822], device='cuda:4'), covar=tensor([0.1151, 0.0840, 0.1043, 0.0712, 0.0740, 0.0978, 0.1003, 0.0921], device='cuda:4'), in_proj_covar=tensor([0.0576, 0.0731, 0.0589, 0.0517, 0.0461, 0.0471, 0.0608, 0.0560], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:48:19,759 INFO [train.py:904] (4/8) Epoch 13, batch 1500, loss[loss=0.1509, simple_loss=0.2349, pruned_loss=0.03342, over 17017.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2561, pruned_loss=0.0466, over 3312162.46 frames. ], batch size: 41, lr: 5.33e-03, grad_scale: 4.0 2023-04-29 17:48:37,122 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=123314.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:48:37,211 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9926, 2.6119, 1.9924, 2.3959, 2.9804, 2.8289, 3.1655, 3.0762], device='cuda:4'), covar=tensor([0.0132, 0.0280, 0.0394, 0.0301, 0.0157, 0.0220, 0.0153, 0.0185], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0218, 0.0209, 0.0209, 0.0218, 0.0216, 0.0223, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:49:30,722 INFO [train.py:904] (4/8) Epoch 13, batch 1550, loss[loss=0.1406, simple_loss=0.2191, pruned_loss=0.03107, over 16789.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2579, pruned_loss=0.04766, over 3301766.62 frames. ], batch size: 39, lr: 5.32e-03, grad_scale: 4.0 2023-04-29 17:49:41,559 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0844, 4.0044, 4.4650, 2.0455, 4.6765, 4.7091, 3.3062, 3.7373], device='cuda:4'), covar=tensor([0.0681, 0.0217, 0.0190, 0.1220, 0.0050, 0.0112, 0.0386, 0.0350], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0102, 0.0091, 0.0139, 0.0070, 0.0111, 0.0122, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 17:50:00,256 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.779e+02 2.426e+02 2.890e+02 3.629e+02 7.930e+02, threshold=5.780e+02, percent-clipped=5.0 2023-04-29 17:50:01,777 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=123375.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:50:31,263 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5062, 1.7278, 2.1830, 2.3526, 2.4751, 2.3542, 1.7979, 2.6542], device='cuda:4'), covar=tensor([0.0148, 0.0335, 0.0236, 0.0201, 0.0213, 0.0211, 0.0358, 0.0094], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0176, 0.0161, 0.0166, 0.0175, 0.0129, 0.0176, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 17:50:39,391 INFO [train.py:904] (4/8) Epoch 13, batch 1600, loss[loss=0.1531, simple_loss=0.2384, pruned_loss=0.03389, over 15956.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2595, pruned_loss=0.0487, over 3309686.33 frames. ], batch size: 35, lr: 5.32e-03, grad_scale: 8.0 2023-04-29 17:51:15,077 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=123428.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:51:43,708 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9036, 4.2913, 3.3953, 2.3487, 2.7934, 2.5085, 4.6328, 3.8215], device='cuda:4'), covar=tensor([0.2537, 0.0603, 0.1372, 0.2454, 0.2716, 0.1835, 0.0357, 0.1116], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0260, 0.0287, 0.0281, 0.0278, 0.0226, 0.0271, 0.0301], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:51:47,304 INFO [train.py:904] (4/8) Epoch 13, batch 1650, loss[loss=0.1714, simple_loss=0.2507, pruned_loss=0.04607, over 15886.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2617, pruned_loss=0.04918, over 3316096.68 frames. ], batch size: 35, lr: 5.32e-03, grad_scale: 8.0 2023-04-29 17:52:18,026 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.667e+02 2.454e+02 2.938e+02 3.518e+02 6.758e+02, threshold=5.875e+02, percent-clipped=2.0 2023-04-29 17:52:38,875 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=123489.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:52:55,813 INFO [train.py:904] (4/8) Epoch 13, batch 1700, loss[loss=0.1703, simple_loss=0.2624, pruned_loss=0.03906, over 15946.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2639, pruned_loss=0.04983, over 3319318.00 frames. ], batch size: 35, lr: 5.32e-03, grad_scale: 8.0 2023-04-29 17:53:31,964 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=123528.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:54:04,797 INFO [train.py:904] (4/8) Epoch 13, batch 1750, loss[loss=0.1628, simple_loss=0.2417, pruned_loss=0.04192, over 16763.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2652, pruned_loss=0.05017, over 3306647.77 frames. ], batch size: 39, lr: 5.32e-03, grad_scale: 8.0 2023-04-29 17:54:34,129 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.793e+02 2.358e+02 2.760e+02 3.262e+02 5.842e+02, threshold=5.520e+02, percent-clipped=0.0 2023-04-29 17:54:37,987 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=123576.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:54:40,364 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.81 vs. limit=5.0 2023-04-29 17:55:14,401 INFO [train.py:904] (4/8) Epoch 13, batch 1800, loss[loss=0.1953, simple_loss=0.2847, pruned_loss=0.05298, over 17114.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2659, pruned_loss=0.04951, over 3318214.34 frames. ], batch size: 48, lr: 5.32e-03, grad_scale: 8.0 2023-04-29 17:55:53,099 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7953, 2.5911, 2.1342, 2.4898, 2.9828, 2.7903, 3.4782, 3.1677], device='cuda:4'), covar=tensor([0.0081, 0.0352, 0.0405, 0.0346, 0.0219, 0.0276, 0.0166, 0.0198], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0216, 0.0208, 0.0207, 0.0216, 0.0215, 0.0222, 0.0205], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:56:23,364 INFO [train.py:904] (4/8) Epoch 13, batch 1850, loss[loss=0.2057, simple_loss=0.28, pruned_loss=0.06564, over 16171.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2663, pruned_loss=0.0493, over 3322939.93 frames. ], batch size: 165, lr: 5.32e-03, grad_scale: 8.0 2023-04-29 17:56:47,526 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=123670.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:56:52,511 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.694e+02 2.459e+02 2.923e+02 3.563e+02 7.345e+02, threshold=5.846e+02, percent-clipped=2.0 2023-04-29 17:57:33,301 INFO [train.py:904] (4/8) Epoch 13, batch 1900, loss[loss=0.1686, simple_loss=0.248, pruned_loss=0.04463, over 16826.00 frames. ], tot_loss[loss=0.181, simple_loss=0.265, pruned_loss=0.04855, over 3317992.42 frames. ], batch size: 124, lr: 5.32e-03, grad_scale: 8.0 2023-04-29 17:57:45,204 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-29 17:58:17,686 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7763, 1.8573, 2.3012, 2.6970, 2.6615, 2.6897, 1.8078, 2.9548], device='cuda:4'), covar=tensor([0.0145, 0.0349, 0.0246, 0.0181, 0.0219, 0.0202, 0.0379, 0.0094], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0176, 0.0161, 0.0166, 0.0176, 0.0129, 0.0177, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:58:35,072 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7004, 3.9068, 2.9668, 2.2379, 2.5897, 2.2876, 3.9120, 3.5279], device='cuda:4'), covar=tensor([0.2384, 0.0572, 0.1540, 0.2598, 0.2356, 0.1828, 0.0499, 0.1132], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0263, 0.0290, 0.0284, 0.0283, 0.0228, 0.0273, 0.0306], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 17:58:39,329 INFO [train.py:904] (4/8) Epoch 13, batch 1950, loss[loss=0.1803, simple_loss=0.2575, pruned_loss=0.05159, over 16902.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2644, pruned_loss=0.048, over 3326502.55 frames. ], batch size: 96, lr: 5.32e-03, grad_scale: 8.0 2023-04-29 17:58:40,958 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7750, 3.7389, 4.2521, 2.0668, 4.3610, 4.4113, 3.0775, 3.3372], device='cuda:4'), covar=tensor([0.0725, 0.0212, 0.0179, 0.1145, 0.0076, 0.0169, 0.0437, 0.0411], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0102, 0.0091, 0.0139, 0.0071, 0.0111, 0.0122, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 17:59:09,937 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.475e+02 2.247e+02 2.704e+02 3.281e+02 7.395e+02, threshold=5.408e+02, percent-clipped=2.0 2023-04-29 17:59:25,303 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=123784.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 17:59:48,916 INFO [train.py:904] (4/8) Epoch 13, batch 2000, loss[loss=0.2166, simple_loss=0.2892, pruned_loss=0.07203, over 16888.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2651, pruned_loss=0.04783, over 3317532.53 frames. ], batch size: 116, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 17:59:52,894 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1146, 5.0520, 4.8828, 4.3793, 5.0014, 1.9529, 4.6892, 4.8428], device='cuda:4'), covar=tensor([0.0080, 0.0085, 0.0163, 0.0338, 0.0083, 0.2353, 0.0130, 0.0174], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0131, 0.0178, 0.0166, 0.0149, 0.0192, 0.0167, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:00:34,178 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.55 vs. limit=2.0 2023-04-29 18:00:54,544 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7367, 3.9227, 2.9972, 2.2761, 2.6940, 2.4019, 3.9806, 3.5374], device='cuda:4'), covar=tensor([0.2232, 0.0560, 0.1392, 0.2540, 0.2058, 0.1680, 0.0447, 0.1113], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0262, 0.0288, 0.0283, 0.0282, 0.0228, 0.0272, 0.0306], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:00:59,626 INFO [train.py:904] (4/8) Epoch 13, batch 2050, loss[loss=0.2106, simple_loss=0.2834, pruned_loss=0.06892, over 16329.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2652, pruned_loss=0.04829, over 3309331.85 frames. ], batch size: 165, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:01:28,750 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=123872.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:01:30,782 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.552e+02 2.367e+02 2.767e+02 3.495e+02 6.657e+02, threshold=5.534e+02, percent-clipped=3.0 2023-04-29 18:01:39,522 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3579, 4.0889, 4.5845, 2.3348, 4.7445, 4.8374, 3.4618, 3.7841], device='cuda:4'), covar=tensor([0.0554, 0.0177, 0.0157, 0.1053, 0.0052, 0.0125, 0.0359, 0.0328], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0101, 0.0090, 0.0138, 0.0070, 0.0111, 0.0121, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 18:02:09,973 INFO [train.py:904] (4/8) Epoch 13, batch 2100, loss[loss=0.2507, simple_loss=0.3235, pruned_loss=0.08896, over 12287.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2667, pruned_loss=0.04935, over 3313486.21 frames. ], batch size: 247, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:02:54,422 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=123933.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:03:14,048 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.87 vs. limit=5.0 2023-04-29 18:03:20,328 INFO [train.py:904] (4/8) Epoch 13, batch 2150, loss[loss=0.1957, simple_loss=0.2897, pruned_loss=0.05088, over 17082.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2672, pruned_loss=0.04947, over 3319047.92 frames. ], batch size: 49, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:03:38,089 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-29 18:03:44,924 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=123970.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:03:50,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.498e+02 2.337e+02 2.985e+02 3.427e+02 6.976e+02, threshold=5.971e+02, percent-clipped=3.0 2023-04-29 18:04:14,061 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3631, 1.6205, 2.0339, 2.1814, 2.4099, 2.3372, 1.6473, 2.4515], device='cuda:4'), covar=tensor([0.0155, 0.0373, 0.0238, 0.0216, 0.0204, 0.0202, 0.0380, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0177, 0.0161, 0.0166, 0.0176, 0.0130, 0.0176, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:04:30,742 INFO [train.py:904] (4/8) Epoch 13, batch 2200, loss[loss=0.1877, simple_loss=0.2714, pruned_loss=0.05195, over 16311.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2681, pruned_loss=0.05001, over 3301292.04 frames. ], batch size: 165, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:04:53,398 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=124018.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:04:54,853 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4568, 2.2330, 2.3400, 4.1960, 2.1700, 2.6196, 2.3147, 2.3954], device='cuda:4'), covar=tensor([0.1151, 0.3404, 0.2373, 0.0532, 0.3715, 0.2313, 0.3222, 0.3081], device='cuda:4'), in_proj_covar=tensor([0.0373, 0.0405, 0.0340, 0.0326, 0.0417, 0.0466, 0.0370, 0.0474], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:05:01,616 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0164, 2.0024, 2.1434, 3.5072, 2.0485, 2.2968, 2.1351, 2.1319], device='cuda:4'), covar=tensor([0.1144, 0.3209, 0.2425, 0.0584, 0.3487, 0.2309, 0.3184, 0.3100], device='cuda:4'), in_proj_covar=tensor([0.0373, 0.0405, 0.0340, 0.0326, 0.0417, 0.0466, 0.0370, 0.0474], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:05:19,589 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124036.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:05:28,736 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8942, 5.2393, 4.9996, 4.9620, 4.7042, 4.6474, 4.6565, 5.3164], device='cuda:4'), covar=tensor([0.1092, 0.0757, 0.0955, 0.0761, 0.0793, 0.0985, 0.1102, 0.0806], device='cuda:4'), in_proj_covar=tensor([0.0579, 0.0733, 0.0588, 0.0519, 0.0463, 0.0468, 0.0609, 0.0558], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:05:40,212 INFO [train.py:904] (4/8) Epoch 13, batch 2250, loss[loss=0.1821, simple_loss=0.2696, pruned_loss=0.04727, over 17123.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2681, pruned_loss=0.04981, over 3311099.28 frames. ], batch size: 47, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:05:53,614 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124062.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:06:09,553 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.722e+02 2.445e+02 2.929e+02 3.478e+02 7.502e+02, threshold=5.857e+02, percent-clipped=2.0 2023-04-29 18:06:22,671 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124084.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:06:42,885 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124097.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:06:46,210 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8248, 4.0365, 2.2522, 4.5277, 2.8410, 4.5109, 2.3432, 3.2191], device='cuda:4'), covar=tensor([0.0252, 0.0305, 0.1548, 0.0235, 0.0834, 0.0519, 0.1449, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0169, 0.0191, 0.0146, 0.0172, 0.0214, 0.0199, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 18:06:48,019 INFO [train.py:904] (4/8) Epoch 13, batch 2300, loss[loss=0.1834, simple_loss=0.2594, pruned_loss=0.05375, over 16512.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2684, pruned_loss=0.04985, over 3309702.23 frames. ], batch size: 146, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:07:17,808 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124123.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:07:29,437 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=124132.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:07:56,749 INFO [train.py:904] (4/8) Epoch 13, batch 2350, loss[loss=0.2183, simple_loss=0.2901, pruned_loss=0.07323, over 16699.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.2687, pruned_loss=0.05042, over 3314425.59 frames. ], batch size: 134, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:08:26,417 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.689e+02 2.342e+02 2.726e+02 3.219e+02 8.801e+02, threshold=5.452e+02, percent-clipped=1.0 2023-04-29 18:09:06,167 INFO [train.py:904] (4/8) Epoch 13, batch 2400, loss[loss=0.1745, simple_loss=0.264, pruned_loss=0.04251, over 16972.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2697, pruned_loss=0.05082, over 3323505.00 frames. ], batch size: 41, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:09:42,796 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124228.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:10:15,688 INFO [train.py:904] (4/8) Epoch 13, batch 2450, loss[loss=0.208, simple_loss=0.2914, pruned_loss=0.0623, over 15483.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2694, pruned_loss=0.05017, over 3328907.20 frames. ], batch size: 190, lr: 5.31e-03, grad_scale: 8.0 2023-04-29 18:10:43,305 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124272.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:10:45,699 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-29 18:10:46,034 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.567e+02 2.400e+02 2.819e+02 3.354e+02 7.582e+02, threshold=5.638e+02, percent-clipped=2.0 2023-04-29 18:11:20,653 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-29 18:11:24,236 INFO [train.py:904] (4/8) Epoch 13, batch 2500, loss[loss=0.1917, simple_loss=0.2781, pruned_loss=0.05268, over 16677.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2692, pruned_loss=0.04968, over 3327748.09 frames. ], batch size: 62, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:12:03,373 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4167, 3.5035, 2.0387, 3.7727, 2.6353, 3.6892, 2.0696, 2.7947], device='cuda:4'), covar=tensor([0.0277, 0.0405, 0.1548, 0.0232, 0.0796, 0.0648, 0.1455, 0.0713], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0170, 0.0191, 0.0146, 0.0172, 0.0215, 0.0200, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 18:12:09,449 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124333.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:12:18,754 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3590, 5.2905, 5.2282, 4.7761, 4.7891, 5.2430, 5.2201, 4.8552], device='cuda:4'), covar=tensor([0.0633, 0.0496, 0.0266, 0.0295, 0.1102, 0.0446, 0.0268, 0.0768], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0367, 0.0324, 0.0303, 0.0346, 0.0346, 0.0219, 0.0379], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 18:12:31,506 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124348.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 18:12:37,093 INFO [train.py:904] (4/8) Epoch 13, batch 2550, loss[loss=0.1752, simple_loss=0.2645, pruned_loss=0.04296, over 17108.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2681, pruned_loss=0.04913, over 3327102.45 frames. ], batch size: 47, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:13:06,416 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.367e+02 2.275e+02 2.738e+02 3.268e+02 6.282e+02, threshold=5.476e+02, percent-clipped=1.0 2023-04-29 18:13:24,605 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.09 vs. limit=5.0 2023-04-29 18:13:30,678 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.40 vs. limit=5.0 2023-04-29 18:13:31,411 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124392.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:13:44,302 INFO [train.py:904] (4/8) Epoch 13, batch 2600, loss[loss=0.2266, simple_loss=0.315, pruned_loss=0.0691, over 16796.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.2677, pruned_loss=0.04877, over 3337058.66 frames. ], batch size: 57, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:13:54,858 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124409.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 18:14:06,950 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124418.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:14:26,642 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124432.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:14:54,847 INFO [train.py:904] (4/8) Epoch 13, batch 2650, loss[loss=0.1947, simple_loss=0.2778, pruned_loss=0.0558, over 16855.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2673, pruned_loss=0.04848, over 3337616.97 frames. ], batch size: 96, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:15:22,979 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1454, 3.7536, 3.7283, 2.2185, 2.9421, 2.5733, 3.6400, 3.8670], device='cuda:4'), covar=tensor([0.0300, 0.0689, 0.0551, 0.1715, 0.0853, 0.0883, 0.0650, 0.0945], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0152, 0.0162, 0.0147, 0.0138, 0.0127, 0.0139, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 18:15:25,988 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.623e+02 2.199e+02 2.542e+02 3.057e+02 5.216e+02, threshold=5.084e+02, percent-clipped=0.0 2023-04-29 18:15:52,740 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124493.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:16:04,556 INFO [train.py:904] (4/8) Epoch 13, batch 2700, loss[loss=0.1826, simple_loss=0.2638, pruned_loss=0.05071, over 16754.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2676, pruned_loss=0.048, over 3339681.13 frames. ], batch size: 89, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:16:40,360 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124528.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:16:43,851 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1490, 5.0632, 5.5664, 5.5739, 5.6660, 5.1518, 5.1710, 4.8637], device='cuda:4'), covar=tensor([0.0278, 0.0438, 0.0385, 0.0352, 0.0393, 0.0325, 0.0879, 0.0409], device='cuda:4'), in_proj_covar=tensor([0.0363, 0.0383, 0.0381, 0.0360, 0.0431, 0.0406, 0.0504, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 18:17:13,324 INFO [train.py:904] (4/8) Epoch 13, batch 2750, loss[loss=0.1403, simple_loss=0.2289, pruned_loss=0.0259, over 16953.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2675, pruned_loss=0.04742, over 3340943.57 frames. ], batch size: 41, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:17:25,356 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-29 18:17:44,014 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.514e+02 2.177e+02 2.567e+02 3.156e+02 4.821e+02, threshold=5.134e+02, percent-clipped=0.0 2023-04-29 18:17:46,591 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=124576.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:17:52,019 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8383, 4.6824, 4.8031, 5.0719, 5.1809, 4.6242, 5.2044, 5.1711], device='cuda:4'), covar=tensor([0.1601, 0.1183, 0.1919, 0.0874, 0.0726, 0.0856, 0.0736, 0.0749], device='cuda:4'), in_proj_covar=tensor([0.0582, 0.0724, 0.0876, 0.0734, 0.0551, 0.0582, 0.0584, 0.0676], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:18:22,959 INFO [train.py:904] (4/8) Epoch 13, batch 2800, loss[loss=0.1988, simple_loss=0.2757, pruned_loss=0.06095, over 16157.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2672, pruned_loss=0.04733, over 3342550.48 frames. ], batch size: 165, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:18:23,491 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8854, 2.2847, 2.3563, 4.7590, 2.3718, 2.8455, 2.4823, 2.4965], device='cuda:4'), covar=tensor([0.0994, 0.3748, 0.2633, 0.0337, 0.3986, 0.2488, 0.3305, 0.3734], device='cuda:4'), in_proj_covar=tensor([0.0374, 0.0406, 0.0339, 0.0325, 0.0417, 0.0468, 0.0369, 0.0476], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:18:45,939 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7407, 4.8039, 4.9874, 4.9142, 4.8074, 5.4938, 4.9972, 4.6842], device='cuda:4'), covar=tensor([0.1242, 0.2049, 0.1964, 0.2083, 0.3054, 0.1115, 0.1607, 0.2506], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0535, 0.0582, 0.0465, 0.0628, 0.0609, 0.0461, 0.0615], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 18:18:59,118 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124628.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:19:08,327 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.99 vs. limit=5.0 2023-04-29 18:19:12,843 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124638.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:19:31,048 INFO [train.py:904] (4/8) Epoch 13, batch 2850, loss[loss=0.1688, simple_loss=0.2653, pruned_loss=0.03611, over 17046.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2668, pruned_loss=0.0476, over 3341121.06 frames. ], batch size: 50, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:19:46,059 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124663.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:20:00,129 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.671e+02 2.347e+02 2.816e+02 3.672e+02 8.891e+02, threshold=5.632e+02, percent-clipped=4.0 2023-04-29 18:20:06,734 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2709, 4.2485, 4.6804, 4.6786, 4.7058, 4.3974, 4.4038, 4.2371], device='cuda:4'), covar=tensor([0.0326, 0.0616, 0.0365, 0.0418, 0.0451, 0.0390, 0.0838, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0365, 0.0384, 0.0384, 0.0363, 0.0433, 0.0408, 0.0506, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 18:20:24,580 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124692.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:20:34,525 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124699.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:20:37,524 INFO [train.py:904] (4/8) Epoch 13, batch 2900, loss[loss=0.1806, simple_loss=0.2671, pruned_loss=0.04711, over 17049.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2662, pruned_loss=0.04782, over 3339353.27 frames. ], batch size: 53, lr: 5.30e-03, grad_scale: 8.0 2023-04-29 18:20:40,109 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124704.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 18:20:58,954 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124718.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:21:07,437 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124724.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:21:29,838 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=124740.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:21:46,234 INFO [train.py:904] (4/8) Epoch 13, batch 2950, loss[loss=0.2138, simple_loss=0.2997, pruned_loss=0.06397, over 17082.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2658, pruned_loss=0.04817, over 3336618.06 frames. ], batch size: 53, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:22:05,796 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=124766.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:22:16,911 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.794e+02 2.554e+02 2.996e+02 3.501e+02 7.058e+02, threshold=5.993e+02, percent-clipped=3.0 2023-04-29 18:22:26,266 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8613, 3.0097, 2.7348, 4.5136, 3.7243, 4.2087, 1.6816, 3.1597], device='cuda:4'), covar=tensor([0.1222, 0.0566, 0.0993, 0.0175, 0.0265, 0.0374, 0.1355, 0.0685], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0185, 0.0159, 0.0200, 0.0211, 0.0185, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 18:22:35,742 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124788.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:22:55,199 INFO [train.py:904] (4/8) Epoch 13, batch 3000, loss[loss=0.1579, simple_loss=0.2457, pruned_loss=0.03502, over 17168.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2661, pruned_loss=0.04883, over 3327933.55 frames. ], batch size: 46, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:22:55,199 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 18:23:04,001 INFO [train.py:938] (4/8) Epoch 13, validation: loss=0.1391, simple_loss=0.2452, pruned_loss=0.01648, over 944034.00 frames. 2023-04-29 18:23:04,002 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 18:23:29,520 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124820.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:23:58,028 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-29 18:24:14,167 INFO [train.py:904] (4/8) Epoch 13, batch 3050, loss[loss=0.1544, simple_loss=0.2453, pruned_loss=0.03178, over 17246.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2651, pruned_loss=0.04811, over 3330079.51 frames. ], batch size: 45, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:24:32,141 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7107, 2.4138, 1.7943, 2.2073, 2.8040, 2.5684, 2.9048, 2.9095], device='cuda:4'), covar=tensor([0.0140, 0.0294, 0.0441, 0.0331, 0.0173, 0.0266, 0.0176, 0.0206], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0217, 0.0209, 0.0209, 0.0218, 0.0216, 0.0227, 0.0207], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:24:44,775 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.534e+02 2.296e+02 3.000e+02 3.565e+02 8.414e+02, threshold=5.999e+02, percent-clipped=2.0 2023-04-29 18:24:55,371 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124881.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:25:25,280 INFO [train.py:904] (4/8) Epoch 13, batch 3100, loss[loss=0.1706, simple_loss=0.2487, pruned_loss=0.04623, over 15882.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2645, pruned_loss=0.04855, over 3335266.24 frames. ], batch size: 35, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:26:01,299 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124928.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:26:30,540 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3700, 5.3463, 5.2158, 4.6881, 5.2437, 2.3010, 5.0065, 5.2307], device='cuda:4'), covar=tensor([0.0063, 0.0059, 0.0141, 0.0293, 0.0068, 0.1959, 0.0104, 0.0116], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0133, 0.0179, 0.0168, 0.0151, 0.0190, 0.0168, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:26:34,832 INFO [train.py:904] (4/8) Epoch 13, batch 3150, loss[loss=0.1506, simple_loss=0.2451, pruned_loss=0.02809, over 17256.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2632, pruned_loss=0.04819, over 3332125.69 frames. ], batch size: 52, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:27:05,836 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.452e+02 2.205e+02 2.626e+02 3.111e+02 5.476e+02, threshold=5.252e+02, percent-clipped=0.0 2023-04-29 18:27:09,072 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=124976.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:27:34,184 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124994.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:27:45,735 INFO [train.py:904] (4/8) Epoch 13, batch 3200, loss[loss=0.1589, simple_loss=0.2472, pruned_loss=0.03532, over 17116.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2631, pruned_loss=0.04806, over 3341109.01 frames. ], batch size: 47, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:27:49,063 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125004.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 18:28:10,531 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125019.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:28:56,247 INFO [train.py:904] (4/8) Epoch 13, batch 3250, loss[loss=0.1946, simple_loss=0.2722, pruned_loss=0.0585, over 16737.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2622, pruned_loss=0.04746, over 3345910.41 frames. ], batch size: 124, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:28:56,530 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=125052.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 18:29:15,386 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0234, 5.0615, 5.5345, 5.5327, 5.5026, 5.1576, 5.0815, 4.9100], device='cuda:4'), covar=tensor([0.0268, 0.0417, 0.0289, 0.0338, 0.0410, 0.0328, 0.0868, 0.0383], device='cuda:4'), in_proj_covar=tensor([0.0368, 0.0387, 0.0386, 0.0365, 0.0435, 0.0409, 0.0513, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 18:29:27,126 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.694e+02 2.287e+02 2.914e+02 3.406e+02 7.385e+02, threshold=5.827e+02, percent-clipped=6.0 2023-04-29 18:29:46,881 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125088.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:30:05,565 INFO [train.py:904] (4/8) Epoch 13, batch 3300, loss[loss=0.1866, simple_loss=0.2768, pruned_loss=0.04816, over 17069.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2633, pruned_loss=0.04782, over 3339186.85 frames. ], batch size: 55, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:30:42,295 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125128.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:30:53,994 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=125136.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:31:15,061 INFO [train.py:904] (4/8) Epoch 13, batch 3350, loss[loss=0.1479, simple_loss=0.2367, pruned_loss=0.02958, over 16955.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2651, pruned_loss=0.04839, over 3326550.02 frames. ], batch size: 41, lr: 5.29e-03, grad_scale: 8.0 2023-04-29 18:31:45,937 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.501e+02 2.375e+02 2.737e+02 3.421e+02 8.798e+02, threshold=5.473e+02, percent-clipped=2.0 2023-04-29 18:31:49,064 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125176.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:32:06,969 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125189.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:32:24,428 INFO [train.py:904] (4/8) Epoch 13, batch 3400, loss[loss=0.1938, simple_loss=0.2642, pruned_loss=0.06164, over 16859.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2657, pruned_loss=0.04867, over 3311118.91 frames. ], batch size: 116, lr: 5.29e-03, grad_scale: 4.0 2023-04-29 18:33:35,398 INFO [train.py:904] (4/8) Epoch 13, batch 3450, loss[loss=0.17, simple_loss=0.2683, pruned_loss=0.03585, over 17141.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2644, pruned_loss=0.04839, over 3307506.49 frames. ], batch size: 48, lr: 5.28e-03, grad_scale: 4.0 2023-04-29 18:34:07,279 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.453e+02 2.393e+02 2.670e+02 3.290e+02 7.203e+02, threshold=5.341e+02, percent-clipped=1.0 2023-04-29 18:34:35,537 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125294.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:34:39,530 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-29 18:34:46,233 INFO [train.py:904] (4/8) Epoch 13, batch 3500, loss[loss=0.1819, simple_loss=0.2773, pruned_loss=0.04332, over 17096.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2634, pruned_loss=0.04803, over 3308933.41 frames. ], batch size: 53, lr: 5.28e-03, grad_scale: 4.0 2023-04-29 18:35:09,900 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125319.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:35:43,532 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=125342.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:35:57,848 INFO [train.py:904] (4/8) Epoch 13, batch 3550, loss[loss=0.1677, simple_loss=0.2392, pruned_loss=0.04807, over 16858.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2628, pruned_loss=0.04818, over 3301264.41 frames. ], batch size: 96, lr: 5.28e-03, grad_scale: 4.0 2023-04-29 18:36:03,833 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125356.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:36:19,305 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=125367.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:36:20,520 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8846, 4.8133, 4.7675, 4.2703, 4.7971, 1.9747, 4.5557, 4.5722], device='cuda:4'), covar=tensor([0.0117, 0.0092, 0.0151, 0.0324, 0.0088, 0.2372, 0.0135, 0.0174], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0134, 0.0181, 0.0170, 0.0152, 0.0191, 0.0170, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:36:30,595 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.719e+02 2.251e+02 2.708e+02 3.279e+02 8.250e+02, threshold=5.415e+02, percent-clipped=4.0 2023-04-29 18:36:32,858 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125376.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:37:08,985 INFO [train.py:904] (4/8) Epoch 13, batch 3600, loss[loss=0.1866, simple_loss=0.2751, pruned_loss=0.04903, over 17040.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2621, pruned_loss=0.04733, over 3311891.77 frames. ], batch size: 55, lr: 5.28e-03, grad_scale: 8.0 2023-04-29 18:37:24,899 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-04-29 18:37:31,305 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125417.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 18:37:35,967 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125421.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:37:59,799 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125437.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:38:21,929 INFO [train.py:904] (4/8) Epoch 13, batch 3650, loss[loss=0.1567, simple_loss=0.232, pruned_loss=0.04072, over 16886.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2614, pruned_loss=0.04799, over 3305561.95 frames. ], batch size: 96, lr: 5.28e-03, grad_scale: 8.0 2023-04-29 18:38:57,386 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.634e+02 2.272e+02 2.828e+02 3.424e+02 6.489e+02, threshold=5.656e+02, percent-clipped=1.0 2023-04-29 18:38:59,572 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125476.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:39:07,755 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125482.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:39:11,350 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125484.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:39:24,882 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5001, 2.3558, 1.7178, 2.0362, 2.6521, 2.4461, 2.7244, 2.7395], device='cuda:4'), covar=tensor([0.0147, 0.0270, 0.0426, 0.0357, 0.0160, 0.0254, 0.0171, 0.0220], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0218, 0.0209, 0.0211, 0.0219, 0.0216, 0.0229, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:39:36,233 INFO [train.py:904] (4/8) Epoch 13, batch 3700, loss[loss=0.1959, simple_loss=0.2696, pruned_loss=0.06113, over 16369.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2602, pruned_loss=0.04927, over 3298886.78 frames. ], batch size: 146, lr: 5.28e-03, grad_scale: 8.0 2023-04-29 18:40:10,905 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=125524.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:40:51,173 INFO [train.py:904] (4/8) Epoch 13, batch 3750, loss[loss=0.1861, simple_loss=0.2683, pruned_loss=0.05191, over 16522.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2604, pruned_loss=0.0504, over 3296222.32 frames. ], batch size: 35, lr: 5.28e-03, grad_scale: 8.0 2023-04-29 18:40:56,541 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7977, 2.5272, 2.4801, 4.6837, 2.3533, 2.9889, 2.6011, 2.8017], device='cuda:4'), covar=tensor([0.0841, 0.2897, 0.2101, 0.0297, 0.3336, 0.2013, 0.2707, 0.2580], device='cuda:4'), in_proj_covar=tensor([0.0374, 0.0409, 0.0340, 0.0327, 0.0421, 0.0472, 0.0372, 0.0479], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:41:24,164 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.582e+02 2.222e+02 2.735e+02 3.211e+02 5.083e+02, threshold=5.471e+02, percent-clipped=0.0 2023-04-29 18:42:05,166 INFO [train.py:904] (4/8) Epoch 13, batch 3800, loss[loss=0.205, simple_loss=0.2832, pruned_loss=0.06343, over 12305.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2615, pruned_loss=0.05174, over 3293269.95 frames. ], batch size: 248, lr: 5.28e-03, grad_scale: 8.0 2023-04-29 18:42:07,412 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125603.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:42:58,375 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.72 vs. limit=5.0 2023-04-29 18:43:18,384 INFO [train.py:904] (4/8) Epoch 13, batch 3850, loss[loss=0.1814, simple_loss=0.255, pruned_loss=0.05392, over 16917.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2611, pruned_loss=0.05212, over 3285952.80 frames. ], batch size: 116, lr: 5.28e-03, grad_scale: 8.0 2023-04-29 18:43:35,062 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125664.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:43:50,560 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 2.444e+02 2.805e+02 3.444e+02 9.574e+02, threshold=5.610e+02, percent-clipped=2.0 2023-04-29 18:44:29,762 INFO [train.py:904] (4/8) Epoch 13, batch 3900, loss[loss=0.1919, simple_loss=0.2662, pruned_loss=0.05885, over 16905.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2611, pruned_loss=0.05313, over 3274575.98 frames. ], batch size: 116, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:44:45,107 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125712.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 18:44:47,112 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5176, 2.9155, 3.0531, 1.9352, 2.6880, 2.2203, 3.1580, 3.1208], device='cuda:4'), covar=tensor([0.0245, 0.0751, 0.0504, 0.1710, 0.0764, 0.0857, 0.0542, 0.0686], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0152, 0.0159, 0.0146, 0.0138, 0.0126, 0.0137, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 18:44:52,288 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.27 vs. limit=5.0 2023-04-29 18:45:03,280 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125725.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:45:14,745 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125732.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:45:42,338 INFO [train.py:904] (4/8) Epoch 13, batch 3950, loss[loss=0.1808, simple_loss=0.2542, pruned_loss=0.05366, over 16820.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.261, pruned_loss=0.05384, over 3273635.46 frames. ], batch size: 102, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:46:16,759 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.736e+02 2.289e+02 2.629e+02 3.210e+02 7.394e+02, threshold=5.259e+02, percent-clipped=3.0 2023-04-29 18:46:20,057 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125777.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:46:20,507 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.94 vs. limit=5.0 2023-04-29 18:46:28,223 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0467, 4.0230, 4.3951, 4.3707, 4.4080, 4.1090, 4.1691, 4.0197], device='cuda:4'), covar=tensor([0.0319, 0.0533, 0.0352, 0.0404, 0.0435, 0.0411, 0.0702, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0356, 0.0374, 0.0371, 0.0356, 0.0423, 0.0394, 0.0493, 0.0317], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 18:46:29,331 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125784.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:46:32,643 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125786.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:46:55,501 INFO [train.py:904] (4/8) Epoch 13, batch 4000, loss[loss=0.1995, simple_loss=0.2717, pruned_loss=0.06365, over 16708.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2615, pruned_loss=0.0549, over 3272764.95 frames. ], batch size: 134, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:47:36,718 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-29 18:47:37,206 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=125832.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:47:49,121 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6661, 2.4583, 2.3952, 4.0309, 3.1130, 3.9602, 1.5043, 2.8394], device='cuda:4'), covar=tensor([0.1322, 0.0793, 0.1203, 0.0145, 0.0245, 0.0353, 0.1541, 0.0819], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0184, 0.0160, 0.0200, 0.0210, 0.0185, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 18:48:05,371 INFO [train.py:904] (4/8) Epoch 13, batch 4050, loss[loss=0.1712, simple_loss=0.2572, pruned_loss=0.04263, over 16802.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2623, pruned_loss=0.05406, over 3267760.64 frames. ], batch size: 39, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:48:36,495 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 1.993e+02 2.274e+02 2.780e+02 5.444e+02, threshold=4.549e+02, percent-clipped=3.0 2023-04-29 18:48:49,464 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-29 18:49:15,115 INFO [train.py:904] (4/8) Epoch 13, batch 4100, loss[loss=0.2197, simple_loss=0.3005, pruned_loss=0.0694, over 16885.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2631, pruned_loss=0.05307, over 3277393.88 frames. ], batch size: 116, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:50:30,152 INFO [train.py:904] (4/8) Epoch 13, batch 4150, loss[loss=0.2025, simple_loss=0.2913, pruned_loss=0.05689, over 16558.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2704, pruned_loss=0.05564, over 3248540.62 frames. ], batch size: 68, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:50:42,629 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125959.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:51:05,850 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.529e+02 2.394e+02 2.865e+02 3.650e+02 6.775e+02, threshold=5.730e+02, percent-clipped=10.0 2023-04-29 18:51:39,644 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.65 vs. limit=5.0 2023-04-29 18:51:40,576 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125998.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:51:48,696 INFO [train.py:904] (4/8) Epoch 13, batch 4200, loss[loss=0.2138, simple_loss=0.3075, pruned_loss=0.06002, over 16729.00 frames. ], tot_loss[loss=0.1962, simple_loss=0.2773, pruned_loss=0.05758, over 3193018.63 frames. ], batch size: 134, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:52:04,529 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126012.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 18:52:08,850 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9223, 4.1738, 3.9875, 4.0709, 3.7564, 3.7191, 3.8186, 4.1521], device='cuda:4'), covar=tensor([0.1098, 0.0949, 0.0962, 0.0745, 0.0817, 0.1812, 0.0992, 0.1124], device='cuda:4'), in_proj_covar=tensor([0.0570, 0.0723, 0.0582, 0.0512, 0.0458, 0.0463, 0.0602, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:52:34,266 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126032.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:52:35,631 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5700, 4.8143, 4.6299, 4.6739, 4.3946, 4.2991, 4.2098, 4.8847], device='cuda:4'), covar=tensor([0.0875, 0.0731, 0.0812, 0.0642, 0.0741, 0.1227, 0.1038, 0.0777], device='cuda:4'), in_proj_covar=tensor([0.0569, 0.0722, 0.0581, 0.0510, 0.0457, 0.0463, 0.0600, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:53:02,529 INFO [train.py:904] (4/8) Epoch 13, batch 4250, loss[loss=0.1945, simple_loss=0.2875, pruned_loss=0.05072, over 17188.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2808, pruned_loss=0.05779, over 3167943.91 frames. ], batch size: 46, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:53:13,094 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=126059.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:53:14,081 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=126060.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:53:36,045 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.600e+02 2.365e+02 2.942e+02 3.644e+02 5.722e+02, threshold=5.884e+02, percent-clipped=0.0 2023-04-29 18:53:39,756 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126077.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:53:39,822 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2536, 3.2859, 3.5867, 1.5551, 3.7607, 3.7686, 2.8304, 2.8218], device='cuda:4'), covar=tensor([0.0829, 0.0220, 0.0189, 0.1249, 0.0055, 0.0123, 0.0411, 0.0456], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0101, 0.0088, 0.0136, 0.0069, 0.0109, 0.0121, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 18:53:43,861 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=126080.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:53:45,084 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=126081.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:54:16,795 INFO [train.py:904] (4/8) Epoch 13, batch 4300, loss[loss=0.2048, simple_loss=0.2892, pruned_loss=0.06024, over 16667.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2819, pruned_loss=0.05643, over 3185370.49 frames. ], batch size: 57, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:54:51,666 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=126125.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:55:00,445 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2905, 4.4173, 4.6121, 4.4151, 4.5264, 5.0004, 4.5590, 4.2382], device='cuda:4'), covar=tensor([0.1405, 0.1574, 0.1718, 0.1745, 0.2152, 0.0922, 0.1253, 0.2282], device='cuda:4'), in_proj_covar=tensor([0.0363, 0.0509, 0.0552, 0.0436, 0.0589, 0.0578, 0.0440, 0.0590], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 18:55:09,296 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-29 18:55:30,694 INFO [train.py:904] (4/8) Epoch 13, batch 4350, loss[loss=0.2017, simple_loss=0.2919, pruned_loss=0.05572, over 17040.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2851, pruned_loss=0.05736, over 3188680.41 frames. ], batch size: 55, lr: 5.27e-03, grad_scale: 8.0 2023-04-29 18:56:06,025 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.671e+02 2.441e+02 2.808e+02 3.338e+02 6.392e+02, threshold=5.616e+02, percent-clipped=2.0 2023-04-29 18:56:46,670 INFO [train.py:904] (4/8) Epoch 13, batch 4400, loss[loss=0.2128, simple_loss=0.295, pruned_loss=0.06529, over 15347.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2869, pruned_loss=0.05818, over 3174505.27 frames. ], batch size: 190, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 18:57:59,272 INFO [train.py:904] (4/8) Epoch 13, batch 4450, loss[loss=0.2295, simple_loss=0.3036, pruned_loss=0.07767, over 11798.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.2908, pruned_loss=0.05978, over 3169434.92 frames. ], batch size: 247, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 18:58:03,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8791, 4.6305, 4.6847, 5.0762, 5.2155, 4.6673, 5.2843, 5.2448], device='cuda:4'), covar=tensor([0.1536, 0.1230, 0.2009, 0.0756, 0.0690, 0.0810, 0.0603, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0546, 0.0689, 0.0822, 0.0692, 0.0522, 0.0545, 0.0550, 0.0634], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:58:10,158 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126259.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 18:58:33,039 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.490e+02 2.129e+02 2.519e+02 3.084e+02 5.887e+02, threshold=5.038e+02, percent-clipped=2.0 2023-04-29 18:59:14,448 INFO [train.py:904] (4/8) Epoch 13, batch 4500, loss[loss=0.1965, simple_loss=0.288, pruned_loss=0.0525, over 16737.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2914, pruned_loss=0.06029, over 3174750.76 frames. ], batch size: 124, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 18:59:19,066 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0149, 5.1097, 4.8627, 4.5391, 4.5172, 4.9866, 4.8142, 4.5975], device='cuda:4'), covar=tensor([0.0477, 0.0193, 0.0193, 0.0216, 0.0771, 0.0218, 0.0305, 0.0537], device='cuda:4'), in_proj_covar=tensor([0.0254, 0.0341, 0.0300, 0.0282, 0.0324, 0.0324, 0.0206, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 18:59:21,401 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=126307.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:00:25,581 INFO [train.py:904] (4/8) Epoch 13, batch 4550, loss[loss=0.2084, simple_loss=0.302, pruned_loss=0.05742, over 16742.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2919, pruned_loss=0.06054, over 3189128.75 frames. ], batch size: 89, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 19:00:28,844 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=126354.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:00:59,195 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.430e+02 2.007e+02 2.258e+02 2.674e+02 4.772e+02, threshold=4.515e+02, percent-clipped=0.0 2023-04-29 19:01:06,174 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=126380.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:01:07,314 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126381.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:01:35,442 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1011, 5.0841, 4.7847, 3.9118, 5.0163, 1.7304, 4.6993, 4.4298], device='cuda:4'), covar=tensor([0.0049, 0.0043, 0.0119, 0.0409, 0.0054, 0.2794, 0.0095, 0.0224], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0126, 0.0171, 0.0161, 0.0144, 0.0182, 0.0160, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:01:36,158 INFO [train.py:904] (4/8) Epoch 13, batch 4600, loss[loss=0.2076, simple_loss=0.291, pruned_loss=0.06204, over 16701.00 frames. ], tot_loss[loss=0.206, simple_loss=0.292, pruned_loss=0.06002, over 3202856.97 frames. ], batch size: 62, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 19:02:00,613 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7874, 3.3850, 3.2567, 5.2125, 4.1551, 4.4646, 2.1996, 3.2502], device='cuda:4'), covar=tensor([0.1294, 0.0597, 0.0922, 0.0111, 0.0367, 0.0320, 0.1292, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0161, 0.0183, 0.0157, 0.0198, 0.0206, 0.0182, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 19:02:22,725 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=126429.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:02:23,034 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3792, 2.3825, 2.3076, 4.4170, 2.2107, 2.8256, 2.4062, 2.4802], device='cuda:4'), covar=tensor([0.0999, 0.2792, 0.2156, 0.0320, 0.3513, 0.1955, 0.2671, 0.3032], device='cuda:4'), in_proj_covar=tensor([0.0372, 0.0405, 0.0336, 0.0319, 0.0419, 0.0470, 0.0369, 0.0473], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:02:26,959 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.79 vs. limit=5.0 2023-04-29 19:02:39,317 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=126441.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:02:53,229 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=126450.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:02:55,164 INFO [train.py:904] (4/8) Epoch 13, batch 4650, loss[loss=0.1772, simple_loss=0.2674, pruned_loss=0.04354, over 16908.00 frames. ], tot_loss[loss=0.205, simple_loss=0.2906, pruned_loss=0.05968, over 3203055.20 frames. ], batch size: 109, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 19:03:27,524 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.404e+02 1.885e+02 2.230e+02 2.796e+02 5.506e+02, threshold=4.460e+02, percent-clipped=0.0 2023-04-29 19:03:36,415 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-29 19:04:07,651 INFO [train.py:904] (4/8) Epoch 13, batch 4700, loss[loss=0.2257, simple_loss=0.2902, pruned_loss=0.08058, over 11896.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.288, pruned_loss=0.05883, over 3200168.92 frames. ], batch size: 246, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 19:04:21,534 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=126511.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:05:13,218 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.62 vs. limit=5.0 2023-04-29 19:05:21,227 INFO [train.py:904] (4/8) Epoch 13, batch 4750, loss[loss=0.1731, simple_loss=0.2586, pruned_loss=0.04382, over 16776.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.2835, pruned_loss=0.05658, over 3210465.89 frames. ], batch size: 124, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 19:05:53,852 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.457e+02 1.907e+02 2.465e+02 2.943e+02 4.218e+02, threshold=4.930e+02, percent-clipped=1.0 2023-04-29 19:06:32,282 INFO [train.py:904] (4/8) Epoch 13, batch 4800, loss[loss=0.158, simple_loss=0.2453, pruned_loss=0.03535, over 17238.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2795, pruned_loss=0.05439, over 3211585.30 frames. ], batch size: 52, lr: 5.26e-03, grad_scale: 8.0 2023-04-29 19:07:26,280 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-29 19:07:43,531 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9450, 3.1946, 2.8874, 5.3088, 4.2126, 4.5526, 2.0303, 3.3095], device='cuda:4'), covar=tensor([0.1283, 0.0692, 0.1115, 0.0099, 0.0374, 0.0347, 0.1360, 0.0804], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0185, 0.0159, 0.0201, 0.0208, 0.0184, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 19:07:46,138 INFO [train.py:904] (4/8) Epoch 13, batch 4850, loss[loss=0.1785, simple_loss=0.2747, pruned_loss=0.04113, over 16505.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.2809, pruned_loss=0.05426, over 3207253.58 frames. ], batch size: 75, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:07:50,102 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126654.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:08:19,805 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.279e+02 1.948e+02 2.386e+02 2.716e+02 6.913e+02, threshold=4.771e+02, percent-clipped=1.0 2023-04-29 19:08:59,235 INFO [train.py:904] (4/8) Epoch 13, batch 4900, loss[loss=0.2104, simple_loss=0.2881, pruned_loss=0.06629, over 12289.00 frames. ], tot_loss[loss=0.1931, simple_loss=0.2801, pruned_loss=0.05307, over 3191418.65 frames. ], batch size: 246, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:08:59,538 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=126702.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:09:20,063 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8901, 3.9232, 4.2709, 4.2423, 4.2132, 3.9758, 3.9559, 3.9295], device='cuda:4'), covar=tensor([0.0308, 0.0485, 0.0339, 0.0426, 0.0486, 0.0350, 0.0825, 0.0470], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0357, 0.0356, 0.0347, 0.0407, 0.0379, 0.0476, 0.0306], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 19:09:22,452 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1526, 3.6398, 3.5814, 1.9168, 2.9285, 2.4758, 3.5180, 3.7658], device='cuda:4'), covar=tensor([0.0236, 0.0637, 0.0550, 0.1922, 0.0806, 0.0842, 0.0632, 0.0762], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0150, 0.0159, 0.0145, 0.0138, 0.0125, 0.0138, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 19:09:49,773 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=126736.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:10:12,983 INFO [train.py:904] (4/8) Epoch 13, batch 4950, loss[loss=0.2007, simple_loss=0.2948, pruned_loss=0.05325, over 16844.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2805, pruned_loss=0.05275, over 3176743.62 frames. ], batch size: 96, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:10:38,184 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5224, 2.1490, 1.6645, 1.9226, 2.5144, 2.1166, 2.2996, 2.6592], device='cuda:4'), covar=tensor([0.0111, 0.0360, 0.0467, 0.0410, 0.0188, 0.0338, 0.0150, 0.0206], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0213, 0.0206, 0.0206, 0.0212, 0.0211, 0.0217, 0.0204], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:10:45,730 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.702e+02 2.239e+02 2.710e+02 3.273e+02 4.681e+02, threshold=5.421e+02, percent-clipped=0.0 2023-04-29 19:11:26,260 INFO [train.py:904] (4/8) Epoch 13, batch 5000, loss[loss=0.1891, simple_loss=0.2804, pruned_loss=0.04891, over 16828.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2823, pruned_loss=0.05287, over 3185309.48 frames. ], batch size: 116, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:11:32,726 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=126806.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:11:47,313 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=126816.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:12:38,196 INFO [train.py:904] (4/8) Epoch 13, batch 5050, loss[loss=0.1934, simple_loss=0.2824, pruned_loss=0.05224, over 15551.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2831, pruned_loss=0.05255, over 3198714.53 frames. ], batch size: 191, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:13:06,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7498, 3.7725, 2.1364, 4.3950, 2.8740, 4.2237, 2.3484, 2.9029], device='cuda:4'), covar=tensor([0.0191, 0.0300, 0.1556, 0.0089, 0.0680, 0.0365, 0.1358, 0.0703], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0165, 0.0188, 0.0136, 0.0167, 0.0207, 0.0194, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 19:13:11,123 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.565e+02 2.161e+02 2.561e+02 3.247e+02 4.827e+02, threshold=5.121e+02, percent-clipped=0.0 2023-04-29 19:13:14,614 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=126877.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:13:40,949 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7756, 3.8127, 3.9956, 3.7913, 3.8288, 4.2848, 4.0129, 3.6821], device='cuda:4'), covar=tensor([0.2105, 0.1776, 0.1663, 0.1954, 0.2642, 0.1564, 0.1336, 0.2483], device='cuda:4'), in_proj_covar=tensor([0.0357, 0.0499, 0.0533, 0.0423, 0.0577, 0.0569, 0.0432, 0.0580], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 19:13:47,914 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-29 19:13:51,464 INFO [train.py:904] (4/8) Epoch 13, batch 5100, loss[loss=0.1737, simple_loss=0.2645, pruned_loss=0.04139, over 17013.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2811, pruned_loss=0.05178, over 3191297.88 frames. ], batch size: 50, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:15:04,900 INFO [train.py:904] (4/8) Epoch 13, batch 5150, loss[loss=0.1817, simple_loss=0.2711, pruned_loss=0.04619, over 16505.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2806, pruned_loss=0.0506, over 3198778.64 frames. ], batch size: 68, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:15:14,499 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6748, 1.7468, 2.1877, 2.6277, 2.5977, 2.9032, 1.7822, 2.9473], device='cuda:4'), covar=tensor([0.0152, 0.0442, 0.0271, 0.0234, 0.0230, 0.0140, 0.0422, 0.0088], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0176, 0.0161, 0.0164, 0.0175, 0.0131, 0.0175, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 19:15:37,790 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.591e+02 2.071e+02 2.483e+02 2.841e+02 7.629e+02, threshold=4.965e+02, percent-clipped=2.0 2023-04-29 19:16:11,672 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8284, 3.6886, 3.9434, 3.6140, 3.8636, 4.2804, 3.9719, 3.6303], device='cuda:4'), covar=tensor([0.1827, 0.2183, 0.1646, 0.2273, 0.2322, 0.1548, 0.1223, 0.2381], device='cuda:4'), in_proj_covar=tensor([0.0353, 0.0493, 0.0527, 0.0419, 0.0571, 0.0565, 0.0427, 0.0575], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 19:16:17,737 INFO [train.py:904] (4/8) Epoch 13, batch 5200, loss[loss=0.2069, simple_loss=0.2913, pruned_loss=0.06125, over 12319.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2792, pruned_loss=0.05025, over 3202003.73 frames. ], batch size: 247, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:17:08,348 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=127036.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:17:31,265 INFO [train.py:904] (4/8) Epoch 13, batch 5250, loss[loss=0.1939, simple_loss=0.273, pruned_loss=0.05745, over 12481.00 frames. ], tot_loss[loss=0.188, simple_loss=0.2761, pruned_loss=0.04998, over 3217179.79 frames. ], batch size: 246, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:18:04,195 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.619e+02 2.082e+02 2.366e+02 2.761e+02 5.233e+02, threshold=4.733e+02, percent-clipped=1.0 2023-04-29 19:18:18,851 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=127084.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:18:44,223 INFO [train.py:904] (4/8) Epoch 13, batch 5300, loss[loss=0.1598, simple_loss=0.2402, pruned_loss=0.0397, over 16621.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2723, pruned_loss=0.04859, over 3216116.26 frames. ], batch size: 57, lr: 5.25e-03, grad_scale: 8.0 2023-04-29 19:18:50,592 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=127106.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:18:55,288 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-29 19:19:43,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3849, 2.1745, 2.2650, 4.2067, 2.0832, 2.5712, 2.2573, 2.3565], device='cuda:4'), covar=tensor([0.1048, 0.3276, 0.2470, 0.0383, 0.3732, 0.2363, 0.3212, 0.3000], device='cuda:4'), in_proj_covar=tensor([0.0368, 0.0402, 0.0335, 0.0318, 0.0415, 0.0466, 0.0368, 0.0468], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:19:58,317 INFO [train.py:904] (4/8) Epoch 13, batch 5350, loss[loss=0.1996, simple_loss=0.2901, pruned_loss=0.05456, over 16605.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.271, pruned_loss=0.04788, over 3211767.21 frames. ], batch size: 57, lr: 5.24e-03, grad_scale: 4.0 2023-04-29 19:20:01,842 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=127154.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:20:11,181 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1511, 2.0744, 2.1383, 3.7870, 2.0674, 2.4804, 2.1941, 2.2448], device='cuda:4'), covar=tensor([0.1115, 0.3273, 0.2442, 0.0441, 0.3573, 0.2238, 0.3215, 0.2950], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0401, 0.0334, 0.0317, 0.0413, 0.0464, 0.0367, 0.0466], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:20:27,694 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=127172.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:20:32,660 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.449e+02 2.172e+02 2.538e+02 2.960e+02 5.660e+02, threshold=5.076e+02, percent-clipped=1.0 2023-04-29 19:21:10,973 INFO [train.py:904] (4/8) Epoch 13, batch 5400, loss[loss=0.1791, simple_loss=0.2717, pruned_loss=0.04321, over 16973.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2734, pruned_loss=0.04864, over 3211140.36 frames. ], batch size: 41, lr: 5.24e-03, grad_scale: 4.0 2023-04-29 19:22:20,868 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127249.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:22:26,302 INFO [train.py:904] (4/8) Epoch 13, batch 5450, loss[loss=0.262, simple_loss=0.3235, pruned_loss=0.1003, over 12151.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2766, pruned_loss=0.05055, over 3203271.48 frames. ], batch size: 248, lr: 5.24e-03, grad_scale: 4.0 2023-04-29 19:22:31,058 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9408, 3.0709, 1.6022, 3.2135, 2.2159, 3.2300, 1.9092, 2.5086], device='cuda:4'), covar=tensor([0.0207, 0.0314, 0.1563, 0.0166, 0.0716, 0.0474, 0.1363, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0165, 0.0189, 0.0136, 0.0167, 0.0207, 0.0195, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 19:23:02,325 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.587e+02 2.258e+02 2.675e+02 3.738e+02 1.100e+03, threshold=5.349e+02, percent-clipped=10.0 2023-04-29 19:23:43,871 INFO [train.py:904] (4/8) Epoch 13, batch 5500, loss[loss=0.2281, simple_loss=0.3125, pruned_loss=0.07178, over 16383.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2856, pruned_loss=0.05629, over 3162679.01 frames. ], batch size: 146, lr: 5.24e-03, grad_scale: 4.0 2023-04-29 19:23:57,092 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127310.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:25:00,717 INFO [train.py:904] (4/8) Epoch 13, batch 5550, loss[loss=0.2316, simple_loss=0.3175, pruned_loss=0.07282, over 16277.00 frames. ], tot_loss[loss=0.2079, simple_loss=0.2929, pruned_loss=0.06152, over 3148966.01 frames. ], batch size: 165, lr: 5.24e-03, grad_scale: 4.0 2023-04-29 19:25:38,573 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.215e+02 3.416e+02 4.230e+02 5.078e+02 9.045e+02, threshold=8.460e+02, percent-clipped=18.0 2023-04-29 19:25:45,195 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3438, 3.2904, 3.3699, 3.4727, 3.4787, 3.2386, 3.4954, 3.5353], device='cuda:4'), covar=tensor([0.1112, 0.0865, 0.1063, 0.0577, 0.0703, 0.2247, 0.0939, 0.0756], device='cuda:4'), in_proj_covar=tensor([0.0546, 0.0688, 0.0821, 0.0691, 0.0524, 0.0542, 0.0553, 0.0634], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:25:55,157 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-29 19:26:20,875 INFO [train.py:904] (4/8) Epoch 13, batch 5600, loss[loss=0.2349, simple_loss=0.3218, pruned_loss=0.07407, over 16729.00 frames. ], tot_loss[loss=0.2152, simple_loss=0.2981, pruned_loss=0.06614, over 3116860.93 frames. ], batch size: 124, lr: 5.24e-03, grad_scale: 8.0 2023-04-29 19:26:43,381 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-29 19:27:02,234 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127427.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 19:27:41,559 INFO [train.py:904] (4/8) Epoch 13, batch 5650, loss[loss=0.229, simple_loss=0.3143, pruned_loss=0.07183, over 16756.00 frames. ], tot_loss[loss=0.2216, simple_loss=0.303, pruned_loss=0.0701, over 3074483.68 frames. ], batch size: 89, lr: 5.24e-03, grad_scale: 8.0 2023-04-29 19:27:55,512 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1013, 2.4114, 2.4154, 2.8675, 2.1355, 3.1953, 1.8654, 2.7205], device='cuda:4'), covar=tensor([0.0961, 0.0496, 0.0919, 0.0167, 0.0137, 0.0379, 0.1191, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0163, 0.0184, 0.0158, 0.0201, 0.0208, 0.0185, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 19:28:11,428 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-29 19:28:12,728 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=127472.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:28:17,740 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.064e+02 3.489e+02 4.637e+02 5.886e+02 1.352e+03, threshold=9.273e+02, percent-clipped=4.0 2023-04-29 19:28:36,978 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127488.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 19:28:58,366 INFO [train.py:904] (4/8) Epoch 13, batch 5700, loss[loss=0.2248, simple_loss=0.3048, pruned_loss=0.07237, over 16622.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.3044, pruned_loss=0.072, over 3071613.08 frames. ], batch size: 57, lr: 5.24e-03, grad_scale: 8.0 2023-04-29 19:29:27,678 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=127520.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:29:41,963 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127529.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:30:00,255 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-29 19:30:18,041 INFO [train.py:904] (4/8) Epoch 13, batch 5750, loss[loss=0.2308, simple_loss=0.315, pruned_loss=0.07331, over 17007.00 frames. ], tot_loss[loss=0.227, simple_loss=0.307, pruned_loss=0.07347, over 3056802.91 frames. ], batch size: 53, lr: 5.24e-03, grad_scale: 8.0 2023-04-29 19:30:56,348 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.152e+02 3.042e+02 3.668e+02 4.519e+02 8.025e+02, threshold=7.337e+02, percent-clipped=0.0 2023-04-29 19:31:21,785 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127590.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:31:39,674 INFO [train.py:904] (4/8) Epoch 13, batch 5800, loss[loss=0.2038, simple_loss=0.2922, pruned_loss=0.0577, over 15476.00 frames. ], tot_loss[loss=0.2264, simple_loss=0.3072, pruned_loss=0.07278, over 3045129.75 frames. ], batch size: 191, lr: 5.24e-03, grad_scale: 4.0 2023-04-29 19:31:44,371 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=127605.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:31:49,450 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8136, 2.5991, 2.6310, 1.9332, 2.4875, 2.7267, 2.6212, 1.8911], device='cuda:4'), covar=tensor([0.0376, 0.0072, 0.0061, 0.0325, 0.0101, 0.0087, 0.0088, 0.0348], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0070, 0.0071, 0.0128, 0.0084, 0.0092, 0.0082, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 19:32:57,618 INFO [train.py:904] (4/8) Epoch 13, batch 5850, loss[loss=0.2386, simple_loss=0.3064, pruned_loss=0.08543, over 11564.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.3054, pruned_loss=0.07153, over 3029573.81 frames. ], batch size: 248, lr: 5.23e-03, grad_scale: 4.0 2023-04-29 19:33:15,136 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127662.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:33:36,796 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.121e+02 3.118e+02 3.715e+02 4.396e+02 1.301e+03, threshold=7.431e+02, percent-clipped=3.0 2023-04-29 19:34:19,532 INFO [train.py:904] (4/8) Epoch 13, batch 5900, loss[loss=0.2119, simple_loss=0.303, pruned_loss=0.0604, over 16238.00 frames. ], tot_loss[loss=0.2236, simple_loss=0.3052, pruned_loss=0.07103, over 3051234.17 frames. ], batch size: 35, lr: 5.23e-03, grad_scale: 4.0 2023-04-29 19:34:30,712 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-29 19:34:57,895 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127723.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:35:36,319 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0664, 5.0313, 4.9066, 4.2220, 4.9818, 1.8732, 4.6844, 4.7267], device='cuda:4'), covar=tensor([0.0086, 0.0072, 0.0140, 0.0338, 0.0082, 0.2276, 0.0115, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0124, 0.0170, 0.0160, 0.0141, 0.0182, 0.0157, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:35:42,725 INFO [train.py:904] (4/8) Epoch 13, batch 5950, loss[loss=0.1942, simple_loss=0.285, pruned_loss=0.0517, over 16258.00 frames. ], tot_loss[loss=0.2216, simple_loss=0.3053, pruned_loss=0.06891, over 3070240.56 frames. ], batch size: 165, lr: 5.23e-03, grad_scale: 4.0 2023-04-29 19:36:16,772 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6239, 4.8133, 5.0168, 4.8467, 4.8704, 5.3933, 4.8690, 4.6492], device='cuda:4'), covar=tensor([0.1133, 0.1822, 0.2251, 0.1783, 0.2218, 0.0958, 0.1656, 0.2511], device='cuda:4'), in_proj_covar=tensor([0.0367, 0.0512, 0.0556, 0.0437, 0.0593, 0.0581, 0.0444, 0.0592], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 19:36:21,828 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.141e+02 2.934e+02 3.631e+02 4.257e+02 8.531e+02, threshold=7.262e+02, percent-clipped=1.0 2023-04-29 19:36:26,899 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127779.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:36:27,125 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-29 19:36:32,789 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=127783.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 19:36:41,763 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-29 19:37:04,075 INFO [train.py:904] (4/8) Epoch 13, batch 6000, loss[loss=0.2339, simple_loss=0.3016, pruned_loss=0.08309, over 11763.00 frames. ], tot_loss[loss=0.2207, simple_loss=0.3046, pruned_loss=0.06842, over 3084623.95 frames. ], batch size: 248, lr: 5.23e-03, grad_scale: 8.0 2023-04-29 19:37:04,075 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 19:37:14,215 INFO [train.py:938] (4/8) Epoch 13, validation: loss=0.1599, simple_loss=0.2726, pruned_loss=0.02359, over 944034.00 frames. 2023-04-29 19:37:14,215 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 19:38:12,165 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9853, 2.3783, 2.2963, 2.8859, 2.1497, 3.2573, 1.7429, 2.6748], device='cuda:4'), covar=tensor([0.1018, 0.0518, 0.0984, 0.0161, 0.0144, 0.0348, 0.1251, 0.0633], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0163, 0.0184, 0.0157, 0.0202, 0.0209, 0.0184, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 19:38:13,851 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127840.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:38:27,290 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4081, 2.1301, 1.6254, 1.8250, 2.4558, 2.1239, 2.3126, 2.6340], device='cuda:4'), covar=tensor([0.0179, 0.0380, 0.0498, 0.0465, 0.0212, 0.0364, 0.0193, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0210, 0.0204, 0.0206, 0.0209, 0.0208, 0.0215, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:38:32,249 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127850.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:38:35,159 INFO [train.py:904] (4/8) Epoch 13, batch 6050, loss[loss=0.2023, simple_loss=0.2929, pruned_loss=0.05584, over 15362.00 frames. ], tot_loss[loss=0.2199, simple_loss=0.3035, pruned_loss=0.0682, over 3087382.62 frames. ], batch size: 190, lr: 5.23e-03, grad_scale: 8.0 2023-04-29 19:38:36,632 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4773, 3.3799, 2.6703, 2.0680, 2.2749, 2.1153, 3.3873, 3.2233], device='cuda:4'), covar=tensor([0.2709, 0.0773, 0.1655, 0.2563, 0.2251, 0.2008, 0.0553, 0.1145], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0259, 0.0288, 0.0285, 0.0283, 0.0227, 0.0273, 0.0305], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:39:00,166 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5872, 3.8028, 2.9670, 2.0832, 2.5900, 2.3793, 3.9367, 3.5151], device='cuda:4'), covar=tensor([0.2845, 0.0637, 0.1512, 0.2683, 0.2459, 0.1803, 0.0467, 0.0969], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0259, 0.0287, 0.0285, 0.0283, 0.0226, 0.0272, 0.0305], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:39:07,452 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127872.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:39:14,937 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.848e+02 2.979e+02 3.574e+02 4.345e+02 1.015e+03, threshold=7.148e+02, percent-clipped=2.0 2023-04-29 19:39:28,467 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=127885.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:39:55,373 INFO [train.py:904] (4/8) Epoch 13, batch 6100, loss[loss=0.2064, simple_loss=0.292, pruned_loss=0.06042, over 15512.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.3023, pruned_loss=0.06697, over 3087060.47 frames. ], batch size: 191, lr: 5.23e-03, grad_scale: 8.0 2023-04-29 19:40:01,613 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=127905.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:40:12,559 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127911.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:40:47,052 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127933.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:41:10,085 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.86 vs. limit=5.0 2023-04-29 19:41:15,198 INFO [train.py:904] (4/8) Epoch 13, batch 6150, loss[loss=0.2073, simple_loss=0.2972, pruned_loss=0.05875, over 16233.00 frames. ], tot_loss[loss=0.2164, simple_loss=0.3002, pruned_loss=0.06625, over 3080746.14 frames. ], batch size: 165, lr: 5.23e-03, grad_scale: 8.0 2023-04-29 19:41:17,425 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=127953.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:41:39,469 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9279, 5.2234, 4.9734, 4.9894, 4.7118, 4.6421, 4.6841, 5.3129], device='cuda:4'), covar=tensor([0.0992, 0.0718, 0.0848, 0.0691, 0.0791, 0.0834, 0.1003, 0.0784], device='cuda:4'), in_proj_covar=tensor([0.0565, 0.0700, 0.0576, 0.0502, 0.0442, 0.0452, 0.0588, 0.0542], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:41:39,656 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8822, 3.3369, 3.2211, 1.9902, 2.7925, 2.2036, 3.5449, 3.4204], device='cuda:4'), covar=tensor([0.0236, 0.0680, 0.0597, 0.1875, 0.0834, 0.0939, 0.0607, 0.0936], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0150, 0.0160, 0.0146, 0.0140, 0.0126, 0.0139, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 19:41:56,814 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.069e+02 3.037e+02 3.583e+02 4.334e+02 1.010e+03, threshold=7.167e+02, percent-clipped=2.0 2023-04-29 19:42:39,406 INFO [train.py:904] (4/8) Epoch 13, batch 6200, loss[loss=0.2041, simple_loss=0.2946, pruned_loss=0.05679, over 16875.00 frames. ], tot_loss[loss=0.2152, simple_loss=0.2987, pruned_loss=0.06583, over 3086332.78 frames. ], batch size: 102, lr: 5.23e-03, grad_scale: 4.0 2023-04-29 19:43:05,194 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128018.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:43:26,175 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-29 19:43:57,666 INFO [train.py:904] (4/8) Epoch 13, batch 6250, loss[loss=0.2004, simple_loss=0.287, pruned_loss=0.05688, over 16845.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.298, pruned_loss=0.06547, over 3091102.41 frames. ], batch size: 116, lr: 5.23e-03, grad_scale: 4.0 2023-04-29 19:44:37,187 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.989e+02 2.817e+02 3.684e+02 4.282e+02 1.153e+03, threshold=7.367e+02, percent-clipped=4.0 2023-04-29 19:44:42,017 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0399, 2.5035, 2.6088, 1.8516, 2.7401, 2.8107, 2.4293, 2.3371], device='cuda:4'), covar=tensor([0.0642, 0.0189, 0.0177, 0.0883, 0.0084, 0.0189, 0.0380, 0.0397], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0101, 0.0088, 0.0137, 0.0068, 0.0108, 0.0120, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 19:44:45,210 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128083.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 19:45:13,032 INFO [train.py:904] (4/8) Epoch 13, batch 6300, loss[loss=0.1876, simple_loss=0.2799, pruned_loss=0.0477, over 16752.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2978, pruned_loss=0.06502, over 3096614.91 frames. ], batch size: 102, lr: 5.23e-03, grad_scale: 4.0 2023-04-29 19:45:34,439 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9829, 2.2835, 2.3435, 2.7424, 2.1644, 3.2225, 1.7326, 2.6796], device='cuda:4'), covar=tensor([0.1086, 0.0552, 0.0942, 0.0148, 0.0138, 0.0371, 0.1286, 0.0629], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0163, 0.0183, 0.0157, 0.0202, 0.0208, 0.0184, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 19:45:38,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5122, 3.5816, 2.7547, 2.1012, 2.3495, 2.2880, 3.7469, 3.3063], device='cuda:4'), covar=tensor([0.2770, 0.0689, 0.1694, 0.2490, 0.2496, 0.1873, 0.0448, 0.1132], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0260, 0.0289, 0.0287, 0.0284, 0.0228, 0.0273, 0.0306], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:46:02,725 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=128131.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 19:46:08,187 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128135.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:46:33,871 INFO [train.py:904] (4/8) Epoch 13, batch 6350, loss[loss=0.2272, simple_loss=0.3201, pruned_loss=0.06718, over 16680.00 frames. ], tot_loss[loss=0.2164, simple_loss=0.2991, pruned_loss=0.06684, over 3068709.81 frames. ], batch size: 134, lr: 5.22e-03, grad_scale: 4.0 2023-04-29 19:46:53,778 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8512, 4.8403, 5.3130, 5.2851, 5.2944, 4.9017, 4.9030, 4.5520], device='cuda:4'), covar=tensor([0.0298, 0.0504, 0.0338, 0.0391, 0.0452, 0.0337, 0.0938, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0351, 0.0368, 0.0368, 0.0352, 0.0417, 0.0389, 0.0487, 0.0315], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 19:47:13,675 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.951e+02 3.177e+02 3.833e+02 4.854e+02 7.993e+02, threshold=7.666e+02, percent-clipped=3.0 2023-04-29 19:47:25,239 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128185.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:47:49,923 INFO [train.py:904] (4/8) Epoch 13, batch 6400, loss[loss=0.2073, simple_loss=0.2941, pruned_loss=0.06028, over 16825.00 frames. ], tot_loss[loss=0.2177, simple_loss=0.2996, pruned_loss=0.06794, over 3065988.56 frames. ], batch size: 116, lr: 5.22e-03, grad_scale: 8.0 2023-04-29 19:47:52,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0151, 3.0406, 1.8371, 3.2771, 2.2762, 3.2847, 1.9911, 2.4813], device='cuda:4'), covar=tensor([0.0249, 0.0405, 0.1577, 0.0187, 0.0772, 0.0563, 0.1496, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0169, 0.0192, 0.0138, 0.0169, 0.0209, 0.0199, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 19:47:56,272 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128206.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:48:30,014 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128228.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:48:30,144 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0133, 2.5105, 2.6116, 1.9639, 2.7285, 2.7946, 2.4090, 2.3566], device='cuda:4'), covar=tensor([0.0705, 0.0202, 0.0195, 0.0870, 0.0089, 0.0207, 0.0383, 0.0439], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0101, 0.0088, 0.0137, 0.0069, 0.0109, 0.0121, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 19:48:36,342 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=128233.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:49:04,023 INFO [train.py:904] (4/8) Epoch 13, batch 6450, loss[loss=0.2186, simple_loss=0.3231, pruned_loss=0.05707, over 16383.00 frames. ], tot_loss[loss=0.2175, simple_loss=0.2998, pruned_loss=0.0676, over 3051599.20 frames. ], batch size: 35, lr: 5.22e-03, grad_scale: 2.0 2023-04-29 19:49:20,548 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4079, 5.6890, 5.4313, 5.4519, 5.1292, 5.0556, 5.1528, 5.7917], device='cuda:4'), covar=tensor([0.0892, 0.0756, 0.0838, 0.0692, 0.0836, 0.0682, 0.0914, 0.0872], device='cuda:4'), in_proj_covar=tensor([0.0562, 0.0695, 0.0573, 0.0502, 0.0439, 0.0450, 0.0585, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:49:48,882 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.913e+02 2.912e+02 3.353e+02 4.052e+02 7.402e+02, threshold=6.705e+02, percent-clipped=0.0 2023-04-29 19:50:21,886 INFO [train.py:904] (4/8) Epoch 13, batch 6500, loss[loss=0.2081, simple_loss=0.2882, pruned_loss=0.064, over 16477.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.2969, pruned_loss=0.06646, over 3042403.45 frames. ], batch size: 68, lr: 5.22e-03, grad_scale: 2.0 2023-04-29 19:50:45,302 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128318.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:50:48,313 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=128320.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:51:39,453 INFO [train.py:904] (4/8) Epoch 13, batch 6550, loss[loss=0.2537, simple_loss=0.3188, pruned_loss=0.09423, over 11413.00 frames. ], tot_loss[loss=0.2183, simple_loss=0.3003, pruned_loss=0.06818, over 3029570.18 frames. ], batch size: 247, lr: 5.22e-03, grad_scale: 2.0 2023-04-29 19:51:50,385 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-04-29 19:52:01,584 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=128366.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:52:22,965 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.909e+02 2.847e+02 3.449e+02 4.250e+02 8.173e+02, threshold=6.898e+02, percent-clipped=4.0 2023-04-29 19:52:25,029 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=128381.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:52:56,140 INFO [train.py:904] (4/8) Epoch 13, batch 6600, loss[loss=0.2834, simple_loss=0.3404, pruned_loss=0.1132, over 11638.00 frames. ], tot_loss[loss=0.219, simple_loss=0.302, pruned_loss=0.06801, over 3037145.78 frames. ], batch size: 248, lr: 5.22e-03, grad_scale: 2.0 2023-04-29 19:53:13,956 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=128413.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:53:48,788 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128435.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:54:01,565 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7639, 4.7166, 4.5372, 3.8899, 4.6408, 1.7593, 4.3817, 4.3754], device='cuda:4'), covar=tensor([0.0065, 0.0059, 0.0162, 0.0360, 0.0067, 0.2493, 0.0117, 0.0188], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0125, 0.0170, 0.0161, 0.0142, 0.0184, 0.0158, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:54:14,487 INFO [train.py:904] (4/8) Epoch 13, batch 6650, loss[loss=0.1871, simple_loss=0.2825, pruned_loss=0.04585, over 16378.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.3022, pruned_loss=0.06896, over 3036577.37 frames. ], batch size: 146, lr: 5.22e-03, grad_scale: 2.0 2023-04-29 19:54:47,954 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=128474.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:54:56,948 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.056e+02 3.058e+02 3.713e+02 4.648e+02 8.477e+02, threshold=7.425e+02, percent-clipped=4.0 2023-04-29 19:55:02,268 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=128483.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:55:30,391 INFO [train.py:904] (4/8) Epoch 13, batch 6700, loss[loss=0.2282, simple_loss=0.2997, pruned_loss=0.07838, over 11383.00 frames. ], tot_loss[loss=0.2187, simple_loss=0.3003, pruned_loss=0.06851, over 3042282.92 frames. ], batch size: 246, lr: 5.22e-03, grad_scale: 2.0 2023-04-29 19:55:36,489 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128506.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:56:09,927 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128528.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:56:46,042 INFO [train.py:904] (4/8) Epoch 13, batch 6750, loss[loss=0.2849, simple_loss=0.3493, pruned_loss=0.1103, over 12176.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.2997, pruned_loss=0.06865, over 3043783.14 frames. ], batch size: 248, lr: 5.22e-03, grad_scale: 2.0 2023-04-29 19:56:49,433 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=128554.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:57:22,973 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=128576.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 19:57:28,178 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.231e+02 3.101e+02 3.923e+02 4.752e+02 1.055e+03, threshold=7.847e+02, percent-clipped=2.0 2023-04-29 19:57:38,543 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.92 vs. limit=5.0 2023-04-29 19:58:01,812 INFO [train.py:904] (4/8) Epoch 13, batch 6800, loss[loss=0.198, simple_loss=0.2897, pruned_loss=0.05316, over 16833.00 frames. ], tot_loss[loss=0.2173, simple_loss=0.2991, pruned_loss=0.06777, over 3050471.15 frames. ], batch size: 83, lr: 5.21e-03, grad_scale: 4.0 2023-04-29 19:58:26,626 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3897, 3.3040, 3.3775, 3.4914, 3.5130, 3.2316, 3.4792, 3.5475], device='cuda:4'), covar=tensor([0.1074, 0.0898, 0.1035, 0.0575, 0.0650, 0.2848, 0.1017, 0.0805], device='cuda:4'), in_proj_covar=tensor([0.0549, 0.0682, 0.0815, 0.0695, 0.0530, 0.0537, 0.0560, 0.0641], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 19:59:19,100 INFO [train.py:904] (4/8) Epoch 13, batch 6850, loss[loss=0.2018, simple_loss=0.3082, pruned_loss=0.04767, over 16890.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.3006, pruned_loss=0.06765, over 3073939.67 frames. ], batch size: 90, lr: 5.21e-03, grad_scale: 4.0 2023-04-29 19:59:40,420 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6641, 3.1489, 2.6835, 5.0876, 3.6588, 4.4243, 1.6529, 3.2299], device='cuda:4'), covar=tensor([0.1442, 0.0627, 0.1132, 0.0106, 0.0425, 0.0334, 0.1575, 0.0766], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0161, 0.0182, 0.0155, 0.0201, 0.0205, 0.0183, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 19:59:54,892 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128676.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:00:00,829 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.804e+02 2.799e+02 3.352e+02 4.100e+02 8.801e+02, threshold=6.704e+02, percent-clipped=1.0 2023-04-29 20:00:34,560 INFO [train.py:904] (4/8) Epoch 13, batch 6900, loss[loss=0.2378, simple_loss=0.3199, pruned_loss=0.07788, over 16542.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.3029, pruned_loss=0.06658, over 3097040.32 frames. ], batch size: 75, lr: 5.21e-03, grad_scale: 4.0 2023-04-29 20:01:16,301 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0069, 3.6111, 3.4423, 2.0175, 2.8168, 2.3566, 3.3248, 3.6836], device='cuda:4'), covar=tensor([0.0360, 0.0653, 0.0545, 0.1866, 0.0872, 0.0916, 0.0877, 0.0920], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0150, 0.0161, 0.0146, 0.0140, 0.0127, 0.0140, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 20:01:52,386 INFO [train.py:904] (4/8) Epoch 13, batch 6950, loss[loss=0.2932, simple_loss=0.3442, pruned_loss=0.1212, over 10983.00 frames. ], tot_loss[loss=0.2215, simple_loss=0.3052, pruned_loss=0.06887, over 3080543.68 frames. ], batch size: 246, lr: 5.21e-03, grad_scale: 4.0 2023-04-29 20:02:18,902 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128769.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:02:22,156 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-29 20:02:36,165 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.242e+02 3.339e+02 3.931e+02 4.735e+02 7.907e+02, threshold=7.862e+02, percent-clipped=3.0 2023-04-29 20:02:38,447 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=128781.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:03:08,750 INFO [train.py:904] (4/8) Epoch 13, batch 7000, loss[loss=0.269, simple_loss=0.3298, pruned_loss=0.1041, over 11474.00 frames. ], tot_loss[loss=0.2211, simple_loss=0.3056, pruned_loss=0.06825, over 3087935.96 frames. ], batch size: 246, lr: 5.21e-03, grad_scale: 4.0 2023-04-29 20:04:07,286 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=128842.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:04:20,571 INFO [train.py:904] (4/8) Epoch 13, batch 7050, loss[loss=0.2331, simple_loss=0.307, pruned_loss=0.07966, over 11588.00 frames. ], tot_loss[loss=0.2211, simple_loss=0.3057, pruned_loss=0.06824, over 3069909.20 frames. ], batch size: 247, lr: 5.21e-03, grad_scale: 4.0 2023-04-29 20:05:02,583 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.996e+02 2.845e+02 3.584e+02 4.333e+02 9.135e+02, threshold=7.167e+02, percent-clipped=3.0 2023-04-29 20:05:19,857 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=128890.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 20:05:38,297 INFO [train.py:904] (4/8) Epoch 13, batch 7100, loss[loss=0.2355, simple_loss=0.3122, pruned_loss=0.07933, over 15298.00 frames. ], tot_loss[loss=0.2207, simple_loss=0.3044, pruned_loss=0.06847, over 3055454.08 frames. ], batch size: 190, lr: 5.21e-03, grad_scale: 4.0 2023-04-29 20:06:08,597 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8923, 3.2275, 3.1879, 2.1227, 3.0172, 3.1743, 3.1151, 1.9514], device='cuda:4'), covar=tensor([0.0478, 0.0042, 0.0046, 0.0374, 0.0084, 0.0091, 0.0070, 0.0383], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0069, 0.0072, 0.0127, 0.0083, 0.0094, 0.0082, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 20:06:56,748 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=128951.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 20:06:57,447 INFO [train.py:904] (4/8) Epoch 13, batch 7150, loss[loss=0.2273, simple_loss=0.3027, pruned_loss=0.07595, over 15195.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.3019, pruned_loss=0.06756, over 3082239.14 frames. ], batch size: 190, lr: 5.21e-03, grad_scale: 4.0 2023-04-29 20:07:20,564 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7981, 4.6189, 4.7996, 4.9774, 5.1452, 4.5651, 5.1135, 5.1205], device='cuda:4'), covar=tensor([0.1636, 0.1054, 0.1532, 0.0629, 0.0511, 0.0809, 0.0530, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0534, 0.0663, 0.0795, 0.0677, 0.0517, 0.0524, 0.0546, 0.0627], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:07:21,741 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7637, 4.8129, 4.6158, 4.3109, 4.1920, 4.6998, 4.5889, 4.4005], device='cuda:4'), covar=tensor([0.0563, 0.0436, 0.0291, 0.0280, 0.1028, 0.0422, 0.0368, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0333, 0.0292, 0.0272, 0.0310, 0.0316, 0.0198, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:07:34,012 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128976.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:07:39,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.000e+02 3.228e+02 3.806e+02 5.170e+02 8.621e+02, threshold=7.612e+02, percent-clipped=3.0 2023-04-29 20:08:12,085 INFO [train.py:904] (4/8) Epoch 13, batch 7200, loss[loss=0.2115, simple_loss=0.2887, pruned_loss=0.0672, over 11970.00 frames. ], tot_loss[loss=0.2151, simple_loss=0.2992, pruned_loss=0.06546, over 3078998.45 frames. ], batch size: 248, lr: 5.21e-03, grad_scale: 8.0 2023-04-29 20:08:45,130 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=129024.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:09:08,254 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9148, 2.2396, 2.2772, 2.8140, 2.0113, 3.1536, 1.6860, 2.7117], device='cuda:4'), covar=tensor([0.1350, 0.0657, 0.1168, 0.0196, 0.0137, 0.0356, 0.1591, 0.0725], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0164, 0.0184, 0.0157, 0.0205, 0.0209, 0.0187, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 20:09:29,929 INFO [train.py:904] (4/8) Epoch 13, batch 7250, loss[loss=0.2171, simple_loss=0.2898, pruned_loss=0.07217, over 11488.00 frames. ], tot_loss[loss=0.2129, simple_loss=0.2968, pruned_loss=0.06449, over 3065431.52 frames. ], batch size: 248, lr: 5.21e-03, grad_scale: 8.0 2023-04-29 20:09:52,189 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-29 20:09:56,915 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=129069.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:10:12,180 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.987e+02 2.885e+02 3.444e+02 4.352e+02 7.294e+02, threshold=6.888e+02, percent-clipped=0.0 2023-04-29 20:10:45,714 INFO [train.py:904] (4/8) Epoch 13, batch 7300, loss[loss=0.2538, simple_loss=0.3126, pruned_loss=0.09751, over 11366.00 frames. ], tot_loss[loss=0.2117, simple_loss=0.2961, pruned_loss=0.06364, over 3086543.93 frames. ], batch size: 246, lr: 5.20e-03, grad_scale: 8.0 2023-04-29 20:11:09,651 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=129117.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:11:40,200 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=129137.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:12:02,352 INFO [train.py:904] (4/8) Epoch 13, batch 7350, loss[loss=0.2581, simple_loss=0.3226, pruned_loss=0.09682, over 10998.00 frames. ], tot_loss[loss=0.2138, simple_loss=0.2975, pruned_loss=0.06506, over 3079981.56 frames. ], batch size: 247, lr: 5.20e-03, grad_scale: 8.0 2023-04-29 20:12:28,607 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-29 20:12:46,221 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.932e+02 2.924e+02 3.417e+02 4.096e+02 1.600e+03, threshold=6.834e+02, percent-clipped=4.0 2023-04-29 20:13:21,034 INFO [train.py:904] (4/8) Epoch 13, batch 7400, loss[loss=0.2031, simple_loss=0.2973, pruned_loss=0.05443, over 16470.00 frames. ], tot_loss[loss=0.2153, simple_loss=0.2988, pruned_loss=0.06593, over 3087063.45 frames. ], batch size: 75, lr: 5.20e-03, grad_scale: 8.0 2023-04-29 20:14:32,917 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=129246.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 20:14:40,220 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.73 vs. limit=2.0 2023-04-29 20:14:41,688 INFO [train.py:904] (4/8) Epoch 13, batch 7450, loss[loss=0.2456, simple_loss=0.3144, pruned_loss=0.08845, over 11668.00 frames. ], tot_loss[loss=0.2176, simple_loss=0.3001, pruned_loss=0.06752, over 3068876.84 frames. ], batch size: 248, lr: 5.20e-03, grad_scale: 8.0 2023-04-29 20:15:30,917 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.914e+02 3.121e+02 3.842e+02 4.432e+02 7.351e+02, threshold=7.685e+02, percent-clipped=1.0 2023-04-29 20:16:05,625 INFO [train.py:904] (4/8) Epoch 13, batch 7500, loss[loss=0.2371, simple_loss=0.3156, pruned_loss=0.07924, over 15157.00 frames. ], tot_loss[loss=0.2177, simple_loss=0.3011, pruned_loss=0.06719, over 3061763.27 frames. ], batch size: 190, lr: 5.20e-03, grad_scale: 8.0 2023-04-29 20:17:24,560 INFO [train.py:904] (4/8) Epoch 13, batch 7550, loss[loss=0.1904, simple_loss=0.2753, pruned_loss=0.05272, over 16821.00 frames. ], tot_loss[loss=0.2167, simple_loss=0.2996, pruned_loss=0.06691, over 3061723.34 frames. ], batch size: 39, lr: 5.20e-03, grad_scale: 8.0 2023-04-29 20:17:33,959 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3378, 3.4075, 1.9320, 3.7718, 2.4663, 3.7310, 1.9519, 2.6114], device='cuda:4'), covar=tensor([0.0247, 0.0333, 0.1651, 0.0145, 0.0836, 0.0501, 0.1599, 0.0777], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0165, 0.0190, 0.0136, 0.0168, 0.0207, 0.0197, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 20:17:56,823 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-29 20:18:07,708 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.866e+02 2.820e+02 3.717e+02 4.924e+02 9.310e+02, threshold=7.434e+02, percent-clipped=3.0 2023-04-29 20:18:41,434 INFO [train.py:904] (4/8) Epoch 13, batch 7600, loss[loss=0.2166, simple_loss=0.2919, pruned_loss=0.07063, over 17011.00 frames. ], tot_loss[loss=0.2167, simple_loss=0.2994, pruned_loss=0.06702, over 3071508.49 frames. ], batch size: 55, lr: 5.20e-03, grad_scale: 8.0 2023-04-29 20:18:45,286 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 20:18:53,111 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8518, 1.9287, 2.2809, 3.1533, 2.1708, 2.2032, 2.2025, 2.0631], device='cuda:4'), covar=tensor([0.1129, 0.3399, 0.2133, 0.0620, 0.3787, 0.2435, 0.2981, 0.3473], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0404, 0.0336, 0.0315, 0.0418, 0.0463, 0.0368, 0.0469], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:18:59,758 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=129413.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 20:19:37,568 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=129437.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:19:41,506 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-29 20:19:43,597 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.43 vs. limit=5.0 2023-04-29 20:20:00,017 INFO [train.py:904] (4/8) Epoch 13, batch 7650, loss[loss=0.2077, simple_loss=0.2936, pruned_loss=0.06085, over 16514.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.3007, pruned_loss=0.06777, over 3084118.36 frames. ], batch size: 75, lr: 5.20e-03, grad_scale: 8.0 2023-04-29 20:20:12,375 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-29 20:20:36,399 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=129474.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 20:20:44,163 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.211e+02 3.428e+02 4.160e+02 4.831e+02 1.059e+03, threshold=8.320e+02, percent-clipped=3.0 2023-04-29 20:20:53,102 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=129485.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:21:06,952 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5467, 4.8219, 4.6229, 4.6040, 4.3558, 4.3472, 4.3418, 4.8649], device='cuda:4'), covar=tensor([0.1013, 0.0801, 0.0944, 0.0758, 0.0780, 0.1106, 0.0994, 0.0858], device='cuda:4'), in_proj_covar=tensor([0.0568, 0.0702, 0.0576, 0.0499, 0.0441, 0.0455, 0.0587, 0.0536], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:21:18,576 INFO [train.py:904] (4/8) Epoch 13, batch 7700, loss[loss=0.2795, simple_loss=0.3253, pruned_loss=0.1169, over 11682.00 frames. ], tot_loss[loss=0.2173, simple_loss=0.2998, pruned_loss=0.06739, over 3094589.66 frames. ], batch size: 248, lr: 5.20e-03, grad_scale: 4.0 2023-04-29 20:21:54,441 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1307, 2.0542, 2.6775, 3.2373, 3.0768, 3.6415, 2.1692, 3.6404], device='cuda:4'), covar=tensor([0.0180, 0.0412, 0.0251, 0.0198, 0.0211, 0.0103, 0.0419, 0.0091], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0174, 0.0158, 0.0162, 0.0173, 0.0128, 0.0175, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 20:22:26,944 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=129546.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 20:22:35,876 INFO [train.py:904] (4/8) Epoch 13, batch 7750, loss[loss=0.1951, simple_loss=0.2874, pruned_loss=0.05141, over 16473.00 frames. ], tot_loss[loss=0.2168, simple_loss=0.2996, pruned_loss=0.06703, over 3105581.45 frames. ], batch size: 75, lr: 5.20e-03, grad_scale: 4.0 2023-04-29 20:23:20,357 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.834e+02 3.197e+02 3.787e+02 4.587e+02 8.661e+02, threshold=7.574e+02, percent-clipped=1.0 2023-04-29 20:23:40,230 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=129594.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 20:23:52,176 INFO [train.py:904] (4/8) Epoch 13, batch 7800, loss[loss=0.1975, simple_loss=0.2853, pruned_loss=0.05483, over 15412.00 frames. ], tot_loss[loss=0.2187, simple_loss=0.301, pruned_loss=0.06823, over 3093800.37 frames. ], batch size: 190, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:24:09,284 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=129612.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:24:48,714 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5403, 4.6623, 4.8420, 4.6615, 4.6811, 5.2301, 4.7912, 4.5620], device='cuda:4'), covar=tensor([0.1142, 0.1962, 0.2294, 0.2037, 0.2616, 0.1045, 0.1495, 0.2306], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0519, 0.0567, 0.0441, 0.0597, 0.0584, 0.0450, 0.0593], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 20:25:09,851 INFO [train.py:904] (4/8) Epoch 13, batch 7850, loss[loss=0.2187, simple_loss=0.2937, pruned_loss=0.0719, over 11736.00 frames. ], tot_loss[loss=0.2194, simple_loss=0.3018, pruned_loss=0.06848, over 3082449.23 frames. ], batch size: 248, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:25:26,603 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6479, 4.7734, 4.9839, 4.7953, 4.8108, 5.3948, 4.8907, 4.6314], device='cuda:4'), covar=tensor([0.1050, 0.2009, 0.2198, 0.1852, 0.2497, 0.0913, 0.1557, 0.2425], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0518, 0.0566, 0.0441, 0.0597, 0.0584, 0.0450, 0.0593], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 20:25:40,322 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=129673.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:25:52,267 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.713e+02 3.020e+02 3.492e+02 4.287e+02 1.158e+03, threshold=6.983e+02, percent-clipped=4.0 2023-04-29 20:26:02,599 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6010, 5.9661, 5.6554, 5.7575, 5.3378, 5.2011, 5.4113, 6.1011], device='cuda:4'), covar=tensor([0.1025, 0.0761, 0.0991, 0.0755, 0.0800, 0.0644, 0.1026, 0.0762], device='cuda:4'), in_proj_covar=tensor([0.0570, 0.0703, 0.0580, 0.0502, 0.0442, 0.0456, 0.0588, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:26:22,709 INFO [train.py:904] (4/8) Epoch 13, batch 7900, loss[loss=0.1943, simple_loss=0.2772, pruned_loss=0.05563, over 16198.00 frames. ], tot_loss[loss=0.2184, simple_loss=0.3007, pruned_loss=0.06808, over 3066091.78 frames. ], batch size: 35, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:27:36,767 INFO [train.py:904] (4/8) Epoch 13, batch 7950, loss[loss=0.2292, simple_loss=0.3065, pruned_loss=0.0759, over 16866.00 frames. ], tot_loss[loss=0.2195, simple_loss=0.3014, pruned_loss=0.06883, over 3067400.85 frames. ], batch size: 109, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:28:02,256 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=129769.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 20:28:18,207 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.009e+02 2.840e+02 3.431e+02 4.053e+02 8.396e+02, threshold=6.863e+02, percent-clipped=2.0 2023-04-29 20:28:33,877 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=129791.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:28:49,446 INFO [train.py:904] (4/8) Epoch 13, batch 8000, loss[loss=0.2102, simple_loss=0.3035, pruned_loss=0.05845, over 16730.00 frames. ], tot_loss[loss=0.2191, simple_loss=0.3014, pruned_loss=0.06838, over 3079833.35 frames. ], batch size: 83, lr: 5.19e-03, grad_scale: 8.0 2023-04-29 20:30:02,332 INFO [train.py:904] (4/8) Epoch 13, batch 8050, loss[loss=0.1961, simple_loss=0.2895, pruned_loss=0.05139, over 16836.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.3013, pruned_loss=0.0679, over 3093367.31 frames. ], batch size: 83, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:30:02,888 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=129852.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:30:25,150 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-29 20:30:37,138 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-29 20:30:45,830 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.353e+02 3.065e+02 3.748e+02 4.862e+02 1.232e+03, threshold=7.497e+02, percent-clipped=5.0 2023-04-29 20:30:53,727 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5052, 3.3483, 3.7829, 1.7439, 3.9729, 4.0101, 2.8107, 2.8284], device='cuda:4'), covar=tensor([0.0713, 0.0238, 0.0163, 0.1207, 0.0053, 0.0131, 0.0436, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0101, 0.0089, 0.0138, 0.0069, 0.0109, 0.0121, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 20:31:15,197 INFO [train.py:904] (4/8) Epoch 13, batch 8100, loss[loss=0.1851, simple_loss=0.2681, pruned_loss=0.05105, over 16529.00 frames. ], tot_loss[loss=0.2165, simple_loss=0.2999, pruned_loss=0.0665, over 3112775.21 frames. ], batch size: 62, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:32:19,157 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1950, 1.5494, 1.9383, 2.1840, 2.2872, 2.4120, 1.6802, 2.4714], device='cuda:4'), covar=tensor([0.0174, 0.0402, 0.0233, 0.0249, 0.0231, 0.0146, 0.0388, 0.0101], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0174, 0.0157, 0.0160, 0.0172, 0.0128, 0.0174, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 20:32:28,790 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5511, 3.4846, 3.4755, 2.8883, 3.3933, 2.0430, 3.1240, 2.9662], device='cuda:4'), covar=tensor([0.0120, 0.0094, 0.0142, 0.0204, 0.0088, 0.2015, 0.0113, 0.0196], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0123, 0.0169, 0.0159, 0.0141, 0.0184, 0.0157, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:32:29,503 INFO [train.py:904] (4/8) Epoch 13, batch 8150, loss[loss=0.2374, simple_loss=0.2985, pruned_loss=0.08811, over 11678.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.2976, pruned_loss=0.06565, over 3106555.85 frames. ], batch size: 247, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:32:53,000 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=129968.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:33:14,447 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.827e+02 2.878e+02 3.437e+02 4.165e+02 9.532e+02, threshold=6.873e+02, percent-clipped=2.0 2023-04-29 20:33:46,554 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130000.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:33:48,621 INFO [train.py:904] (4/8) Epoch 13, batch 8200, loss[loss=0.1969, simple_loss=0.2873, pruned_loss=0.05323, over 16287.00 frames. ], tot_loss[loss=0.2125, simple_loss=0.2948, pruned_loss=0.06509, over 3113670.95 frames. ], batch size: 165, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:33:49,154 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130002.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:35:09,259 INFO [train.py:904] (4/8) Epoch 13, batch 8250, loss[loss=0.1864, simple_loss=0.2832, pruned_loss=0.04485, over 16716.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2938, pruned_loss=0.06302, over 3093895.66 frames. ], batch size: 89, lr: 5.19e-03, grad_scale: 4.0 2023-04-29 20:35:23,987 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130061.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:35:27,857 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130063.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 20:35:37,721 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130069.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 20:35:57,078 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.490e+02 2.705e+02 3.349e+02 4.037e+02 8.257e+02, threshold=6.697e+02, percent-clipped=3.0 2023-04-29 20:36:29,983 INFO [train.py:904] (4/8) Epoch 13, batch 8300, loss[loss=0.179, simple_loss=0.2731, pruned_loss=0.04244, over 15269.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.291, pruned_loss=0.06038, over 3058927.71 frames. ], batch size: 190, lr: 5.18e-03, grad_scale: 4.0 2023-04-29 20:36:55,890 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=130117.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 20:37:05,173 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5158, 3.0562, 3.0895, 1.9411, 2.7774, 2.1713, 3.0914, 3.1776], device='cuda:4'), covar=tensor([0.0295, 0.0732, 0.0492, 0.1904, 0.0824, 0.0967, 0.0677, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0147, 0.0158, 0.0143, 0.0137, 0.0124, 0.0136, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 20:37:43,810 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130147.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:37:51,333 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0209, 4.0596, 3.9169, 3.7079, 3.6287, 4.0322, 3.6885, 3.7846], device='cuda:4'), covar=tensor([0.0583, 0.0718, 0.0259, 0.0237, 0.0713, 0.0523, 0.0944, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0335, 0.0293, 0.0271, 0.0306, 0.0314, 0.0200, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:37:51,995 INFO [train.py:904] (4/8) Epoch 13, batch 8350, loss[loss=0.2011, simple_loss=0.2959, pruned_loss=0.05317, over 15421.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2901, pruned_loss=0.05798, over 3062206.65 frames. ], batch size: 191, lr: 5.18e-03, grad_scale: 4.0 2023-04-29 20:38:39,963 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.464e+02 2.345e+02 2.793e+02 3.218e+02 5.548e+02, threshold=5.587e+02, percent-clipped=0.0 2023-04-29 20:39:12,413 INFO [train.py:904] (4/8) Epoch 13, batch 8400, loss[loss=0.2006, simple_loss=0.2886, pruned_loss=0.05626, over 16688.00 frames. ], tot_loss[loss=0.1994, simple_loss=0.287, pruned_loss=0.05595, over 3034122.30 frames. ], batch size: 124, lr: 5.18e-03, grad_scale: 8.0 2023-04-29 20:40:29,223 INFO [train.py:904] (4/8) Epoch 13, batch 8450, loss[loss=0.1876, simple_loss=0.2787, pruned_loss=0.04823, over 16404.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2851, pruned_loss=0.05398, over 3045917.02 frames. ], batch size: 146, lr: 5.18e-03, grad_scale: 8.0 2023-04-29 20:40:40,760 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5243, 4.5464, 4.3932, 4.0986, 4.0526, 4.4971, 4.3251, 4.1826], device='cuda:4'), covar=tensor([0.0535, 0.0438, 0.0311, 0.0266, 0.0894, 0.0398, 0.0514, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0334, 0.0293, 0.0271, 0.0305, 0.0314, 0.0200, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:40:55,923 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130268.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:41:17,434 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.448e+02 2.422e+02 2.811e+02 3.625e+02 1.309e+03, threshold=5.622e+02, percent-clipped=6.0 2023-04-29 20:41:49,404 INFO [train.py:904] (4/8) Epoch 13, batch 8500, loss[loss=0.1588, simple_loss=0.2561, pruned_loss=0.03073, over 16688.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2814, pruned_loss=0.05155, over 3035174.19 frames. ], batch size: 76, lr: 5.18e-03, grad_scale: 8.0 2023-04-29 20:42:13,190 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=130316.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:43:10,836 INFO [train.py:904] (4/8) Epoch 13, batch 8550, loss[loss=0.1991, simple_loss=0.2947, pruned_loss=0.05168, over 16664.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2792, pruned_loss=0.05003, over 3043517.62 frames. ], batch size: 134, lr: 5.18e-03, grad_scale: 8.0 2023-04-29 20:43:19,716 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130356.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:43:23,659 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130358.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 20:44:07,711 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.521e+02 2.204e+02 2.928e+02 3.716e+02 7.366e+02, threshold=5.857e+02, percent-clipped=6.0 2023-04-29 20:44:12,600 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130384.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:44:50,594 INFO [train.py:904] (4/8) Epoch 13, batch 8600, loss[loss=0.1871, simple_loss=0.282, pruned_loss=0.04611, over 16784.00 frames. ], tot_loss[loss=0.189, simple_loss=0.2797, pruned_loss=0.0492, over 3047976.34 frames. ], batch size: 124, lr: 5.18e-03, grad_scale: 4.0 2023-04-29 20:45:44,166 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8179, 3.6544, 3.8824, 3.9725, 4.0877, 3.6682, 4.0549, 4.0778], device='cuda:4'), covar=tensor([0.1383, 0.1099, 0.1201, 0.0645, 0.0512, 0.1619, 0.0538, 0.0642], device='cuda:4'), in_proj_covar=tensor([0.0525, 0.0650, 0.0776, 0.0668, 0.0508, 0.0515, 0.0531, 0.0620], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:45:51,458 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130432.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:46:04,217 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130438.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:46:16,940 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130445.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:46:20,581 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130447.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:46:29,996 INFO [train.py:904] (4/8) Epoch 13, batch 8650, loss[loss=0.1956, simple_loss=0.2988, pruned_loss=0.04616, over 16308.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2777, pruned_loss=0.04744, over 3048022.03 frames. ], batch size: 146, lr: 5.18e-03, grad_scale: 4.0 2023-04-29 20:47:20,219 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7034, 3.5521, 3.5594, 3.8706, 3.9676, 3.5871, 3.9107, 3.9451], device='cuda:4'), covar=tensor([0.1531, 0.1335, 0.2285, 0.0988, 0.0837, 0.2322, 0.1039, 0.1069], device='cuda:4'), in_proj_covar=tensor([0.0524, 0.0648, 0.0773, 0.0664, 0.0505, 0.0513, 0.0529, 0.0617], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:47:40,991 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.566e+02 2.301e+02 2.678e+02 3.280e+02 8.282e+02, threshold=5.356e+02, percent-clipped=3.0 2023-04-29 20:48:01,218 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130493.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:48:05,351 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=130495.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:48:12,886 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130499.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:48:17,482 INFO [train.py:904] (4/8) Epoch 13, batch 8700, loss[loss=0.1853, simple_loss=0.2815, pruned_loss=0.04457, over 16738.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2751, pruned_loss=0.0463, over 3050304.84 frames. ], batch size: 134, lr: 5.18e-03, grad_scale: 4.0 2023-04-29 20:49:47,028 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130548.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:49:54,307 INFO [train.py:904] (4/8) Epoch 13, batch 8750, loss[loss=0.1783, simple_loss=0.279, pruned_loss=0.03882, over 16713.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2748, pruned_loss=0.04563, over 3063575.42 frames. ], batch size: 83, lr: 5.18e-03, grad_scale: 4.0 2023-04-29 20:50:30,335 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130565.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:50:45,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5804, 3.7091, 2.8059, 2.1307, 2.3520, 2.3370, 3.9034, 3.1954], device='cuda:4'), covar=tensor([0.2772, 0.0635, 0.1598, 0.2446, 0.2601, 0.1860, 0.0419, 0.1250], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0252, 0.0283, 0.0279, 0.0269, 0.0224, 0.0266, 0.0293], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 20:51:07,892 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.672e+02 2.544e+02 3.252e+02 4.020e+02 9.168e+02, threshold=6.504e+02, percent-clipped=9.0 2023-04-29 20:51:48,203 INFO [train.py:904] (4/8) Epoch 13, batch 8800, loss[loss=0.1826, simple_loss=0.2698, pruned_loss=0.04772, over 16963.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2735, pruned_loss=0.04494, over 3073618.69 frames. ], batch size: 109, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 20:52:02,431 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130609.0, num_to_drop=1, layers_to_drop={3} 2023-04-29 20:52:39,097 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130626.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:53:21,507 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2788, 1.4826, 1.9973, 2.3142, 2.2189, 2.4533, 1.7532, 2.4338], device='cuda:4'), covar=tensor([0.0208, 0.0487, 0.0290, 0.0255, 0.0277, 0.0220, 0.0428, 0.0129], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0171, 0.0154, 0.0156, 0.0168, 0.0125, 0.0171, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 20:53:32,459 INFO [train.py:904] (4/8) Epoch 13, batch 8850, loss[loss=0.1808, simple_loss=0.2809, pruned_loss=0.04036, over 15533.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2759, pruned_loss=0.04461, over 3048940.15 frames. ], batch size: 192, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 20:53:41,389 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130656.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:53:45,803 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130658.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 20:54:37,528 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.421e+02 2.288e+02 2.852e+02 3.513e+02 7.607e+02, threshold=5.705e+02, percent-clipped=2.0 2023-04-29 20:55:17,114 INFO [train.py:904] (4/8) Epoch 13, batch 8900, loss[loss=0.2227, simple_loss=0.3169, pruned_loss=0.06425, over 16265.00 frames. ], tot_loss[loss=0.182, simple_loss=0.2761, pruned_loss=0.04395, over 3048909.91 frames. ], batch size: 165, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 20:55:22,704 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=130704.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:55:26,794 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=130706.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:55:49,051 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6848, 2.7121, 2.4241, 4.1511, 2.8881, 4.0638, 1.4239, 2.9298], device='cuda:4'), covar=tensor([0.1355, 0.0687, 0.1217, 0.0161, 0.0167, 0.0323, 0.1626, 0.0719], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0163, 0.0184, 0.0155, 0.0197, 0.0206, 0.0187, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 20:56:54,229 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130740.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:57:21,720 INFO [train.py:904] (4/8) Epoch 13, batch 8950, loss[loss=0.1677, simple_loss=0.259, pruned_loss=0.03823, over 16244.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2758, pruned_loss=0.04403, over 3081000.04 frames. ], batch size: 165, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 20:58:29,462 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.201e+02 2.684e+02 3.167e+02 7.959e+02, threshold=5.368e+02, percent-clipped=1.0 2023-04-29 20:58:37,145 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0315, 2.5539, 2.5977, 1.8447, 2.7978, 2.8919, 2.4491, 2.3612], device='cuda:4'), covar=tensor([0.0685, 0.0199, 0.0181, 0.0995, 0.0080, 0.0154, 0.0431, 0.0425], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0097, 0.0084, 0.0133, 0.0065, 0.0103, 0.0116, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 20:58:41,745 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130788.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:58:54,912 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130794.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 20:59:11,360 INFO [train.py:904] (4/8) Epoch 13, batch 9000, loss[loss=0.1793, simple_loss=0.2601, pruned_loss=0.04926, over 12682.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2727, pruned_loss=0.04266, over 3083174.92 frames. ], batch size: 248, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 20:59:11,361 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 20:59:22,054 INFO [train.py:938] (4/8) Epoch 13, validation: loss=0.1517, simple_loss=0.2561, pruned_loss=0.02371, over 944034.00 frames. 2023-04-29 20:59:22,055 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 21:00:51,508 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4331, 3.4129, 3.5125, 3.5926, 3.6500, 3.3114, 3.6298, 3.6940], device='cuda:4'), covar=tensor([0.1294, 0.0937, 0.1069, 0.0673, 0.0612, 0.2559, 0.0776, 0.0666], device='cuda:4'), in_proj_covar=tensor([0.0522, 0.0644, 0.0768, 0.0662, 0.0501, 0.0506, 0.0525, 0.0608], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:01:06,016 INFO [train.py:904] (4/8) Epoch 13, batch 9050, loss[loss=0.1819, simple_loss=0.2709, pruned_loss=0.04646, over 16882.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2731, pruned_loss=0.04303, over 3084066.07 frames. ], batch size: 96, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 21:01:54,628 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2916, 2.4213, 1.8910, 2.2285, 2.8611, 2.6062, 3.0131, 3.1220], device='cuda:4'), covar=tensor([0.0116, 0.0379, 0.0512, 0.0412, 0.0218, 0.0321, 0.0211, 0.0190], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0206, 0.0200, 0.0200, 0.0204, 0.0204, 0.0202, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:02:07,079 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.631e+02 2.403e+02 2.798e+02 3.395e+02 5.131e+02, threshold=5.596e+02, percent-clipped=0.0 2023-04-29 21:02:52,476 INFO [train.py:904] (4/8) Epoch 13, batch 9100, loss[loss=0.1851, simple_loss=0.2849, pruned_loss=0.04262, over 16354.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2726, pruned_loss=0.04342, over 3068127.64 frames. ], batch size: 146, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 21:02:58,159 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130904.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 21:03:33,377 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130921.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:04:49,576 INFO [train.py:904] (4/8) Epoch 13, batch 9150, loss[loss=0.1768, simple_loss=0.273, pruned_loss=0.04033, over 16196.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2732, pruned_loss=0.04353, over 3050180.71 frames. ], batch size: 165, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 21:05:52,920 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.726e+02 2.558e+02 2.960e+02 3.606e+02 6.781e+02, threshold=5.920e+02, percent-clipped=4.0 2023-04-29 21:06:10,236 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.87 vs. limit=5.0 2023-04-29 21:06:31,913 INFO [train.py:904] (4/8) Epoch 13, batch 9200, loss[loss=0.1572, simple_loss=0.2386, pruned_loss=0.0379, over 12330.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2689, pruned_loss=0.04293, over 3052402.49 frames. ], batch size: 248, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 21:07:10,636 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8321, 2.1859, 1.8204, 1.9817, 2.5734, 2.2738, 2.5670, 2.7379], device='cuda:4'), covar=tensor([0.0124, 0.0368, 0.0422, 0.0370, 0.0204, 0.0282, 0.0200, 0.0201], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0208, 0.0201, 0.0201, 0.0205, 0.0204, 0.0203, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:07:43,467 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131040.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:08:07,048 INFO [train.py:904] (4/8) Epoch 13, batch 9250, loss[loss=0.1813, simple_loss=0.2719, pruned_loss=0.04534, over 16867.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2689, pruned_loss=0.04297, over 3065413.60 frames. ], batch size: 125, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 21:09:12,849 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.311e+02 2.375e+02 2.887e+02 3.471e+02 9.563e+02, threshold=5.774e+02, percent-clipped=3.0 2023-04-29 21:09:24,455 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=131088.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:09:24,515 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131088.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:09:39,708 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131094.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:09:55,037 INFO [train.py:904] (4/8) Epoch 13, batch 9300, loss[loss=0.1467, simple_loss=0.2328, pruned_loss=0.03026, over 12172.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2677, pruned_loss=0.04237, over 3073305.72 frames. ], batch size: 250, lr: 5.17e-03, grad_scale: 8.0 2023-04-29 21:10:00,713 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2595, 1.5762, 1.8992, 2.3697, 2.2974, 2.5083, 1.7281, 2.4688], device='cuda:4'), covar=tensor([0.0173, 0.0433, 0.0291, 0.0259, 0.0259, 0.0162, 0.0397, 0.0120], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0169, 0.0153, 0.0154, 0.0167, 0.0124, 0.0170, 0.0115], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 21:11:11,075 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=131136.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:11:24,001 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=131142.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:11:40,550 INFO [train.py:904] (4/8) Epoch 13, batch 9350, loss[loss=0.1824, simple_loss=0.2801, pruned_loss=0.04239, over 16926.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2672, pruned_loss=0.04226, over 3072805.52 frames. ], batch size: 116, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:12:25,590 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131174.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:12:32,487 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-29 21:12:41,633 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.534e+02 2.118e+02 2.389e+02 2.842e+02 4.349e+02, threshold=4.777e+02, percent-clipped=0.0 2023-04-29 21:13:20,691 INFO [train.py:904] (4/8) Epoch 13, batch 9400, loss[loss=0.1528, simple_loss=0.236, pruned_loss=0.03477, over 12652.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2671, pruned_loss=0.04165, over 3079919.73 frames. ], batch size: 247, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:13:25,618 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131204.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 21:13:59,480 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131221.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:14:28,007 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131235.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 21:14:59,708 INFO [train.py:904] (4/8) Epoch 13, batch 9450, loss[loss=0.1844, simple_loss=0.2704, pruned_loss=0.04924, over 12108.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2681, pruned_loss=0.0418, over 3053068.37 frames. ], batch size: 247, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:15:00,864 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=131252.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:15:34,983 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=131269.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:15:37,529 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.58 vs. limit=5.0 2023-04-29 21:16:03,421 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.757e+02 2.351e+02 2.704e+02 3.479e+02 5.453e+02, threshold=5.408e+02, percent-clipped=2.0 2023-04-29 21:16:40,123 INFO [train.py:904] (4/8) Epoch 13, batch 9500, loss[loss=0.1653, simple_loss=0.2589, pruned_loss=0.03583, over 15276.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2676, pruned_loss=0.04148, over 3061201.89 frames. ], batch size: 191, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:18:25,254 INFO [train.py:904] (4/8) Epoch 13, batch 9550, loss[loss=0.2036, simple_loss=0.302, pruned_loss=0.05265, over 16342.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2672, pruned_loss=0.04146, over 3073655.75 frames. ], batch size: 146, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:19:29,578 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.636e+02 2.317e+02 2.711e+02 3.253e+02 6.482e+02, threshold=5.422e+02, percent-clipped=4.0 2023-04-29 21:20:03,674 INFO [train.py:904] (4/8) Epoch 13, batch 9600, loss[loss=0.1849, simple_loss=0.2864, pruned_loss=0.04166, over 16288.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.269, pruned_loss=0.04213, over 3070435.70 frames. ], batch size: 165, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:20:17,942 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131409.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:20:29,771 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131416.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:21:12,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8770, 4.1988, 4.0218, 4.0578, 3.6603, 3.7396, 3.8136, 4.1847], device='cuda:4'), covar=tensor([0.1063, 0.0978, 0.1010, 0.0715, 0.0809, 0.1777, 0.0948, 0.1064], device='cuda:4'), in_proj_covar=tensor([0.0547, 0.0686, 0.0549, 0.0486, 0.0428, 0.0440, 0.0569, 0.0521], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:21:49,828 INFO [train.py:904] (4/8) Epoch 13, batch 9650, loss[loss=0.1723, simple_loss=0.2667, pruned_loss=0.03892, over 16893.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2706, pruned_loss=0.04261, over 3066917.23 frames. ], batch size: 102, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:22:34,325 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131470.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:22:48,219 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131477.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:22:58,034 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.598e+02 2.434e+02 2.934e+02 3.737e+02 7.668e+02, threshold=5.867e+02, percent-clipped=4.0 2023-04-29 21:23:18,189 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131492.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:23:38,319 INFO [train.py:904] (4/8) Epoch 13, batch 9700, loss[loss=0.1862, simple_loss=0.2761, pruned_loss=0.04813, over 15424.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2699, pruned_loss=0.04247, over 3066414.15 frames. ], batch size: 190, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:24:36,284 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=131530.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 21:25:20,235 INFO [train.py:904] (4/8) Epoch 13, batch 9750, loss[loss=0.167, simple_loss=0.2654, pruned_loss=0.03426, over 15181.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2682, pruned_loss=0.04167, over 3077589.98 frames. ], batch size: 190, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:25:22,324 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131553.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:26:22,846 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 2.333e+02 2.960e+02 3.674e+02 6.318e+02, threshold=5.921e+02, percent-clipped=2.0 2023-04-29 21:26:57,813 INFO [train.py:904] (4/8) Epoch 13, batch 9800, loss[loss=0.1648, simple_loss=0.2667, pruned_loss=0.0315, over 16589.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2682, pruned_loss=0.04095, over 3075070.71 frames. ], batch size: 62, lr: 5.16e-03, grad_scale: 8.0 2023-04-29 21:28:39,148 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3507, 3.9938, 4.1291, 2.9661, 3.5760, 4.1220, 3.8928, 2.4036], device='cuda:4'), covar=tensor([0.0441, 0.0028, 0.0024, 0.0266, 0.0082, 0.0052, 0.0048, 0.0363], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0067, 0.0069, 0.0125, 0.0081, 0.0088, 0.0078, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 21:28:39,811 INFO [train.py:904] (4/8) Epoch 13, batch 9850, loss[loss=0.1673, simple_loss=0.2638, pruned_loss=0.03535, over 16787.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2698, pruned_loss=0.04103, over 3080563.38 frames. ], batch size: 76, lr: 5.15e-03, grad_scale: 4.0 2023-04-29 21:28:46,239 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5176, 4.5439, 4.3963, 4.0519, 4.0818, 4.4450, 4.2934, 4.1511], device='cuda:4'), covar=tensor([0.0502, 0.0457, 0.0274, 0.0283, 0.0753, 0.0421, 0.0419, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0320, 0.0286, 0.0263, 0.0293, 0.0306, 0.0193, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-29 21:29:46,600 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.743e+02 2.305e+02 2.925e+02 3.439e+02 6.228e+02, threshold=5.850e+02, percent-clipped=2.0 2023-04-29 21:30:18,591 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-29 21:30:19,961 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3900, 4.6781, 4.5158, 4.5357, 4.2296, 4.1833, 4.2101, 4.7070], device='cuda:4'), covar=tensor([0.0963, 0.0831, 0.0800, 0.0638, 0.0677, 0.1413, 0.0916, 0.0875], device='cuda:4'), in_proj_covar=tensor([0.0548, 0.0688, 0.0552, 0.0486, 0.0431, 0.0441, 0.0569, 0.0524], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:30:32,565 INFO [train.py:904] (4/8) Epoch 13, batch 9900, loss[loss=0.1728, simple_loss=0.2593, pruned_loss=0.0431, over 12746.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2701, pruned_loss=0.04127, over 3057440.06 frames. ], batch size: 247, lr: 5.15e-03, grad_scale: 4.0 2023-04-29 21:32:30,668 INFO [train.py:904] (4/8) Epoch 13, batch 9950, loss[loss=0.1836, simple_loss=0.28, pruned_loss=0.04353, over 16231.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2728, pruned_loss=0.0419, over 3057143.25 frames. ], batch size: 165, lr: 5.15e-03, grad_scale: 4.0 2023-04-29 21:32:51,916 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6708, 4.0398, 3.6579, 3.4271, 3.0176, 3.9115, 3.5425, 3.5777], device='cuda:4'), covar=tensor([0.1028, 0.0570, 0.0606, 0.0469, 0.1640, 0.0566, 0.1485, 0.0661], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0319, 0.0286, 0.0262, 0.0292, 0.0305, 0.0193, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-29 21:33:02,061 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=131765.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:33:20,073 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=131772.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:33:40,958 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9425, 2.9508, 2.4702, 4.9125, 3.6242, 4.4131, 1.5672, 3.2720], device='cuda:4'), covar=tensor([0.1207, 0.0665, 0.1157, 0.0110, 0.0151, 0.0273, 0.1464, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0159, 0.0179, 0.0150, 0.0187, 0.0201, 0.0184, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-29 21:33:47,202 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.482e+02 2.347e+02 2.831e+02 3.741e+02 5.862e+02, threshold=5.662e+02, percent-clipped=1.0 2023-04-29 21:34:31,155 INFO [train.py:904] (4/8) Epoch 13, batch 10000, loss[loss=0.1802, simple_loss=0.2831, pruned_loss=0.03866, over 16256.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.271, pruned_loss=0.04118, over 3063837.63 frames. ], batch size: 165, lr: 5.15e-03, grad_scale: 8.0 2023-04-29 21:35:04,677 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.40 vs. limit=5.0 2023-04-29 21:35:27,559 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131830.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:36:06,299 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=131848.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:36:11,632 INFO [train.py:904] (4/8) Epoch 13, batch 10050, loss[loss=0.1777, simple_loss=0.2729, pruned_loss=0.0413, over 16356.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2706, pruned_loss=0.04076, over 3074220.84 frames. ], batch size: 146, lr: 5.15e-03, grad_scale: 8.0 2023-04-29 21:36:23,374 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131857.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:37:04,198 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=131878.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:37:14,088 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.561e+02 2.136e+02 2.458e+02 2.977e+02 5.348e+02, threshold=4.916e+02, percent-clipped=0.0 2023-04-29 21:37:21,438 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9227, 3.9833, 4.3397, 4.2995, 4.2921, 4.0593, 4.0149, 4.0477], device='cuda:4'), covar=tensor([0.0352, 0.0735, 0.0351, 0.0389, 0.0413, 0.0415, 0.0762, 0.0408], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0340, 0.0340, 0.0330, 0.0386, 0.0366, 0.0445, 0.0293], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-29 21:37:45,821 INFO [train.py:904] (4/8) Epoch 13, batch 10100, loss[loss=0.1705, simple_loss=0.2581, pruned_loss=0.04151, over 15387.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2713, pruned_loss=0.04141, over 3057868.25 frames. ], batch size: 191, lr: 5.15e-03, grad_scale: 8.0 2023-04-29 21:38:16,891 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131918.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:39:29,405 INFO [train.py:904] (4/8) Epoch 14, batch 0, loss[loss=0.1807, simple_loss=0.2726, pruned_loss=0.04445, over 17131.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2726, pruned_loss=0.04445, over 17131.00 frames. ], batch size: 48, lr: 4.96e-03, grad_scale: 8.0 2023-04-29 21:39:29,405 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 21:39:36,899 INFO [train.py:938] (4/8) Epoch 14, validation: loss=0.1515, simple_loss=0.2551, pruned_loss=0.024, over 944034.00 frames. 2023-04-29 21:39:36,900 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 21:40:03,497 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0228, 4.4716, 4.5315, 3.4826, 3.8840, 4.5477, 4.1417, 2.6668], device='cuda:4'), covar=tensor([0.0375, 0.0032, 0.0023, 0.0238, 0.0079, 0.0059, 0.0058, 0.0383], device='cuda:4'), in_proj_covar=tensor([0.0129, 0.0069, 0.0070, 0.0127, 0.0082, 0.0089, 0.0080, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 21:40:04,647 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4105, 3.3780, 1.8265, 3.5692, 2.5385, 3.5285, 1.7992, 2.7077], device='cuda:4'), covar=tensor([0.0197, 0.0331, 0.1404, 0.0196, 0.0696, 0.0498, 0.1435, 0.0623], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0159, 0.0185, 0.0130, 0.0164, 0.0196, 0.0194, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-29 21:40:22,373 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.657e+02 2.510e+02 3.083e+02 4.151e+02 9.991e+02, threshold=6.165e+02, percent-clipped=13.0 2023-04-29 21:40:46,683 INFO [train.py:904] (4/8) Epoch 14, batch 50, loss[loss=0.1743, simple_loss=0.2635, pruned_loss=0.04252, over 15895.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.279, pruned_loss=0.05926, over 754115.79 frames. ], batch size: 35, lr: 4.96e-03, grad_scale: 1.0 2023-04-29 21:41:27,207 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=132031.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:41:48,864 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6855, 2.1930, 2.1955, 3.2864, 2.3214, 3.5368, 1.3631, 2.6385], device='cuda:4'), covar=tensor([0.1479, 0.0833, 0.1218, 0.0158, 0.0182, 0.0432, 0.1690, 0.0875], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0162, 0.0183, 0.0154, 0.0190, 0.0205, 0.0188, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 21:41:55,078 INFO [train.py:904] (4/8) Epoch 14, batch 100, loss[loss=0.1619, simple_loss=0.2429, pruned_loss=0.04048, over 16773.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.2743, pruned_loss=0.05439, over 1329108.56 frames. ], batch size: 39, lr: 4.96e-03, grad_scale: 1.0 2023-04-29 21:41:58,484 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=132054.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:42:13,230 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132065.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:42:22,311 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-29 21:42:23,844 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132072.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:42:44,827 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.577e+02 2.362e+02 2.775e+02 3.378e+02 5.633e+02, threshold=5.550e+02, percent-clipped=0.0 2023-04-29 21:42:51,105 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=132092.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:43:04,249 INFO [train.py:904] (4/8) Epoch 14, batch 150, loss[loss=0.168, simple_loss=0.2665, pruned_loss=0.0347, over 17264.00 frames. ], tot_loss[loss=0.1898, simple_loss=0.2732, pruned_loss=0.05324, over 1766238.16 frames. ], batch size: 52, lr: 4.96e-03, grad_scale: 1.0 2023-04-29 21:43:20,790 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=132113.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:43:23,984 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=132115.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 21:43:30,561 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=132120.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:44:07,543 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8503, 4.4940, 4.8519, 5.0060, 5.2242, 4.6428, 5.1917, 5.1987], device='cuda:4'), covar=tensor([0.1543, 0.1287, 0.1562, 0.0741, 0.0495, 0.0802, 0.0533, 0.0602], device='cuda:4'), in_proj_covar=tensor([0.0543, 0.0666, 0.0799, 0.0682, 0.0515, 0.0523, 0.0545, 0.0628], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:44:08,577 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132148.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:44:15,119 INFO [train.py:904] (4/8) Epoch 14, batch 200, loss[loss=0.1574, simple_loss=0.2415, pruned_loss=0.03664, over 16811.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2714, pruned_loss=0.05184, over 2115851.95 frames. ], batch size: 42, lr: 4.96e-03, grad_scale: 1.0 2023-04-29 21:44:15,660 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8125, 1.7349, 2.2962, 2.6360, 2.6397, 2.6219, 1.6963, 2.9378], device='cuda:4'), covar=tensor([0.0145, 0.0437, 0.0282, 0.0214, 0.0221, 0.0217, 0.0448, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0177, 0.0159, 0.0162, 0.0173, 0.0129, 0.0178, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 21:44:27,911 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-29 21:45:02,144 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.381e+02 2.781e+02 3.514e+02 5.258e+02, threshold=5.562e+02, percent-clipped=0.0 2023-04-29 21:45:15,516 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=132196.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:45:21,486 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-29 21:45:23,095 INFO [train.py:904] (4/8) Epoch 14, batch 250, loss[loss=0.1903, simple_loss=0.2641, pruned_loss=0.05818, over 16707.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2703, pruned_loss=0.05222, over 2388260.05 frames. ], batch size: 134, lr: 4.96e-03, grad_scale: 1.0 2023-04-29 21:45:38,789 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=132213.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:46:33,334 INFO [train.py:904] (4/8) Epoch 14, batch 300, loss[loss=0.1819, simple_loss=0.2534, pruned_loss=0.05519, over 16749.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.267, pruned_loss=0.05043, over 2598084.62 frames. ], batch size: 124, lr: 4.96e-03, grad_scale: 1.0 2023-04-29 21:47:16,156 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7637, 2.8004, 2.2017, 2.7025, 3.2249, 3.0462, 3.5462, 3.4355], device='cuda:4'), covar=tensor([0.0105, 0.0365, 0.0521, 0.0357, 0.0230, 0.0305, 0.0249, 0.0197], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0216, 0.0209, 0.0208, 0.0214, 0.0215, 0.0216, 0.0203], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:47:22,674 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.582e+02 2.346e+02 2.777e+02 3.394e+02 6.651e+02, threshold=5.554e+02, percent-clipped=2.0 2023-04-29 21:47:43,454 INFO [train.py:904] (4/8) Epoch 14, batch 350, loss[loss=0.1567, simple_loss=0.2433, pruned_loss=0.03507, over 16974.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2644, pruned_loss=0.04961, over 2753069.36 frames. ], batch size: 41, lr: 4.95e-03, grad_scale: 1.0 2023-04-29 21:48:00,507 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9314, 1.9879, 2.4308, 2.9513, 2.6674, 3.2160, 2.1430, 3.2473], device='cuda:4'), covar=tensor([0.0182, 0.0375, 0.0258, 0.0218, 0.0232, 0.0147, 0.0369, 0.0111], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0176, 0.0159, 0.0162, 0.0173, 0.0129, 0.0177, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 21:48:51,145 INFO [train.py:904] (4/8) Epoch 14, batch 400, loss[loss=0.1898, simple_loss=0.2702, pruned_loss=0.05471, over 16811.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2643, pruned_loss=0.04967, over 2885511.55 frames. ], batch size: 102, lr: 4.95e-03, grad_scale: 2.0 2023-04-29 21:49:38,425 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.217e+02 2.653e+02 3.223e+02 5.439e+02, threshold=5.306e+02, percent-clipped=0.0 2023-04-29 21:49:39,466 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=132387.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:49:58,664 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6499, 4.6978, 5.1323, 5.0989, 5.1068, 4.7600, 4.7596, 4.5297], device='cuda:4'), covar=tensor([0.0365, 0.0597, 0.0384, 0.0414, 0.0401, 0.0395, 0.0810, 0.0500], device='cuda:4'), in_proj_covar=tensor([0.0352, 0.0367, 0.0368, 0.0352, 0.0411, 0.0395, 0.0483, 0.0315], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 21:50:00,122 INFO [train.py:904] (4/8) Epoch 14, batch 450, loss[loss=0.1833, simple_loss=0.2797, pruned_loss=0.04347, over 16558.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2631, pruned_loss=0.04865, over 2984924.17 frames. ], batch size: 68, lr: 4.95e-03, grad_scale: 2.0 2023-04-29 21:50:11,962 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=132410.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 21:50:35,482 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7184, 2.8735, 2.4467, 2.6642, 3.0884, 2.9516, 3.5035, 3.2608], device='cuda:4'), covar=tensor([0.0097, 0.0295, 0.0403, 0.0328, 0.0236, 0.0291, 0.0203, 0.0219], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0216, 0.0210, 0.0209, 0.0214, 0.0215, 0.0218, 0.0205], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:51:11,603 INFO [train.py:904] (4/8) Epoch 14, batch 500, loss[loss=0.1783, simple_loss=0.2736, pruned_loss=0.04149, over 17112.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2623, pruned_loss=0.04837, over 3051147.86 frames. ], batch size: 49, lr: 4.95e-03, grad_scale: 2.0 2023-04-29 21:51:35,546 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0897, 3.2123, 3.3301, 2.2433, 2.9582, 2.3165, 3.5695, 3.4641], device='cuda:4'), covar=tensor([0.0217, 0.0771, 0.0525, 0.1577, 0.0711, 0.0951, 0.0501, 0.0865], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0144, 0.0157, 0.0144, 0.0136, 0.0124, 0.0134, 0.0156], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 21:51:58,582 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.766e+02 2.377e+02 2.811e+02 3.534e+02 1.282e+03, threshold=5.622e+02, percent-clipped=6.0 2023-04-29 21:52:19,293 INFO [train.py:904] (4/8) Epoch 14, batch 550, loss[loss=0.1699, simple_loss=0.2645, pruned_loss=0.03767, over 17134.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2608, pruned_loss=0.04774, over 3102751.36 frames. ], batch size: 48, lr: 4.95e-03, grad_scale: 2.0 2023-04-29 21:52:34,307 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132513.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:52:46,504 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-29 21:53:28,147 INFO [train.py:904] (4/8) Epoch 14, batch 600, loss[loss=0.1699, simple_loss=0.2675, pruned_loss=0.03614, over 17034.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2599, pruned_loss=0.04744, over 3160106.92 frames. ], batch size: 50, lr: 4.95e-03, grad_scale: 2.0 2023-04-29 21:53:41,855 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=132561.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:53:46,346 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9323, 2.2518, 2.3326, 2.8455, 2.3611, 3.2256, 1.7665, 2.6782], device='cuda:4'), covar=tensor([0.1170, 0.0668, 0.1024, 0.0175, 0.0167, 0.0404, 0.1352, 0.0724], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0162, 0.0182, 0.0157, 0.0194, 0.0207, 0.0187, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 21:54:17,820 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.404e+02 2.200e+02 2.653e+02 3.173e+02 6.384e+02, threshold=5.306e+02, percent-clipped=1.0 2023-04-29 21:54:39,140 INFO [train.py:904] (4/8) Epoch 14, batch 650, loss[loss=0.1638, simple_loss=0.2439, pruned_loss=0.04192, over 16671.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.259, pruned_loss=0.0468, over 3191356.05 frames. ], batch size: 134, lr: 4.95e-03, grad_scale: 2.0 2023-04-29 21:55:49,508 INFO [train.py:904] (4/8) Epoch 14, batch 700, loss[loss=0.1807, simple_loss=0.2559, pruned_loss=0.05273, over 16783.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2585, pruned_loss=0.04656, over 3226719.63 frames. ], batch size: 102, lr: 4.95e-03, grad_scale: 2.0 2023-04-29 21:56:01,344 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=132660.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:56:38,037 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.642e+02 2.258e+02 2.690e+02 3.168e+02 6.172e+02, threshold=5.381e+02, percent-clipped=1.0 2023-04-29 21:56:39,019 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132687.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:56:58,534 INFO [train.py:904] (4/8) Epoch 14, batch 750, loss[loss=0.2133, simple_loss=0.2853, pruned_loss=0.07069, over 16436.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2592, pruned_loss=0.04632, over 3257339.80 frames. ], batch size: 146, lr: 4.95e-03, grad_scale: 2.0 2023-04-29 21:57:09,217 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132710.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 21:57:24,379 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=132721.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:57:43,135 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=132735.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:57:59,758 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=132746.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:58:07,420 INFO [train.py:904] (4/8) Epoch 14, batch 800, loss[loss=0.1792, simple_loss=0.2729, pruned_loss=0.04272, over 16715.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2586, pruned_loss=0.0462, over 3277457.10 frames. ], batch size: 57, lr: 4.95e-03, grad_scale: 4.0 2023-04-29 21:58:16,055 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=132758.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:58:54,911 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.513e+02 2.382e+02 2.716e+02 3.142e+02 6.467e+02, threshold=5.432e+02, percent-clipped=2.0 2023-04-29 21:59:05,788 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-29 21:59:15,856 INFO [train.py:904] (4/8) Epoch 14, batch 850, loss[loss=0.1645, simple_loss=0.2609, pruned_loss=0.03411, over 16714.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2586, pruned_loss=0.04627, over 3282401.18 frames. ], batch size: 57, lr: 4.94e-03, grad_scale: 4.0 2023-04-29 21:59:23,465 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=132807.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 21:59:26,717 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4208, 5.3714, 5.2515, 4.7440, 4.8345, 5.2828, 5.2888, 4.8773], device='cuda:4'), covar=tensor([0.0531, 0.0408, 0.0271, 0.0282, 0.1079, 0.0376, 0.0291, 0.0683], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0357, 0.0316, 0.0294, 0.0328, 0.0342, 0.0214, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 21:59:37,232 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5811, 4.6394, 4.7942, 4.6788, 4.6819, 5.2904, 4.8276, 4.5414], device='cuda:4'), covar=tensor([0.1533, 0.1884, 0.2232, 0.2215, 0.3233, 0.1086, 0.1450, 0.2595], device='cuda:4'), in_proj_covar=tensor([0.0370, 0.0524, 0.0577, 0.0444, 0.0605, 0.0601, 0.0457, 0.0598], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 22:00:24,478 INFO [train.py:904] (4/8) Epoch 14, batch 900, loss[loss=0.2033, simple_loss=0.2733, pruned_loss=0.06664, over 16971.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2581, pruned_loss=0.04626, over 3286712.00 frames. ], batch size: 109, lr: 4.94e-03, grad_scale: 4.0 2023-04-29 22:01:14,368 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.740e+02 2.299e+02 2.623e+02 3.116e+02 6.405e+02, threshold=5.245e+02, percent-clipped=2.0 2023-04-29 22:01:34,126 INFO [train.py:904] (4/8) Epoch 14, batch 950, loss[loss=0.1803, simple_loss=0.2546, pruned_loss=0.05303, over 16778.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2582, pruned_loss=0.04611, over 3293098.56 frames. ], batch size: 102, lr: 4.94e-03, grad_scale: 4.0 2023-04-29 22:02:28,746 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-29 22:02:41,756 INFO [train.py:904] (4/8) Epoch 14, batch 1000, loss[loss=0.1633, simple_loss=0.2539, pruned_loss=0.03635, over 17235.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2573, pruned_loss=0.04563, over 3306587.90 frames. ], batch size: 52, lr: 4.94e-03, grad_scale: 4.0 2023-04-29 22:03:29,372 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.118e+02 2.524e+02 3.195e+02 6.689e+02, threshold=5.047e+02, percent-clipped=1.0 2023-04-29 22:03:49,998 INFO [train.py:904] (4/8) Epoch 14, batch 1050, loss[loss=0.1886, simple_loss=0.2595, pruned_loss=0.05888, over 16886.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2573, pruned_loss=0.04592, over 3316427.09 frames. ], batch size: 109, lr: 4.94e-03, grad_scale: 4.0 2023-04-29 22:04:06,886 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8016, 2.8868, 2.4416, 2.7504, 3.1470, 2.9933, 3.6255, 3.4301], device='cuda:4'), covar=tensor([0.0086, 0.0284, 0.0369, 0.0303, 0.0205, 0.0266, 0.0162, 0.0176], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0218, 0.0211, 0.0210, 0.0217, 0.0218, 0.0222, 0.0210], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:04:10,852 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133016.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:04:10,975 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8049, 3.7147, 4.1299, 1.9276, 4.3401, 4.3319, 3.1332, 3.2291], device='cuda:4'), covar=tensor([0.0697, 0.0215, 0.0197, 0.1174, 0.0063, 0.0151, 0.0399, 0.0415], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0103, 0.0090, 0.0138, 0.0070, 0.0113, 0.0122, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 22:04:59,792 INFO [train.py:904] (4/8) Epoch 14, batch 1100, loss[loss=0.1676, simple_loss=0.2584, pruned_loss=0.0384, over 17185.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2565, pruned_loss=0.04524, over 3313526.22 frames. ], batch size: 46, lr: 4.94e-03, grad_scale: 4.0 2023-04-29 22:05:26,419 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6691, 2.6957, 2.3756, 2.5034, 2.9403, 2.7872, 3.4362, 3.1995], device='cuda:4'), covar=tensor([0.0109, 0.0319, 0.0391, 0.0331, 0.0246, 0.0329, 0.0193, 0.0229], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0217, 0.0210, 0.0208, 0.0216, 0.0218, 0.0222, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:05:39,990 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9492, 3.2824, 2.7356, 5.2293, 4.4761, 4.4326, 1.7397, 3.2744], device='cuda:4'), covar=tensor([0.1112, 0.0561, 0.1068, 0.0131, 0.0199, 0.0433, 0.1298, 0.0714], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0162, 0.0182, 0.0159, 0.0195, 0.0207, 0.0185, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 22:05:47,408 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.639e+02 2.252e+02 2.587e+02 3.047e+02 6.829e+02, threshold=5.174e+02, percent-clipped=2.0 2023-04-29 22:05:47,925 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0280, 1.9371, 2.4691, 2.9679, 2.6797, 3.3461, 2.2568, 3.3365], device='cuda:4'), covar=tensor([0.0194, 0.0407, 0.0289, 0.0236, 0.0277, 0.0143, 0.0371, 0.0124], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0177, 0.0160, 0.0165, 0.0175, 0.0131, 0.0177, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-29 22:06:08,349 INFO [train.py:904] (4/8) Epoch 14, batch 1150, loss[loss=0.1822, simple_loss=0.2558, pruned_loss=0.05428, over 16885.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2555, pruned_loss=0.04466, over 3309205.74 frames. ], batch size: 116, lr: 4.94e-03, grad_scale: 4.0 2023-04-29 22:06:08,668 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133102.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:06:20,906 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8933, 4.9768, 5.4773, 5.4509, 5.3994, 5.0469, 5.0256, 4.7882], device='cuda:4'), covar=tensor([0.0325, 0.0594, 0.0330, 0.0403, 0.0515, 0.0435, 0.0974, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0372, 0.0391, 0.0391, 0.0371, 0.0434, 0.0417, 0.0509, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 22:07:00,056 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-29 22:07:14,162 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6359, 5.0342, 5.3874, 5.3546, 5.3538, 5.0087, 4.6669, 4.6922], device='cuda:4'), covar=tensor([0.0595, 0.0697, 0.0519, 0.0661, 0.0698, 0.0649, 0.1427, 0.0541], device='cuda:4'), in_proj_covar=tensor([0.0373, 0.0392, 0.0392, 0.0372, 0.0435, 0.0418, 0.0510, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 22:07:16,503 INFO [train.py:904] (4/8) Epoch 14, batch 1200, loss[loss=0.1617, simple_loss=0.2504, pruned_loss=0.03652, over 17042.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2546, pruned_loss=0.04441, over 3310320.46 frames. ], batch size: 50, lr: 4.94e-03, grad_scale: 8.0 2023-04-29 22:08:05,294 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7603, 2.7305, 2.4184, 2.7012, 3.0863, 2.9891, 3.5595, 3.3132], device='cuda:4'), covar=tensor([0.0109, 0.0326, 0.0383, 0.0316, 0.0220, 0.0269, 0.0206, 0.0213], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0218, 0.0212, 0.0210, 0.0218, 0.0219, 0.0224, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:08:05,977 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.586e+02 2.506e+02 2.966e+02 3.654e+02 5.614e+02, threshold=5.932e+02, percent-clipped=4.0 2023-04-29 22:08:12,968 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0290, 4.4371, 4.2530, 3.2310, 3.7226, 4.2028, 4.0098, 2.2053], device='cuda:4'), covar=tensor([0.0425, 0.0063, 0.0048, 0.0318, 0.0130, 0.0124, 0.0098, 0.0535], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0073, 0.0073, 0.0128, 0.0085, 0.0094, 0.0082, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 22:08:28,179 INFO [train.py:904] (4/8) Epoch 14, batch 1250, loss[loss=0.1273, simple_loss=0.211, pruned_loss=0.02175, over 15926.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2546, pruned_loss=0.0452, over 3317048.27 frames. ], batch size: 35, lr: 4.94e-03, grad_scale: 8.0 2023-04-29 22:09:37,994 INFO [train.py:904] (4/8) Epoch 14, batch 1300, loss[loss=0.1817, simple_loss=0.2703, pruned_loss=0.04656, over 17104.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2551, pruned_loss=0.04521, over 3314492.76 frames. ], batch size: 47, lr: 4.94e-03, grad_scale: 8.0 2023-04-29 22:10:27,126 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.451e+02 2.160e+02 2.631e+02 3.090e+02 7.401e+02, threshold=5.263e+02, percent-clipped=1.0 2023-04-29 22:10:48,345 INFO [train.py:904] (4/8) Epoch 14, batch 1350, loss[loss=0.1957, simple_loss=0.2669, pruned_loss=0.06228, over 16762.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2551, pruned_loss=0.04499, over 3308763.61 frames. ], batch size: 83, lr: 4.94e-03, grad_scale: 8.0 2023-04-29 22:11:06,800 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=133316.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:11:27,648 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0817, 5.0425, 4.9156, 4.4812, 4.4985, 4.9966, 4.9034, 4.6271], device='cuda:4'), covar=tensor([0.0549, 0.0452, 0.0293, 0.0299, 0.1012, 0.0385, 0.0382, 0.0676], device='cuda:4'), in_proj_covar=tensor([0.0269, 0.0366, 0.0325, 0.0303, 0.0338, 0.0350, 0.0220, 0.0375], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:11:27,741 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4753, 2.6241, 2.0872, 2.3004, 2.9421, 2.6371, 3.1742, 3.1232], device='cuda:4'), covar=tensor([0.0121, 0.0336, 0.0470, 0.0422, 0.0224, 0.0321, 0.0207, 0.0209], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0218, 0.0211, 0.0210, 0.0216, 0.0219, 0.0224, 0.0210], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:11:56,451 INFO [train.py:904] (4/8) Epoch 14, batch 1400, loss[loss=0.1608, simple_loss=0.2524, pruned_loss=0.03465, over 17197.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2549, pruned_loss=0.04508, over 3318198.37 frames. ], batch size: 46, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:12:13,483 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=133364.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:12:26,441 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133374.0, num_to_drop=1, layers_to_drop={0} 2023-04-29 22:12:30,051 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133376.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:12:44,804 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.632e+02 2.235e+02 2.580e+02 3.131e+02 5.456e+02, threshold=5.161e+02, percent-clipped=1.0 2023-04-29 22:13:06,248 INFO [train.py:904] (4/8) Epoch 14, batch 1450, loss[loss=0.1652, simple_loss=0.2405, pruned_loss=0.04499, over 15585.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.254, pruned_loss=0.04455, over 3315096.51 frames. ], batch size: 191, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:13:06,591 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=133402.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:13:49,401 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2146, 5.1855, 4.9964, 4.4704, 5.0628, 1.8882, 4.7883, 4.9769], device='cuda:4'), covar=tensor([0.0081, 0.0074, 0.0159, 0.0333, 0.0083, 0.2460, 0.0131, 0.0159], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0134, 0.0179, 0.0165, 0.0150, 0.0194, 0.0169, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:13:52,419 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=133435.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 22:13:55,264 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=133437.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:14:12,863 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=133450.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:14:14,808 INFO [train.py:904] (4/8) Epoch 14, batch 1500, loss[loss=0.1795, simple_loss=0.2747, pruned_loss=0.0422, over 17045.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2542, pruned_loss=0.04493, over 3320884.35 frames. ], batch size: 50, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:14:46,635 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133474.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:14:57,941 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2130, 2.1754, 2.7282, 3.1642, 2.9878, 3.4570, 2.3677, 3.6145], device='cuda:4'), covar=tensor([0.0173, 0.0377, 0.0245, 0.0216, 0.0221, 0.0171, 0.0371, 0.0118], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0179, 0.0163, 0.0167, 0.0177, 0.0133, 0.0179, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:15:03,755 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.705e+02 2.267e+02 2.734e+02 3.162e+02 6.055e+02, threshold=5.467e+02, percent-clipped=3.0 2023-04-29 22:15:18,612 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9520, 4.5178, 4.7779, 5.1503, 5.2857, 4.7035, 5.3327, 5.2820], device='cuda:4'), covar=tensor([0.1709, 0.1578, 0.2275, 0.0902, 0.0818, 0.0864, 0.0687, 0.0831], device='cuda:4'), in_proj_covar=tensor([0.0592, 0.0731, 0.0877, 0.0749, 0.0562, 0.0575, 0.0588, 0.0692], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:15:23,303 INFO [train.py:904] (4/8) Epoch 14, batch 1550, loss[loss=0.1793, simple_loss=0.2743, pruned_loss=0.04215, over 17076.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2558, pruned_loss=0.04593, over 3318265.99 frames. ], batch size: 53, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:16:10,848 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=133535.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:16:33,559 INFO [train.py:904] (4/8) Epoch 14, batch 1600, loss[loss=0.1597, simple_loss=0.2557, pruned_loss=0.0319, over 17193.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2576, pruned_loss=0.04607, over 3322831.07 frames. ], batch size: 46, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:16:59,171 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0597, 3.1113, 1.8922, 3.2740, 2.3403, 3.2870, 2.0759, 2.5593], device='cuda:4'), covar=tensor([0.0272, 0.0365, 0.1441, 0.0251, 0.0796, 0.0527, 0.1240, 0.0655], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0170, 0.0191, 0.0147, 0.0170, 0.0212, 0.0199, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 22:17:21,876 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.622e+02 2.312e+02 2.830e+02 3.166e+02 6.146e+02, threshold=5.659e+02, percent-clipped=1.0 2023-04-29 22:17:29,418 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0578, 4.4556, 3.2743, 2.2509, 2.8928, 2.6428, 4.7432, 3.8797], device='cuda:4'), covar=tensor([0.2386, 0.0575, 0.1474, 0.2678, 0.2665, 0.1793, 0.0313, 0.1136], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0261, 0.0289, 0.0287, 0.0279, 0.0231, 0.0275, 0.0311], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 22:17:43,027 INFO [train.py:904] (4/8) Epoch 14, batch 1650, loss[loss=0.2153, simple_loss=0.3049, pruned_loss=0.06285, over 16662.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2593, pruned_loss=0.04706, over 3322481.80 frames. ], batch size: 62, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:17:47,492 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-29 22:18:05,754 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-29 22:18:52,510 INFO [train.py:904] (4/8) Epoch 14, batch 1700, loss[loss=0.1933, simple_loss=0.2671, pruned_loss=0.05978, over 16778.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2617, pruned_loss=0.04806, over 3327510.11 frames. ], batch size: 134, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:18:53,012 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133652.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:19:16,187 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-29 22:19:21,760 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-29 22:19:42,412 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.666e+02 2.547e+02 3.320e+02 4.199e+02 1.315e+03, threshold=6.641e+02, percent-clipped=10.0 2023-04-29 22:20:03,031 INFO [train.py:904] (4/8) Epoch 14, batch 1750, loss[loss=0.1564, simple_loss=0.2464, pruned_loss=0.0332, over 17260.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2618, pruned_loss=0.04793, over 3325189.23 frames. ], batch size: 45, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:20:19,360 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=133713.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:20:42,446 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133730.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 22:20:45,407 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133732.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:21:12,608 INFO [train.py:904] (4/8) Epoch 14, batch 1800, loss[loss=0.1836, simple_loss=0.2784, pruned_loss=0.04439, over 16720.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2633, pruned_loss=0.04822, over 3327247.54 frames. ], batch size: 62, lr: 4.93e-03, grad_scale: 8.0 2023-04-29 22:22:00,848 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.275e+02 2.807e+02 3.369e+02 7.697e+02, threshold=5.615e+02, percent-clipped=1.0 2023-04-29 22:22:07,637 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1134, 3.7971, 4.3077, 2.0401, 4.5206, 4.5578, 3.1553, 3.5133], device='cuda:4'), covar=tensor([0.0623, 0.0226, 0.0207, 0.1129, 0.0059, 0.0149, 0.0412, 0.0343], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0102, 0.0089, 0.0136, 0.0070, 0.0113, 0.0121, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 22:22:22,390 INFO [train.py:904] (4/8) Epoch 14, batch 1850, loss[loss=0.1532, simple_loss=0.2479, pruned_loss=0.02925, over 17029.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2641, pruned_loss=0.04812, over 3324407.81 frames. ], batch size: 41, lr: 4.93e-03, grad_scale: 4.0 2023-04-29 22:22:51,334 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5391, 4.5089, 4.4424, 3.9352, 4.4690, 1.6633, 4.2131, 4.1835], device='cuda:4'), covar=tensor([0.0109, 0.0098, 0.0158, 0.0317, 0.0102, 0.2573, 0.0146, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0134, 0.0181, 0.0168, 0.0152, 0.0194, 0.0171, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:23:02,956 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133830.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:23:14,112 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3839, 4.2040, 4.5506, 2.3025, 4.7642, 4.7905, 3.5005, 3.7907], device='cuda:4'), covar=tensor([0.0592, 0.0192, 0.0232, 0.1038, 0.0063, 0.0142, 0.0372, 0.0314], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0103, 0.0089, 0.0137, 0.0070, 0.0113, 0.0122, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 22:23:32,975 INFO [train.py:904] (4/8) Epoch 14, batch 1900, loss[loss=0.1978, simple_loss=0.2779, pruned_loss=0.05886, over 15402.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2631, pruned_loss=0.04752, over 3320832.16 frames. ], batch size: 190, lr: 4.93e-03, grad_scale: 4.0 2023-04-29 22:24:11,809 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9763, 1.7628, 2.4648, 2.8572, 2.7201, 3.2021, 2.0003, 3.2280], device='cuda:4'), covar=tensor([0.0176, 0.0425, 0.0278, 0.0228, 0.0242, 0.0152, 0.0431, 0.0125], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0180, 0.0163, 0.0168, 0.0177, 0.0135, 0.0180, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:24:23,478 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.361e+02 2.126e+02 2.452e+02 2.920e+02 9.757e+02, threshold=4.904e+02, percent-clipped=2.0 2023-04-29 22:24:42,153 INFO [train.py:904] (4/8) Epoch 14, batch 1950, loss[loss=0.2056, simple_loss=0.2815, pruned_loss=0.06485, over 16878.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2632, pruned_loss=0.04734, over 3331285.55 frames. ], batch size: 109, lr: 4.92e-03, grad_scale: 4.0 2023-04-29 22:25:52,904 INFO [train.py:904] (4/8) Epoch 14, batch 2000, loss[loss=0.2295, simple_loss=0.3064, pruned_loss=0.07625, over 11899.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2626, pruned_loss=0.04686, over 3328733.46 frames. ], batch size: 247, lr: 4.92e-03, grad_scale: 8.0 2023-04-29 22:26:00,322 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-29 22:26:43,618 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.687e+02 2.429e+02 2.777e+02 3.533e+02 6.577e+02, threshold=5.554e+02, percent-clipped=5.0 2023-04-29 22:26:57,104 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133997.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:27:04,351 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6960, 4.4654, 4.7485, 4.9218, 5.0414, 4.4777, 4.9736, 5.0370], device='cuda:4'), covar=tensor([0.1535, 0.1142, 0.1364, 0.0664, 0.0483, 0.0970, 0.0865, 0.0540], device='cuda:4'), in_proj_covar=tensor([0.0596, 0.0738, 0.0886, 0.0756, 0.0567, 0.0584, 0.0595, 0.0699], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:27:06,903 INFO [train.py:904] (4/8) Epoch 14, batch 2050, loss[loss=0.1834, simple_loss=0.2767, pruned_loss=0.04501, over 16798.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2629, pruned_loss=0.04706, over 3321763.84 frames. ], batch size: 57, lr: 4.92e-03, grad_scale: 8.0 2023-04-29 22:27:15,545 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=134008.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:27:46,379 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134030.0, num_to_drop=1, layers_to_drop={2} 2023-04-29 22:27:49,233 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134032.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:28:16,827 INFO [train.py:904] (4/8) Epoch 14, batch 2100, loss[loss=0.2268, simple_loss=0.3051, pruned_loss=0.07423, over 15590.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2629, pruned_loss=0.04693, over 3321803.78 frames. ], batch size: 190, lr: 4.92e-03, grad_scale: 8.0 2023-04-29 22:28:25,645 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=134058.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:28:54,050 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=134078.0, num_to_drop=1, layers_to_drop={1} 2023-04-29 22:28:56,312 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=134080.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:29:09,011 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.453e+02 2.284e+02 2.721e+02 3.276e+02 7.722e+02, threshold=5.442e+02, percent-clipped=1.0 2023-04-29 22:29:26,681 INFO [train.py:904] (4/8) Epoch 14, batch 2150, loss[loss=0.2161, simple_loss=0.289, pruned_loss=0.07158, over 16437.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2647, pruned_loss=0.04805, over 3307517.34 frames. ], batch size: 146, lr: 4.92e-03, grad_scale: 4.0 2023-04-29 22:30:06,080 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134130.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:30:35,960 INFO [train.py:904] (4/8) Epoch 14, batch 2200, loss[loss=0.1902, simple_loss=0.2719, pruned_loss=0.05421, over 15549.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2648, pruned_loss=0.0483, over 3316422.40 frames. ], batch size: 190, lr: 4.92e-03, grad_scale: 4.0 2023-04-29 22:31:12,755 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=134178.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:31:20,096 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1830, 5.1231, 5.0098, 4.5460, 4.6695, 5.0211, 4.9867, 4.6214], device='cuda:4'), covar=tensor([0.0529, 0.0448, 0.0263, 0.0337, 0.0956, 0.0468, 0.0315, 0.0709], device='cuda:4'), in_proj_covar=tensor([0.0273, 0.0373, 0.0331, 0.0310, 0.0345, 0.0359, 0.0224, 0.0384], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 22:31:22,442 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4537, 3.5491, 1.8725, 3.7565, 2.6092, 3.6623, 2.0700, 2.8285], device='cuda:4'), covar=tensor([0.0216, 0.0288, 0.1588, 0.0249, 0.0747, 0.0596, 0.1398, 0.0622], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0172, 0.0191, 0.0147, 0.0170, 0.0213, 0.0199, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 22:31:27,737 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.617e+02 2.496e+02 2.998e+02 3.803e+02 8.330e+02, threshold=5.996e+02, percent-clipped=6.0 2023-04-29 22:31:40,905 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-29 22:31:45,970 INFO [train.py:904] (4/8) Epoch 14, batch 2250, loss[loss=0.1887, simple_loss=0.262, pruned_loss=0.05769, over 16881.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2653, pruned_loss=0.0488, over 3322412.43 frames. ], batch size: 109, lr: 4.92e-03, grad_scale: 4.0 2023-04-29 22:32:43,622 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8506, 1.9187, 2.3836, 2.6808, 2.6643, 2.6278, 1.8204, 2.8819], device='cuda:4'), covar=tensor([0.0153, 0.0380, 0.0267, 0.0250, 0.0246, 0.0241, 0.0443, 0.0127], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0181, 0.0165, 0.0171, 0.0179, 0.0136, 0.0182, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:32:54,979 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-29 22:32:56,453 INFO [train.py:904] (4/8) Epoch 14, batch 2300, loss[loss=0.1613, simple_loss=0.241, pruned_loss=0.04083, over 15866.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2654, pruned_loss=0.04887, over 3315188.55 frames. ], batch size: 35, lr: 4.92e-03, grad_scale: 4.0 2023-04-29 22:33:06,773 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9460, 3.0181, 2.5109, 2.7907, 3.2516, 2.9906, 3.7297, 3.4836], device='cuda:4'), covar=tensor([0.0088, 0.0280, 0.0410, 0.0329, 0.0212, 0.0289, 0.0186, 0.0190], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0220, 0.0213, 0.0212, 0.0219, 0.0221, 0.0228, 0.0214], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:33:48,101 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.417e+02 2.438e+02 2.937e+02 3.292e+02 7.337e+02, threshold=5.875e+02, percent-clipped=1.0 2023-04-29 22:34:06,356 INFO [train.py:904] (4/8) Epoch 14, batch 2350, loss[loss=0.1699, simple_loss=0.2678, pruned_loss=0.03597, over 17142.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2659, pruned_loss=0.04896, over 3322509.00 frames. ], batch size: 48, lr: 4.92e-03, grad_scale: 4.0 2023-04-29 22:34:09,674 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=134304.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:34:15,129 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134308.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:34:16,991 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=134309.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:34:45,076 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5080, 2.3147, 2.3979, 4.2971, 2.1929, 2.6945, 2.3698, 2.4663], device='cuda:4'), covar=tensor([0.1105, 0.3327, 0.2505, 0.0466, 0.3841, 0.2343, 0.3275, 0.3080], device='cuda:4'), in_proj_covar=tensor([0.0379, 0.0412, 0.0345, 0.0328, 0.0420, 0.0477, 0.0378, 0.0484], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:35:17,012 INFO [train.py:904] (4/8) Epoch 14, batch 2400, loss[loss=0.1816, simple_loss=0.2768, pruned_loss=0.04322, over 17063.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2665, pruned_loss=0.04897, over 3321283.92 frames. ], batch size: 55, lr: 4.92e-03, grad_scale: 8.0 2023-04-29 22:35:19,203 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=134353.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:35:19,309 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2657, 4.0734, 4.2691, 4.4723, 4.5958, 4.0947, 4.2815, 4.5398], device='cuda:4'), covar=tensor([0.1465, 0.1094, 0.1455, 0.0689, 0.0578, 0.1358, 0.2142, 0.0693], device='cuda:4'), in_proj_covar=tensor([0.0609, 0.0751, 0.0901, 0.0771, 0.0578, 0.0597, 0.0609, 0.0710], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:35:19,396 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5304, 3.6420, 2.0833, 3.8086, 2.7280, 3.7324, 2.2235, 2.9880], device='cuda:4'), covar=tensor([0.0223, 0.0298, 0.1372, 0.0238, 0.0697, 0.0705, 0.1267, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0172, 0.0191, 0.0148, 0.0171, 0.0213, 0.0200, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 22:35:23,354 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=134356.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:35:23,558 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7636, 1.8502, 2.3177, 2.6962, 2.6118, 2.5287, 1.7680, 2.8156], device='cuda:4'), covar=tensor([0.0141, 0.0373, 0.0255, 0.0170, 0.0209, 0.0239, 0.0408, 0.0115], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0181, 0.0166, 0.0171, 0.0180, 0.0136, 0.0181, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:35:35,539 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=134365.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:35:42,553 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=134370.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:35:54,660 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.64 vs. limit=5.0 2023-04-29 22:36:08,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.471e+02 2.303e+02 2.657e+02 3.180e+02 6.909e+02, threshold=5.314e+02, percent-clipped=1.0 2023-04-29 22:36:25,718 INFO [train.py:904] (4/8) Epoch 14, batch 2450, loss[loss=0.1654, simple_loss=0.2469, pruned_loss=0.042, over 16785.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2666, pruned_loss=0.04835, over 3327421.47 frames. ], batch size: 102, lr: 4.92e-03, grad_scale: 4.0 2023-04-29 22:36:47,391 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6287, 6.0884, 5.6986, 5.8820, 5.3813, 5.1093, 5.4021, 6.1839], device='cuda:4'), covar=tensor([0.1129, 0.0780, 0.1171, 0.0681, 0.0859, 0.0722, 0.1024, 0.0832], device='cuda:4'), in_proj_covar=tensor([0.0601, 0.0751, 0.0612, 0.0539, 0.0477, 0.0478, 0.0626, 0.0575], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:37:35,055 INFO [train.py:904] (4/8) Epoch 14, batch 2500, loss[loss=0.1819, simple_loss=0.2576, pruned_loss=0.05313, over 16815.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2664, pruned_loss=0.04787, over 3324999.06 frames. ], batch size: 83, lr: 4.91e-03, grad_scale: 4.0 2023-04-29 22:38:08,081 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2824, 5.2527, 5.1369, 4.7207, 4.7091, 5.1907, 5.1707, 4.7724], device='cuda:4'), covar=tensor([0.0562, 0.0449, 0.0259, 0.0287, 0.1129, 0.0379, 0.0288, 0.0669], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0373, 0.0329, 0.0309, 0.0345, 0.0356, 0.0224, 0.0384], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 22:38:28,207 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.634e+02 2.356e+02 2.745e+02 3.374e+02 4.761e+02, threshold=5.489e+02, percent-clipped=0.0 2023-04-29 22:38:45,444 INFO [train.py:904] (4/8) Epoch 14, batch 2550, loss[loss=0.1849, simple_loss=0.2677, pruned_loss=0.05106, over 16543.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2666, pruned_loss=0.04775, over 3326106.11 frames. ], batch size: 75, lr: 4.91e-03, grad_scale: 4.0 2023-04-29 22:39:00,820 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-29 22:39:55,150 INFO [train.py:904] (4/8) Epoch 14, batch 2600, loss[loss=0.1948, simple_loss=0.273, pruned_loss=0.0583, over 16488.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2669, pruned_loss=0.04822, over 3326692.40 frames. ], batch size: 75, lr: 4.91e-03, grad_scale: 4.0 2023-04-29 22:40:37,379 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4926, 2.2738, 2.2374, 4.3054, 2.2244, 2.6823, 2.3601, 2.5143], device='cuda:4'), covar=tensor([0.1061, 0.3389, 0.2525, 0.0440, 0.3760, 0.2338, 0.3247, 0.3189], device='cuda:4'), in_proj_covar=tensor([0.0379, 0.0412, 0.0345, 0.0327, 0.0420, 0.0477, 0.0378, 0.0484], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:40:45,092 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2020, 5.1768, 4.9686, 4.4378, 5.0980, 1.9932, 4.7785, 4.9397], device='cuda:4'), covar=tensor([0.0082, 0.0077, 0.0180, 0.0346, 0.0077, 0.2353, 0.0120, 0.0161], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0135, 0.0182, 0.0169, 0.0153, 0.0194, 0.0172, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:40:46,904 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.457e+02 2.364e+02 2.744e+02 3.247e+02 7.768e+02, threshold=5.488e+02, percent-clipped=4.0 2023-04-29 22:41:03,691 INFO [train.py:904] (4/8) Epoch 14, batch 2650, loss[loss=0.2014, simple_loss=0.287, pruned_loss=0.05795, over 16792.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2681, pruned_loss=0.04841, over 3321973.70 frames. ], batch size: 83, lr: 4.91e-03, grad_scale: 4.0 2023-04-29 22:42:12,181 INFO [train.py:904] (4/8) Epoch 14, batch 2700, loss[loss=0.159, simple_loss=0.2463, pruned_loss=0.03588, over 16836.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2674, pruned_loss=0.04737, over 3328673.73 frames. ], batch size: 39, lr: 4.91e-03, grad_scale: 4.0 2023-04-29 22:42:13,612 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134653.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:42:23,622 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=134660.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:42:24,949 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4795, 2.1965, 1.6025, 1.9552, 2.5958, 2.3706, 2.6281, 2.7377], device='cuda:4'), covar=tensor([0.0160, 0.0362, 0.0534, 0.0410, 0.0213, 0.0305, 0.0216, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0222, 0.0216, 0.0214, 0.0221, 0.0222, 0.0231, 0.0216], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:42:31,341 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=134665.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:43:04,243 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.644e+02 2.219e+02 2.559e+02 3.067e+02 4.567e+02, threshold=5.119e+02, percent-clipped=0.0 2023-04-29 22:43:04,801 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6339, 2.2050, 2.3200, 4.5336, 2.2005, 2.5789, 2.3164, 2.3875], device='cuda:4'), covar=tensor([0.1001, 0.3615, 0.2644, 0.0378, 0.4026, 0.2758, 0.3308, 0.3667], device='cuda:4'), in_proj_covar=tensor([0.0378, 0.0412, 0.0345, 0.0327, 0.0420, 0.0477, 0.0377, 0.0483], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:43:20,430 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=134701.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:43:21,290 INFO [train.py:904] (4/8) Epoch 14, batch 2750, loss[loss=0.1693, simple_loss=0.2641, pruned_loss=0.03726, over 17190.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2678, pruned_loss=0.04718, over 3333221.68 frames. ], batch size: 46, lr: 4.91e-03, grad_scale: 4.0 2023-04-29 22:43:56,445 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-29 22:44:00,669 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4927, 4.3071, 4.5851, 4.7137, 4.8419, 4.3487, 4.7168, 4.8248], device='cuda:4'), covar=tensor([0.1573, 0.1186, 0.1360, 0.0714, 0.0626, 0.1134, 0.1552, 0.0657], device='cuda:4'), in_proj_covar=tensor([0.0602, 0.0744, 0.0894, 0.0764, 0.0574, 0.0590, 0.0602, 0.0702], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:44:29,010 INFO [train.py:904] (4/8) Epoch 14, batch 2800, loss[loss=0.1659, simple_loss=0.2565, pruned_loss=0.03768, over 17181.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2673, pruned_loss=0.04728, over 3329998.57 frames. ], batch size: 46, lr: 4.91e-03, grad_scale: 8.0 2023-04-29 22:44:29,947 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-29 22:44:42,363 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-29 22:45:20,153 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.581e+02 2.461e+02 2.848e+02 3.331e+02 6.185e+02, threshold=5.697e+02, percent-clipped=2.0 2023-04-29 22:45:37,177 INFO [train.py:904] (4/8) Epoch 14, batch 2850, loss[loss=0.1896, simple_loss=0.2786, pruned_loss=0.05033, over 16662.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2672, pruned_loss=0.04732, over 3327504.69 frames. ], batch size: 62, lr: 4.91e-03, grad_scale: 8.0 2023-04-29 22:46:45,299 INFO [train.py:904] (4/8) Epoch 14, batch 2900, loss[loss=0.1553, simple_loss=0.2428, pruned_loss=0.03391, over 17226.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2659, pruned_loss=0.0476, over 3329780.00 frames. ], batch size: 44, lr: 4.91e-03, grad_scale: 8.0 2023-04-29 22:46:58,774 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=134862.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:47:04,307 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0950, 3.2732, 2.7805, 4.7031, 3.8827, 4.4126, 1.6422, 3.1217], device='cuda:4'), covar=tensor([0.1147, 0.0558, 0.0958, 0.0166, 0.0219, 0.0366, 0.1389, 0.0732], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0162, 0.0182, 0.0163, 0.0199, 0.0210, 0.0186, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 22:47:18,741 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.59 vs. limit=5.0 2023-04-29 22:47:36,338 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.583e+02 2.493e+02 2.912e+02 3.442e+02 5.810e+02, threshold=5.824e+02, percent-clipped=2.0 2023-04-29 22:47:54,149 INFO [train.py:904] (4/8) Epoch 14, batch 2950, loss[loss=0.1811, simple_loss=0.263, pruned_loss=0.04959, over 16566.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2651, pruned_loss=0.04832, over 3317471.31 frames. ], batch size: 68, lr: 4.91e-03, grad_scale: 8.0 2023-04-29 22:48:24,082 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=134923.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:49:02,685 INFO [train.py:904] (4/8) Epoch 14, batch 3000, loss[loss=0.2082, simple_loss=0.273, pruned_loss=0.07169, over 16892.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2649, pruned_loss=0.04882, over 3320398.96 frames. ], batch size: 109, lr: 4.91e-03, grad_scale: 8.0 2023-04-29 22:49:02,686 INFO [train.py:929] (4/8) Computing validation loss 2023-04-29 22:49:11,374 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4919, 4.5102, 4.8930, 4.8704, 4.8750, 4.5819, 4.5344, 4.3879], device='cuda:4'), covar=tensor([0.0324, 0.0479, 0.0369, 0.0388, 0.0376, 0.0333, 0.0848, 0.0399], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0392, 0.0394, 0.0374, 0.0437, 0.0418, 0.0511, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 22:49:12,420 INFO [train.py:938] (4/8) Epoch 14, validation: loss=0.1382, simple_loss=0.244, pruned_loss=0.01621, over 944034.00 frames. 2023-04-29 22:49:12,421 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-29 22:49:24,935 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134960.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:49:31,406 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134965.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:50:06,983 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.405e+02 2.394e+02 2.911e+02 3.370e+02 7.406e+02, threshold=5.821e+02, percent-clipped=1.0 2023-04-29 22:50:24,227 INFO [train.py:904] (4/8) Epoch 14, batch 3050, loss[loss=0.1908, simple_loss=0.2736, pruned_loss=0.05396, over 16478.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2643, pruned_loss=0.04884, over 3317531.77 frames. ], batch size: 75, lr: 4.90e-03, grad_scale: 8.0 2023-04-29 22:50:33,222 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=135008.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:50:40,114 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=135013.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:50:40,489 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-29 22:50:46,656 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-29 22:51:32,216 INFO [train.py:904] (4/8) Epoch 14, batch 3100, loss[loss=0.1912, simple_loss=0.2546, pruned_loss=0.06395, over 16851.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2633, pruned_loss=0.04829, over 3323974.35 frames. ], batch size: 90, lr: 4.90e-03, grad_scale: 8.0 2023-04-29 22:51:35,483 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0346, 5.3949, 5.0985, 5.1217, 4.8318, 4.7880, 4.8391, 5.4593], device='cuda:4'), covar=tensor([0.1205, 0.0904, 0.1107, 0.0873, 0.0953, 0.0968, 0.1149, 0.1021], device='cuda:4'), in_proj_covar=tensor([0.0603, 0.0752, 0.0616, 0.0542, 0.0480, 0.0478, 0.0628, 0.0582], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:52:24,972 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.254e+02 2.683e+02 3.212e+02 6.924e+02, threshold=5.366e+02, percent-clipped=1.0 2023-04-29 22:52:40,973 INFO [train.py:904] (4/8) Epoch 14, batch 3150, loss[loss=0.225, simple_loss=0.2909, pruned_loss=0.07957, over 16818.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2617, pruned_loss=0.04776, over 3330054.65 frames. ], batch size: 124, lr: 4.90e-03, grad_scale: 8.0 2023-04-29 22:52:52,053 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-29 22:53:29,020 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-29 22:53:50,739 INFO [train.py:904] (4/8) Epoch 14, batch 3200, loss[loss=0.1537, simple_loss=0.2361, pruned_loss=0.03562, over 16767.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2619, pruned_loss=0.0474, over 3319956.18 frames. ], batch size: 102, lr: 4.90e-03, grad_scale: 8.0 2023-04-29 22:54:17,387 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=135171.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:54:41,146 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.733e+02 2.470e+02 2.854e+02 3.612e+02 7.577e+02, threshold=5.709e+02, percent-clipped=5.0 2023-04-29 22:54:59,015 INFO [train.py:904] (4/8) Epoch 14, batch 3250, loss[loss=0.1847, simple_loss=0.2709, pruned_loss=0.04923, over 16420.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2627, pruned_loss=0.04808, over 3314315.64 frames. ], batch size: 68, lr: 4.90e-03, grad_scale: 8.0 2023-04-29 22:55:22,124 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=135218.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:55:41,805 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=135232.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 22:56:07,846 INFO [train.py:904] (4/8) Epoch 14, batch 3300, loss[loss=0.176, simple_loss=0.2709, pruned_loss=0.04058, over 17065.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2632, pruned_loss=0.04812, over 3315713.23 frames. ], batch size: 55, lr: 4.90e-03, grad_scale: 4.0 2023-04-29 22:56:23,780 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.27 vs. limit=5.0 2023-04-29 22:57:01,617 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.361e+02 2.198e+02 2.720e+02 3.178e+02 6.678e+02, threshold=5.440e+02, percent-clipped=1.0 2023-04-29 22:57:16,993 INFO [train.py:904] (4/8) Epoch 14, batch 3350, loss[loss=0.1642, simple_loss=0.2452, pruned_loss=0.04162, over 15950.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2637, pruned_loss=0.04765, over 3318454.91 frames. ], batch size: 35, lr: 4.90e-03, grad_scale: 4.0 2023-04-29 22:58:24,247 INFO [train.py:904] (4/8) Epoch 14, batch 3400, loss[loss=0.1598, simple_loss=0.2528, pruned_loss=0.03344, over 17194.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2642, pruned_loss=0.04759, over 3320145.65 frames. ], batch size: 46, lr: 4.90e-03, grad_scale: 4.0 2023-04-29 22:59:01,441 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7749, 4.0575, 4.2207, 2.9000, 3.5362, 4.1849, 3.7867, 2.4322], device='cuda:4'), covar=tensor([0.0361, 0.0075, 0.0035, 0.0307, 0.0104, 0.0072, 0.0070, 0.0392], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0073, 0.0073, 0.0127, 0.0085, 0.0094, 0.0083, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 22:59:16,983 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.551e+02 2.205e+02 2.551e+02 3.042e+02 5.815e+02, threshold=5.103e+02, percent-clipped=2.0 2023-04-29 22:59:20,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0204, 4.7449, 4.9909, 5.2349, 5.4188, 4.7511, 5.4375, 5.3765], device='cuda:4'), covar=tensor([0.1803, 0.1188, 0.1778, 0.0700, 0.0522, 0.0902, 0.0453, 0.0591], device='cuda:4'), in_proj_covar=tensor([0.0601, 0.0745, 0.0899, 0.0765, 0.0576, 0.0588, 0.0602, 0.0704], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:59:32,657 INFO [train.py:904] (4/8) Epoch 14, batch 3450, loss[loss=0.1975, simple_loss=0.2865, pruned_loss=0.05426, over 16521.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2637, pruned_loss=0.04768, over 3299235.19 frames. ], batch size: 68, lr: 4.90e-03, grad_scale: 4.0 2023-04-29 22:59:35,423 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9491, 2.9441, 2.6429, 4.5294, 3.6250, 4.2888, 1.6441, 3.0942], device='cuda:4'), covar=tensor([0.1237, 0.0707, 0.1067, 0.0229, 0.0257, 0.0382, 0.1507, 0.0742], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0164, 0.0183, 0.0164, 0.0202, 0.0212, 0.0188, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 22:59:43,693 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2516, 5.7796, 5.8731, 5.6818, 5.7055, 6.2573, 5.8604, 5.5671], device='cuda:4'), covar=tensor([0.0770, 0.1916, 0.2207, 0.2002, 0.2651, 0.0921, 0.1292, 0.2157], device='cuda:4'), in_proj_covar=tensor([0.0382, 0.0545, 0.0595, 0.0463, 0.0627, 0.0624, 0.0472, 0.0617], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 22:59:48,832 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9193, 5.3288, 5.0290, 5.0663, 4.7867, 4.8149, 4.7563, 5.3892], device='cuda:4'), covar=tensor([0.1242, 0.0869, 0.1029, 0.0757, 0.0837, 0.0883, 0.1050, 0.0876], device='cuda:4'), in_proj_covar=tensor([0.0609, 0.0759, 0.0622, 0.0547, 0.0485, 0.0481, 0.0632, 0.0584], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 22:59:57,244 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-29 23:00:41,082 INFO [train.py:904] (4/8) Epoch 14, batch 3500, loss[loss=0.1511, simple_loss=0.2357, pruned_loss=0.03328, over 16992.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2622, pruned_loss=0.04736, over 3302624.21 frames. ], batch size: 41, lr: 4.90e-03, grad_scale: 4.0 2023-04-29 23:00:47,536 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3294, 3.6972, 3.8216, 2.0873, 3.0086, 2.4452, 3.7679, 3.8183], device='cuda:4'), covar=tensor([0.0319, 0.0735, 0.0500, 0.1862, 0.0796, 0.0967, 0.0651, 0.0975], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0156, 0.0163, 0.0148, 0.0140, 0.0128, 0.0141, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 23:01:04,943 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-29 23:01:33,341 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9440, 3.7354, 4.1275, 1.7895, 4.1817, 4.3467, 3.3094, 3.3574], device='cuda:4'), covar=tensor([0.0654, 0.0225, 0.0201, 0.1291, 0.0125, 0.0185, 0.0367, 0.0378], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0104, 0.0090, 0.0137, 0.0072, 0.0115, 0.0122, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 23:01:37,086 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.468e+02 2.212e+02 2.743e+02 3.344e+02 6.080e+02, threshold=5.486e+02, percent-clipped=2.0 2023-04-29 23:01:51,741 INFO [train.py:904] (4/8) Epoch 14, batch 3550, loss[loss=0.1618, simple_loss=0.258, pruned_loss=0.03276, over 17249.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2616, pruned_loss=0.04653, over 3318823.35 frames. ], batch size: 52, lr: 4.90e-03, grad_scale: 4.0 2023-04-29 23:02:15,165 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=135518.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:02:27,914 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=135527.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:03:02,251 INFO [train.py:904] (4/8) Epoch 14, batch 3600, loss[loss=0.1393, simple_loss=0.2214, pruned_loss=0.02861, over 16833.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2601, pruned_loss=0.04614, over 3314862.53 frames. ], batch size: 42, lr: 4.89e-03, grad_scale: 8.0 2023-04-29 23:03:22,564 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=135566.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:03:58,636 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.573e+02 2.279e+02 2.683e+02 3.136e+02 7.106e+02, threshold=5.366e+02, percent-clipped=1.0 2023-04-29 23:04:14,252 INFO [train.py:904] (4/8) Epoch 14, batch 3650, loss[loss=0.1727, simple_loss=0.2446, pruned_loss=0.05041, over 16879.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2589, pruned_loss=0.04671, over 3308848.35 frames. ], batch size: 116, lr: 4.89e-03, grad_scale: 8.0 2023-04-29 23:04:20,161 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-29 23:04:57,255 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=135630.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:05:31,030 INFO [train.py:904] (4/8) Epoch 14, batch 3700, loss[loss=0.1653, simple_loss=0.2432, pruned_loss=0.04372, over 15568.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2582, pruned_loss=0.04824, over 3279029.92 frames. ], batch size: 191, lr: 4.89e-03, grad_scale: 4.0 2023-04-29 23:06:30,611 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=135691.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:06:32,146 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.568e+02 2.199e+02 2.631e+02 3.126e+02 8.071e+02, threshold=5.262e+02, percent-clipped=3.0 2023-04-29 23:06:46,317 INFO [train.py:904] (4/8) Epoch 14, batch 3750, loss[loss=0.1962, simple_loss=0.2632, pruned_loss=0.06456, over 16728.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2586, pruned_loss=0.04989, over 3273363.90 frames. ], batch size: 134, lr: 4.89e-03, grad_scale: 4.0 2023-04-29 23:07:30,554 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7871, 4.0524, 3.1151, 2.3184, 2.8085, 2.5572, 4.2183, 3.7096], device='cuda:4'), covar=tensor([0.2563, 0.0592, 0.1591, 0.2567, 0.2394, 0.1745, 0.0461, 0.1048], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0264, 0.0291, 0.0291, 0.0289, 0.0234, 0.0277, 0.0316], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:08:01,576 INFO [train.py:904] (4/8) Epoch 14, batch 3800, loss[loss=0.1788, simple_loss=0.2548, pruned_loss=0.05136, over 16882.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2605, pruned_loss=0.05134, over 3256955.65 frames. ], batch size: 116, lr: 4.89e-03, grad_scale: 4.0 2023-04-29 23:09:00,855 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.772e+02 2.313e+02 2.548e+02 3.011e+02 5.523e+02, threshold=5.096e+02, percent-clipped=2.0 2023-04-29 23:09:06,431 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.50 vs. limit=5.0 2023-04-29 23:09:16,026 INFO [train.py:904] (4/8) Epoch 14, batch 3850, loss[loss=0.1944, simple_loss=0.2633, pruned_loss=0.06279, over 16702.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2608, pruned_loss=0.05197, over 3259989.80 frames. ], batch size: 124, lr: 4.89e-03, grad_scale: 4.0 2023-04-29 23:09:31,449 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7405, 4.6945, 4.7080, 4.1758, 4.7088, 1.8997, 4.5025, 4.4073], device='cuda:4'), covar=tensor([0.0115, 0.0102, 0.0151, 0.0317, 0.0088, 0.2339, 0.0129, 0.0177], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0135, 0.0182, 0.0170, 0.0154, 0.0193, 0.0173, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:09:43,646 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4012, 2.2101, 2.3391, 4.2723, 2.2451, 2.5878, 2.2638, 2.4465], device='cuda:4'), covar=tensor([0.1097, 0.3302, 0.2399, 0.0440, 0.3469, 0.2318, 0.3453, 0.2555], device='cuda:4'), in_proj_covar=tensor([0.0380, 0.0417, 0.0347, 0.0329, 0.0423, 0.0481, 0.0381, 0.0490], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:09:53,681 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=135827.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:10:30,025 INFO [train.py:904] (4/8) Epoch 14, batch 3900, loss[loss=0.1648, simple_loss=0.2436, pruned_loss=0.04302, over 16864.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2597, pruned_loss=0.052, over 3266555.35 frames. ], batch size: 116, lr: 4.89e-03, grad_scale: 4.0 2023-04-29 23:11:05,160 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=135875.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:11:30,216 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.696e+02 2.144e+02 2.448e+02 3.272e+02 6.091e+02, threshold=4.896e+02, percent-clipped=2.0 2023-04-29 23:11:40,133 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6125, 4.5971, 4.7495, 4.6461, 4.6166, 5.2308, 4.7720, 4.4955], device='cuda:4'), covar=tensor([0.1514, 0.2175, 0.2420, 0.2214, 0.2902, 0.1037, 0.1526, 0.2431], device='cuda:4'), in_proj_covar=tensor([0.0379, 0.0541, 0.0585, 0.0457, 0.0618, 0.0612, 0.0467, 0.0608], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:11:45,241 INFO [train.py:904] (4/8) Epoch 14, batch 3950, loss[loss=0.2057, simple_loss=0.281, pruned_loss=0.06525, over 12808.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2597, pruned_loss=0.053, over 3261526.41 frames. ], batch size: 246, lr: 4.89e-03, grad_scale: 4.0 2023-04-29 23:12:58,425 INFO [train.py:904] (4/8) Epoch 14, batch 4000, loss[loss=0.217, simple_loss=0.2821, pruned_loss=0.07597, over 16483.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2594, pruned_loss=0.05321, over 3272058.45 frames. ], batch size: 146, lr: 4.89e-03, grad_scale: 8.0 2023-04-29 23:13:35,962 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8072, 4.0669, 2.2775, 4.7789, 3.1263, 4.7683, 2.6853, 3.0650], device='cuda:4'), covar=tensor([0.0235, 0.0295, 0.1685, 0.0077, 0.0696, 0.0180, 0.1287, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0172, 0.0191, 0.0149, 0.0171, 0.0215, 0.0201, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 23:13:41,751 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=135982.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:13:47,683 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=135986.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:13:55,963 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.530e+02 2.249e+02 2.823e+02 3.453e+02 5.778e+02, threshold=5.647e+02, percent-clipped=7.0 2023-04-29 23:13:58,962 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=135994.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:14:13,530 INFO [train.py:904] (4/8) Epoch 14, batch 4050, loss[loss=0.1674, simple_loss=0.2546, pruned_loss=0.04003, over 16317.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2598, pruned_loss=0.05217, over 3265030.89 frames. ], batch size: 165, lr: 4.89e-03, grad_scale: 8.0 2023-04-29 23:15:11,292 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-29 23:15:12,419 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136043.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:15:24,936 INFO [train.py:904] (4/8) Epoch 14, batch 4100, loss[loss=0.1803, simple_loss=0.2691, pruned_loss=0.04572, over 17023.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2616, pruned_loss=0.05168, over 3251895.11 frames. ], batch size: 50, lr: 4.89e-03, grad_scale: 8.0 2023-04-29 23:15:30,464 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136055.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:16:11,128 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-29 23:16:24,202 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.443e+02 1.983e+02 2.350e+02 2.941e+02 4.738e+02, threshold=4.699e+02, percent-clipped=0.0 2023-04-29 23:16:40,524 INFO [train.py:904] (4/8) Epoch 14, batch 4150, loss[loss=0.2353, simple_loss=0.3132, pruned_loss=0.07875, over 17063.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2687, pruned_loss=0.05401, over 3244202.13 frames. ], batch size: 53, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:17:32,359 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8282, 3.7597, 3.9179, 3.6979, 3.8010, 4.2333, 3.9316, 3.6114], device='cuda:4'), covar=tensor([0.1861, 0.1985, 0.1897, 0.2367, 0.2810, 0.1590, 0.1362, 0.2446], device='cuda:4'), in_proj_covar=tensor([0.0372, 0.0532, 0.0574, 0.0449, 0.0607, 0.0602, 0.0460, 0.0598], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:17:41,301 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=136142.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:17:55,762 INFO [train.py:904] (4/8) Epoch 14, batch 4200, loss[loss=0.224, simple_loss=0.3114, pruned_loss=0.06826, over 15561.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2764, pruned_loss=0.05603, over 3213684.84 frames. ], batch size: 191, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:18:10,150 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8378, 3.7459, 3.8630, 4.0040, 4.0652, 3.6097, 3.9904, 4.1099], device='cuda:4'), covar=tensor([0.1355, 0.1031, 0.1314, 0.0692, 0.0552, 0.1918, 0.0839, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0577, 0.0713, 0.0852, 0.0732, 0.0551, 0.0562, 0.0572, 0.0676], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:18:55,501 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.477e+02 2.476e+02 2.870e+02 3.638e+02 7.412e+02, threshold=5.739e+02, percent-clipped=7.0 2023-04-29 23:19:10,103 INFO [train.py:904] (4/8) Epoch 14, batch 4250, loss[loss=0.1841, simple_loss=0.2807, pruned_loss=0.04371, over 16763.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2806, pruned_loss=0.05669, over 3184849.42 frames. ], batch size: 134, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:19:12,600 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136203.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:19:14,418 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9524, 2.6922, 2.6048, 1.9704, 2.5817, 2.7045, 2.6165, 1.6873], device='cuda:4'), covar=tensor([0.0312, 0.0080, 0.0082, 0.0291, 0.0097, 0.0110, 0.0096, 0.0443], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0073, 0.0074, 0.0130, 0.0086, 0.0095, 0.0084, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:19:17,631 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3162, 4.1766, 4.2372, 4.5008, 4.6270, 4.2309, 4.5427, 4.6852], device='cuda:4'), covar=tensor([0.1485, 0.1282, 0.1819, 0.0896, 0.0625, 0.1078, 0.0927, 0.0603], device='cuda:4'), in_proj_covar=tensor([0.0576, 0.0713, 0.0849, 0.0732, 0.0551, 0.0563, 0.0573, 0.0675], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:19:41,760 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-29 23:20:23,644 INFO [train.py:904] (4/8) Epoch 14, batch 4300, loss[loss=0.1956, simple_loss=0.2941, pruned_loss=0.04859, over 16869.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.2815, pruned_loss=0.05568, over 3173673.71 frames. ], batch size: 102, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:20:48,225 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2539, 2.2253, 2.2434, 4.1189, 2.0521, 2.5933, 2.3198, 2.3485], device='cuda:4'), covar=tensor([0.1108, 0.2948, 0.2361, 0.0394, 0.3636, 0.2078, 0.2835, 0.2972], device='cuda:4'), in_proj_covar=tensor([0.0378, 0.0415, 0.0345, 0.0325, 0.0420, 0.0478, 0.0378, 0.0486], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:21:14,589 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136286.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:21:23,665 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.749e+02 2.283e+02 2.653e+02 3.073e+02 4.214e+02, threshold=5.307e+02, percent-clipped=0.0 2023-04-29 23:21:38,131 INFO [train.py:904] (4/8) Epoch 14, batch 4350, loss[loss=0.1891, simple_loss=0.2776, pruned_loss=0.05029, over 17283.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2838, pruned_loss=0.05583, over 3174283.54 frames. ], batch size: 52, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:22:26,307 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=136334.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:22:29,175 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=136335.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:22:33,161 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136338.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:22:51,135 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136350.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:22:53,919 INFO [train.py:904] (4/8) Epoch 14, batch 4400, loss[loss=0.2163, simple_loss=0.2993, pruned_loss=0.06669, over 15411.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2856, pruned_loss=0.05712, over 3168688.77 frames. ], batch size: 191, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:23:23,038 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8143, 5.0541, 5.3144, 5.0775, 5.1016, 5.6860, 5.2184, 4.9009], device='cuda:4'), covar=tensor([0.0986, 0.1478, 0.1646, 0.1672, 0.2132, 0.0793, 0.1150, 0.2061], device='cuda:4'), in_proj_covar=tensor([0.0371, 0.0528, 0.0568, 0.0446, 0.0602, 0.0597, 0.0456, 0.0595], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:23:51,148 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.596e+02 2.242e+02 2.686e+02 3.350e+02 5.764e+02, threshold=5.372e+02, percent-clipped=1.0 2023-04-29 23:23:57,347 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136396.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:24:06,498 INFO [train.py:904] (4/8) Epoch 14, batch 4450, loss[loss=0.2016, simple_loss=0.2881, pruned_loss=0.05754, over 16781.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2887, pruned_loss=0.058, over 3176914.60 frames. ], batch size: 39, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:25:18,607 INFO [train.py:904] (4/8) Epoch 14, batch 4500, loss[loss=0.2214, simple_loss=0.3041, pruned_loss=0.06934, over 16708.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2889, pruned_loss=0.05851, over 3171941.13 frames. ], batch size: 134, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:25:27,291 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8637, 3.9064, 4.2450, 2.2719, 3.3197, 2.5707, 4.0917, 4.1140], device='cuda:4'), covar=tensor([0.0162, 0.0625, 0.0399, 0.1880, 0.0741, 0.0907, 0.0501, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0152, 0.0161, 0.0146, 0.0139, 0.0126, 0.0139, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-29 23:25:55,227 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.78 vs. limit=5.0 2023-04-29 23:26:18,140 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.582e+02 2.032e+02 2.340e+02 2.822e+02 5.483e+02, threshold=4.680e+02, percent-clipped=1.0 2023-04-29 23:26:27,082 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136498.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:26:31,838 INFO [train.py:904] (4/8) Epoch 14, batch 4550, loss[loss=0.1903, simple_loss=0.274, pruned_loss=0.0533, over 17149.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2893, pruned_loss=0.05912, over 3190095.50 frames. ], batch size: 46, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:26:54,786 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-29 23:27:12,662 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5333, 3.5878, 1.5805, 4.0553, 2.5707, 3.8922, 1.7092, 2.6848], device='cuda:4'), covar=tensor([0.0203, 0.0268, 0.1884, 0.0132, 0.0749, 0.0368, 0.1833, 0.0698], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0169, 0.0188, 0.0141, 0.0168, 0.0210, 0.0196, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 23:27:45,693 INFO [train.py:904] (4/8) Epoch 14, batch 4600, loss[loss=0.2152, simple_loss=0.2934, pruned_loss=0.06852, over 17049.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2907, pruned_loss=0.05939, over 3205307.92 frames. ], batch size: 41, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:28:42,904 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.386e+02 2.145e+02 2.405e+02 2.956e+02 4.611e+02, threshold=4.810e+02, percent-clipped=0.0 2023-04-29 23:28:57,101 INFO [train.py:904] (4/8) Epoch 14, batch 4650, loss[loss=0.1942, simple_loss=0.2836, pruned_loss=0.05234, over 16749.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2902, pruned_loss=0.05958, over 3198413.18 frames. ], batch size: 124, lr: 4.88e-03, grad_scale: 8.0 2023-04-29 23:29:52,214 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136638.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:29:58,842 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=136643.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:30:07,989 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136650.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:30:10,584 INFO [train.py:904] (4/8) Epoch 14, batch 4700, loss[loss=0.2047, simple_loss=0.2945, pruned_loss=0.05749, over 16372.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.2878, pruned_loss=0.05821, over 3215044.13 frames. ], batch size: 146, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:31:01,303 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=136686.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:31:09,100 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136691.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:31:09,907 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.403e+02 2.115e+02 2.410e+02 3.015e+02 5.394e+02, threshold=4.821e+02, percent-clipped=2.0 2023-04-29 23:31:14,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6858, 3.7240, 2.9484, 2.2785, 2.6277, 2.3915, 4.0491, 3.4321], device='cuda:4'), covar=tensor([0.2717, 0.0806, 0.1540, 0.2258, 0.2285, 0.1790, 0.0496, 0.1096], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0260, 0.0290, 0.0289, 0.0285, 0.0231, 0.0276, 0.0311], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:31:18,046 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=136698.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:31:23,574 INFO [train.py:904] (4/8) Epoch 14, batch 4750, loss[loss=0.1609, simple_loss=0.2386, pruned_loss=0.04158, over 17029.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.283, pruned_loss=0.05582, over 3218100.83 frames. ], batch size: 55, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:31:27,821 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136704.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:32:36,926 INFO [train.py:904] (4/8) Epoch 14, batch 4800, loss[loss=0.1744, simple_loss=0.2762, pruned_loss=0.03627, over 16830.00 frames. ], tot_loss[loss=0.1931, simple_loss=0.2792, pruned_loss=0.05346, over 3223908.75 frames. ], batch size: 96, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:33:03,268 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5211, 1.8146, 2.1508, 2.4976, 2.5256, 2.8215, 1.6776, 2.7210], device='cuda:4'), covar=tensor([0.0165, 0.0407, 0.0304, 0.0248, 0.0245, 0.0155, 0.0478, 0.0113], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0177, 0.0162, 0.0170, 0.0178, 0.0134, 0.0179, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:33:24,807 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-29 23:33:36,432 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.299e+02 1.970e+02 2.266e+02 2.773e+02 4.879e+02, threshold=4.532e+02, percent-clipped=1.0 2023-04-29 23:33:46,377 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136798.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:33:52,184 INFO [train.py:904] (4/8) Epoch 14, batch 4850, loss[loss=0.2019, simple_loss=0.2793, pruned_loss=0.06223, over 16353.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2798, pruned_loss=0.0531, over 3213796.09 frames. ], batch size: 35, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:33:52,953 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-04-29 23:33:58,242 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=136806.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:34:56,904 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=136846.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:35:05,045 INFO [train.py:904] (4/8) Epoch 14, batch 4900, loss[loss=0.185, simple_loss=0.269, pruned_loss=0.0505, over 16755.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.2788, pruned_loss=0.05182, over 3200170.80 frames. ], batch size: 83, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:35:28,829 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136867.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:35:48,990 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8053, 3.9747, 1.9810, 4.7468, 2.8702, 4.5655, 2.4362, 2.9787], device='cuda:4'), covar=tensor([0.0244, 0.0296, 0.1876, 0.0087, 0.0834, 0.0313, 0.1429, 0.0717], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0170, 0.0191, 0.0141, 0.0171, 0.0211, 0.0198, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 23:36:02,817 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.552e+02 2.076e+02 2.406e+02 2.771e+02 5.710e+02, threshold=4.811e+02, percent-clipped=1.0 2023-04-29 23:36:18,496 INFO [train.py:904] (4/8) Epoch 14, batch 4950, loss[loss=0.2245, simple_loss=0.3218, pruned_loss=0.06364, over 16658.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2789, pruned_loss=0.05147, over 3212249.31 frames. ], batch size: 134, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:36:48,228 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9913, 4.0671, 3.8724, 3.6188, 3.5643, 3.9549, 3.6419, 3.7242], device='cuda:4'), covar=tensor([0.0497, 0.0454, 0.0276, 0.0267, 0.0769, 0.0395, 0.0917, 0.0528], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0346, 0.0304, 0.0284, 0.0320, 0.0330, 0.0206, 0.0356], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:37:04,034 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4334, 5.7783, 5.4168, 5.5725, 5.1881, 5.0123, 5.0944, 5.8588], device='cuda:4'), covar=tensor([0.1160, 0.0697, 0.1059, 0.0682, 0.0816, 0.0603, 0.0969, 0.0825], device='cuda:4'), in_proj_covar=tensor([0.0574, 0.0708, 0.0582, 0.0512, 0.0455, 0.0458, 0.0596, 0.0553], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:37:24,309 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.09 vs. limit=5.0 2023-04-29 23:37:30,153 INFO [train.py:904] (4/8) Epoch 14, batch 5000, loss[loss=0.1934, simple_loss=0.2882, pruned_loss=0.04925, over 15313.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2812, pruned_loss=0.05191, over 3227375.68 frames. ], batch size: 190, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:37:44,616 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-29 23:37:48,282 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4661, 1.6728, 2.0725, 2.4964, 2.4595, 2.7913, 1.6211, 2.6282], device='cuda:4'), covar=tensor([0.0188, 0.0437, 0.0310, 0.0298, 0.0262, 0.0166, 0.0486, 0.0109], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0177, 0.0162, 0.0170, 0.0177, 0.0134, 0.0179, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:38:24,355 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136991.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:38:25,066 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.688e+02 2.271e+02 2.523e+02 3.075e+02 5.093e+02, threshold=5.046e+02, percent-clipped=1.0 2023-04-29 23:38:29,215 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4171, 2.9354, 2.6162, 2.2507, 2.2468, 2.1890, 2.9089, 2.8221], device='cuda:4'), covar=tensor([0.2171, 0.0768, 0.1445, 0.2092, 0.1915, 0.1731, 0.0506, 0.1074], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0261, 0.0291, 0.0290, 0.0286, 0.0231, 0.0277, 0.0311], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:38:36,204 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136999.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:38:39,688 INFO [train.py:904] (4/8) Epoch 14, batch 5050, loss[loss=0.1896, simple_loss=0.281, pruned_loss=0.04909, over 16471.00 frames. ], tot_loss[loss=0.1927, simple_loss=0.2817, pruned_loss=0.05187, over 3215174.20 frames. ], batch size: 75, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:38:41,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3477, 4.4240, 4.2136, 3.9000, 3.8193, 4.3029, 4.0809, 4.0421], device='cuda:4'), covar=tensor([0.0516, 0.0347, 0.0280, 0.0281, 0.0958, 0.0408, 0.0532, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0251, 0.0350, 0.0306, 0.0287, 0.0323, 0.0334, 0.0207, 0.0359], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:38:49,725 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2021, 2.0760, 2.1560, 3.9675, 2.0054, 2.4552, 2.1946, 2.2667], device='cuda:4'), covar=tensor([0.1148, 0.3281, 0.2561, 0.0439, 0.3829, 0.2511, 0.3144, 0.2988], device='cuda:4'), in_proj_covar=tensor([0.0374, 0.0414, 0.0343, 0.0322, 0.0418, 0.0476, 0.0375, 0.0482], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:39:02,051 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-29 23:39:04,537 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-29 23:39:31,296 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=137039.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:39:32,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5490, 2.1950, 1.7992, 2.0448, 2.5348, 2.2669, 2.4542, 2.7387], device='cuda:4'), covar=tensor([0.0117, 0.0329, 0.0410, 0.0360, 0.0168, 0.0283, 0.0136, 0.0195], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0212, 0.0207, 0.0206, 0.0212, 0.0213, 0.0218, 0.0207], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:39:49,977 INFO [train.py:904] (4/8) Epoch 14, batch 5100, loss[loss=0.1713, simple_loss=0.2494, pruned_loss=0.04655, over 17013.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2805, pruned_loss=0.05134, over 3210013.76 frames. ], batch size: 55, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:40:13,671 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9033, 2.7154, 2.7396, 1.9021, 2.5129, 2.7180, 2.5980, 1.8255], device='cuda:4'), covar=tensor([0.0348, 0.0052, 0.0051, 0.0312, 0.0090, 0.0084, 0.0086, 0.0359], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0073, 0.0073, 0.0129, 0.0086, 0.0095, 0.0084, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:40:45,953 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 2.206e+02 2.593e+02 3.193e+02 6.541e+02, threshold=5.187e+02, percent-clipped=2.0 2023-04-29 23:41:00,264 INFO [train.py:904] (4/8) Epoch 14, batch 5150, loss[loss=0.1675, simple_loss=0.2712, pruned_loss=0.03192, over 16869.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2803, pruned_loss=0.05073, over 3193138.31 frames. ], batch size: 96, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:41:30,866 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2601, 4.2850, 4.4610, 4.2423, 4.3089, 4.8159, 4.3387, 3.9839], device='cuda:4'), covar=tensor([0.1604, 0.1822, 0.1838, 0.1961, 0.2508, 0.1014, 0.1317, 0.2568], device='cuda:4'), in_proj_covar=tensor([0.0370, 0.0518, 0.0558, 0.0439, 0.0591, 0.0587, 0.0447, 0.0589], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-29 23:41:41,030 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-29 23:42:08,571 INFO [train.py:904] (4/8) Epoch 14, batch 5200, loss[loss=0.2166, simple_loss=0.2959, pruned_loss=0.0686, over 12277.00 frames. ], tot_loss[loss=0.1897, simple_loss=0.2788, pruned_loss=0.05027, over 3193179.63 frames. ], batch size: 246, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:42:22,359 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=137162.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:42:27,884 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-29 23:42:37,930 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-29 23:43:03,326 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.418e+02 2.223e+02 2.609e+02 3.158e+02 5.537e+02, threshold=5.218e+02, percent-clipped=1.0 2023-04-29 23:43:18,146 INFO [train.py:904] (4/8) Epoch 14, batch 5250, loss[loss=0.1929, simple_loss=0.2698, pruned_loss=0.05795, over 12348.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2756, pruned_loss=0.04977, over 3189177.70 frames. ], batch size: 246, lr: 4.87e-03, grad_scale: 8.0 2023-04-29 23:44:00,729 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6163, 2.5655, 2.3055, 3.6214, 2.5680, 3.8813, 1.4157, 2.8410], device='cuda:4'), covar=tensor([0.1383, 0.0710, 0.1262, 0.0131, 0.0185, 0.0344, 0.1607, 0.0749], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0162, 0.0184, 0.0160, 0.0199, 0.0207, 0.0186, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 23:44:26,543 INFO [train.py:904] (4/8) Epoch 14, batch 5300, loss[loss=0.1527, simple_loss=0.2359, pruned_loss=0.03477, over 16571.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.2722, pruned_loss=0.04828, over 3196696.88 frames. ], batch size: 68, lr: 4.86e-03, grad_scale: 8.0 2023-04-29 23:44:32,770 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8849, 4.6119, 4.5838, 5.1148, 5.2760, 4.7231, 5.2801, 5.2702], device='cuda:4'), covar=tensor([0.1423, 0.1102, 0.2421, 0.0807, 0.0629, 0.0978, 0.0577, 0.0757], device='cuda:4'), in_proj_covar=tensor([0.0562, 0.0691, 0.0829, 0.0707, 0.0533, 0.0549, 0.0551, 0.0653], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:44:49,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3417, 4.1965, 4.4324, 4.6095, 4.7491, 4.3130, 4.7155, 4.7592], device='cuda:4'), covar=tensor([0.1642, 0.1234, 0.1562, 0.0668, 0.0468, 0.1008, 0.0507, 0.0566], device='cuda:4'), in_proj_covar=tensor([0.0563, 0.0693, 0.0831, 0.0708, 0.0534, 0.0549, 0.0551, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:45:23,603 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 2.002e+02 2.371e+02 2.932e+02 5.680e+02, threshold=4.741e+02, percent-clipped=1.0 2023-04-29 23:45:33,746 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=137299.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:45:35,139 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-29 23:45:36,785 INFO [train.py:904] (4/8) Epoch 14, batch 5350, loss[loss=0.1648, simple_loss=0.2653, pruned_loss=0.03212, over 16868.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2698, pruned_loss=0.04721, over 3206578.72 frames. ], batch size: 102, lr: 4.86e-03, grad_scale: 8.0 2023-04-29 23:46:19,234 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1645, 4.3931, 4.6820, 4.6160, 4.5866, 4.2778, 3.9518, 4.1753], device='cuda:4'), covar=tensor([0.0501, 0.0639, 0.0510, 0.0640, 0.0833, 0.0554, 0.1558, 0.0579], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0375, 0.0372, 0.0357, 0.0420, 0.0397, 0.0492, 0.0318], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-29 23:46:40,460 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=137347.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:46:46,692 INFO [train.py:904] (4/8) Epoch 14, batch 5400, loss[loss=0.189, simple_loss=0.2826, pruned_loss=0.04776, over 15488.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2725, pruned_loss=0.0479, over 3199146.47 frames. ], batch size: 191, lr: 4.86e-03, grad_scale: 8.0 2023-04-29 23:47:17,193 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-29 23:47:37,435 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8146, 3.7922, 4.1865, 1.9909, 4.4288, 4.4687, 3.1349, 3.0487], device='cuda:4'), covar=tensor([0.0691, 0.0215, 0.0185, 0.1185, 0.0041, 0.0064, 0.0372, 0.0435], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0105, 0.0089, 0.0139, 0.0072, 0.0113, 0.0124, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-29 23:47:46,867 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.624e+02 2.156e+02 2.492e+02 3.031e+02 4.895e+02, threshold=4.984e+02, percent-clipped=1.0 2023-04-29 23:47:55,326 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9065, 2.2943, 2.2806, 2.7174, 2.1250, 3.3204, 1.6099, 2.6919], device='cuda:4'), covar=tensor([0.1157, 0.0622, 0.1018, 0.0116, 0.0118, 0.0309, 0.1428, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0163, 0.0185, 0.0160, 0.0200, 0.0208, 0.0187, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-29 23:48:02,464 INFO [train.py:904] (4/8) Epoch 14, batch 5450, loss[loss=0.2195, simple_loss=0.2981, pruned_loss=0.07041, over 12135.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2755, pruned_loss=0.04959, over 3186803.63 frames. ], batch size: 248, lr: 4.86e-03, grad_scale: 4.0 2023-04-29 23:49:17,934 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-29 23:49:19,969 INFO [train.py:904] (4/8) Epoch 14, batch 5500, loss[loss=0.2321, simple_loss=0.3196, pruned_loss=0.07229, over 16886.00 frames. ], tot_loss[loss=0.196, simple_loss=0.2835, pruned_loss=0.05419, over 3182545.27 frames. ], batch size: 109, lr: 4.86e-03, grad_scale: 4.0 2023-04-29 23:49:35,217 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=137462.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:49:49,645 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7942, 1.8729, 2.4051, 2.8092, 2.6279, 3.2584, 1.8922, 3.2213], device='cuda:4'), covar=tensor([0.0177, 0.0370, 0.0241, 0.0233, 0.0217, 0.0112, 0.0399, 0.0076], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0176, 0.0160, 0.0168, 0.0176, 0.0132, 0.0178, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:50:23,834 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.042e+02 3.226e+02 3.760e+02 4.566e+02 8.180e+02, threshold=7.520e+02, percent-clipped=15.0 2023-04-29 23:50:36,991 INFO [train.py:904] (4/8) Epoch 14, batch 5550, loss[loss=0.2676, simple_loss=0.3363, pruned_loss=0.09941, over 16670.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2907, pruned_loss=0.05973, over 3160816.40 frames. ], batch size: 134, lr: 4.86e-03, grad_scale: 4.0 2023-04-29 23:50:51,277 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=137510.0, num_to_drop=0, layers_to_drop=set() 2023-04-29 23:51:55,242 INFO [train.py:904] (4/8) Epoch 14, batch 5600, loss[loss=0.2116, simple_loss=0.2921, pruned_loss=0.06558, over 16714.00 frames. ], tot_loss[loss=0.2125, simple_loss=0.2959, pruned_loss=0.06452, over 3116559.49 frames. ], batch size: 124, lr: 4.86e-03, grad_scale: 8.0 2023-04-29 23:53:03,057 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.471e+02 3.513e+02 4.095e+02 5.435e+02 1.324e+03, threshold=8.190e+02, percent-clipped=8.0 2023-04-29 23:53:17,473 INFO [train.py:904] (4/8) Epoch 14, batch 5650, loss[loss=0.2427, simple_loss=0.3327, pruned_loss=0.07638, over 15501.00 frames. ], tot_loss[loss=0.2189, simple_loss=0.3009, pruned_loss=0.06845, over 3088771.72 frames. ], batch size: 190, lr: 4.86e-03, grad_scale: 8.0 2023-04-29 23:54:32,229 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3619, 4.6377, 4.4204, 4.4283, 4.1579, 4.0773, 4.1926, 4.6969], device='cuda:4'), covar=tensor([0.1130, 0.0821, 0.0997, 0.0809, 0.0772, 0.1381, 0.0986, 0.0849], device='cuda:4'), in_proj_covar=tensor([0.0576, 0.0712, 0.0590, 0.0518, 0.0454, 0.0460, 0.0598, 0.0552], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:54:38,723 INFO [train.py:904] (4/8) Epoch 14, batch 5700, loss[loss=0.2864, simple_loss=0.3453, pruned_loss=0.1138, over 11332.00 frames. ], tot_loss[loss=0.22, simple_loss=0.3015, pruned_loss=0.06926, over 3090447.25 frames. ], batch size: 247, lr: 4.86e-03, grad_scale: 8.0 2023-04-29 23:54:57,542 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4987, 3.5615, 3.2920, 2.9970, 3.1679, 3.4708, 3.2652, 3.2477], device='cuda:4'), covar=tensor([0.0567, 0.0498, 0.0243, 0.0247, 0.0560, 0.0398, 0.1227, 0.0460], device='cuda:4'), in_proj_covar=tensor([0.0256, 0.0351, 0.0306, 0.0288, 0.0325, 0.0337, 0.0209, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:55:45,156 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.087e+02 3.198e+02 3.940e+02 4.824e+02 8.425e+02, threshold=7.880e+02, percent-clipped=1.0 2023-04-29 23:55:49,913 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7669, 1.3380, 1.6879, 1.6445, 1.7345, 1.9416, 1.5717, 1.7561], device='cuda:4'), covar=tensor([0.0180, 0.0277, 0.0148, 0.0176, 0.0173, 0.0130, 0.0281, 0.0090], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0176, 0.0161, 0.0167, 0.0175, 0.0133, 0.0178, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-29 23:55:59,936 INFO [train.py:904] (4/8) Epoch 14, batch 5750, loss[loss=0.2413, simple_loss=0.3284, pruned_loss=0.07707, over 16837.00 frames. ], tot_loss[loss=0.2237, simple_loss=0.3052, pruned_loss=0.07113, over 3070272.94 frames. ], batch size: 116, lr: 4.86e-03, grad_scale: 8.0 2023-04-29 23:57:21,152 INFO [train.py:904] (4/8) Epoch 14, batch 5800, loss[loss=0.2167, simple_loss=0.3153, pruned_loss=0.05906, over 16245.00 frames. ], tot_loss[loss=0.2214, simple_loss=0.3043, pruned_loss=0.0693, over 3080152.83 frames. ], batch size: 165, lr: 4.86e-03, grad_scale: 8.0 2023-04-29 23:58:26,263 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.930e+02 3.076e+02 3.658e+02 4.339e+02 6.692e+02, threshold=7.316e+02, percent-clipped=0.0 2023-04-29 23:58:41,181 INFO [train.py:904] (4/8) Epoch 14, batch 5850, loss[loss=0.2165, simple_loss=0.2904, pruned_loss=0.07126, over 11707.00 frames. ], tot_loss[loss=0.2192, simple_loss=0.3025, pruned_loss=0.06799, over 3074419.67 frames. ], batch size: 248, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:00:03,509 INFO [train.py:904] (4/8) Epoch 14, batch 5900, loss[loss=0.2112, simple_loss=0.3007, pruned_loss=0.06079, over 16650.00 frames. ], tot_loss[loss=0.2193, simple_loss=0.3023, pruned_loss=0.06809, over 3088822.59 frames. ], batch size: 57, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:01:00,836 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1383, 2.1006, 2.1698, 3.8455, 1.9930, 2.5531, 2.1486, 2.2827], device='cuda:4'), covar=tensor([0.1173, 0.3368, 0.2485, 0.0453, 0.3935, 0.2167, 0.3223, 0.2936], device='cuda:4'), in_proj_covar=tensor([0.0370, 0.0407, 0.0338, 0.0317, 0.0415, 0.0467, 0.0371, 0.0473], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:01:04,115 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-30 00:01:10,326 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=137892.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:01:10,915 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.882e+02 2.749e+02 3.330e+02 4.231e+02 8.888e+02, threshold=6.660e+02, percent-clipped=2.0 2023-04-30 00:01:15,047 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=137895.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:01:24,299 INFO [train.py:904] (4/8) Epoch 14, batch 5950, loss[loss=0.2152, simple_loss=0.2953, pruned_loss=0.06756, over 16892.00 frames. ], tot_loss[loss=0.2186, simple_loss=0.3026, pruned_loss=0.06726, over 3071166.00 frames. ], batch size: 109, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:02:01,531 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6904, 3.6134, 4.1842, 1.8777, 4.3095, 4.3311, 3.0930, 2.9971], device='cuda:4'), covar=tensor([0.0736, 0.0211, 0.0131, 0.1162, 0.0046, 0.0103, 0.0374, 0.0420], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0104, 0.0089, 0.0139, 0.0072, 0.0114, 0.0123, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 00:02:01,624 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6366, 3.9846, 2.8839, 2.2343, 2.8439, 2.5103, 4.1353, 3.6572], device='cuda:4'), covar=tensor([0.2774, 0.0642, 0.1659, 0.2302, 0.2321, 0.1703, 0.0459, 0.1004], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0261, 0.0289, 0.0289, 0.0284, 0.0230, 0.0276, 0.0307], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:02:41,528 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3197, 5.6555, 5.3536, 5.3877, 5.0263, 5.0168, 4.9870, 5.7101], device='cuda:4'), covar=tensor([0.1039, 0.0682, 0.0916, 0.0720, 0.0780, 0.0689, 0.0963, 0.0847], device='cuda:4'), in_proj_covar=tensor([0.0577, 0.0713, 0.0590, 0.0521, 0.0455, 0.0463, 0.0598, 0.0551], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:02:44,551 INFO [train.py:904] (4/8) Epoch 14, batch 6000, loss[loss=0.1999, simple_loss=0.2849, pruned_loss=0.05749, over 16842.00 frames. ], tot_loss[loss=0.2174, simple_loss=0.3017, pruned_loss=0.06653, over 3088773.75 frames. ], batch size: 96, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:02:44,552 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 00:02:52,325 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.8580, 5.8543, 5.8016, 5.4120, 5.4285, 5.7987, 5.9256, 5.5429], device='cuda:4'), covar=tensor([0.0352, 0.0208, 0.0155, 0.0212, 0.0767, 0.0276, 0.0110, 0.0488], device='cuda:4'), in_proj_covar=tensor([0.0255, 0.0350, 0.0304, 0.0285, 0.0323, 0.0334, 0.0207, 0.0360], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:02:55,350 INFO [train.py:938] (4/8) Epoch 14, validation: loss=0.1574, simple_loss=0.2706, pruned_loss=0.0221, over 944034.00 frames. 2023-04-30 00:02:55,351 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 00:02:57,535 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=137953.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:03:01,637 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=137956.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:03:47,761 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.81 vs. limit=5.0 2023-04-30 00:03:55,629 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7917, 3.7974, 4.2316, 1.9580, 4.3735, 4.4279, 3.1913, 3.2241], device='cuda:4'), covar=tensor([0.0671, 0.0178, 0.0112, 0.1103, 0.0040, 0.0081, 0.0317, 0.0355], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0103, 0.0088, 0.0137, 0.0071, 0.0113, 0.0122, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 00:03:59,081 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.967e+02 2.895e+02 3.490e+02 4.190e+02 7.002e+02, threshold=6.980e+02, percent-clipped=3.0 2023-04-30 00:04:18,772 INFO [train.py:904] (4/8) Epoch 14, batch 6050, loss[loss=0.2108, simple_loss=0.3007, pruned_loss=0.06039, over 16544.00 frames. ], tot_loss[loss=0.2154, simple_loss=0.2999, pruned_loss=0.06547, over 3115930.18 frames. ], batch size: 62, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:05:03,146 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-30 00:05:29,669 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-30 00:05:40,154 INFO [train.py:904] (4/8) Epoch 14, batch 6100, loss[loss=0.2474, simple_loss=0.3109, pruned_loss=0.09199, over 11692.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2992, pruned_loss=0.06433, over 3122979.03 frames. ], batch size: 250, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:06:45,097 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.655e+02 2.919e+02 3.731e+02 4.841e+02 8.594e+02, threshold=7.462e+02, percent-clipped=5.0 2023-04-30 00:06:59,047 INFO [train.py:904] (4/8) Epoch 14, batch 6150, loss[loss=0.2699, simple_loss=0.3271, pruned_loss=0.1063, over 11656.00 frames. ], tot_loss[loss=0.2141, simple_loss=0.2986, pruned_loss=0.06483, over 3102015.59 frames. ], batch size: 248, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:08:16,192 INFO [train.py:904] (4/8) Epoch 14, batch 6200, loss[loss=0.2421, simple_loss=0.3062, pruned_loss=0.08895, over 11674.00 frames. ], tot_loss[loss=0.2131, simple_loss=0.2972, pruned_loss=0.06455, over 3089303.66 frames. ], batch size: 248, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:09:02,593 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.70 vs. limit=2.0 2023-04-30 00:09:17,888 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.837e+02 2.947e+02 3.424e+02 4.172e+02 1.390e+03, threshold=6.848e+02, percent-clipped=3.0 2023-04-30 00:09:32,459 INFO [train.py:904] (4/8) Epoch 14, batch 6250, loss[loss=0.2042, simple_loss=0.2847, pruned_loss=0.06188, over 16265.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.2966, pruned_loss=0.06399, over 3095191.59 frames. ], batch size: 35, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:10:42,286 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=138248.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:10:46,670 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=138251.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:10:47,496 INFO [train.py:904] (4/8) Epoch 14, batch 6300, loss[loss=0.2137, simple_loss=0.2864, pruned_loss=0.07044, over 11337.00 frames. ], tot_loss[loss=0.2118, simple_loss=0.2965, pruned_loss=0.0636, over 3094255.89 frames. ], batch size: 248, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:11:15,919 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4726, 2.3182, 2.4211, 4.2938, 2.1960, 2.6881, 2.3819, 2.4514], device='cuda:4'), covar=tensor([0.1039, 0.3414, 0.2463, 0.0382, 0.3888, 0.2288, 0.3221, 0.3010], device='cuda:4'), in_proj_covar=tensor([0.0371, 0.0408, 0.0339, 0.0317, 0.0417, 0.0470, 0.0373, 0.0477], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:11:17,685 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138270.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:11:52,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.960e+02 2.966e+02 3.499e+02 4.315e+02 9.401e+02, threshold=6.998e+02, percent-clipped=7.0 2023-04-30 00:12:00,170 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.18 vs. limit=5.0 2023-04-30 00:12:05,888 INFO [train.py:904] (4/8) Epoch 14, batch 6350, loss[loss=0.2716, simple_loss=0.3297, pruned_loss=0.1068, over 11571.00 frames. ], tot_loss[loss=0.2136, simple_loss=0.2972, pruned_loss=0.06498, over 3079996.06 frames. ], batch size: 247, lr: 4.85e-03, grad_scale: 8.0 2023-04-30 00:12:37,638 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138322.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:12:50,030 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138331.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:13:22,144 INFO [train.py:904] (4/8) Epoch 14, batch 6400, loss[loss=0.1821, simple_loss=0.2754, pruned_loss=0.04438, over 16833.00 frames. ], tot_loss[loss=0.2147, simple_loss=0.2972, pruned_loss=0.06611, over 3062545.61 frames. ], batch size: 102, lr: 4.84e-03, grad_scale: 8.0 2023-04-30 00:14:09,310 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138383.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:14:23,619 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.151e+02 3.125e+02 3.740e+02 4.766e+02 1.292e+03, threshold=7.481e+02, percent-clipped=6.0 2023-04-30 00:14:37,172 INFO [train.py:904] (4/8) Epoch 14, batch 6450, loss[loss=0.1926, simple_loss=0.281, pruned_loss=0.05211, over 16854.00 frames. ], tot_loss[loss=0.2138, simple_loss=0.2973, pruned_loss=0.06513, over 3082267.22 frames. ], batch size: 116, lr: 4.84e-03, grad_scale: 8.0 2023-04-30 00:15:34,298 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4732, 3.5161, 1.8567, 4.0625, 2.4975, 4.0010, 2.1921, 2.7844], device='cuda:4'), covar=tensor([0.0258, 0.0391, 0.1831, 0.0136, 0.0859, 0.0509, 0.1528, 0.0749], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0168, 0.0189, 0.0138, 0.0166, 0.0208, 0.0196, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 00:15:54,740 INFO [train.py:904] (4/8) Epoch 14, batch 6500, loss[loss=0.2426, simple_loss=0.3215, pruned_loss=0.08182, over 16717.00 frames. ], tot_loss[loss=0.2113, simple_loss=0.2946, pruned_loss=0.06405, over 3086055.12 frames. ], batch size: 124, lr: 4.84e-03, grad_scale: 4.0 2023-04-30 00:16:29,996 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0752, 2.3611, 2.3889, 2.7527, 2.1046, 3.2033, 1.7493, 2.7569], device='cuda:4'), covar=tensor([0.1107, 0.0558, 0.0994, 0.0147, 0.0132, 0.0417, 0.1422, 0.0619], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0165, 0.0187, 0.0163, 0.0204, 0.0211, 0.0189, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 00:16:39,017 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4573, 2.3705, 2.2998, 4.4547, 2.1684, 2.8249, 2.4191, 2.5887], device='cuda:4'), covar=tensor([0.1053, 0.3210, 0.2483, 0.0365, 0.3742, 0.2214, 0.3065, 0.2840], device='cuda:4'), in_proj_covar=tensor([0.0369, 0.0406, 0.0338, 0.0316, 0.0415, 0.0468, 0.0372, 0.0474], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:16:59,452 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.094e+02 2.847e+02 3.325e+02 4.301e+02 7.980e+02, threshold=6.650e+02, percent-clipped=1.0 2023-04-30 00:17:04,995 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3824, 3.5679, 3.8463, 1.7626, 3.9492, 4.0764, 3.1277, 2.8992], device='cuda:4'), covar=tensor([0.0891, 0.0206, 0.0159, 0.1223, 0.0058, 0.0120, 0.0334, 0.0469], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0105, 0.0090, 0.0139, 0.0073, 0.0114, 0.0124, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 00:17:12,085 INFO [train.py:904] (4/8) Epoch 14, batch 6550, loss[loss=0.2124, simple_loss=0.3082, pruned_loss=0.05827, over 15584.00 frames. ], tot_loss[loss=0.2137, simple_loss=0.2972, pruned_loss=0.0651, over 3070733.88 frames. ], batch size: 191, lr: 4.84e-03, grad_scale: 4.0 2023-04-30 00:18:22,267 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=138548.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:18:25,704 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=138551.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:18:26,459 INFO [train.py:904] (4/8) Epoch 14, batch 6600, loss[loss=0.2764, simple_loss=0.335, pruned_loss=0.1089, over 11280.00 frames. ], tot_loss[loss=0.2167, simple_loss=0.3005, pruned_loss=0.06649, over 3051646.56 frames. ], batch size: 247, lr: 4.84e-03, grad_scale: 4.0 2023-04-30 00:19:29,570 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.028e+02 2.801e+02 3.547e+02 4.436e+02 1.538e+03, threshold=7.095e+02, percent-clipped=5.0 2023-04-30 00:19:33,986 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=138596.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:19:38,208 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=138599.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:19:42,709 INFO [train.py:904] (4/8) Epoch 14, batch 6650, loss[loss=0.2002, simple_loss=0.2935, pruned_loss=0.05343, over 16549.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.301, pruned_loss=0.06738, over 3056176.00 frames. ], batch size: 75, lr: 4.84e-03, grad_scale: 4.0 2023-04-30 00:20:19,649 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=138626.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:20:58,168 INFO [train.py:904] (4/8) Epoch 14, batch 6700, loss[loss=0.2157, simple_loss=0.3005, pruned_loss=0.06542, over 16890.00 frames. ], tot_loss[loss=0.217, simple_loss=0.2995, pruned_loss=0.06723, over 3062433.91 frames. ], batch size: 116, lr: 4.84e-03, grad_scale: 4.0 2023-04-30 00:21:16,454 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3909, 1.5591, 1.9735, 2.2875, 2.4100, 2.6774, 1.5410, 2.5406], device='cuda:4'), covar=tensor([0.0179, 0.0443, 0.0283, 0.0265, 0.0253, 0.0158, 0.0488, 0.0106], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0174, 0.0159, 0.0165, 0.0172, 0.0131, 0.0176, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-30 00:21:40,467 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=138678.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:21:44,858 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138681.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:21:53,058 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138686.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:22:03,781 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.231e+02 3.354e+02 4.045e+02 5.286e+02 1.404e+03, threshold=8.090e+02, percent-clipped=8.0 2023-04-30 00:22:09,111 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138697.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:22:15,571 INFO [train.py:904] (4/8) Epoch 14, batch 6750, loss[loss=0.2329, simple_loss=0.3078, pruned_loss=0.07899, over 15377.00 frames. ], tot_loss[loss=0.2166, simple_loss=0.2987, pruned_loss=0.06722, over 3074642.21 frames. ], batch size: 190, lr: 4.84e-03, grad_scale: 4.0 2023-04-30 00:23:14,921 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138742.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:23:22,381 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138747.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:23:24,887 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1786, 1.9288, 2.4400, 2.9652, 2.9962, 3.4202, 1.8786, 3.3370], device='cuda:4'), covar=tensor([0.0150, 0.0394, 0.0269, 0.0238, 0.0210, 0.0129, 0.0446, 0.0122], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0174, 0.0158, 0.0164, 0.0172, 0.0131, 0.0176, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-30 00:23:29,043 INFO [train.py:904] (4/8) Epoch 14, batch 6800, loss[loss=0.2072, simple_loss=0.2993, pruned_loss=0.05752, over 16769.00 frames. ], tot_loss[loss=0.2159, simple_loss=0.2984, pruned_loss=0.06669, over 3093697.16 frames. ], batch size: 83, lr: 4.84e-03, grad_scale: 8.0 2023-04-30 00:23:38,446 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138758.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:24:34,694 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.741e+02 2.812e+02 3.433e+02 4.059e+02 1.030e+03, threshold=6.866e+02, percent-clipped=2.0 2023-04-30 00:24:46,910 INFO [train.py:904] (4/8) Epoch 14, batch 6850, loss[loss=0.2122, simple_loss=0.3121, pruned_loss=0.05612, over 16756.00 frames. ], tot_loss[loss=0.2172, simple_loss=0.3002, pruned_loss=0.06711, over 3084477.58 frames. ], batch size: 124, lr: 4.84e-03, grad_scale: 8.0 2023-04-30 00:25:32,876 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-30 00:25:45,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2041, 1.4927, 1.8762, 2.0008, 2.2558, 2.3448, 1.5912, 2.2584], device='cuda:4'), covar=tensor([0.0137, 0.0416, 0.0215, 0.0275, 0.0216, 0.0169, 0.0409, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0174, 0.0159, 0.0165, 0.0172, 0.0131, 0.0176, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-30 00:26:02,477 INFO [train.py:904] (4/8) Epoch 14, batch 6900, loss[loss=0.2086, simple_loss=0.3052, pruned_loss=0.05602, over 16836.00 frames. ], tot_loss[loss=0.2189, simple_loss=0.3031, pruned_loss=0.06735, over 3101415.07 frames. ], batch size: 102, lr: 4.84e-03, grad_scale: 8.0 2023-04-30 00:26:37,072 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-30 00:27:03,357 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.65 vs. limit=5.0 2023-04-30 00:27:10,347 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.725e+02 2.815e+02 3.844e+02 4.824e+02 1.140e+03, threshold=7.689e+02, percent-clipped=2.0 2023-04-30 00:27:22,851 INFO [train.py:904] (4/8) Epoch 14, batch 6950, loss[loss=0.1908, simple_loss=0.2795, pruned_loss=0.05108, over 16657.00 frames. ], tot_loss[loss=0.2205, simple_loss=0.304, pruned_loss=0.06845, over 3092985.63 frames. ], batch size: 134, lr: 4.84e-03, grad_scale: 8.0 2023-04-30 00:27:33,809 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138909.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:27:41,252 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0869, 5.8395, 6.1142, 5.7402, 5.8070, 6.3842, 5.9643, 5.7169], device='cuda:4'), covar=tensor([0.0922, 0.1580, 0.1862, 0.1782, 0.2289, 0.0817, 0.1283, 0.2068], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0528, 0.0581, 0.0449, 0.0600, 0.0601, 0.0458, 0.0606], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 00:28:00,045 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=138926.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:28:38,216 INFO [train.py:904] (4/8) Epoch 14, batch 7000, loss[loss=0.199, simple_loss=0.2948, pruned_loss=0.05154, over 16266.00 frames. ], tot_loss[loss=0.219, simple_loss=0.3034, pruned_loss=0.06731, over 3105317.05 frames. ], batch size: 35, lr: 4.83e-03, grad_scale: 8.0 2023-04-30 00:29:06,796 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138970.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 00:29:09,986 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4870, 3.5342, 1.9299, 3.9777, 2.5110, 3.9010, 2.1858, 2.7416], device='cuda:4'), covar=tensor([0.0279, 0.0349, 0.1873, 0.0155, 0.0971, 0.0559, 0.1586, 0.0810], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0169, 0.0191, 0.0140, 0.0167, 0.0211, 0.0199, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 00:29:10,157 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.61 vs. limit=5.0 2023-04-30 00:29:12,849 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=138974.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:29:18,531 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=138978.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:29:42,519 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.027e+02 2.958e+02 3.558e+02 4.270e+02 7.816e+02, threshold=7.117e+02, percent-clipped=1.0 2023-04-30 00:29:55,148 INFO [train.py:904] (4/8) Epoch 14, batch 7050, loss[loss=0.2051, simple_loss=0.2925, pruned_loss=0.0588, over 16247.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.3037, pruned_loss=0.06625, over 3126032.76 frames. ], batch size: 165, lr: 4.83e-03, grad_scale: 8.0 2023-04-30 00:30:33,407 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=139026.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:30:49,913 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139037.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:30:52,155 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6629, 2.6786, 2.3272, 3.8973, 2.8144, 3.8872, 1.4044, 2.8424], device='cuda:4'), covar=tensor([0.1374, 0.0713, 0.1278, 0.0181, 0.0257, 0.0422, 0.1663, 0.0809], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0164, 0.0186, 0.0162, 0.0203, 0.0209, 0.0188, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 00:30:58,271 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139042.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:30:59,675 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1035, 3.3653, 3.3424, 2.1304, 2.7530, 2.3297, 3.6227, 3.5610], device='cuda:4'), covar=tensor([0.0257, 0.0749, 0.0627, 0.1919, 0.0909, 0.0934, 0.0598, 0.1002], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0151, 0.0161, 0.0147, 0.0138, 0.0126, 0.0138, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 00:31:13,407 INFO [train.py:904] (4/8) Epoch 14, batch 7100, loss[loss=0.2195, simple_loss=0.3064, pruned_loss=0.06631, over 16241.00 frames. ], tot_loss[loss=0.2159, simple_loss=0.3016, pruned_loss=0.06511, over 3138014.52 frames. ], batch size: 165, lr: 4.83e-03, grad_scale: 8.0 2023-04-30 00:31:15,012 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139053.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:32:14,633 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.88 vs. limit=5.0 2023-04-30 00:32:18,348 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.051e+02 2.828e+02 3.507e+02 4.047e+02 5.820e+02, threshold=7.014e+02, percent-clipped=0.0 2023-04-30 00:32:31,526 INFO [train.py:904] (4/8) Epoch 14, batch 7150, loss[loss=0.2437, simple_loss=0.303, pruned_loss=0.09218, over 11480.00 frames. ], tot_loss[loss=0.2147, simple_loss=0.2995, pruned_loss=0.06496, over 3134498.72 frames. ], batch size: 246, lr: 4.83e-03, grad_scale: 8.0 2023-04-30 00:32:44,446 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3632, 4.3677, 4.2249, 3.5865, 4.2833, 1.5953, 4.0731, 3.9760], device='cuda:4'), covar=tensor([0.0114, 0.0078, 0.0166, 0.0334, 0.0093, 0.2640, 0.0132, 0.0211], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0129, 0.0174, 0.0163, 0.0147, 0.0189, 0.0163, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:32:58,294 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7543, 1.7635, 1.5711, 1.4990, 1.8593, 1.5420, 1.7210, 1.8969], device='cuda:4'), covar=tensor([0.0125, 0.0228, 0.0333, 0.0291, 0.0160, 0.0220, 0.0144, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0213, 0.0206, 0.0208, 0.0213, 0.0213, 0.0217, 0.0204], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:33:47,875 INFO [train.py:904] (4/8) Epoch 14, batch 7200, loss[loss=0.2077, simple_loss=0.2919, pruned_loss=0.0617, over 11802.00 frames. ], tot_loss[loss=0.2125, simple_loss=0.2975, pruned_loss=0.0637, over 3110201.86 frames. ], batch size: 246, lr: 4.83e-03, grad_scale: 8.0 2023-04-30 00:34:55,352 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.878e+02 2.722e+02 3.151e+02 3.906e+02 1.100e+03, threshold=6.302e+02, percent-clipped=3.0 2023-04-30 00:35:07,112 INFO [train.py:904] (4/8) Epoch 14, batch 7250, loss[loss=0.2398, simple_loss=0.302, pruned_loss=0.08879, over 11529.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2953, pruned_loss=0.06224, over 3126898.39 frames. ], batch size: 247, lr: 4.83e-03, grad_scale: 4.0 2023-04-30 00:35:16,884 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8492, 1.3209, 1.6370, 1.6155, 1.7879, 1.8838, 1.5272, 1.7190], device='cuda:4'), covar=tensor([0.0172, 0.0310, 0.0159, 0.0249, 0.0193, 0.0132, 0.0318, 0.0091], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0174, 0.0158, 0.0164, 0.0172, 0.0130, 0.0176, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-30 00:36:01,930 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139238.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:36:22,184 INFO [train.py:904] (4/8) Epoch 14, batch 7300, loss[loss=0.1955, simple_loss=0.2826, pruned_loss=0.05416, over 16997.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2952, pruned_loss=0.0629, over 3092097.68 frames. ], batch size: 53, lr: 4.83e-03, grad_scale: 4.0 2023-04-30 00:36:43,285 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139265.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 00:37:00,570 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139276.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:37:03,575 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0230, 4.9841, 4.7249, 4.1373, 4.9394, 1.5923, 4.6931, 4.5959], device='cuda:4'), covar=tensor([0.0052, 0.0042, 0.0127, 0.0273, 0.0046, 0.2603, 0.0085, 0.0125], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0127, 0.0172, 0.0160, 0.0144, 0.0186, 0.0161, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:37:06,781 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3896, 3.1533, 2.5446, 2.0682, 2.2499, 2.1909, 3.2015, 3.0229], device='cuda:4'), covar=tensor([0.2884, 0.0810, 0.1807, 0.2459, 0.2407, 0.2034, 0.0575, 0.1077], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0262, 0.0291, 0.0292, 0.0286, 0.0232, 0.0277, 0.0309], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:37:20,403 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139289.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:37:29,626 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.078e+02 2.984e+02 3.521e+02 4.495e+02 8.579e+02, threshold=7.042e+02, percent-clipped=5.0 2023-04-30 00:37:37,017 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139299.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:37:40,536 INFO [train.py:904] (4/8) Epoch 14, batch 7350, loss[loss=0.2612, simple_loss=0.3207, pruned_loss=0.1008, over 10689.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.296, pruned_loss=0.06361, over 3067287.03 frames. ], batch size: 246, lr: 4.83e-03, grad_scale: 4.0 2023-04-30 00:38:36,849 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139337.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:38:36,908 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139337.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 00:38:44,958 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139342.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:38:57,367 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139350.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:38:59,353 INFO [train.py:904] (4/8) Epoch 14, batch 7400, loss[loss=0.1987, simple_loss=0.2892, pruned_loss=0.05412, over 17062.00 frames. ], tot_loss[loss=0.2128, simple_loss=0.297, pruned_loss=0.06426, over 3075539.70 frames. ], batch size: 53, lr: 4.83e-03, grad_scale: 4.0 2023-04-30 00:39:01,634 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139353.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:39:52,860 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=139385.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:40:00,701 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=139390.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:40:08,330 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.005e+02 2.794e+02 3.378e+02 4.039e+02 6.978e+02, threshold=6.757e+02, percent-clipped=0.0 2023-04-30 00:40:18,576 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=139401.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:40:19,476 INFO [train.py:904] (4/8) Epoch 14, batch 7450, loss[loss=0.2261, simple_loss=0.3176, pruned_loss=0.06736, over 15465.00 frames. ], tot_loss[loss=0.2158, simple_loss=0.2991, pruned_loss=0.06629, over 3055489.94 frames. ], batch size: 191, lr: 4.83e-03, grad_scale: 4.0 2023-04-30 00:40:24,527 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2592, 3.5931, 3.5700, 2.0374, 3.0715, 2.3752, 3.5851, 3.8888], device='cuda:4'), covar=tensor([0.0284, 0.0670, 0.0534, 0.1886, 0.0737, 0.0896, 0.0646, 0.0734], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0152, 0.0162, 0.0147, 0.0139, 0.0127, 0.0139, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 00:41:43,385 INFO [train.py:904] (4/8) Epoch 14, batch 7500, loss[loss=0.2482, simple_loss=0.3145, pruned_loss=0.09096, over 11096.00 frames. ], tot_loss[loss=0.2159, simple_loss=0.2997, pruned_loss=0.06604, over 3042761.96 frames. ], batch size: 246, lr: 4.83e-03, grad_scale: 4.0 2023-04-30 00:42:53,286 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.056e+02 3.028e+02 3.537e+02 4.351e+02 9.459e+02, threshold=7.073e+02, percent-clipped=3.0 2023-04-30 00:43:02,943 INFO [train.py:904] (4/8) Epoch 14, batch 7550, loss[loss=0.1959, simple_loss=0.278, pruned_loss=0.0569, over 16626.00 frames. ], tot_loss[loss=0.2159, simple_loss=0.2989, pruned_loss=0.06644, over 3039016.70 frames. ], batch size: 62, lr: 4.82e-03, grad_scale: 4.0 2023-04-30 00:43:50,281 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2776, 4.3434, 4.1094, 3.8909, 3.8425, 4.2293, 3.9726, 3.9368], device='cuda:4'), covar=tensor([0.0594, 0.0475, 0.0285, 0.0286, 0.0896, 0.0449, 0.0686, 0.0651], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0342, 0.0299, 0.0279, 0.0312, 0.0326, 0.0204, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:44:19,040 INFO [train.py:904] (4/8) Epoch 14, batch 7600, loss[loss=0.1998, simple_loss=0.2899, pruned_loss=0.05483, over 16470.00 frames. ], tot_loss[loss=0.2153, simple_loss=0.2979, pruned_loss=0.0663, over 3047952.20 frames. ], batch size: 75, lr: 4.82e-03, grad_scale: 8.0 2023-04-30 00:44:25,028 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-30 00:44:29,356 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-30 00:44:39,725 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139565.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 00:44:49,011 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139571.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:45:06,408 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139582.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:45:23,451 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139594.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:45:24,243 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.790e+02 2.745e+02 3.435e+02 4.599e+02 7.932e+02, threshold=6.869e+02, percent-clipped=2.0 2023-04-30 00:45:34,159 INFO [train.py:904] (4/8) Epoch 14, batch 7650, loss[loss=0.2775, simple_loss=0.3347, pruned_loss=0.1102, over 10996.00 frames. ], tot_loss[loss=0.2157, simple_loss=0.2981, pruned_loss=0.06664, over 3039022.24 frames. ], batch size: 247, lr: 4.82e-03, grad_scale: 8.0 2023-04-30 00:45:38,001 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3747, 3.3873, 2.6266, 2.0431, 2.2981, 2.2045, 3.4913, 3.1831], device='cuda:4'), covar=tensor([0.3022, 0.0746, 0.1868, 0.2658, 0.2537, 0.2068, 0.0489, 0.1120], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0263, 0.0291, 0.0293, 0.0287, 0.0234, 0.0276, 0.0311], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 00:45:50,718 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=139613.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:46:19,134 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139632.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 00:46:19,264 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139632.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:46:37,184 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139643.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:46:39,467 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139645.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:46:49,447 INFO [train.py:904] (4/8) Epoch 14, batch 7700, loss[loss=0.2223, simple_loss=0.304, pruned_loss=0.07026, over 16244.00 frames. ], tot_loss[loss=0.2161, simple_loss=0.2983, pruned_loss=0.06694, over 3052229.14 frames. ], batch size: 165, lr: 4.82e-03, grad_scale: 8.0 2023-04-30 00:47:57,375 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.201e+02 3.391e+02 3.997e+02 4.581e+02 8.153e+02, threshold=7.994e+02, percent-clipped=6.0 2023-04-30 00:48:06,798 INFO [train.py:904] (4/8) Epoch 14, batch 7750, loss[loss=0.2578, simple_loss=0.325, pruned_loss=0.09523, over 11351.00 frames. ], tot_loss[loss=0.2153, simple_loss=0.298, pruned_loss=0.06627, over 3072271.90 frames. ], batch size: 248, lr: 4.82e-03, grad_scale: 4.0 2023-04-30 00:49:10,204 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4969, 3.5973, 2.0946, 3.9508, 2.5987, 3.9438, 2.1546, 2.8376], device='cuda:4'), covar=tensor([0.0248, 0.0353, 0.1594, 0.0175, 0.0766, 0.0562, 0.1556, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0169, 0.0190, 0.0139, 0.0166, 0.0209, 0.0196, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 00:49:24,607 INFO [train.py:904] (4/8) Epoch 14, batch 7800, loss[loss=0.2078, simple_loss=0.2951, pruned_loss=0.06028, over 16910.00 frames. ], tot_loss[loss=0.2161, simple_loss=0.2988, pruned_loss=0.06673, over 3078687.31 frames. ], batch size: 116, lr: 4.82e-03, grad_scale: 4.0 2023-04-30 00:49:51,698 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139769.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:50:22,495 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139789.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:50:32,559 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.773e+02 2.870e+02 3.646e+02 4.608e+02 1.262e+03, threshold=7.291e+02, percent-clipped=3.0 2023-04-30 00:50:41,156 INFO [train.py:904] (4/8) Epoch 14, batch 7850, loss[loss=0.2352, simple_loss=0.3189, pruned_loss=0.07579, over 15272.00 frames. ], tot_loss[loss=0.2166, simple_loss=0.2995, pruned_loss=0.0668, over 3064224.98 frames. ], batch size: 190, lr: 4.82e-03, grad_scale: 4.0 2023-04-30 00:50:46,076 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139804.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:50:48,468 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139806.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:51:24,544 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139830.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:51:42,885 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7191, 1.8132, 2.2436, 2.6871, 2.7181, 3.0516, 1.9424, 3.0618], device='cuda:4'), covar=tensor([0.0168, 0.0429, 0.0254, 0.0238, 0.0216, 0.0156, 0.0438, 0.0102], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0174, 0.0158, 0.0164, 0.0173, 0.0131, 0.0177, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-30 00:51:53,572 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139850.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:51:53,609 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139850.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:51:56,192 INFO [train.py:904] (4/8) Epoch 14, batch 7900, loss[loss=0.2312, simple_loss=0.3203, pruned_loss=0.07106, over 16769.00 frames. ], tot_loss[loss=0.2157, simple_loss=0.2988, pruned_loss=0.06634, over 3069298.33 frames. ], batch size: 124, lr: 4.82e-03, grad_scale: 4.0 2023-04-30 00:52:15,801 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139865.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:52:18,337 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139867.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:53:01,517 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139894.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:53:03,656 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.073e+02 2.924e+02 3.560e+02 4.104e+02 6.335e+02, threshold=7.120e+02, percent-clipped=0.0 2023-04-30 00:53:12,903 INFO [train.py:904] (4/8) Epoch 14, batch 7950, loss[loss=0.1741, simple_loss=0.2626, pruned_loss=0.04276, over 16318.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.299, pruned_loss=0.06679, over 3077743.85 frames. ], batch size: 35, lr: 4.82e-03, grad_scale: 4.0 2023-04-30 00:53:19,581 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5332, 3.6053, 3.3275, 3.1053, 3.1674, 3.4703, 3.3121, 3.2354], device='cuda:4'), covar=tensor([0.0577, 0.0565, 0.0253, 0.0244, 0.0543, 0.0420, 0.1020, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0341, 0.0297, 0.0276, 0.0310, 0.0325, 0.0204, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:53:26,151 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3579, 4.3478, 4.7629, 4.7423, 4.7433, 4.4618, 4.4577, 4.3017], device='cuda:4'), covar=tensor([0.0294, 0.0574, 0.0356, 0.0393, 0.0437, 0.0348, 0.0847, 0.0431], device='cuda:4'), in_proj_covar=tensor([0.0352, 0.0370, 0.0368, 0.0352, 0.0417, 0.0394, 0.0481, 0.0312], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 00:53:28,024 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139911.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:53:51,345 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139927.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:53:59,038 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139932.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 00:54:02,453 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.39 vs. limit=5.0 2023-04-30 00:54:08,357 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139938.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:54:13,651 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=139942.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:54:18,367 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139945.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:54:28,354 INFO [train.py:904] (4/8) Epoch 14, batch 8000, loss[loss=0.215, simple_loss=0.3069, pruned_loss=0.06158, over 16362.00 frames. ], tot_loss[loss=0.2174, simple_loss=0.2999, pruned_loss=0.06745, over 3068873.27 frames. ], batch size: 146, lr: 4.82e-03, grad_scale: 8.0 2023-04-30 00:55:10,467 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=139980.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:55:30,334 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=139993.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:55:34,105 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.098e+02 3.021e+02 3.514e+02 4.508e+02 9.326e+02, threshold=7.028e+02, percent-clipped=1.0 2023-04-30 00:55:46,024 INFO [train.py:904] (4/8) Epoch 14, batch 8050, loss[loss=0.2394, simple_loss=0.3068, pruned_loss=0.08597, over 11547.00 frames. ], tot_loss[loss=0.2169, simple_loss=0.2996, pruned_loss=0.06704, over 3079744.69 frames. ], batch size: 248, lr: 4.82e-03, grad_scale: 8.0 2023-04-30 00:56:59,040 INFO [train.py:904] (4/8) Epoch 14, batch 8100, loss[loss=0.2404, simple_loss=0.3036, pruned_loss=0.08862, over 11488.00 frames. ], tot_loss[loss=0.2169, simple_loss=0.2998, pruned_loss=0.06701, over 3084653.93 frames. ], batch size: 248, lr: 4.82e-03, grad_scale: 8.0 2023-04-30 00:58:06,570 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.001e+02 3.052e+02 3.743e+02 5.058e+02 1.059e+03, threshold=7.487e+02, percent-clipped=6.0 2023-04-30 00:58:17,065 INFO [train.py:904] (4/8) Epoch 14, batch 8150, loss[loss=0.175, simple_loss=0.2656, pruned_loss=0.04221, over 16833.00 frames. ], tot_loss[loss=0.2141, simple_loss=0.2969, pruned_loss=0.06562, over 3096458.94 frames. ], batch size: 102, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 00:58:18,965 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140103.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:58:29,421 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140110.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:58:52,071 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140125.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:59:09,368 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8859, 4.8760, 4.7145, 4.4005, 4.3009, 4.7716, 4.6038, 4.4758], device='cuda:4'), covar=tensor([0.0698, 0.0561, 0.0317, 0.0320, 0.1107, 0.0538, 0.0485, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0343, 0.0299, 0.0277, 0.0311, 0.0327, 0.0205, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:59:23,723 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140145.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:59:34,235 INFO [train.py:904] (4/8) Epoch 14, batch 8200, loss[loss=0.2336, simple_loss=0.3226, pruned_loss=0.07231, over 15304.00 frames. ], tot_loss[loss=0.2125, simple_loss=0.2951, pruned_loss=0.06491, over 3102214.89 frames. ], batch size: 190, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 00:59:44,391 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9545, 4.9048, 4.7507, 3.8306, 4.7620, 1.8876, 4.4955, 4.5008], device='cuda:4'), covar=tensor([0.0106, 0.0095, 0.0173, 0.0488, 0.0111, 0.2558, 0.0165, 0.0237], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0127, 0.0174, 0.0162, 0.0146, 0.0189, 0.0162, 0.0156], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 00:59:47,416 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140160.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:59:50,572 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140162.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 00:59:53,806 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140164.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:00:05,611 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140171.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:00:44,842 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.704e+02 2.531e+02 3.267e+02 4.112e+02 9.606e+02, threshold=6.535e+02, percent-clipped=4.0 2023-04-30 01:00:55,629 INFO [train.py:904] (4/8) Epoch 14, batch 8250, loss[loss=0.1829, simple_loss=0.279, pruned_loss=0.04343, over 15250.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.294, pruned_loss=0.06231, over 3092585.20 frames. ], batch size: 190, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:01:02,320 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140206.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:01:37,263 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140227.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:01:49,635 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4118, 4.6668, 4.5041, 4.5085, 4.2097, 4.1931, 4.2053, 4.6986], device='cuda:4'), covar=tensor([0.1032, 0.0855, 0.0850, 0.0705, 0.0746, 0.1305, 0.0935, 0.0844], device='cuda:4'), in_proj_covar=tensor([0.0578, 0.0710, 0.0585, 0.0511, 0.0449, 0.0463, 0.0594, 0.0542], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:01:55,113 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140238.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:02:16,387 INFO [train.py:904] (4/8) Epoch 14, batch 8300, loss[loss=0.1758, simple_loss=0.2756, pruned_loss=0.03804, over 16717.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.291, pruned_loss=0.05931, over 3084457.93 frames. ], batch size: 89, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:02:28,950 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140259.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:02:46,499 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9881, 1.9103, 2.2961, 3.2068, 2.0872, 2.1155, 2.1535, 2.0449], device='cuda:4'), covar=tensor([0.1065, 0.4109, 0.2446, 0.0638, 0.4632, 0.3008, 0.3705, 0.3779], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0405, 0.0337, 0.0313, 0.0415, 0.0463, 0.0369, 0.0471], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:02:54,964 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140275.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:03:12,771 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140286.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:03:27,883 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.670e+02 2.340e+02 2.728e+02 3.506e+02 7.480e+02, threshold=5.455e+02, percent-clipped=2.0 2023-04-30 01:03:36,676 INFO [train.py:904] (4/8) Epoch 14, batch 8350, loss[loss=0.2144, simple_loss=0.2906, pruned_loss=0.06911, over 11976.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.291, pruned_loss=0.05789, over 3082110.45 frames. ], batch size: 247, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:04:07,931 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140320.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:04:41,093 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7717, 3.8397, 2.9493, 2.2178, 2.4322, 2.3854, 3.9930, 3.3983], device='cuda:4'), covar=tensor([0.2485, 0.0552, 0.1589, 0.2669, 0.2860, 0.1922, 0.0354, 0.1099], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0253, 0.0282, 0.0283, 0.0275, 0.0227, 0.0267, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:04:57,945 INFO [train.py:904] (4/8) Epoch 14, batch 8400, loss[loss=0.1873, simple_loss=0.2781, pruned_loss=0.04825, over 16366.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.288, pruned_loss=0.0557, over 3068105.67 frames. ], batch size: 146, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:05:54,624 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140387.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:05:56,099 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2610, 3.4071, 3.6289, 3.6094, 3.6055, 3.4581, 3.4616, 3.4890], device='cuda:4'), covar=tensor([0.0369, 0.0644, 0.0430, 0.0473, 0.0464, 0.0466, 0.0772, 0.0448], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0368, 0.0366, 0.0351, 0.0414, 0.0393, 0.0479, 0.0310], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 01:06:08,144 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.599e+02 2.349e+02 2.839e+02 3.572e+02 6.227e+02, threshold=5.678e+02, percent-clipped=3.0 2023-04-30 01:06:18,701 INFO [train.py:904] (4/8) Epoch 14, batch 8450, loss[loss=0.1687, simple_loss=0.2686, pruned_loss=0.03442, over 16276.00 frames. ], tot_loss[loss=0.1962, simple_loss=0.2853, pruned_loss=0.0536, over 3053893.05 frames. ], batch size: 165, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:06:55,833 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140425.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:07:12,249 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2693, 3.4167, 3.6424, 3.6095, 3.6181, 3.4572, 3.4722, 3.5076], device='cuda:4'), covar=tensor([0.0413, 0.0698, 0.0428, 0.0491, 0.0507, 0.0534, 0.0815, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0366, 0.0364, 0.0349, 0.0412, 0.0391, 0.0476, 0.0308], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 01:07:26,083 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6294, 2.6303, 1.9015, 2.7549, 2.0994, 2.7077, 2.1791, 2.4818], device='cuda:4'), covar=tensor([0.0238, 0.0269, 0.1114, 0.0227, 0.0578, 0.0367, 0.0990, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0163, 0.0184, 0.0134, 0.0163, 0.0201, 0.0191, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 01:07:27,296 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140445.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:07:32,303 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140448.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:07:37,958 INFO [train.py:904] (4/8) Epoch 14, batch 8500, loss[loss=0.196, simple_loss=0.2893, pruned_loss=0.05135, over 16720.00 frames. ], tot_loss[loss=0.1919, simple_loss=0.2814, pruned_loss=0.05118, over 3062677.28 frames. ], batch size: 134, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:07:50,011 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140459.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:07:51,411 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140460.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:07:54,679 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140462.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:08:01,374 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140466.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:08:13,349 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140473.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:08:38,536 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140489.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:08:46,067 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140493.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:08:51,330 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.625e+02 2.295e+02 2.729e+02 3.409e+02 8.118e+02, threshold=5.459e+02, percent-clipped=4.0 2023-04-30 01:09:01,488 INFO [train.py:904] (4/8) Epoch 14, batch 8550, loss[loss=0.1955, simple_loss=0.2776, pruned_loss=0.05668, over 16645.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.2794, pruned_loss=0.05037, over 3034840.15 frames. ], batch size: 57, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:09:11,904 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140506.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:09:13,746 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.09 vs. limit=5.0 2023-04-30 01:09:15,777 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140508.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:09:17,248 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140509.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:09:19,129 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140510.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:09:52,287 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9409, 2.2364, 2.3101, 2.9078, 1.9387, 3.3319, 1.6272, 2.7467], device='cuda:4'), covar=tensor([0.1264, 0.0694, 0.1049, 0.0136, 0.0095, 0.0409, 0.1496, 0.0698], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0160, 0.0182, 0.0158, 0.0197, 0.0207, 0.0185, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-30 01:10:39,032 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140550.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:10:41,516 INFO [train.py:904] (4/8) Epoch 14, batch 8600, loss[loss=0.2032, simple_loss=0.2979, pruned_loss=0.05428, over 16393.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2794, pruned_loss=0.04937, over 3028734.70 frames. ], batch size: 146, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:10:46,487 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140554.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:11:19,609 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140570.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:12:06,768 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.419e+02 2.341e+02 2.953e+02 3.525e+02 9.042e+02, threshold=5.906e+02, percent-clipped=3.0 2023-04-30 01:12:18,451 INFO [train.py:904] (4/8) Epoch 14, batch 8650, loss[loss=0.1781, simple_loss=0.2629, pruned_loss=0.04666, over 12088.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2775, pruned_loss=0.04797, over 3017482.76 frames. ], batch size: 247, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:12:52,267 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140615.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:14:05,319 INFO [train.py:904] (4/8) Epoch 14, batch 8700, loss[loss=0.1828, simple_loss=0.2743, pruned_loss=0.04558, over 16247.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2745, pruned_loss=0.04656, over 3054790.08 frames. ], batch size: 165, lr: 4.81e-03, grad_scale: 8.0 2023-04-30 01:15:24,088 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.417e+02 2.380e+02 2.814e+02 3.680e+02 6.881e+02, threshold=5.627e+02, percent-clipped=2.0 2023-04-30 01:15:38,752 INFO [train.py:904] (4/8) Epoch 14, batch 8750, loss[loss=0.1759, simple_loss=0.2737, pruned_loss=0.03908, over 16565.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2741, pruned_loss=0.04605, over 3050468.91 frames. ], batch size: 62, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:15:49,861 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9848, 1.8597, 2.4360, 2.9151, 2.7286, 3.2651, 2.0531, 3.2188], device='cuda:4'), covar=tensor([0.0167, 0.0447, 0.0278, 0.0225, 0.0246, 0.0161, 0.0413, 0.0115], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0174, 0.0157, 0.0162, 0.0173, 0.0129, 0.0176, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-30 01:16:30,720 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140723.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:16:40,222 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7980, 4.0891, 3.7785, 3.5436, 3.2866, 3.9944, 3.6321, 3.6899], device='cuda:4'), covar=tensor([0.0937, 0.0445, 0.0462, 0.0405, 0.1470, 0.0438, 0.1245, 0.0657], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0337, 0.0296, 0.0274, 0.0306, 0.0322, 0.0204, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:17:11,077 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140743.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:17:28,574 INFO [train.py:904] (4/8) Epoch 14, batch 8800, loss[loss=0.2257, simple_loss=0.3106, pruned_loss=0.0704, over 12498.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2725, pruned_loss=0.04479, over 3049317.90 frames. ], batch size: 248, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:17:33,770 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-30 01:17:43,136 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140759.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:17:56,959 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140766.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:18:35,780 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140784.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:19:00,060 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.423e+02 2.211e+02 2.894e+02 3.689e+02 7.406e+02, threshold=5.789e+02, percent-clipped=2.0 2023-04-30 01:19:12,350 INFO [train.py:904] (4/8) Epoch 14, batch 8850, loss[loss=0.1808, simple_loss=0.2852, pruned_loss=0.03813, over 16211.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2751, pruned_loss=0.04431, over 3046732.02 frames. ], batch size: 165, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:19:23,456 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140807.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:19:27,245 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140809.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:19:37,091 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140814.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:20:44,014 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140845.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:20:58,215 INFO [train.py:904] (4/8) Epoch 14, batch 8900, loss[loss=0.1827, simple_loss=0.2761, pruned_loss=0.04459, over 16277.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2742, pruned_loss=0.04333, over 3023970.85 frames. ], batch size: 165, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:21:25,581 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140865.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:21:34,615 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140870.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:22:39,989 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-30 01:22:42,214 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2870, 3.8366, 3.6668, 2.6601, 3.4185, 3.7835, 3.5383, 1.7581], device='cuda:4'), covar=tensor([0.0505, 0.0038, 0.0061, 0.0360, 0.0086, 0.0097, 0.0086, 0.0598], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0069, 0.0070, 0.0126, 0.0083, 0.0092, 0.0081, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 01:22:46,890 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.640e+02 2.342e+02 2.775e+02 3.236e+02 6.448e+02, threshold=5.549e+02, percent-clipped=1.0 2023-04-30 01:23:01,545 INFO [train.py:904] (4/8) Epoch 14, batch 8950, loss[loss=0.1729, simple_loss=0.2658, pruned_loss=0.04001, over 16973.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2744, pruned_loss=0.04353, over 3056887.53 frames. ], batch size: 116, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:23:14,731 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9787, 4.2276, 4.0723, 4.0941, 3.7448, 3.8305, 3.8631, 4.2301], device='cuda:4'), covar=tensor([0.1014, 0.0949, 0.0924, 0.0757, 0.0831, 0.1436, 0.0929, 0.0957], device='cuda:4'), in_proj_covar=tensor([0.0568, 0.0704, 0.0573, 0.0507, 0.0445, 0.0458, 0.0590, 0.0535], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:23:29,064 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140915.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:23:32,811 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140917.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:24:02,258 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1580, 4.2246, 4.5775, 4.5536, 4.5491, 4.2870, 4.2789, 4.2156], device='cuda:4'), covar=tensor([0.0333, 0.0589, 0.0386, 0.0414, 0.0443, 0.0352, 0.0754, 0.0387], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0353, 0.0356, 0.0337, 0.0399, 0.0376, 0.0457, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 01:24:38,414 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140947.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:24:48,129 INFO [train.py:904] (4/8) Epoch 14, batch 9000, loss[loss=0.1433, simple_loss=0.2375, pruned_loss=0.02453, over 16812.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2711, pruned_loss=0.04217, over 3066355.97 frames. ], batch size: 83, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:24:48,129 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 01:24:58,094 INFO [train.py:938] (4/8) Epoch 14, validation: loss=0.1514, simple_loss=0.2553, pruned_loss=0.0237, over 944034.00 frames. 2023-04-30 01:24:58,095 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 01:25:12,426 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1179, 5.3847, 5.1952, 5.1685, 4.8679, 4.8234, 4.8324, 5.5029], device='cuda:4'), covar=tensor([0.1206, 0.0887, 0.0939, 0.0722, 0.0817, 0.0789, 0.1164, 0.0785], device='cuda:4'), in_proj_covar=tensor([0.0568, 0.0704, 0.0574, 0.0507, 0.0445, 0.0458, 0.0589, 0.0534], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:25:21,238 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=140963.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:25:52,413 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140978.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:26:29,789 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.243e+02 2.199e+02 2.701e+02 3.272e+02 7.252e+02, threshold=5.402e+02, percent-clipped=4.0 2023-04-30 01:26:39,921 INFO [train.py:904] (4/8) Epoch 14, batch 9050, loss[loss=0.1769, simple_loss=0.261, pruned_loss=0.0464, over 16696.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2721, pruned_loss=0.04292, over 3056622.63 frames. ], batch size: 134, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:26:48,676 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.26 vs. limit=5.0 2023-04-30 01:26:55,111 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141008.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:28:08,240 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141043.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:28:25,840 INFO [train.py:904] (4/8) Epoch 14, batch 9100, loss[loss=0.1926, simple_loss=0.291, pruned_loss=0.04715, over 16357.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2715, pruned_loss=0.04315, over 3066234.61 frames. ], batch size: 146, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:28:36,560 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 01:28:47,634 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8951, 1.3454, 1.7291, 1.6901, 1.8503, 1.8548, 1.6197, 1.8386], device='cuda:4'), covar=tensor([0.0215, 0.0324, 0.0171, 0.0246, 0.0227, 0.0188, 0.0318, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0174, 0.0156, 0.0161, 0.0172, 0.0129, 0.0175, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-30 01:29:30,788 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141079.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:29:58,702 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=141091.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:30:13,181 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.729e+02 2.366e+02 2.876e+02 3.628e+02 8.857e+02, threshold=5.752e+02, percent-clipped=2.0 2023-04-30 01:30:22,842 INFO [train.py:904] (4/8) Epoch 14, batch 9150, loss[loss=0.1585, simple_loss=0.2559, pruned_loss=0.03057, over 16874.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2721, pruned_loss=0.04272, over 3056818.10 frames. ], batch size: 96, lr: 4.80e-03, grad_scale: 4.0 2023-04-30 01:31:51,308 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141143.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:31:54,398 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141145.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:32:04,830 INFO [train.py:904] (4/8) Epoch 14, batch 9200, loss[loss=0.1873, simple_loss=0.2779, pruned_loss=0.04837, over 15415.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2676, pruned_loss=0.04166, over 3035672.33 frames. ], batch size: 191, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:32:28,725 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141165.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:32:28,763 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141165.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:33:23,826 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=141193.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:33:31,524 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.484e+02 2.509e+02 2.979e+02 4.006e+02 1.027e+03, threshold=5.958e+02, percent-clipped=7.0 2023-04-30 01:33:40,990 INFO [train.py:904] (4/8) Epoch 14, batch 9250, loss[loss=0.1723, simple_loss=0.2659, pruned_loss=0.03936, over 16641.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2675, pruned_loss=0.04172, over 3041183.22 frames. ], batch size: 134, lr: 4.80e-03, grad_scale: 8.0 2023-04-30 01:33:46,098 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141204.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:34:03,407 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=141213.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:35:30,114 INFO [train.py:904] (4/8) Epoch 14, batch 9300, loss[loss=0.1645, simple_loss=0.2562, pruned_loss=0.03642, over 16675.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2658, pruned_loss=0.04104, over 3059798.54 frames. ], batch size: 134, lr: 4.79e-03, grad_scale: 8.0 2023-04-30 01:36:20,225 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141273.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:36:57,191 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.75 vs. limit=5.0 2023-04-30 01:37:05,691 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.463e+02 2.032e+02 2.564e+02 3.502e+02 9.246e+02, threshold=5.129e+02, percent-clipped=2.0 2023-04-30 01:37:14,521 INFO [train.py:904] (4/8) Epoch 14, batch 9350, loss[loss=0.1719, simple_loss=0.2659, pruned_loss=0.03893, over 16663.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2656, pruned_loss=0.04106, over 3065281.92 frames. ], batch size: 89, lr: 4.79e-03, grad_scale: 8.0 2023-04-30 01:37:16,951 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141303.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:38:08,683 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141329.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:38:54,829 INFO [train.py:904] (4/8) Epoch 14, batch 9400, loss[loss=0.174, simple_loss=0.2826, pruned_loss=0.03269, over 16624.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2661, pruned_loss=0.04086, over 3070981.74 frames. ], batch size: 89, lr: 4.79e-03, grad_scale: 8.0 2023-04-30 01:39:16,176 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8190, 3.7204, 3.8885, 3.7244, 3.8831, 4.2530, 3.9911, 3.7003], device='cuda:4'), covar=tensor([0.1922, 0.2424, 0.2081, 0.2408, 0.2799, 0.1670, 0.1469, 0.2482], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0495, 0.0545, 0.0418, 0.0563, 0.0572, 0.0434, 0.0568], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 01:39:49,375 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141379.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:40:12,100 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141390.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:40:17,735 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-30 01:40:25,324 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 2.208e+02 2.590e+02 3.171e+02 6.862e+02, threshold=5.180e+02, percent-clipped=4.0 2023-04-30 01:40:33,409 INFO [train.py:904] (4/8) Epoch 14, batch 9450, loss[loss=0.1629, simple_loss=0.2628, pruned_loss=0.03147, over 16458.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2676, pruned_loss=0.04104, over 3057167.22 frames. ], batch size: 75, lr: 4.79e-03, grad_scale: 8.0 2023-04-30 01:41:03,132 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141417.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 01:41:24,367 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=141427.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:41:44,372 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-30 01:41:54,145 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141441.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:42:14,326 INFO [train.py:904] (4/8) Epoch 14, batch 9500, loss[loss=0.1566, simple_loss=0.2488, pruned_loss=0.03215, over 16375.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2664, pruned_loss=0.04038, over 3064070.44 frames. ], batch size: 68, lr: 4.79e-03, grad_scale: 8.0 2023-04-30 01:42:43,680 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141465.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:43:08,373 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5946, 3.6683, 3.4312, 3.1552, 3.2448, 3.5596, 3.3581, 3.3781], device='cuda:4'), covar=tensor([0.0532, 0.0494, 0.0265, 0.0250, 0.0527, 0.0404, 0.1070, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0328, 0.0290, 0.0269, 0.0300, 0.0314, 0.0199, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-30 01:43:08,416 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141478.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 01:43:47,717 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.502e+02 2.299e+02 2.699e+02 3.093e+02 4.944e+02, threshold=5.398e+02, percent-clipped=0.0 2023-04-30 01:43:55,182 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141499.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:43:59,690 INFO [train.py:904] (4/8) Epoch 14, batch 9550, loss[loss=0.1871, simple_loss=0.2752, pruned_loss=0.04945, over 12340.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2659, pruned_loss=0.04031, over 3067659.88 frames. ], batch size: 246, lr: 4.79e-03, grad_scale: 8.0 2023-04-30 01:44:01,032 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141502.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:44:23,849 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=141513.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:44:51,318 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8585, 2.8135, 2.5578, 1.9764, 2.4213, 2.8324, 2.6997, 1.8905], device='cuda:4'), covar=tensor([0.0407, 0.0049, 0.0052, 0.0312, 0.0124, 0.0082, 0.0087, 0.0394], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0069, 0.0071, 0.0127, 0.0084, 0.0092, 0.0081, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 01:45:41,072 INFO [train.py:904] (4/8) Epoch 14, batch 9600, loss[loss=0.2148, simple_loss=0.3191, pruned_loss=0.05521, over 15402.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2678, pruned_loss=0.04146, over 3044262.37 frames. ], batch size: 191, lr: 4.79e-03, grad_scale: 8.0 2023-04-30 01:46:22,387 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141573.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:47:19,265 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.635e+02 2.296e+02 2.650e+02 3.286e+02 9.159e+02, threshold=5.299e+02, percent-clipped=3.0 2023-04-30 01:47:30,277 INFO [train.py:904] (4/8) Epoch 14, batch 9650, loss[loss=0.181, simple_loss=0.275, pruned_loss=0.04348, over 15496.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2699, pruned_loss=0.04186, over 3049046.30 frames. ], batch size: 190, lr: 4.79e-03, grad_scale: 4.0 2023-04-30 01:47:34,539 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141603.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:48:02,738 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3228, 3.3189, 3.5450, 1.6704, 3.7224, 3.7970, 2.8380, 2.7944], device='cuda:4'), covar=tensor([0.0838, 0.0245, 0.0206, 0.1321, 0.0067, 0.0137, 0.0481, 0.0484], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0100, 0.0085, 0.0135, 0.0069, 0.0107, 0.0120, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0002, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 01:48:16,939 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=141621.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:49:16,990 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=141651.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:49:17,847 INFO [train.py:904] (4/8) Epoch 14, batch 9700, loss[loss=0.1818, simple_loss=0.271, pruned_loss=0.04628, over 16786.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2696, pruned_loss=0.04196, over 3055974.04 frames. ], batch size: 124, lr: 4.79e-03, grad_scale: 4.0 2023-04-30 01:50:27,992 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141685.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:50:53,324 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.703e+02 2.310e+02 2.782e+02 3.506e+02 1.114e+03, threshold=5.564e+02, percent-clipped=5.0 2023-04-30 01:51:00,294 INFO [train.py:904] (4/8) Epoch 14, batch 9750, loss[loss=0.1542, simple_loss=0.2488, pruned_loss=0.02981, over 16856.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2674, pruned_loss=0.04175, over 3042921.49 frames. ], batch size: 102, lr: 4.79e-03, grad_scale: 4.0 2023-04-30 01:52:38,197 INFO [train.py:904] (4/8) Epoch 14, batch 9800, loss[loss=0.1816, simple_loss=0.2972, pruned_loss=0.03297, over 16772.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2684, pruned_loss=0.04104, over 3064940.82 frames. ], batch size: 83, lr: 4.79e-03, grad_scale: 4.0 2023-04-30 01:53:14,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4560, 4.4558, 4.2564, 3.7543, 4.3517, 1.6706, 4.1269, 4.0682], device='cuda:4'), covar=tensor([0.0065, 0.0061, 0.0146, 0.0233, 0.0070, 0.2525, 0.0107, 0.0196], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0124, 0.0166, 0.0150, 0.0141, 0.0185, 0.0155, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:53:18,068 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141773.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 01:53:40,028 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6390, 3.9013, 2.8527, 2.1816, 2.4262, 2.4558, 4.1119, 3.4167], device='cuda:4'), covar=tensor([0.2673, 0.0587, 0.1607, 0.2677, 0.2669, 0.1854, 0.0347, 0.1140], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0252, 0.0283, 0.0280, 0.0264, 0.0227, 0.0263, 0.0294], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:54:13,229 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141797.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:54:13,943 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.610e+02 2.347e+02 2.793e+02 3.311e+02 9.229e+02, threshold=5.586e+02, percent-clipped=3.0 2023-04-30 01:54:17,230 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141799.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:54:21,447 INFO [train.py:904] (4/8) Epoch 14, batch 9850, loss[loss=0.1573, simple_loss=0.2465, pruned_loss=0.034, over 12223.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2692, pruned_loss=0.0405, over 3074307.49 frames. ], batch size: 248, lr: 4.79e-03, grad_scale: 4.0 2023-04-30 01:56:04,789 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=141847.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:56:14,788 INFO [train.py:904] (4/8) Epoch 14, batch 9900, loss[loss=0.1693, simple_loss=0.2672, pruned_loss=0.03573, over 16739.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2698, pruned_loss=0.04061, over 3075515.35 frames. ], batch size: 76, lr: 4.78e-03, grad_scale: 4.0 2023-04-30 01:56:17,993 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2815, 4.1081, 4.3807, 4.4640, 4.6176, 4.2029, 4.6154, 4.6254], device='cuda:4'), covar=tensor([0.1553, 0.1017, 0.1250, 0.0669, 0.0467, 0.0965, 0.0487, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0534, 0.0657, 0.0774, 0.0677, 0.0508, 0.0525, 0.0531, 0.0628], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 01:57:44,782 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8699, 4.8542, 5.3040, 5.2915, 5.2641, 4.9746, 4.9140, 4.6829], device='cuda:4'), covar=tensor([0.0247, 0.0469, 0.0299, 0.0329, 0.0390, 0.0294, 0.0982, 0.0386], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0349, 0.0351, 0.0333, 0.0394, 0.0373, 0.0450, 0.0296], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0002], device='cuda:4') 2023-04-30 01:58:03,748 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.629e+02 2.142e+02 2.825e+02 3.311e+02 8.255e+02, threshold=5.650e+02, percent-clipped=2.0 2023-04-30 01:58:08,114 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3598, 4.0572, 4.0198, 2.7074, 3.6004, 4.0261, 3.7966, 2.4894], device='cuda:4'), covar=tensor([0.0447, 0.0023, 0.0030, 0.0329, 0.0065, 0.0055, 0.0041, 0.0376], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0069, 0.0071, 0.0126, 0.0083, 0.0092, 0.0081, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 01:58:13,950 INFO [train.py:904] (4/8) Epoch 14, batch 9950, loss[loss=0.1693, simple_loss=0.2687, pruned_loss=0.03493, over 16910.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2723, pruned_loss=0.04119, over 3086918.73 frames. ], batch size: 102, lr: 4.78e-03, grad_scale: 4.0 2023-04-30 01:58:48,367 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141916.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 01:59:21,124 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141928.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:00:16,381 INFO [train.py:904] (4/8) Epoch 14, batch 10000, loss[loss=0.1665, simple_loss=0.2643, pruned_loss=0.03436, over 16541.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2704, pruned_loss=0.04061, over 3107731.48 frames. ], batch size: 75, lr: 4.78e-03, grad_scale: 8.0 2023-04-30 02:01:07,122 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141977.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:01:25,909 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141985.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:01:34,654 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141989.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:01:50,846 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.061e+02 2.551e+02 3.100e+02 8.241e+02, threshold=5.102e+02, percent-clipped=1.0 2023-04-30 02:02:01,160 INFO [train.py:904] (4/8) Epoch 14, batch 10050, loss[loss=0.2006, simple_loss=0.2919, pruned_loss=0.05468, over 16675.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2704, pruned_loss=0.04025, over 3100354.07 frames. ], batch size: 134, lr: 4.78e-03, grad_scale: 8.0 2023-04-30 02:02:09,379 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3397, 5.6524, 5.4181, 5.4277, 5.1488, 5.0602, 5.0472, 5.6836], device='cuda:4'), covar=tensor([0.0951, 0.0730, 0.0925, 0.0590, 0.0678, 0.0678, 0.0987, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0555, 0.0687, 0.0562, 0.0497, 0.0437, 0.0448, 0.0575, 0.0525], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:03:00,954 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=142033.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:03:21,462 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1963, 2.0840, 2.1859, 3.7267, 2.0830, 2.4121, 2.2262, 2.2231], device='cuda:4'), covar=tensor([0.1039, 0.3384, 0.2624, 0.0439, 0.3964, 0.2393, 0.3109, 0.3358], device='cuda:4'), in_proj_covar=tensor([0.0359, 0.0395, 0.0333, 0.0306, 0.0406, 0.0448, 0.0360, 0.0457], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:03:34,640 INFO [train.py:904] (4/8) Epoch 14, batch 10100, loss[loss=0.1675, simple_loss=0.2447, pruned_loss=0.04518, over 12352.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2702, pruned_loss=0.04014, over 3112635.91 frames. ], batch size: 248, lr: 4.78e-03, grad_scale: 8.0 2023-04-30 02:04:19,988 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=142073.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 02:04:33,510 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-30 02:04:49,686 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=142097.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:04:50,457 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.741e+02 2.492e+02 2.995e+02 3.648e+02 7.218e+02, threshold=5.990e+02, percent-clipped=8.0 2023-04-30 02:05:19,913 INFO [train.py:904] (4/8) Epoch 15, batch 0, loss[loss=0.1999, simple_loss=0.2896, pruned_loss=0.05514, over 17128.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2896, pruned_loss=0.05514, over 17128.00 frames. ], batch size: 49, lr: 4.62e-03, grad_scale: 8.0 2023-04-30 02:05:19,914 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 02:05:27,351 INFO [train.py:938] (4/8) Epoch 15, validation: loss=0.1501, simple_loss=0.2536, pruned_loss=0.02333, over 944034.00 frames. 2023-04-30 02:05:27,352 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 02:05:53,971 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=142121.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 02:06:27,844 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=142145.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:06:37,638 INFO [train.py:904] (4/8) Epoch 15, batch 50, loss[loss=0.161, simple_loss=0.2528, pruned_loss=0.03458, over 17019.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2806, pruned_loss=0.05829, over 740037.18 frames. ], batch size: 41, lr: 4.62e-03, grad_scale: 2.0 2023-04-30 02:07:24,362 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.00 vs. limit=5.0 2023-04-30 02:07:44,550 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.745e+02 2.683e+02 3.202e+02 3.967e+02 8.381e+02, threshold=6.403e+02, percent-clipped=5.0 2023-04-30 02:07:47,915 INFO [train.py:904] (4/8) Epoch 15, batch 100, loss[loss=0.155, simple_loss=0.2387, pruned_loss=0.03564, over 16830.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2768, pruned_loss=0.05496, over 1309296.95 frames. ], batch size: 42, lr: 4.62e-03, grad_scale: 2.0 2023-04-30 02:08:25,960 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.51 vs. limit=5.0 2023-04-30 02:08:56,686 INFO [train.py:904] (4/8) Epoch 15, batch 150, loss[loss=0.1732, simple_loss=0.2541, pruned_loss=0.0462, over 16773.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2751, pruned_loss=0.0535, over 1751766.30 frames. ], batch size: 39, lr: 4.62e-03, grad_scale: 2.0 2023-04-30 02:09:25,348 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=142272.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:09:42,880 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=142284.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:10:04,593 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.689e+02 2.354e+02 2.662e+02 3.206e+02 6.929e+02, threshold=5.325e+02, percent-clipped=1.0 2023-04-30 02:10:07,770 INFO [train.py:904] (4/8) Epoch 15, batch 200, loss[loss=0.1618, simple_loss=0.2529, pruned_loss=0.03534, over 17196.00 frames. ], tot_loss[loss=0.1894, simple_loss=0.2739, pruned_loss=0.05245, over 2106360.65 frames. ], batch size: 46, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:10:48,028 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-30 02:11:17,109 INFO [train.py:904] (4/8) Epoch 15, batch 250, loss[loss=0.1933, simple_loss=0.278, pruned_loss=0.0543, over 17033.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2721, pruned_loss=0.05236, over 2376853.27 frames. ], batch size: 53, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:11:17,845 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-30 02:12:22,521 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.243e+02 2.346e+02 2.797e+02 3.590e+02 6.679e+02, threshold=5.594e+02, percent-clipped=6.0 2023-04-30 02:12:25,436 INFO [train.py:904] (4/8) Epoch 15, batch 300, loss[loss=0.1859, simple_loss=0.2638, pruned_loss=0.05402, over 16818.00 frames. ], tot_loss[loss=0.1862, simple_loss=0.2693, pruned_loss=0.0515, over 2589771.63 frames. ], batch size: 83, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:13:36,231 INFO [train.py:904] (4/8) Epoch 15, batch 350, loss[loss=0.1981, simple_loss=0.285, pruned_loss=0.05562, over 16674.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.2666, pruned_loss=0.05091, over 2743186.47 frames. ], batch size: 62, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:14:42,901 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.497e+02 2.253e+02 2.722e+02 3.272e+02 5.602e+02, threshold=5.443e+02, percent-clipped=1.0 2023-04-30 02:14:45,198 INFO [train.py:904] (4/8) Epoch 15, batch 400, loss[loss=0.1819, simple_loss=0.2634, pruned_loss=0.05017, over 15555.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2646, pruned_loss=0.04959, over 2876351.73 frames. ], batch size: 190, lr: 4.61e-03, grad_scale: 4.0 2023-04-30 02:15:09,477 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2770, 4.3050, 4.7571, 2.5502, 4.9560, 4.9507, 3.5692, 3.9581], device='cuda:4'), covar=tensor([0.0604, 0.0157, 0.0139, 0.1037, 0.0048, 0.0124, 0.0349, 0.0308], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0101, 0.0086, 0.0138, 0.0071, 0.0111, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 02:15:54,199 INFO [train.py:904] (4/8) Epoch 15, batch 450, loss[loss=0.1821, simple_loss=0.2831, pruned_loss=0.04052, over 17029.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2629, pruned_loss=0.04882, over 2961246.55 frames. ], batch size: 50, lr: 4.61e-03, grad_scale: 4.0 2023-04-30 02:16:23,354 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=142572.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:16:24,628 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=142573.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:16:41,583 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=142584.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:17:03,318 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.426e+02 2.083e+02 2.627e+02 3.131e+02 6.454e+02, threshold=5.253e+02, percent-clipped=1.0 2023-04-30 02:17:05,222 INFO [train.py:904] (4/8) Epoch 15, batch 500, loss[loss=0.1679, simple_loss=0.2638, pruned_loss=0.03598, over 17130.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2615, pruned_loss=0.04777, over 3042266.33 frames. ], batch size: 47, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:17:28,778 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=142620.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:17:46,434 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=142632.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:17:49,278 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=142634.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:17:52,873 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8850, 3.8194, 4.3389, 2.1453, 4.5654, 4.5959, 3.2237, 3.4623], device='cuda:4'), covar=tensor([0.0703, 0.0235, 0.0227, 0.1119, 0.0063, 0.0161, 0.0425, 0.0374], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0102, 0.0088, 0.0139, 0.0072, 0.0113, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 02:18:13,784 INFO [train.py:904] (4/8) Epoch 15, batch 550, loss[loss=0.1677, simple_loss=0.2572, pruned_loss=0.03913, over 17216.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2601, pruned_loss=0.04655, over 3104712.49 frames. ], batch size: 45, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:19:22,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 2.274e+02 2.667e+02 3.434e+02 1.165e+03, threshold=5.335e+02, percent-clipped=9.0 2023-04-30 02:19:23,199 INFO [train.py:904] (4/8) Epoch 15, batch 600, loss[loss=0.1851, simple_loss=0.2538, pruned_loss=0.05827, over 16517.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2604, pruned_loss=0.04756, over 3153622.94 frames. ], batch size: 75, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:19:53,054 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=142724.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:20:00,180 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0933, 2.1113, 1.7423, 1.8118, 2.3080, 2.0212, 2.2254, 2.4270], device='cuda:4'), covar=tensor([0.0244, 0.0324, 0.0412, 0.0416, 0.0195, 0.0310, 0.0216, 0.0219], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0220, 0.0210, 0.0212, 0.0219, 0.0219, 0.0224, 0.0210], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:20:32,860 INFO [train.py:904] (4/8) Epoch 15, batch 650, loss[loss=0.1743, simple_loss=0.2716, pruned_loss=0.03851, over 17118.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2591, pruned_loss=0.04728, over 3174898.24 frames. ], batch size: 49, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:21:17,846 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=142785.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:21:25,448 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7318, 2.2907, 2.3584, 4.5181, 2.2073, 2.7118, 2.4056, 2.4786], device='cuda:4'), covar=tensor([0.0980, 0.3547, 0.2686, 0.0388, 0.4049, 0.2536, 0.3176, 0.3443], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0413, 0.0346, 0.0324, 0.0422, 0.0473, 0.0378, 0.0482], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:21:27,578 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-30 02:21:40,878 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.547e+02 2.460e+02 3.000e+02 3.809e+02 2.810e+03, threshold=6.001e+02, percent-clipped=15.0 2023-04-30 02:21:42,123 INFO [train.py:904] (4/8) Epoch 15, batch 700, loss[loss=0.2171, simple_loss=0.286, pruned_loss=0.07405, over 16344.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2581, pruned_loss=0.04693, over 3196282.68 frames. ], batch size: 165, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:22:13,465 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-30 02:22:50,284 INFO [train.py:904] (4/8) Epoch 15, batch 750, loss[loss=0.1535, simple_loss=0.2343, pruned_loss=0.03642, over 15865.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2586, pruned_loss=0.04771, over 3217896.82 frames. ], batch size: 35, lr: 4.61e-03, grad_scale: 2.0 2023-04-30 02:23:57,714 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.460e+02 2.171e+02 2.529e+02 2.994e+02 6.163e+02, threshold=5.058e+02, percent-clipped=1.0 2023-04-30 02:23:59,534 INFO [train.py:904] (4/8) Epoch 15, batch 800, loss[loss=0.1474, simple_loss=0.236, pruned_loss=0.02939, over 17182.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.259, pruned_loss=0.04778, over 3238157.61 frames. ], batch size: 46, lr: 4.60e-03, grad_scale: 4.0 2023-04-30 02:24:36,442 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=142929.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:24:36,583 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8628, 4.8522, 4.6861, 4.2121, 4.7583, 2.0167, 4.5091, 4.5298], device='cuda:4'), covar=tensor([0.0082, 0.0075, 0.0166, 0.0287, 0.0081, 0.2312, 0.0116, 0.0181], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0134, 0.0180, 0.0164, 0.0152, 0.0196, 0.0168, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:25:08,641 INFO [train.py:904] (4/8) Epoch 15, batch 850, loss[loss=0.167, simple_loss=0.2414, pruned_loss=0.04633, over 16810.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2585, pruned_loss=0.04638, over 3259899.25 frames. ], batch size: 83, lr: 4.60e-03, grad_scale: 4.0 2023-04-30 02:25:38,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5757, 3.6276, 3.3923, 3.0178, 3.1626, 3.4841, 3.2886, 3.2944], device='cuda:4'), covar=tensor([0.0581, 0.0678, 0.0306, 0.0312, 0.0634, 0.0422, 0.1226, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0366, 0.0318, 0.0299, 0.0334, 0.0349, 0.0219, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:26:15,122 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.443e+02 2.170e+02 2.499e+02 2.974e+02 6.190e+02, threshold=4.998e+02, percent-clipped=2.0 2023-04-30 02:26:16,305 INFO [train.py:904] (4/8) Epoch 15, batch 900, loss[loss=0.1749, simple_loss=0.271, pruned_loss=0.0394, over 17270.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2575, pruned_loss=0.04557, over 3273195.28 frames. ], batch size: 52, lr: 4.60e-03, grad_scale: 4.0 2023-04-30 02:26:43,171 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.40 vs. limit=5.0 2023-04-30 02:26:47,001 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-30 02:27:24,768 INFO [train.py:904] (4/8) Epoch 15, batch 950, loss[loss=0.1459, simple_loss=0.2331, pruned_loss=0.02928, over 16796.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2572, pruned_loss=0.04573, over 3274239.42 frames. ], batch size: 39, lr: 4.60e-03, grad_scale: 4.0 2023-04-30 02:27:46,638 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0573, 4.9770, 4.8778, 4.5359, 4.4993, 4.9544, 4.8365, 4.6671], device='cuda:4'), covar=tensor([0.0670, 0.0846, 0.0375, 0.0355, 0.1128, 0.0526, 0.0443, 0.0796], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0368, 0.0319, 0.0300, 0.0336, 0.0350, 0.0220, 0.0380], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:28:02,486 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=143080.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:28:30,229 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.522e+02 2.127e+02 2.534e+02 2.926e+02 6.813e+02, threshold=5.068e+02, percent-clipped=2.0 2023-04-30 02:28:31,471 INFO [train.py:904] (4/8) Epoch 15, batch 1000, loss[loss=0.1456, simple_loss=0.2346, pruned_loss=0.02828, over 16834.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2554, pruned_loss=0.04536, over 3286225.88 frames. ], batch size: 42, lr: 4.60e-03, grad_scale: 4.0 2023-04-30 02:28:43,496 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7156, 3.5252, 3.7574, 2.8955, 3.4258, 3.9218, 3.6927, 2.3140], device='cuda:4'), covar=tensor([0.0389, 0.0212, 0.0050, 0.0287, 0.0103, 0.0087, 0.0073, 0.0396], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0075, 0.0075, 0.0131, 0.0088, 0.0098, 0.0085, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 02:28:55,778 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6526, 4.5444, 4.5208, 4.2427, 4.2297, 4.5584, 4.3944, 4.3377], device='cuda:4'), covar=tensor([0.0636, 0.0809, 0.0279, 0.0301, 0.0812, 0.0484, 0.0466, 0.0676], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0366, 0.0318, 0.0299, 0.0334, 0.0348, 0.0218, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:29:21,802 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5374, 2.9267, 2.9934, 1.9682, 2.6521, 2.2603, 2.9641, 3.1981], device='cuda:4'), covar=tensor([0.0348, 0.0850, 0.0570, 0.1886, 0.0892, 0.0887, 0.0769, 0.0950], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0152, 0.0162, 0.0148, 0.0139, 0.0126, 0.0139, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 02:29:41,403 INFO [train.py:904] (4/8) Epoch 15, batch 1050, loss[loss=0.1484, simple_loss=0.2314, pruned_loss=0.0327, over 16722.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2557, pruned_loss=0.04498, over 3297175.59 frames. ], batch size: 39, lr: 4.60e-03, grad_scale: 4.0 2023-04-30 02:30:46,963 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.748e+02 2.203e+02 2.568e+02 3.063e+02 1.237e+03, threshold=5.135e+02, percent-clipped=3.0 2023-04-30 02:30:49,004 INFO [train.py:904] (4/8) Epoch 15, batch 1100, loss[loss=0.1419, simple_loss=0.2268, pruned_loss=0.02849, over 16998.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2543, pruned_loss=0.04409, over 3308031.95 frames. ], batch size: 41, lr: 4.60e-03, grad_scale: 4.0 2023-04-30 02:31:26,711 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=143229.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:31:40,448 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2457, 5.5565, 5.3236, 5.3924, 5.0642, 4.9866, 4.9867, 5.6755], device='cuda:4'), covar=tensor([0.1157, 0.0904, 0.1029, 0.0782, 0.0813, 0.0767, 0.1076, 0.0849], device='cuda:4'), in_proj_covar=tensor([0.0613, 0.0764, 0.0622, 0.0548, 0.0482, 0.0489, 0.0641, 0.0580], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:31:58,384 INFO [train.py:904] (4/8) Epoch 15, batch 1150, loss[loss=0.1605, simple_loss=0.2409, pruned_loss=0.04001, over 16853.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2536, pruned_loss=0.04359, over 3309305.11 frames. ], batch size: 96, lr: 4.60e-03, grad_scale: 4.0 2023-04-30 02:32:02,096 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7196, 2.5242, 2.4157, 3.9881, 3.1612, 3.9955, 1.5740, 2.8499], device='cuda:4'), covar=tensor([0.1507, 0.0757, 0.1156, 0.0183, 0.0182, 0.0376, 0.1569, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0166, 0.0187, 0.0167, 0.0198, 0.0214, 0.0191, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 02:32:06,993 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8072, 2.6730, 2.5926, 4.8478, 3.8963, 4.3900, 1.6389, 3.1960], device='cuda:4'), covar=tensor([0.1377, 0.0793, 0.1211, 0.0168, 0.0284, 0.0397, 0.1532, 0.0741], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0166, 0.0187, 0.0167, 0.0198, 0.0214, 0.0191, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 02:32:34,543 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=143277.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:33:01,280 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8792, 3.8509, 4.3207, 2.0558, 4.4576, 4.4889, 3.1600, 3.4915], device='cuda:4'), covar=tensor([0.0662, 0.0219, 0.0194, 0.1163, 0.0061, 0.0150, 0.0412, 0.0362], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0105, 0.0090, 0.0141, 0.0073, 0.0116, 0.0126, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 02:33:07,200 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.631e+02 2.162e+02 2.475e+02 3.140e+02 4.983e+02, threshold=4.950e+02, percent-clipped=0.0 2023-04-30 02:33:08,295 INFO [train.py:904] (4/8) Epoch 15, batch 1200, loss[loss=0.171, simple_loss=0.2589, pruned_loss=0.0415, over 17172.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2535, pruned_loss=0.04293, over 3324294.48 frames. ], batch size: 46, lr: 4.60e-03, grad_scale: 8.0 2023-04-30 02:33:11,215 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9845, 4.0227, 2.4506, 4.6789, 3.0339, 4.5883, 2.8382, 3.3062], device='cuda:4'), covar=tensor([0.0217, 0.0325, 0.1424, 0.0237, 0.0746, 0.0431, 0.1256, 0.0598], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0171, 0.0192, 0.0147, 0.0170, 0.0213, 0.0200, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 02:33:41,576 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.26 vs. limit=5.0 2023-04-30 02:33:43,178 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9792, 2.4689, 2.6240, 1.8502, 2.7432, 2.7657, 2.4426, 2.3234], device='cuda:4'), covar=tensor([0.0688, 0.0240, 0.0230, 0.0984, 0.0107, 0.0256, 0.0480, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0104, 0.0089, 0.0140, 0.0073, 0.0116, 0.0126, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 02:33:44,410 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8637, 4.6134, 4.8616, 5.0540, 5.2419, 4.6009, 5.1941, 5.2276], device='cuda:4'), covar=tensor([0.1674, 0.1352, 0.1640, 0.0708, 0.0533, 0.1018, 0.0499, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0594, 0.0730, 0.0872, 0.0746, 0.0559, 0.0579, 0.0591, 0.0695], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:34:16,292 INFO [train.py:904] (4/8) Epoch 15, batch 1250, loss[loss=0.1773, simple_loss=0.2548, pruned_loss=0.04992, over 16857.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2532, pruned_loss=0.04313, over 3304093.80 frames. ], batch size: 83, lr: 4.60e-03, grad_scale: 8.0 2023-04-30 02:34:56,619 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=143380.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:35:25,388 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=143400.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 02:35:26,163 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.456e+02 2.199e+02 2.560e+02 2.965e+02 4.851e+02, threshold=5.120e+02, percent-clipped=0.0 2023-04-30 02:35:27,819 INFO [train.py:904] (4/8) Epoch 15, batch 1300, loss[loss=0.1551, simple_loss=0.2383, pruned_loss=0.03599, over 17033.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2532, pruned_loss=0.04333, over 3312360.80 frames. ], batch size: 41, lr: 4.60e-03, grad_scale: 8.0 2023-04-30 02:36:03,860 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=143428.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:36:05,510 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-30 02:36:36,226 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-30 02:36:37,208 INFO [train.py:904] (4/8) Epoch 15, batch 1350, loss[loss=0.1851, simple_loss=0.2573, pruned_loss=0.05647, over 16777.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2543, pruned_loss=0.04316, over 3319025.06 frames. ], batch size: 83, lr: 4.60e-03, grad_scale: 8.0 2023-04-30 02:36:39,935 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8865, 4.6142, 4.7270, 5.1180, 5.3210, 4.7336, 5.3485, 5.2973], device='cuda:4'), covar=tensor([0.1993, 0.1698, 0.2354, 0.1005, 0.0853, 0.0826, 0.0691, 0.0860], device='cuda:4'), in_proj_covar=tensor([0.0604, 0.0742, 0.0889, 0.0759, 0.0568, 0.0589, 0.0599, 0.0707], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:36:50,750 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=143461.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 02:37:06,534 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=143473.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 02:37:45,745 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.631e+02 2.266e+02 2.683e+02 3.217e+02 5.542e+02, threshold=5.366e+02, percent-clipped=1.0 2023-04-30 02:37:47,548 INFO [train.py:904] (4/8) Epoch 15, batch 1400, loss[loss=0.187, simple_loss=0.2554, pruned_loss=0.05932, over 16453.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2538, pruned_loss=0.04346, over 3306409.46 frames. ], batch size: 146, lr: 4.59e-03, grad_scale: 8.0 2023-04-30 02:38:31,331 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=143534.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 02:38:46,920 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-30 02:38:56,002 INFO [train.py:904] (4/8) Epoch 15, batch 1450, loss[loss=0.16, simple_loss=0.2496, pruned_loss=0.03519, over 17126.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2543, pruned_loss=0.04383, over 3312940.69 frames. ], batch size: 47, lr: 4.59e-03, grad_scale: 8.0 2023-04-30 02:39:59,444 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.67 vs. limit=5.0 2023-04-30 02:40:05,565 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.540e+02 2.183e+02 2.517e+02 3.188e+02 7.484e+02, threshold=5.034e+02, percent-clipped=1.0 2023-04-30 02:40:06,052 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3871, 3.3429, 3.6312, 1.8637, 3.7114, 3.6895, 2.9809, 2.7714], device='cuda:4'), covar=tensor([0.0758, 0.0210, 0.0171, 0.1122, 0.0092, 0.0186, 0.0405, 0.0417], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0104, 0.0091, 0.0140, 0.0073, 0.0117, 0.0126, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 02:40:06,727 INFO [train.py:904] (4/8) Epoch 15, batch 1500, loss[loss=0.196, simple_loss=0.2637, pruned_loss=0.0642, over 16713.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2541, pruned_loss=0.04389, over 3312409.44 frames. ], batch size: 134, lr: 4.59e-03, grad_scale: 8.0 2023-04-30 02:41:14,536 INFO [train.py:904] (4/8) Epoch 15, batch 1550, loss[loss=0.1635, simple_loss=0.2593, pruned_loss=0.0338, over 17288.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2556, pruned_loss=0.0454, over 3305929.44 frames. ], batch size: 52, lr: 4.59e-03, grad_scale: 8.0 2023-04-30 02:41:53,849 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0919, 4.5781, 4.5628, 3.4418, 3.7523, 4.5562, 4.0077, 2.5283], device='cuda:4'), covar=tensor([0.0392, 0.0061, 0.0032, 0.0281, 0.0118, 0.0068, 0.0072, 0.0423], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0075, 0.0075, 0.0131, 0.0088, 0.0098, 0.0086, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 02:41:57,691 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1005, 5.6183, 5.7870, 5.4914, 5.4981, 6.1412, 5.7110, 5.4499], device='cuda:4'), covar=tensor([0.0805, 0.1868, 0.1859, 0.2074, 0.2808, 0.0944, 0.1325, 0.2143], device='cuda:4'), in_proj_covar=tensor([0.0387, 0.0554, 0.0608, 0.0470, 0.0634, 0.0637, 0.0482, 0.0624], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 02:41:59,479 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-30 02:42:22,872 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.594e+02 2.352e+02 2.733e+02 3.255e+02 5.583e+02, threshold=5.465e+02, percent-clipped=1.0 2023-04-30 02:42:24,071 INFO [train.py:904] (4/8) Epoch 15, batch 1600, loss[loss=0.1921, simple_loss=0.2881, pruned_loss=0.04804, over 17021.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2584, pruned_loss=0.04669, over 3313487.45 frames. ], batch size: 50, lr: 4.59e-03, grad_scale: 8.0 2023-04-30 02:43:13,898 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8231, 2.8520, 2.5657, 4.3795, 3.6002, 4.2116, 1.6776, 3.0460], device='cuda:4'), covar=tensor([0.1329, 0.0648, 0.1074, 0.0172, 0.0171, 0.0367, 0.1428, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0165, 0.0185, 0.0167, 0.0198, 0.0214, 0.0189, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 02:43:33,399 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.96 vs. limit=5.0 2023-04-30 02:43:35,419 INFO [train.py:904] (4/8) Epoch 15, batch 1650, loss[loss=0.1951, simple_loss=0.2594, pruned_loss=0.0654, over 16869.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2592, pruned_loss=0.04645, over 3323793.14 frames. ], batch size: 109, lr: 4.59e-03, grad_scale: 8.0 2023-04-30 02:43:40,913 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=143756.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 02:43:48,924 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-30 02:44:46,123 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.543e+02 2.448e+02 2.826e+02 3.842e+02 9.668e+02, threshold=5.653e+02, percent-clipped=8.0 2023-04-30 02:44:46,138 INFO [train.py:904] (4/8) Epoch 15, batch 1700, loss[loss=0.1743, simple_loss=0.2578, pruned_loss=0.04545, over 16741.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2606, pruned_loss=0.0469, over 3321754.64 frames. ], batch size: 102, lr: 4.59e-03, grad_scale: 4.0 2023-04-30 02:44:58,165 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4631, 4.4612, 4.6750, 4.4895, 4.5642, 5.1246, 4.7230, 4.3790], device='cuda:4'), covar=tensor([0.1587, 0.2189, 0.2354, 0.2377, 0.2898, 0.1153, 0.1535, 0.2725], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0552, 0.0608, 0.0467, 0.0631, 0.0633, 0.0481, 0.0622], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 02:45:22,437 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=143829.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 02:45:54,349 INFO [train.py:904] (4/8) Epoch 15, batch 1750, loss[loss=0.1617, simple_loss=0.2578, pruned_loss=0.03281, over 16854.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2613, pruned_loss=0.04675, over 3315031.04 frames. ], batch size: 42, lr: 4.59e-03, grad_scale: 4.0 2023-04-30 02:46:04,362 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-30 02:47:05,601 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.550e+02 2.250e+02 2.638e+02 3.028e+02 6.204e+02, threshold=5.275e+02, percent-clipped=1.0 2023-04-30 02:47:05,620 INFO [train.py:904] (4/8) Epoch 15, batch 1800, loss[loss=0.2099, simple_loss=0.2868, pruned_loss=0.06648, over 16886.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2623, pruned_loss=0.04658, over 3314283.53 frames. ], batch size: 109, lr: 4.59e-03, grad_scale: 4.0 2023-04-30 02:47:16,668 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-30 02:47:17,099 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2201, 5.8358, 6.0275, 5.7325, 5.8076, 6.3365, 5.9557, 5.6323], device='cuda:4'), covar=tensor([0.0782, 0.1734, 0.1994, 0.1888, 0.2490, 0.0889, 0.1220, 0.2156], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0551, 0.0604, 0.0465, 0.0628, 0.0631, 0.0478, 0.0620], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 02:47:59,790 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-30 02:48:15,536 INFO [train.py:904] (4/8) Epoch 15, batch 1850, loss[loss=0.2068, simple_loss=0.2928, pruned_loss=0.0604, over 12376.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.263, pruned_loss=0.04639, over 3321817.53 frames. ], batch size: 246, lr: 4.59e-03, grad_scale: 2.0 2023-04-30 02:48:18,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9300, 4.6935, 4.9118, 5.1631, 5.3642, 4.7139, 5.2851, 5.3547], device='cuda:4'), covar=tensor([0.1767, 0.1449, 0.2073, 0.0854, 0.0590, 0.0926, 0.0635, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0602, 0.0741, 0.0890, 0.0763, 0.0568, 0.0591, 0.0598, 0.0707], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:48:36,076 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2516, 4.1661, 4.4712, 2.3358, 4.7652, 4.7579, 3.3096, 3.8371], device='cuda:4'), covar=tensor([0.0594, 0.0190, 0.0207, 0.1040, 0.0057, 0.0130, 0.0382, 0.0323], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0105, 0.0092, 0.0141, 0.0074, 0.0118, 0.0126, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 02:49:30,761 INFO [train.py:904] (4/8) Epoch 15, batch 1900, loss[loss=0.1647, simple_loss=0.2469, pruned_loss=0.04121, over 16715.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2629, pruned_loss=0.04577, over 3329022.57 frames. ], batch size: 89, lr: 4.59e-03, grad_scale: 2.0 2023-04-30 02:49:31,843 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.452e+02 2.211e+02 2.636e+02 2.995e+02 6.158e+02, threshold=5.272e+02, percent-clipped=2.0 2023-04-30 02:49:40,875 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7547, 2.6247, 2.3530, 2.5590, 2.9433, 2.7758, 3.4080, 3.2472], device='cuda:4'), covar=tensor([0.0104, 0.0351, 0.0395, 0.0349, 0.0248, 0.0335, 0.0222, 0.0223], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0223, 0.0215, 0.0215, 0.0224, 0.0223, 0.0230, 0.0216], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:50:07,536 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7074, 2.4796, 2.2438, 3.7028, 2.9478, 3.8633, 1.3818, 2.7794], device='cuda:4'), covar=tensor([0.1350, 0.0709, 0.1234, 0.0179, 0.0160, 0.0351, 0.1546, 0.0809], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0164, 0.0184, 0.0166, 0.0198, 0.0212, 0.0188, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 02:50:39,887 INFO [train.py:904] (4/8) Epoch 15, batch 1950, loss[loss=0.1836, simple_loss=0.2707, pruned_loss=0.04823, over 16259.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2632, pruned_loss=0.04547, over 3331519.91 frames. ], batch size: 165, lr: 4.59e-03, grad_scale: 2.0 2023-04-30 02:50:46,764 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=144056.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 02:51:04,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7820, 3.8478, 4.1048, 4.0852, 4.1215, 3.8704, 3.9064, 3.8572], device='cuda:4'), covar=tensor([0.0371, 0.0566, 0.0393, 0.0446, 0.0585, 0.0448, 0.0817, 0.0544], device='cuda:4'), in_proj_covar=tensor([0.0375, 0.0398, 0.0395, 0.0375, 0.0439, 0.0419, 0.0508, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 02:51:25,948 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9884, 3.0703, 2.6159, 4.9465, 4.0989, 4.5879, 1.6779, 3.2769], device='cuda:4'), covar=tensor([0.1227, 0.0707, 0.1231, 0.0196, 0.0227, 0.0379, 0.1494, 0.0725], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0164, 0.0185, 0.0166, 0.0198, 0.0212, 0.0188, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 02:51:49,544 INFO [train.py:904] (4/8) Epoch 15, batch 2000, loss[loss=0.1718, simple_loss=0.2646, pruned_loss=0.03946, over 17105.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2623, pruned_loss=0.04497, over 3328001.89 frames. ], batch size: 47, lr: 4.59e-03, grad_scale: 4.0 2023-04-30 02:51:51,361 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.567e+02 2.196e+02 2.584e+02 2.966e+02 4.475e+02, threshold=5.169e+02, percent-clipped=0.0 2023-04-30 02:51:52,651 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=144104.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 02:52:23,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0287, 4.5002, 3.3992, 2.3958, 2.8911, 2.6179, 4.8715, 3.8318], device='cuda:4'), covar=tensor([0.2556, 0.0576, 0.1475, 0.2675, 0.2828, 0.1905, 0.0354, 0.1280], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0264, 0.0294, 0.0293, 0.0286, 0.0239, 0.0278, 0.0318], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 02:52:27,252 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=144129.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 02:52:28,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0978, 5.0864, 4.8750, 4.3643, 4.9226, 1.9166, 4.6769, 4.8253], device='cuda:4'), covar=tensor([0.0084, 0.0078, 0.0172, 0.0355, 0.0092, 0.2535, 0.0126, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0138, 0.0186, 0.0170, 0.0157, 0.0198, 0.0173, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 02:52:30,156 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-30 02:52:58,033 INFO [train.py:904] (4/8) Epoch 15, batch 2050, loss[loss=0.1506, simple_loss=0.2341, pruned_loss=0.0336, over 16837.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2622, pruned_loss=0.0457, over 3322161.81 frames. ], batch size: 42, lr: 4.58e-03, grad_scale: 4.0 2023-04-30 02:53:32,905 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=144177.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 02:54:07,934 INFO [train.py:904] (4/8) Epoch 15, batch 2100, loss[loss=0.1587, simple_loss=0.2529, pruned_loss=0.03223, over 17216.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2624, pruned_loss=0.04568, over 3317955.58 frames. ], batch size: 45, lr: 4.58e-03, grad_scale: 4.0 2023-04-30 02:54:08,981 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.706e+02 2.513e+02 2.931e+02 3.819e+02 1.829e+03, threshold=5.862e+02, percent-clipped=10.0 2023-04-30 02:55:17,937 INFO [train.py:904] (4/8) Epoch 15, batch 2150, loss[loss=0.1879, simple_loss=0.2704, pruned_loss=0.05272, over 16599.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2636, pruned_loss=0.04649, over 3317730.19 frames. ], batch size: 68, lr: 4.58e-03, grad_scale: 4.0 2023-04-30 02:55:19,633 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8857, 4.3108, 3.1673, 2.3059, 2.8655, 2.7620, 4.7349, 3.7788], device='cuda:4'), covar=tensor([0.2760, 0.0626, 0.1712, 0.2672, 0.2589, 0.1788, 0.0383, 0.1229], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0264, 0.0294, 0.0292, 0.0286, 0.0239, 0.0277, 0.0317], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 02:55:31,820 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.27 vs. limit=5.0 2023-04-30 02:56:20,050 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3236, 5.3387, 5.1756, 4.7824, 4.5981, 5.3044, 5.2036, 4.7900], device='cuda:4'), covar=tensor([0.0603, 0.0553, 0.0326, 0.0332, 0.1246, 0.0447, 0.0263, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0277, 0.0380, 0.0330, 0.0311, 0.0343, 0.0359, 0.0225, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 02:56:25,352 INFO [train.py:904] (4/8) Epoch 15, batch 2200, loss[loss=0.2492, simple_loss=0.3131, pruned_loss=0.09262, over 11942.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2644, pruned_loss=0.04691, over 3310042.63 frames. ], batch size: 246, lr: 4.58e-03, grad_scale: 4.0 2023-04-30 02:56:27,073 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.566e+02 2.322e+02 2.712e+02 3.377e+02 6.214e+02, threshold=5.423e+02, percent-clipped=1.0 2023-04-30 02:56:52,880 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 02:57:36,224 INFO [train.py:904] (4/8) Epoch 15, batch 2250, loss[loss=0.2049, simple_loss=0.2746, pruned_loss=0.06757, over 16695.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2657, pruned_loss=0.04837, over 3300391.51 frames. ], batch size: 134, lr: 4.58e-03, grad_scale: 4.0 2023-04-30 02:58:37,682 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-30 02:58:46,690 INFO [train.py:904] (4/8) Epoch 15, batch 2300, loss[loss=0.2061, simple_loss=0.3041, pruned_loss=0.0541, over 17057.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2651, pruned_loss=0.04787, over 3309586.64 frames. ], batch size: 53, lr: 4.58e-03, grad_scale: 4.0 2023-04-30 02:58:47,877 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.638e+02 2.350e+02 2.907e+02 3.506e+02 6.150e+02, threshold=5.814e+02, percent-clipped=3.0 2023-04-30 02:58:58,083 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=144411.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 02:59:53,214 INFO [train.py:904] (4/8) Epoch 15, batch 2350, loss[loss=0.1791, simple_loss=0.256, pruned_loss=0.05109, over 16817.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2667, pruned_loss=0.04898, over 3298535.41 frames. ], batch size: 42, lr: 4.58e-03, grad_scale: 4.0 2023-04-30 03:00:20,672 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=144472.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:01:02,756 INFO [train.py:904] (4/8) Epoch 15, batch 2400, loss[loss=0.1786, simple_loss=0.2727, pruned_loss=0.04229, over 17088.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2671, pruned_loss=0.04862, over 3310659.17 frames. ], batch size: 55, lr: 4.58e-03, grad_scale: 8.0 2023-04-30 03:01:04,724 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.656e+02 2.391e+02 2.805e+02 3.317e+02 7.772e+02, threshold=5.609e+02, percent-clipped=1.0 2023-04-30 03:01:22,894 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-30 03:02:09,960 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-30 03:02:10,317 INFO [train.py:904] (4/8) Epoch 15, batch 2450, loss[loss=0.1632, simple_loss=0.2454, pruned_loss=0.04046, over 16824.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2665, pruned_loss=0.04803, over 3317596.70 frames. ], batch size: 96, lr: 4.58e-03, grad_scale: 8.0 2023-04-30 03:02:44,170 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5262, 2.2393, 2.3848, 4.5402, 2.1594, 2.6969, 2.3455, 2.4297], device='cuda:4'), covar=tensor([0.1098, 0.3790, 0.2696, 0.0415, 0.4068, 0.2588, 0.3362, 0.3453], device='cuda:4'), in_proj_covar=tensor([0.0379, 0.0417, 0.0349, 0.0327, 0.0423, 0.0480, 0.0382, 0.0488], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:03:17,689 INFO [train.py:904] (4/8) Epoch 15, batch 2500, loss[loss=0.1795, simple_loss=0.2535, pruned_loss=0.05274, over 16918.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2659, pruned_loss=0.04783, over 3321985.23 frames. ], batch size: 116, lr: 4.58e-03, grad_scale: 8.0 2023-04-30 03:03:18,674 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.596e+02 2.258e+02 2.678e+02 3.424e+02 5.626e+02, threshold=5.355e+02, percent-clipped=1.0 2023-04-30 03:03:19,132 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1139, 1.8583, 2.5341, 3.0111, 2.7278, 3.3824, 2.0302, 3.3970], device='cuda:4'), covar=tensor([0.0166, 0.0477, 0.0279, 0.0216, 0.0265, 0.0165, 0.0493, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0183, 0.0169, 0.0172, 0.0181, 0.0138, 0.0183, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:04:26,856 INFO [train.py:904] (4/8) Epoch 15, batch 2550, loss[loss=0.1619, simple_loss=0.2582, pruned_loss=0.03276, over 17127.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2655, pruned_loss=0.04786, over 3330185.85 frames. ], batch size: 47, lr: 4.58e-03, grad_scale: 8.0 2023-04-30 03:05:34,896 INFO [train.py:904] (4/8) Epoch 15, batch 2600, loss[loss=0.17, simple_loss=0.2642, pruned_loss=0.03792, over 15915.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2658, pruned_loss=0.04718, over 3328089.72 frames. ], batch size: 35, lr: 4.58e-03, grad_scale: 8.0 2023-04-30 03:05:36,052 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.651e+02 2.335e+02 2.561e+02 3.164e+02 7.288e+02, threshold=5.122e+02, percent-clipped=2.0 2023-04-30 03:05:36,403 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6519, 4.4043, 4.6682, 4.8558, 5.0064, 4.4326, 4.9697, 4.9790], device='cuda:4'), covar=tensor([0.1615, 0.1429, 0.1721, 0.0776, 0.0623, 0.1093, 0.0757, 0.0730], device='cuda:4'), in_proj_covar=tensor([0.0611, 0.0755, 0.0903, 0.0775, 0.0581, 0.0604, 0.0610, 0.0716], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:06:43,554 INFO [train.py:904] (4/8) Epoch 15, batch 2650, loss[loss=0.1846, simple_loss=0.277, pruned_loss=0.04615, over 17026.00 frames. ], tot_loss[loss=0.1794, simple_loss=0.2653, pruned_loss=0.04672, over 3330019.98 frames. ], batch size: 53, lr: 4.58e-03, grad_scale: 8.0 2023-04-30 03:07:05,951 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=144767.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:07:50,678 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5739, 3.5349, 3.8701, 1.9759, 3.9509, 3.9351, 3.0083, 2.8686], device='cuda:4'), covar=tensor([0.0751, 0.0207, 0.0145, 0.1164, 0.0086, 0.0148, 0.0433, 0.0448], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0105, 0.0091, 0.0139, 0.0074, 0.0118, 0.0125, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 03:07:53,565 INFO [train.py:904] (4/8) Epoch 15, batch 2700, loss[loss=0.174, simple_loss=0.2742, pruned_loss=0.0369, over 16694.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2663, pruned_loss=0.04696, over 3326948.21 frames. ], batch size: 57, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:07:54,733 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.560e+02 2.168e+02 2.529e+02 3.023e+02 4.642e+02, threshold=5.059e+02, percent-clipped=0.0 2023-04-30 03:08:56,998 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1736, 2.0596, 2.2501, 3.9288, 2.1092, 2.4346, 2.1413, 2.2569], device='cuda:4'), covar=tensor([0.1284, 0.3613, 0.2704, 0.0506, 0.3682, 0.2451, 0.3647, 0.2924], device='cuda:4'), in_proj_covar=tensor([0.0380, 0.0416, 0.0347, 0.0327, 0.0422, 0.0480, 0.0380, 0.0488], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:09:02,445 INFO [train.py:904] (4/8) Epoch 15, batch 2750, loss[loss=0.1702, simple_loss=0.2587, pruned_loss=0.04084, over 17110.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.266, pruned_loss=0.04657, over 3325722.42 frames. ], batch size: 47, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:09:04,151 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6103, 2.5869, 2.2185, 3.9372, 3.0317, 3.9273, 1.4779, 2.6390], device='cuda:4'), covar=tensor([0.1644, 0.0724, 0.1344, 0.0217, 0.0241, 0.0387, 0.1713, 0.0870], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0165, 0.0186, 0.0170, 0.0201, 0.0214, 0.0189, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 03:10:11,019 INFO [train.py:904] (4/8) Epoch 15, batch 2800, loss[loss=0.241, simple_loss=0.3044, pruned_loss=0.08875, over 12040.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2655, pruned_loss=0.04628, over 3319692.56 frames. ], batch size: 247, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:10:12,143 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.381e+02 2.181e+02 2.488e+02 3.014e+02 5.995e+02, threshold=4.976e+02, percent-clipped=2.0 2023-04-30 03:10:26,076 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1969, 2.0469, 1.7093, 1.7636, 2.3136, 2.0300, 2.1761, 2.4452], device='cuda:4'), covar=tensor([0.0181, 0.0382, 0.0442, 0.0435, 0.0246, 0.0306, 0.0192, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0224, 0.0213, 0.0215, 0.0225, 0.0223, 0.0230, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:10:46,073 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-30 03:11:21,054 INFO [train.py:904] (4/8) Epoch 15, batch 2850, loss[loss=0.1775, simple_loss=0.2707, pruned_loss=0.04212, over 16748.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2649, pruned_loss=0.0463, over 3308088.33 frames. ], batch size: 57, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:12:22,184 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=144996.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:12:31,819 INFO [train.py:904] (4/8) Epoch 15, batch 2900, loss[loss=0.1673, simple_loss=0.2552, pruned_loss=0.03969, over 16552.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2643, pruned_loss=0.04645, over 3300545.06 frames. ], batch size: 68, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:12:33,009 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.445e+02 2.844e+02 3.300e+02 6.709e+02, threshold=5.687e+02, percent-clipped=6.0 2023-04-30 03:13:06,879 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145027.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:13:40,920 INFO [train.py:904] (4/8) Epoch 15, batch 2950, loss[loss=0.1608, simple_loss=0.2418, pruned_loss=0.03991, over 17001.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2625, pruned_loss=0.04648, over 3307409.20 frames. ], batch size: 41, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:13:44,054 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-30 03:13:47,740 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145057.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 03:13:58,438 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7677, 3.9482, 2.5040, 4.6486, 3.0579, 4.5090, 2.4885, 3.1905], device='cuda:4'), covar=tensor([0.0275, 0.0344, 0.1413, 0.0188, 0.0757, 0.0460, 0.1427, 0.0670], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0173, 0.0192, 0.0151, 0.0170, 0.0217, 0.0201, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 03:14:01,082 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145066.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:14:02,154 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=145067.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:14:31,753 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0846, 3.9270, 4.1590, 4.2908, 4.3782, 3.9668, 4.1654, 4.3663], device='cuda:4'), covar=tensor([0.1456, 0.1139, 0.1268, 0.0618, 0.0564, 0.1495, 0.1557, 0.0649], device='cuda:4'), in_proj_covar=tensor([0.0611, 0.0757, 0.0907, 0.0777, 0.0581, 0.0604, 0.0613, 0.0717], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:14:31,801 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145088.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:14:49,683 INFO [train.py:904] (4/8) Epoch 15, batch 3000, loss[loss=0.1993, simple_loss=0.267, pruned_loss=0.06577, over 16906.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2623, pruned_loss=0.04702, over 3308959.89 frames. ], batch size: 109, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:14:49,683 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 03:14:58,806 INFO [train.py:938] (4/8) Epoch 15, validation: loss=0.138, simple_loss=0.2438, pruned_loss=0.01616, over 944034.00 frames. 2023-04-30 03:14:58,807 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 03:15:00,800 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.691e+02 2.402e+02 2.841e+02 3.286e+02 6.614e+02, threshold=5.681e+02, percent-clipped=1.0 2023-04-30 03:15:17,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=145115.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:15:34,506 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145127.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:16:07,544 INFO [train.py:904] (4/8) Epoch 15, batch 3050, loss[loss=0.155, simple_loss=0.2416, pruned_loss=0.0342, over 16807.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2627, pruned_loss=0.04719, over 3314489.26 frames. ], batch size: 42, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:17:11,374 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-30 03:17:18,166 INFO [train.py:904] (4/8) Epoch 15, batch 3100, loss[loss=0.1539, simple_loss=0.245, pruned_loss=0.03142, over 17126.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2626, pruned_loss=0.04677, over 3311690.03 frames. ], batch size: 47, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:17:19,336 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.571e+02 2.445e+02 2.806e+02 3.389e+02 5.168e+02, threshold=5.611e+02, percent-clipped=0.0 2023-04-30 03:18:28,445 INFO [train.py:904] (4/8) Epoch 15, batch 3150, loss[loss=0.1671, simple_loss=0.2679, pruned_loss=0.03319, over 17117.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.262, pruned_loss=0.04677, over 3309806.84 frames. ], batch size: 48, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:19:01,967 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0754, 4.7375, 5.0171, 5.2370, 5.4785, 4.7446, 5.4084, 5.4180], device='cuda:4'), covar=tensor([0.1595, 0.1440, 0.1930, 0.0761, 0.0516, 0.0885, 0.0478, 0.0584], device='cuda:4'), in_proj_covar=tensor([0.0615, 0.0762, 0.0911, 0.0778, 0.0582, 0.0608, 0.0612, 0.0718], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:19:37,246 INFO [train.py:904] (4/8) Epoch 15, batch 3200, loss[loss=0.1768, simple_loss=0.2595, pruned_loss=0.04709, over 16562.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2612, pruned_loss=0.04608, over 3306392.92 frames. ], batch size: 68, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:19:38,469 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.650e+02 2.241e+02 2.735e+02 3.234e+02 5.514e+02, threshold=5.469e+02, percent-clipped=0.0 2023-04-30 03:20:46,509 INFO [train.py:904] (4/8) Epoch 15, batch 3250, loss[loss=0.1922, simple_loss=0.2913, pruned_loss=0.04661, over 16700.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2616, pruned_loss=0.04569, over 3316283.77 frames. ], batch size: 57, lr: 4.57e-03, grad_scale: 8.0 2023-04-30 03:20:46,748 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145352.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 03:21:13,276 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145371.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:21:30,555 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145383.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:21:38,011 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3891, 5.3301, 5.2061, 4.7008, 4.7954, 5.2817, 5.1587, 4.8770], device='cuda:4'), covar=tensor([0.0581, 0.0455, 0.0271, 0.0317, 0.1136, 0.0419, 0.0307, 0.0788], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0390, 0.0335, 0.0320, 0.0352, 0.0366, 0.0227, 0.0398], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 03:21:38,144 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4725, 3.7323, 4.0258, 2.1377, 3.2518, 2.5116, 3.9617, 3.8997], device='cuda:4'), covar=tensor([0.0268, 0.0836, 0.0440, 0.1802, 0.0721, 0.0909, 0.0536, 0.0960], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0154, 0.0162, 0.0146, 0.0139, 0.0126, 0.0140, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 03:21:57,383 INFO [train.py:904] (4/8) Epoch 15, batch 3300, loss[loss=0.1854, simple_loss=0.2667, pruned_loss=0.05199, over 16900.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2628, pruned_loss=0.04659, over 3318032.37 frames. ], batch size: 96, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:21:58,629 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.241e+02 2.723e+02 3.266e+02 5.157e+02, threshold=5.447e+02, percent-clipped=0.0 2023-04-30 03:22:25,242 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145422.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:22:38,829 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145432.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:23:06,148 INFO [train.py:904] (4/8) Epoch 15, batch 3350, loss[loss=0.1569, simple_loss=0.2411, pruned_loss=0.03633, over 16975.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2626, pruned_loss=0.04598, over 3331495.68 frames. ], batch size: 41, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:23:55,378 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7983, 3.9015, 2.2058, 4.5205, 2.9658, 4.4515, 2.4712, 3.0623], device='cuda:4'), covar=tensor([0.0262, 0.0347, 0.1609, 0.0202, 0.0770, 0.0467, 0.1456, 0.0731], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0172, 0.0191, 0.0151, 0.0169, 0.0216, 0.0200, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 03:24:07,872 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.14 vs. limit=5.0 2023-04-30 03:24:17,515 INFO [train.py:904] (4/8) Epoch 15, batch 3400, loss[loss=0.188, simple_loss=0.2657, pruned_loss=0.05513, over 16869.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.263, pruned_loss=0.04644, over 3316066.46 frames. ], batch size: 90, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:24:18,608 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.624e+02 2.153e+02 2.635e+02 3.209e+02 5.771e+02, threshold=5.270e+02, percent-clipped=2.0 2023-04-30 03:25:28,517 INFO [train.py:904] (4/8) Epoch 15, batch 3450, loss[loss=0.1414, simple_loss=0.2362, pruned_loss=0.02331, over 17226.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2612, pruned_loss=0.04568, over 3319527.54 frames. ], batch size: 45, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:25:35,552 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-30 03:26:27,267 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145594.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:26:38,102 INFO [train.py:904] (4/8) Epoch 15, batch 3500, loss[loss=0.1763, simple_loss=0.2728, pruned_loss=0.03988, over 17095.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2604, pruned_loss=0.04567, over 3312698.11 frames. ], batch size: 53, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:26:39,231 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.386e+02 2.268e+02 2.638e+02 3.199e+02 5.613e+02, threshold=5.276e+02, percent-clipped=1.0 2023-04-30 03:27:36,019 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145643.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:27:37,592 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-30 03:27:47,266 INFO [train.py:904] (4/8) Epoch 15, batch 3550, loss[loss=0.1774, simple_loss=0.2532, pruned_loss=0.05075, over 16807.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.259, pruned_loss=0.04497, over 3315059.73 frames. ], batch size: 102, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:27:48,309 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=145652.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 03:27:51,980 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-30 03:27:53,466 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145655.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 03:28:31,736 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=145683.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:28:54,822 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=145700.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:28:57,584 INFO [train.py:904] (4/8) Epoch 15, batch 3600, loss[loss=0.1575, simple_loss=0.2473, pruned_loss=0.0338, over 17173.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2585, pruned_loss=0.045, over 3301494.08 frames. ], batch size: 46, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:28:58,725 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 2.263e+02 2.507e+02 3.007e+02 5.256e+02, threshold=5.015e+02, percent-clipped=0.0 2023-04-30 03:29:00,361 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145704.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:29:26,931 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=145722.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:29:34,302 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145727.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:29:39,860 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=145731.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:30:10,693 INFO [train.py:904] (4/8) Epoch 15, batch 3650, loss[loss=0.171, simple_loss=0.2414, pruned_loss=0.05029, over 16851.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2575, pruned_loss=0.04514, over 3287110.91 frames. ], batch size: 83, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:30:32,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4481, 4.2682, 4.4416, 4.6134, 4.7479, 4.3251, 4.5472, 4.7125], device='cuda:4'), covar=tensor([0.1550, 0.1199, 0.1449, 0.0704, 0.0611, 0.0994, 0.2412, 0.0735], device='cuda:4'), in_proj_covar=tensor([0.0617, 0.0765, 0.0919, 0.0786, 0.0585, 0.0613, 0.0618, 0.0725], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:30:38,219 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=145770.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:31:00,809 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-04-30 03:31:13,903 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9960, 3.2164, 3.2533, 2.1228, 2.7484, 2.3583, 3.4968, 3.5462], device='cuda:4'), covar=tensor([0.0227, 0.0755, 0.0544, 0.1554, 0.0778, 0.0890, 0.0476, 0.0702], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0155, 0.0161, 0.0146, 0.0139, 0.0125, 0.0139, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 03:31:24,489 INFO [train.py:904] (4/8) Epoch 15, batch 3700, loss[loss=0.177, simple_loss=0.2466, pruned_loss=0.0537, over 16811.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2562, pruned_loss=0.04656, over 3272762.32 frames. ], batch size: 116, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:31:26,279 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.472e+02 2.174e+02 2.701e+02 3.170e+02 5.249e+02, threshold=5.403e+02, percent-clipped=2.0 2023-04-30 03:31:49,215 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1544, 5.0069, 5.1823, 5.3808, 5.5997, 4.8327, 5.5136, 5.5707], device='cuda:4'), covar=tensor([0.1794, 0.1138, 0.1718, 0.0750, 0.0446, 0.0798, 0.0567, 0.0555], device='cuda:4'), in_proj_covar=tensor([0.0611, 0.0757, 0.0909, 0.0777, 0.0578, 0.0606, 0.0611, 0.0717], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:31:53,379 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145821.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 03:32:33,351 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-04-30 03:32:38,909 INFO [train.py:904] (4/8) Epoch 15, batch 3750, loss[loss=0.1693, simple_loss=0.2406, pruned_loss=0.04896, over 16827.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.257, pruned_loss=0.04796, over 3274320.69 frames. ], batch size: 83, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:33:23,396 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145882.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 03:33:23,512 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5842, 2.3303, 2.3840, 4.5865, 2.2867, 2.6554, 2.4881, 2.5358], device='cuda:4'), covar=tensor([0.1028, 0.3354, 0.2386, 0.0354, 0.3753, 0.2396, 0.2990, 0.3065], device='cuda:4'), in_proj_covar=tensor([0.0383, 0.0422, 0.0349, 0.0328, 0.0426, 0.0485, 0.0385, 0.0495], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:33:51,795 INFO [train.py:904] (4/8) Epoch 15, batch 3800, loss[loss=0.1706, simple_loss=0.2534, pruned_loss=0.04389, over 16720.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2577, pruned_loss=0.04936, over 3274555.05 frames. ], batch size: 76, lr: 4.56e-03, grad_scale: 8.0 2023-04-30 03:33:53,679 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.566e+02 2.221e+02 2.643e+02 3.115e+02 5.182e+02, threshold=5.286e+02, percent-clipped=0.0 2023-04-30 03:35:01,670 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145950.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 03:35:04,374 INFO [train.py:904] (4/8) Epoch 15, batch 3850, loss[loss=0.1616, simple_loss=0.2413, pruned_loss=0.04096, over 16875.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2574, pruned_loss=0.04974, over 3282914.07 frames. ], batch size: 42, lr: 4.56e-03, grad_scale: 16.0 2023-04-30 03:35:58,524 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8482, 3.7775, 3.9282, 3.7144, 3.7878, 4.2726, 3.9482, 3.6107], device='cuda:4'), covar=tensor([0.2019, 0.2016, 0.2027, 0.2424, 0.2878, 0.1656, 0.1400, 0.2590], device='cuda:4'), in_proj_covar=tensor([0.0385, 0.0548, 0.0600, 0.0469, 0.0624, 0.0623, 0.0478, 0.0619], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 03:36:13,419 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145999.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:36:21,005 INFO [train.py:904] (4/8) Epoch 15, batch 3900, loss[loss=0.1547, simple_loss=0.2362, pruned_loss=0.03661, over 16719.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2575, pruned_loss=0.05018, over 3286223.51 frames. ], batch size: 83, lr: 4.56e-03, grad_scale: 16.0 2023-04-30 03:36:22,203 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.721e+02 2.255e+02 2.647e+02 3.185e+02 6.041e+02, threshold=5.295e+02, percent-clipped=2.0 2023-04-30 03:36:57,896 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146027.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:37:32,699 INFO [train.py:904] (4/8) Epoch 15, batch 3950, loss[loss=0.1725, simple_loss=0.2395, pruned_loss=0.05276, over 16677.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2569, pruned_loss=0.05084, over 3278776.28 frames. ], batch size: 134, lr: 4.55e-03, grad_scale: 16.0 2023-04-30 03:38:06,891 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=146075.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:38:18,066 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8474, 4.7184, 4.9230, 5.0737, 5.2594, 4.6445, 5.1491, 5.2533], device='cuda:4'), covar=tensor([0.1706, 0.1049, 0.1439, 0.0655, 0.0443, 0.0897, 0.0553, 0.0449], device='cuda:4'), in_proj_covar=tensor([0.0606, 0.0747, 0.0893, 0.0766, 0.0573, 0.0601, 0.0606, 0.0708], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:38:46,136 INFO [train.py:904] (4/8) Epoch 15, batch 4000, loss[loss=0.1574, simple_loss=0.2419, pruned_loss=0.03643, over 17023.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2574, pruned_loss=0.05139, over 3285777.12 frames. ], batch size: 50, lr: 4.55e-03, grad_scale: 16.0 2023-04-30 03:38:47,409 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.562e+02 2.290e+02 2.701e+02 3.084e+02 7.730e+02, threshold=5.403e+02, percent-clipped=2.0 2023-04-30 03:38:54,551 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-04-30 03:39:56,268 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=146149.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:39:59,983 INFO [train.py:904] (4/8) Epoch 15, batch 4050, loss[loss=0.1823, simple_loss=0.2672, pruned_loss=0.04873, over 17134.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2584, pruned_loss=0.05106, over 3270415.39 frames. ], batch size: 47, lr: 4.55e-03, grad_scale: 16.0 2023-04-30 03:40:34,543 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3729, 3.3183, 3.4265, 3.4844, 3.5750, 3.2797, 3.4888, 3.6253], device='cuda:4'), covar=tensor([0.1249, 0.0874, 0.1006, 0.0563, 0.0547, 0.2522, 0.1035, 0.0630], device='cuda:4'), in_proj_covar=tensor([0.0604, 0.0743, 0.0890, 0.0762, 0.0569, 0.0597, 0.0601, 0.0704], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:40:36,969 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=146177.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 03:41:13,945 INFO [train.py:904] (4/8) Epoch 15, batch 4100, loss[loss=0.1995, simple_loss=0.2872, pruned_loss=0.05595, over 15372.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2601, pruned_loss=0.05053, over 3260322.35 frames. ], batch size: 191, lr: 4.55e-03, grad_scale: 16.0 2023-04-30 03:41:15,744 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.382e+02 1.975e+02 2.401e+02 2.875e+02 5.931e+02, threshold=4.803e+02, percent-clipped=1.0 2023-04-30 03:41:26,656 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=146210.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:41:49,298 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-30 03:42:30,099 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146250.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:42:32,939 INFO [train.py:904] (4/8) Epoch 15, batch 4150, loss[loss=0.1973, simple_loss=0.2871, pruned_loss=0.05375, over 16751.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2678, pruned_loss=0.05342, over 3226773.58 frames. ], batch size: 89, lr: 4.55e-03, grad_scale: 8.0 2023-04-30 03:43:02,996 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-30 03:43:45,579 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=146298.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:43:46,699 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146299.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:43:50,401 INFO [train.py:904] (4/8) Epoch 15, batch 4200, loss[loss=0.2255, simple_loss=0.3168, pruned_loss=0.06711, over 16816.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2751, pruned_loss=0.05479, over 3217236.66 frames. ], batch size: 102, lr: 4.55e-03, grad_scale: 8.0 2023-04-30 03:43:53,464 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.634e+02 2.342e+02 2.800e+02 3.448e+02 4.997e+02, threshold=5.600e+02, percent-clipped=3.0 2023-04-30 03:44:26,276 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0136, 4.1136, 4.3954, 4.4011, 4.4000, 4.1510, 4.0696, 4.0799], device='cuda:4'), covar=tensor([0.0280, 0.0427, 0.0372, 0.0352, 0.0365, 0.0331, 0.1033, 0.0426], device='cuda:4'), in_proj_covar=tensor([0.0367, 0.0396, 0.0390, 0.0369, 0.0437, 0.0409, 0.0505, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 03:44:58,660 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=146347.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:45:06,384 INFO [train.py:904] (4/8) Epoch 15, batch 4250, loss[loss=0.1732, simple_loss=0.2699, pruned_loss=0.03818, over 16437.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2779, pruned_loss=0.05407, over 3218805.30 frames. ], batch size: 146, lr: 4.55e-03, grad_scale: 4.0 2023-04-30 03:45:18,089 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9049, 3.8376, 3.9714, 3.7029, 3.8195, 4.3001, 3.9314, 3.6074], device='cuda:4'), covar=tensor([0.1956, 0.2057, 0.1839, 0.2611, 0.2859, 0.1579, 0.1383, 0.2921], device='cuda:4'), in_proj_covar=tensor([0.0382, 0.0542, 0.0590, 0.0461, 0.0612, 0.0613, 0.0469, 0.0614], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 03:46:19,356 INFO [train.py:904] (4/8) Epoch 15, batch 4300, loss[loss=0.1929, simple_loss=0.2899, pruned_loss=0.04802, over 17217.00 frames. ], tot_loss[loss=0.1928, simple_loss=0.2789, pruned_loss=0.05336, over 3201323.23 frames. ], batch size: 44, lr: 4.55e-03, grad_scale: 4.0 2023-04-30 03:46:23,353 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.530e+02 2.412e+02 2.971e+02 3.359e+02 7.082e+02, threshold=5.941e+02, percent-clipped=4.0 2023-04-30 03:47:07,049 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1307, 2.0617, 2.1678, 3.7018, 2.0328, 2.4040, 2.2189, 2.2433], device='cuda:4'), covar=tensor([0.1167, 0.3367, 0.2612, 0.0490, 0.3817, 0.2280, 0.3123, 0.3226], device='cuda:4'), in_proj_covar=tensor([0.0381, 0.0421, 0.0348, 0.0323, 0.0423, 0.0484, 0.0384, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:47:31,145 INFO [train.py:904] (4/8) Epoch 15, batch 4350, loss[loss=0.198, simple_loss=0.2805, pruned_loss=0.05779, over 17051.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2824, pruned_loss=0.05458, over 3199116.64 frames. ], batch size: 53, lr: 4.55e-03, grad_scale: 4.0 2023-04-30 03:48:08,836 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146477.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 03:48:45,757 INFO [train.py:904] (4/8) Epoch 15, batch 4400, loss[loss=0.2034, simple_loss=0.2898, pruned_loss=0.05855, over 16469.00 frames. ], tot_loss[loss=0.1981, simple_loss=0.2845, pruned_loss=0.05582, over 3194416.82 frames. ], batch size: 146, lr: 4.55e-03, grad_scale: 8.0 2023-04-30 03:48:50,393 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.716e+02 2.574e+02 2.972e+02 3.574e+02 6.742e+02, threshold=5.944e+02, percent-clipped=2.0 2023-04-30 03:48:51,427 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=146505.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:49:21,076 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=146525.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 03:49:58,634 INFO [train.py:904] (4/8) Epoch 15, batch 4450, loss[loss=0.2028, simple_loss=0.2979, pruned_loss=0.05389, over 16937.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2879, pruned_loss=0.05687, over 3198422.08 frames. ], batch size: 109, lr: 4.55e-03, grad_scale: 8.0 2023-04-30 03:50:49,010 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6337, 4.9255, 4.7280, 4.7621, 4.4778, 4.3845, 4.4468, 5.0257], device='cuda:4'), covar=tensor([0.1114, 0.0830, 0.0891, 0.0733, 0.0729, 0.1104, 0.0984, 0.0775], device='cuda:4'), in_proj_covar=tensor([0.0597, 0.0744, 0.0614, 0.0537, 0.0469, 0.0478, 0.0619, 0.0565], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:50:57,061 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=146591.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:51:06,062 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0824, 2.1018, 2.1341, 3.7486, 2.0027, 2.4862, 2.2132, 2.2311], device='cuda:4'), covar=tensor([0.1163, 0.3247, 0.2625, 0.0496, 0.3896, 0.2181, 0.2854, 0.3207], device='cuda:4'), in_proj_covar=tensor([0.0380, 0.0419, 0.0346, 0.0323, 0.0424, 0.0483, 0.0383, 0.0491], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:51:12,434 INFO [train.py:904] (4/8) Epoch 15, batch 4500, loss[loss=0.1875, simple_loss=0.2832, pruned_loss=0.04587, over 16789.00 frames. ], tot_loss[loss=0.2013, simple_loss=0.2883, pruned_loss=0.05718, over 3213411.09 frames. ], batch size: 76, lr: 4.55e-03, grad_scale: 8.0 2023-04-30 03:51:16,080 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.360e+02 1.915e+02 2.182e+02 2.470e+02 4.715e+02, threshold=4.363e+02, percent-clipped=0.0 2023-04-30 03:51:32,805 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-30 03:51:48,029 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=146626.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:52:25,474 INFO [train.py:904] (4/8) Epoch 15, batch 4550, loss[loss=0.203, simple_loss=0.2912, pruned_loss=0.05739, over 16585.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2886, pruned_loss=0.05786, over 3220229.47 frames. ], batch size: 68, lr: 4.55e-03, grad_scale: 8.0 2023-04-30 03:52:25,945 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=146652.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:53:06,026 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-30 03:53:16,736 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=146687.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 03:53:23,431 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1176, 4.9403, 5.1359, 5.3168, 5.4779, 4.8272, 5.4597, 5.5081], device='cuda:4'), covar=tensor([0.1456, 0.0965, 0.1313, 0.0504, 0.0382, 0.0623, 0.0405, 0.0374], device='cuda:4'), in_proj_covar=tensor([0.0561, 0.0696, 0.0837, 0.0714, 0.0532, 0.0559, 0.0565, 0.0658], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:53:37,390 INFO [train.py:904] (4/8) Epoch 15, batch 4600, loss[loss=0.1974, simple_loss=0.278, pruned_loss=0.05841, over 16428.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2893, pruned_loss=0.05812, over 3232257.79 frames. ], batch size: 68, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 03:53:41,728 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.606e+02 1.940e+02 2.192e+02 2.697e+02 4.440e+02, threshold=4.384e+02, percent-clipped=1.0 2023-04-30 03:54:15,060 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3938, 2.3897, 2.3268, 4.2118, 2.2787, 2.7664, 2.3634, 2.4851], device='cuda:4'), covar=tensor([0.1079, 0.2897, 0.2423, 0.0409, 0.3669, 0.2027, 0.2943, 0.2975], device='cuda:4'), in_proj_covar=tensor([0.0378, 0.0417, 0.0345, 0.0321, 0.0423, 0.0481, 0.0382, 0.0488], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:54:49,172 INFO [train.py:904] (4/8) Epoch 15, batch 4650, loss[loss=0.1835, simple_loss=0.2695, pruned_loss=0.04879, over 16514.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2887, pruned_loss=0.05848, over 3235570.82 frames. ], batch size: 68, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 03:55:22,626 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5585, 3.5040, 3.6933, 1.9904, 3.0044, 2.2994, 3.8513, 3.5604], device='cuda:4'), covar=tensor([0.0198, 0.0779, 0.0571, 0.2133, 0.0831, 0.1031, 0.0585, 0.1150], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0154, 0.0161, 0.0146, 0.0138, 0.0126, 0.0139, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 03:55:52,767 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2090, 2.1474, 2.6595, 3.1727, 2.9686, 3.5905, 2.0599, 3.5689], device='cuda:4'), covar=tensor([0.0160, 0.0426, 0.0272, 0.0224, 0.0237, 0.0126, 0.0470, 0.0088], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0184, 0.0168, 0.0175, 0.0184, 0.0139, 0.0185, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 03:56:03,161 INFO [train.py:904] (4/8) Epoch 15, batch 4700, loss[loss=0.1936, simple_loss=0.2732, pruned_loss=0.05705, over 16804.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2857, pruned_loss=0.05704, over 3234791.50 frames. ], batch size: 39, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 03:56:07,885 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.363e+02 1.906e+02 2.324e+02 2.777e+02 7.777e+02, threshold=4.648e+02, percent-clipped=3.0 2023-04-30 03:56:08,882 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146805.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:56:40,219 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-30 03:57:07,543 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=146844.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:57:18,584 INFO [train.py:904] (4/8) Epoch 15, batch 4750, loss[loss=0.178, simple_loss=0.259, pruned_loss=0.04849, over 16998.00 frames. ], tot_loss[loss=0.1957, simple_loss=0.2814, pruned_loss=0.055, over 3230796.74 frames. ], batch size: 50, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 03:57:20,527 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=146853.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:58:31,082 INFO [train.py:904] (4/8) Epoch 15, batch 4800, loss[loss=0.2178, simple_loss=0.2905, pruned_loss=0.07252, over 11694.00 frames. ], tot_loss[loss=0.1917, simple_loss=0.2777, pruned_loss=0.05288, over 3232313.23 frames. ], batch size: 248, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 03:58:36,177 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.856e+02 2.167e+02 2.629e+02 4.962e+02, threshold=4.334e+02, percent-clipped=1.0 2023-04-30 03:58:36,822 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=146905.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:58:44,632 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6748, 3.8005, 2.8989, 2.2491, 2.6222, 2.4223, 3.9467, 3.4071], device='cuda:4'), covar=tensor([0.2797, 0.0713, 0.1785, 0.2640, 0.2514, 0.1865, 0.0528, 0.1170], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0264, 0.0292, 0.0293, 0.0287, 0.0235, 0.0277, 0.0314], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 03:59:40,718 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=146947.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 03:59:47,622 INFO [train.py:904] (4/8) Epoch 15, batch 4850, loss[loss=0.186, simple_loss=0.288, pruned_loss=0.042, over 16900.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.2783, pruned_loss=0.05202, over 3214222.02 frames. ], batch size: 96, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 04:00:10,701 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1131, 3.6503, 3.5421, 2.1308, 2.9745, 2.5313, 3.4273, 3.7106], device='cuda:4'), covar=tensor([0.0323, 0.0637, 0.0564, 0.1752, 0.0820, 0.0821, 0.0755, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0153, 0.0160, 0.0146, 0.0138, 0.0125, 0.0138, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 04:00:34,193 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=146982.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 04:01:03,676 INFO [train.py:904] (4/8) Epoch 15, batch 4900, loss[loss=0.1901, simple_loss=0.2772, pruned_loss=0.05148, over 17103.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.2779, pruned_loss=0.05112, over 3199041.65 frames. ], batch size: 49, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 04:01:08,003 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.422e+02 1.989e+02 2.229e+02 2.704e+02 6.823e+02, threshold=4.458e+02, percent-clipped=4.0 2023-04-30 04:01:43,273 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.85 vs. limit=5.0 2023-04-30 04:02:16,312 INFO [train.py:904] (4/8) Epoch 15, batch 4950, loss[loss=0.2099, simple_loss=0.298, pruned_loss=0.06086, over 12140.00 frames. ], tot_loss[loss=0.1897, simple_loss=0.278, pruned_loss=0.0507, over 3198676.19 frames. ], batch size: 246, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 04:03:28,741 INFO [train.py:904] (4/8) Epoch 15, batch 5000, loss[loss=0.2137, simple_loss=0.3004, pruned_loss=0.0635, over 12031.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2802, pruned_loss=0.05125, over 3189200.10 frames. ], batch size: 247, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 04:03:32,271 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.607e+02 2.177e+02 2.680e+02 3.002e+02 5.827e+02, threshold=5.360e+02, percent-clipped=3.0 2023-04-30 04:03:41,389 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-30 04:04:39,304 INFO [train.py:904] (4/8) Epoch 15, batch 5050, loss[loss=0.1932, simple_loss=0.2773, pruned_loss=0.05456, over 11951.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2802, pruned_loss=0.05092, over 3188490.06 frames. ], batch size: 248, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 04:05:15,058 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.72 vs. limit=5.0 2023-04-30 04:05:46,872 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147200.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:05:48,951 INFO [train.py:904] (4/8) Epoch 15, batch 5100, loss[loss=0.1767, simple_loss=0.2668, pruned_loss=0.04333, over 16714.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2783, pruned_loss=0.04998, over 3204633.80 frames. ], batch size: 89, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 04:05:52,961 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.025e+02 2.374e+02 2.786e+02 3.985e+02, threshold=4.748e+02, percent-clipped=0.0 2023-04-30 04:06:16,086 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-30 04:06:45,534 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147241.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:06:53,114 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147247.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:06:59,912 INFO [train.py:904] (4/8) Epoch 15, batch 5150, loss[loss=0.1878, simple_loss=0.2851, pruned_loss=0.04524, over 16755.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2782, pruned_loss=0.04924, over 3200090.53 frames. ], batch size: 89, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 04:07:09,786 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9018, 3.9362, 2.3264, 4.6055, 2.9356, 4.4304, 2.4324, 2.9395], device='cuda:4'), covar=tensor([0.0240, 0.0338, 0.1660, 0.0147, 0.0856, 0.0445, 0.1567, 0.0837], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0169, 0.0189, 0.0141, 0.0167, 0.0208, 0.0196, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 04:07:45,081 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147282.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:08:03,057 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=147295.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:08:14,126 INFO [train.py:904] (4/8) Epoch 15, batch 5200, loss[loss=0.2088, simple_loss=0.3027, pruned_loss=0.05742, over 15434.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.277, pruned_loss=0.04892, over 3177376.28 frames. ], batch size: 190, lr: 4.54e-03, grad_scale: 8.0 2023-04-30 04:08:14,632 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147302.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:08:17,787 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.539e+02 2.189e+02 2.515e+02 3.060e+02 5.067e+02, threshold=5.031e+02, percent-clipped=2.0 2023-04-30 04:08:25,228 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147310.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:08:55,677 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=147330.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:09:27,024 INFO [train.py:904] (4/8) Epoch 15, batch 5250, loss[loss=0.174, simple_loss=0.2682, pruned_loss=0.03992, over 16897.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2744, pruned_loss=0.04841, over 3184698.53 frames. ], batch size: 96, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:09:54,968 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147371.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:10:02,489 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0035, 1.8914, 2.4938, 2.9740, 2.8356, 3.4402, 2.0861, 3.3484], device='cuda:4'), covar=tensor([0.0177, 0.0466, 0.0312, 0.0268, 0.0271, 0.0132, 0.0471, 0.0097], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0182, 0.0168, 0.0173, 0.0181, 0.0138, 0.0185, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:10:37,510 INFO [train.py:904] (4/8) Epoch 15, batch 5300, loss[loss=0.1751, simple_loss=0.2603, pruned_loss=0.04494, over 12436.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2702, pruned_loss=0.0468, over 3206586.88 frames. ], batch size: 246, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:10:40,970 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.533e+02 1.977e+02 2.284e+02 2.754e+02 4.909e+02, threshold=4.569e+02, percent-clipped=0.0 2023-04-30 04:10:45,214 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147407.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:11:22,219 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.24 vs. limit=5.0 2023-04-30 04:11:23,856 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147434.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:11:49,965 INFO [train.py:904] (4/8) Epoch 15, batch 5350, loss[loss=0.1851, simple_loss=0.2782, pruned_loss=0.046, over 16689.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2686, pruned_loss=0.04631, over 3213524.39 frames. ], batch size: 89, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:12:00,648 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.76 vs. limit=5.0 2023-04-30 04:12:14,739 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147468.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:12:53,017 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147495.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:13:01,205 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147500.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:13:03,281 INFO [train.py:904] (4/8) Epoch 15, batch 5400, loss[loss=0.2004, simple_loss=0.2939, pruned_loss=0.05349, over 16231.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2712, pruned_loss=0.04686, over 3228650.80 frames. ], batch size: 165, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:13:07,675 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.412e+02 2.104e+02 2.571e+02 3.291e+02 5.861e+02, threshold=5.143e+02, percent-clipped=4.0 2023-04-30 04:14:13,937 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=147548.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:14:20,734 INFO [train.py:904] (4/8) Epoch 15, batch 5450, loss[loss=0.2255, simple_loss=0.2974, pruned_loss=0.07675, over 11978.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2738, pruned_loss=0.04808, over 3226091.63 frames. ], batch size: 248, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:14:42,627 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-04-30 04:15:32,508 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147597.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:15:40,036 INFO [train.py:904] (4/8) Epoch 15, batch 5500, loss[loss=0.2168, simple_loss=0.3078, pruned_loss=0.06285, over 17106.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2815, pruned_loss=0.05297, over 3190613.02 frames. ], batch size: 47, lr: 4.53e-03, grad_scale: 4.0 2023-04-30 04:15:45,690 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.855e+02 2.456e+02 3.016e+02 4.198e+02 6.055e+02, threshold=6.032e+02, percent-clipped=7.0 2023-04-30 04:16:58,368 INFO [train.py:904] (4/8) Epoch 15, batch 5550, loss[loss=0.3032, simple_loss=0.3535, pruned_loss=0.1265, over 10863.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2893, pruned_loss=0.05866, over 3156786.88 frames. ], batch size: 248, lr: 4.53e-03, grad_scale: 4.0 2023-04-30 04:17:10,688 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147659.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:17:22,924 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147666.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:18:21,668 INFO [train.py:904] (4/8) Epoch 15, batch 5600, loss[loss=0.2927, simple_loss=0.3463, pruned_loss=0.1195, over 11247.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.2952, pruned_loss=0.06379, over 3126186.14 frames. ], batch size: 248, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:18:28,273 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.312e+02 3.387e+02 4.165e+02 5.212e+02 1.017e+03, threshold=8.330e+02, percent-clipped=15.0 2023-04-30 04:18:40,921 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5419, 3.2145, 2.9153, 1.9045, 2.5733, 2.2876, 3.0049, 3.3793], device='cuda:4'), covar=tensor([0.0434, 0.0708, 0.0734, 0.1955, 0.1016, 0.0948, 0.0909, 0.0865], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0153, 0.0161, 0.0146, 0.0137, 0.0126, 0.0139, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 04:18:52,992 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147720.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:19:03,325 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9935, 5.2815, 5.0065, 4.9965, 4.7488, 4.6154, 4.6762, 5.3839], device='cuda:4'), covar=tensor([0.1035, 0.0770, 0.1052, 0.0868, 0.0724, 0.0901, 0.1126, 0.0733], device='cuda:4'), in_proj_covar=tensor([0.0588, 0.0730, 0.0607, 0.0524, 0.0460, 0.0470, 0.0604, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:19:46,188 INFO [train.py:904] (4/8) Epoch 15, batch 5650, loss[loss=0.2491, simple_loss=0.3234, pruned_loss=0.08746, over 15356.00 frames. ], tot_loss[loss=0.22, simple_loss=0.3017, pruned_loss=0.06921, over 3079838.43 frames. ], batch size: 190, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:19:47,202 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5256, 4.8060, 4.5728, 4.5733, 4.3589, 4.2479, 4.2445, 4.8875], device='cuda:4'), covar=tensor([0.1130, 0.0869, 0.1029, 0.0870, 0.0759, 0.1229, 0.1118, 0.0814], device='cuda:4'), in_proj_covar=tensor([0.0587, 0.0730, 0.0606, 0.0523, 0.0459, 0.0469, 0.0603, 0.0553], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:20:04,644 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147763.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:20:12,276 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-30 04:20:47,700 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147790.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:21:05,458 INFO [train.py:904] (4/8) Epoch 15, batch 5700, loss[loss=0.2617, simple_loss=0.3274, pruned_loss=0.09799, over 11288.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.3033, pruned_loss=0.07091, over 3063612.82 frames. ], batch size: 249, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:21:11,569 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.489e+02 3.643e+02 4.297e+02 4.973e+02 1.168e+03, threshold=8.593e+02, percent-clipped=2.0 2023-04-30 04:21:26,321 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147815.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:22:23,096 INFO [train.py:904] (4/8) Epoch 15, batch 5750, loss[loss=0.2274, simple_loss=0.3156, pruned_loss=0.06961, over 17005.00 frames. ], tot_loss[loss=0.2247, simple_loss=0.3053, pruned_loss=0.07207, over 3042664.32 frames. ], batch size: 53, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:22:30,992 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-30 04:22:32,713 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147857.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:23:03,179 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147876.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:23:36,714 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147897.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:23:44,668 INFO [train.py:904] (4/8) Epoch 15, batch 5800, loss[loss=0.2185, simple_loss=0.3091, pruned_loss=0.06397, over 16862.00 frames. ], tot_loss[loss=0.2227, simple_loss=0.305, pruned_loss=0.0702, over 3062326.45 frames. ], batch size: 83, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:23:51,419 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.104e+02 2.959e+02 3.306e+02 4.369e+02 1.266e+03, threshold=6.612e+02, percent-clipped=1.0 2023-04-30 04:24:00,139 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147911.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:24:12,246 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147918.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:24:55,226 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=147945.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:25:05,474 INFO [train.py:904] (4/8) Epoch 15, batch 5850, loss[loss=0.2045, simple_loss=0.2919, pruned_loss=0.05851, over 16649.00 frames. ], tot_loss[loss=0.2194, simple_loss=0.3024, pruned_loss=0.06823, over 3069966.03 frames. ], batch size: 134, lr: 4.53e-03, grad_scale: 8.0 2023-04-30 04:25:16,258 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8044, 2.1944, 1.8113, 2.1144, 2.6137, 2.3638, 2.6413, 2.8773], device='cuda:4'), covar=tensor([0.0141, 0.0371, 0.0461, 0.0383, 0.0229, 0.0292, 0.0204, 0.0200], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0215, 0.0208, 0.0209, 0.0214, 0.0214, 0.0218, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:25:29,184 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147966.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:25:38,363 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147972.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:26:30,250 INFO [train.py:904] (4/8) Epoch 15, batch 5900, loss[loss=0.2373, simple_loss=0.3019, pruned_loss=0.0863, over 11439.00 frames. ], tot_loss[loss=0.2193, simple_loss=0.3029, pruned_loss=0.06791, over 3080572.46 frames. ], batch size: 248, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:26:39,373 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.034e+02 2.736e+02 3.208e+02 4.005e+02 8.372e+02, threshold=6.416e+02, percent-clipped=2.0 2023-04-30 04:26:52,639 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=148014.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:26:54,010 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148015.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:27:50,061 INFO [train.py:904] (4/8) Epoch 15, batch 5950, loss[loss=0.2254, simple_loss=0.3149, pruned_loss=0.06797, over 16771.00 frames. ], tot_loss[loss=0.218, simple_loss=0.3035, pruned_loss=0.06629, over 3096652.56 frames. ], batch size: 83, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:28:08,799 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148063.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:28:48,994 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148090.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:29:08,854 INFO [train.py:904] (4/8) Epoch 15, batch 6000, loss[loss=0.1902, simple_loss=0.2771, pruned_loss=0.05164, over 16774.00 frames. ], tot_loss[loss=0.216, simple_loss=0.3015, pruned_loss=0.06524, over 3108339.70 frames. ], batch size: 124, lr: 4.52e-03, grad_scale: 8.0 2023-04-30 04:29:08,854 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 04:29:19,436 INFO [train.py:938] (4/8) Epoch 15, validation: loss=0.1559, simple_loss=0.2691, pruned_loss=0.0214, over 944034.00 frames. 2023-04-30 04:29:19,437 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 04:29:26,129 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.824e+02 2.753e+02 3.314e+02 4.296e+02 7.936e+02, threshold=6.628e+02, percent-clipped=2.0 2023-04-30 04:29:33,717 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=148111.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:29:38,357 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-30 04:30:15,356 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=148138.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:30:36,869 INFO [train.py:904] (4/8) Epoch 15, batch 6050, loss[loss=0.1945, simple_loss=0.2866, pruned_loss=0.05118, over 17033.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2996, pruned_loss=0.06404, over 3130696.41 frames. ], batch size: 55, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:31:08,627 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148171.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:31:59,322 INFO [train.py:904] (4/8) Epoch 15, batch 6100, loss[loss=0.2016, simple_loss=0.2861, pruned_loss=0.05855, over 17215.00 frames. ], tot_loss[loss=0.213, simple_loss=0.2993, pruned_loss=0.06335, over 3130881.31 frames. ], batch size: 45, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:32:08,602 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.685e+02 2.651e+02 3.170e+02 3.946e+02 8.387e+02, threshold=6.339e+02, percent-clipped=2.0 2023-04-30 04:32:18,650 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148213.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:33:18,245 INFO [train.py:904] (4/8) Epoch 15, batch 6150, loss[loss=0.2398, simple_loss=0.3061, pruned_loss=0.0867, over 11482.00 frames. ], tot_loss[loss=0.2112, simple_loss=0.2969, pruned_loss=0.06278, over 3115156.32 frames. ], batch size: 246, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:33:21,674 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.88 vs. limit=5.0 2023-04-30 04:33:43,189 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148267.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:34:15,485 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148287.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:34:37,597 INFO [train.py:904] (4/8) Epoch 15, batch 6200, loss[loss=0.2506, simple_loss=0.3082, pruned_loss=0.09647, over 11481.00 frames. ], tot_loss[loss=0.2107, simple_loss=0.2956, pruned_loss=0.06284, over 3103104.87 frames. ], batch size: 248, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:34:46,165 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.083e+02 3.031e+02 3.685e+02 4.615e+02 1.155e+03, threshold=7.370e+02, percent-clipped=8.0 2023-04-30 04:34:58,565 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148315.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:35:48,293 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148348.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:35:54,858 INFO [train.py:904] (4/8) Epoch 15, batch 6250, loss[loss=0.2113, simple_loss=0.2961, pruned_loss=0.06329, over 16892.00 frames. ], tot_loss[loss=0.2108, simple_loss=0.2955, pruned_loss=0.06301, over 3098210.69 frames. ], batch size: 116, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:36:11,758 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=148363.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:36:23,426 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-30 04:37:05,048 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7762, 4.1287, 3.0612, 2.3341, 2.8162, 2.5476, 4.4314, 3.7077], device='cuda:4'), covar=tensor([0.2526, 0.0656, 0.1566, 0.2442, 0.2290, 0.1709, 0.0402, 0.1020], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0262, 0.0290, 0.0291, 0.0283, 0.0233, 0.0275, 0.0311], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:37:11,937 INFO [train.py:904] (4/8) Epoch 15, batch 6300, loss[loss=0.2035, simple_loss=0.2911, pruned_loss=0.058, over 16751.00 frames. ], tot_loss[loss=0.2107, simple_loss=0.2957, pruned_loss=0.06279, over 3109656.89 frames. ], batch size: 89, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:37:21,867 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.698e+02 2.809e+02 3.245e+02 4.323e+02 1.479e+03, threshold=6.491e+02, percent-clipped=2.0 2023-04-30 04:37:57,323 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148429.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:38:28,239 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7888, 1.2892, 1.7013, 1.6658, 1.7501, 1.8482, 1.5527, 1.7512], device='cuda:4'), covar=tensor([0.0187, 0.0313, 0.0163, 0.0187, 0.0209, 0.0144, 0.0309, 0.0096], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0179, 0.0164, 0.0168, 0.0177, 0.0135, 0.0182, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:38:33,530 INFO [train.py:904] (4/8) Epoch 15, batch 6350, loss[loss=0.1918, simple_loss=0.2822, pruned_loss=0.0507, over 16446.00 frames. ], tot_loss[loss=0.2113, simple_loss=0.2956, pruned_loss=0.06349, over 3103751.12 frames. ], batch size: 146, lr: 4.52e-03, grad_scale: 4.0 2023-04-30 04:38:47,272 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8287, 1.3606, 1.7251, 1.6964, 1.7950, 1.9087, 1.5894, 1.8123], device='cuda:4'), covar=tensor([0.0188, 0.0321, 0.0166, 0.0191, 0.0208, 0.0139, 0.0316, 0.0094], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0180, 0.0164, 0.0168, 0.0177, 0.0136, 0.0182, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:38:48,952 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.63 vs. limit=5.0 2023-04-30 04:38:56,112 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0283, 4.0174, 3.9465, 3.2990, 3.9759, 1.8573, 3.7652, 3.5477], device='cuda:4'), covar=tensor([0.0110, 0.0093, 0.0159, 0.0267, 0.0083, 0.2499, 0.0125, 0.0200], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0132, 0.0181, 0.0168, 0.0152, 0.0190, 0.0169, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:39:03,808 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148471.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:39:34,199 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148490.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:39:45,769 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148498.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:39:51,469 INFO [train.py:904] (4/8) Epoch 15, batch 6400, loss[loss=0.18, simple_loss=0.2621, pruned_loss=0.04894, over 16587.00 frames. ], tot_loss[loss=0.2131, simple_loss=0.2963, pruned_loss=0.06496, over 3088893.03 frames. ], batch size: 57, lr: 4.52e-03, grad_scale: 8.0 2023-04-30 04:39:59,983 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.110e+02 3.263e+02 3.782e+02 4.494e+02 8.205e+02, threshold=7.565e+02, percent-clipped=3.0 2023-04-30 04:40:07,862 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5174, 4.5759, 4.3847, 4.0968, 4.0461, 4.4971, 4.2736, 4.2034], device='cuda:4'), covar=tensor([0.0590, 0.0489, 0.0283, 0.0292, 0.0956, 0.0430, 0.0469, 0.0675], device='cuda:4'), in_proj_covar=tensor([0.0262, 0.0363, 0.0310, 0.0292, 0.0324, 0.0339, 0.0209, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:40:09,616 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148513.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:40:15,848 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1106, 5.7119, 5.8843, 5.6117, 5.6131, 6.2045, 5.7594, 5.5469], device='cuda:4'), covar=tensor([0.0798, 0.1663, 0.2168, 0.1858, 0.2454, 0.0929, 0.1377, 0.2225], device='cuda:4'), in_proj_covar=tensor([0.0382, 0.0540, 0.0587, 0.0458, 0.0607, 0.0616, 0.0467, 0.0612], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 04:40:18,145 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=148519.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:40:42,400 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1442, 3.1476, 3.3733, 1.7437, 3.4800, 3.5207, 2.7511, 2.6671], device='cuda:4'), covar=tensor([0.0805, 0.0213, 0.0159, 0.1184, 0.0079, 0.0181, 0.0414, 0.0463], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0105, 0.0091, 0.0138, 0.0073, 0.0115, 0.0124, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 04:41:07,997 INFO [train.py:904] (4/8) Epoch 15, batch 6450, loss[loss=0.181, simple_loss=0.2776, pruned_loss=0.04224, over 17249.00 frames. ], tot_loss[loss=0.2125, simple_loss=0.2963, pruned_loss=0.06434, over 3083285.06 frames. ], batch size: 52, lr: 4.52e-03, grad_scale: 8.0 2023-04-30 04:41:19,827 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148559.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:41:22,168 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=148561.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:41:32,433 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148567.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:41:43,341 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4966, 4.3250, 4.5421, 4.6980, 4.8348, 4.4220, 4.8016, 4.8149], device='cuda:4'), covar=tensor([0.1517, 0.1143, 0.1400, 0.0677, 0.0506, 0.0892, 0.0557, 0.0587], device='cuda:4'), in_proj_covar=tensor([0.0564, 0.0697, 0.0837, 0.0708, 0.0535, 0.0560, 0.0566, 0.0662], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:41:56,869 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.53 vs. limit=2.0 2023-04-30 04:42:27,001 INFO [train.py:904] (4/8) Epoch 15, batch 6500, loss[loss=0.2012, simple_loss=0.288, pruned_loss=0.05721, over 17063.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2945, pruned_loss=0.06322, over 3097453.08 frames. ], batch size: 50, lr: 4.52e-03, grad_scale: 8.0 2023-04-30 04:42:36,999 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.948e+02 2.869e+02 3.331e+02 4.011e+02 8.248e+02, threshold=6.661e+02, percent-clipped=1.0 2023-04-30 04:42:47,764 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=148615.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:43:27,632 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-30 04:43:33,198 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148643.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:43:47,248 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148651.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:43:47,987 INFO [train.py:904] (4/8) Epoch 15, batch 6550, loss[loss=0.2123, simple_loss=0.3106, pruned_loss=0.05699, over 16980.00 frames. ], tot_loss[loss=0.2135, simple_loss=0.2973, pruned_loss=0.06481, over 3086752.91 frames. ], batch size: 41, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:44:09,151 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9339, 5.3149, 5.5461, 5.2407, 5.3106, 5.8834, 5.3774, 5.1129], device='cuda:4'), covar=tensor([0.0916, 0.1703, 0.2290, 0.1748, 0.2313, 0.0837, 0.1390, 0.2310], device='cuda:4'), in_proj_covar=tensor([0.0384, 0.0541, 0.0589, 0.0459, 0.0611, 0.0619, 0.0467, 0.0613], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 04:44:16,998 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9163, 4.1923, 3.9654, 4.0524, 3.7474, 3.7993, 3.8600, 4.1725], device='cuda:4'), covar=tensor([0.1069, 0.0853, 0.1022, 0.0734, 0.0819, 0.1463, 0.0864, 0.0899], device='cuda:4'), in_proj_covar=tensor([0.0601, 0.0746, 0.0619, 0.0536, 0.0470, 0.0484, 0.0619, 0.0564], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:44:58,460 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2548, 2.1956, 2.2030, 4.0591, 2.1342, 2.6375, 2.3144, 2.3780], device='cuda:4'), covar=tensor([0.1162, 0.3236, 0.2568, 0.0424, 0.3738, 0.2164, 0.3124, 0.3040], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0414, 0.0344, 0.0317, 0.0422, 0.0476, 0.0381, 0.0484], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:45:05,100 INFO [train.py:904] (4/8) Epoch 15, batch 6600, loss[loss=0.2281, simple_loss=0.3073, pruned_loss=0.07439, over 16305.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2989, pruned_loss=0.06447, over 3094907.64 frames. ], batch size: 165, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:45:13,945 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.578e+02 3.041e+02 3.810e+02 4.934e+02 1.256e+03, threshold=7.620e+02, percent-clipped=11.0 2023-04-30 04:45:19,935 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148712.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:45:26,308 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148716.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:46:12,160 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-30 04:46:22,112 INFO [train.py:904] (4/8) Epoch 15, batch 6650, loss[loss=0.2265, simple_loss=0.314, pruned_loss=0.06952, over 15402.00 frames. ], tot_loss[loss=0.214, simple_loss=0.2985, pruned_loss=0.06481, over 3089913.52 frames. ], batch size: 191, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:46:56,684 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9747, 4.9765, 4.8149, 4.2005, 4.8913, 2.0261, 4.5943, 4.6613], device='cuda:4'), covar=tensor([0.0095, 0.0067, 0.0149, 0.0353, 0.0077, 0.2352, 0.0127, 0.0182], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0131, 0.0179, 0.0167, 0.0151, 0.0190, 0.0168, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:47:00,925 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148777.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:47:02,085 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7631, 4.7777, 5.2492, 5.2034, 5.1991, 4.8168, 4.8191, 4.5494], device='cuda:4'), covar=tensor([0.0293, 0.0423, 0.0319, 0.0374, 0.0461, 0.0352, 0.0994, 0.0479], device='cuda:4'), in_proj_covar=tensor([0.0370, 0.0394, 0.0387, 0.0372, 0.0439, 0.0412, 0.0509, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 04:47:13,455 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148785.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:47:19,246 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5357, 1.5291, 2.1298, 2.4138, 2.3877, 2.7168, 1.7539, 2.6329], device='cuda:4'), covar=tensor([0.0180, 0.0510, 0.0264, 0.0258, 0.0266, 0.0179, 0.0457, 0.0143], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0180, 0.0164, 0.0168, 0.0178, 0.0136, 0.0182, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:47:39,115 INFO [train.py:904] (4/8) Epoch 15, batch 6700, loss[loss=0.2489, simple_loss=0.3119, pruned_loss=0.09297, over 11664.00 frames. ], tot_loss[loss=0.2132, simple_loss=0.2966, pruned_loss=0.06496, over 3087175.94 frames. ], batch size: 248, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:47:47,145 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.925e+02 2.803e+02 3.506e+02 4.354e+02 9.571e+02, threshold=7.012e+02, percent-clipped=2.0 2023-04-30 04:48:54,918 INFO [train.py:904] (4/8) Epoch 15, batch 6750, loss[loss=0.2004, simple_loss=0.284, pruned_loss=0.05838, over 16761.00 frames. ], tot_loss[loss=0.212, simple_loss=0.2952, pruned_loss=0.06439, over 3110366.83 frames. ], batch size: 124, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:48:58,475 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148854.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:49:00,345 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6941, 4.6403, 4.4571, 3.8100, 4.5978, 1.7983, 4.2944, 4.2298], device='cuda:4'), covar=tensor([0.0069, 0.0058, 0.0155, 0.0311, 0.0062, 0.2552, 0.0113, 0.0205], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0132, 0.0179, 0.0167, 0.0151, 0.0191, 0.0168, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:49:05,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5029, 3.4886, 3.4143, 2.5238, 3.3266, 1.9746, 3.1352, 2.6688], device='cuda:4'), covar=tensor([0.0186, 0.0142, 0.0214, 0.0411, 0.0120, 0.2908, 0.0188, 0.0332], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0132, 0.0179, 0.0167, 0.0151, 0.0191, 0.0168, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:49:45,903 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3746, 3.3438, 3.3979, 3.4930, 3.5174, 3.2525, 3.4880, 3.5533], device='cuda:4'), covar=tensor([0.1065, 0.0870, 0.0954, 0.0556, 0.0672, 0.2347, 0.0928, 0.0698], device='cuda:4'), in_proj_covar=tensor([0.0566, 0.0700, 0.0837, 0.0709, 0.0534, 0.0561, 0.0569, 0.0662], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 04:49:46,514 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.72 vs. limit=2.0 2023-04-30 04:50:09,781 INFO [train.py:904] (4/8) Epoch 15, batch 6800, loss[loss=0.2455, simple_loss=0.3049, pruned_loss=0.09308, over 11459.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.2953, pruned_loss=0.0646, over 3095191.89 frames. ], batch size: 247, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:50:21,236 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.768e+02 2.837e+02 3.542e+02 4.429e+02 7.153e+02, threshold=7.083e+02, percent-clipped=1.0 2023-04-30 04:51:15,796 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148943.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:51:28,455 INFO [train.py:904] (4/8) Epoch 15, batch 6850, loss[loss=0.1838, simple_loss=0.2888, pruned_loss=0.03935, over 17167.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.2965, pruned_loss=0.06507, over 3088185.37 frames. ], batch size: 46, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:51:29,850 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-30 04:52:26,495 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=148991.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:52:29,692 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7454, 4.7517, 5.1462, 5.1403, 5.1165, 4.8337, 4.7789, 4.6004], device='cuda:4'), covar=tensor([0.0341, 0.0618, 0.0477, 0.0469, 0.0489, 0.0457, 0.0951, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0371, 0.0396, 0.0387, 0.0373, 0.0441, 0.0413, 0.0512, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 04:52:29,975 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-30 04:52:42,625 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0404, 3.2802, 3.2943, 2.1431, 3.0794, 3.3364, 3.1937, 1.8109], device='cuda:4'), covar=tensor([0.0514, 0.0061, 0.0061, 0.0427, 0.0099, 0.0119, 0.0083, 0.0463], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0076, 0.0075, 0.0134, 0.0089, 0.0099, 0.0086, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 04:52:43,264 INFO [train.py:904] (4/8) Epoch 15, batch 6900, loss[loss=0.2325, simple_loss=0.3146, pruned_loss=0.07524, over 16198.00 frames. ], tot_loss[loss=0.2129, simple_loss=0.2982, pruned_loss=0.0638, over 3118527.79 frames. ], batch size: 165, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:52:52,140 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=149007.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:52:54,700 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.886e+02 2.687e+02 3.079e+02 3.976e+02 7.116e+02, threshold=6.157e+02, percent-clipped=1.0 2023-04-30 04:53:26,739 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-30 04:53:50,659 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-30 04:53:58,309 INFO [train.py:904] (4/8) Epoch 15, batch 6950, loss[loss=0.2699, simple_loss=0.3355, pruned_loss=0.1022, over 11565.00 frames. ], tot_loss[loss=0.2154, simple_loss=0.3002, pruned_loss=0.06524, over 3114505.16 frames. ], batch size: 248, lr: 4.51e-03, grad_scale: 8.0 2023-04-30 04:54:27,424 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=149072.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:54:45,875 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=149085.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:55:10,868 INFO [train.py:904] (4/8) Epoch 15, batch 7000, loss[loss=0.242, simple_loss=0.3194, pruned_loss=0.08232, over 11673.00 frames. ], tot_loss[loss=0.214, simple_loss=0.2997, pruned_loss=0.0642, over 3116817.71 frames. ], batch size: 248, lr: 4.51e-03, grad_scale: 2.0 2023-04-30 04:55:23,348 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.638e+02 2.893e+02 3.767e+02 4.998e+02 1.151e+03, threshold=7.533e+02, percent-clipped=10.0 2023-04-30 04:55:58,208 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=149133.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:56:01,587 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3704, 2.8726, 2.6639, 2.2574, 2.2678, 2.2786, 2.9215, 2.8378], device='cuda:4'), covar=tensor([0.2302, 0.0766, 0.1383, 0.2194, 0.2049, 0.1852, 0.0523, 0.1092], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0263, 0.0292, 0.0292, 0.0285, 0.0235, 0.0277, 0.0312], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 04:56:25,563 INFO [train.py:904] (4/8) Epoch 15, batch 7050, loss[loss=0.2066, simple_loss=0.294, pruned_loss=0.05966, over 16920.00 frames. ], tot_loss[loss=0.214, simple_loss=0.3004, pruned_loss=0.0638, over 3126055.16 frames. ], batch size: 109, lr: 4.51e-03, grad_scale: 2.0 2023-04-30 04:56:28,523 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=149154.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:57:40,343 INFO [train.py:904] (4/8) Epoch 15, batch 7100, loss[loss=0.2777, simple_loss=0.325, pruned_loss=0.1152, over 11189.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.2999, pruned_loss=0.06452, over 3105908.37 frames. ], batch size: 248, lr: 4.51e-03, grad_scale: 2.0 2023-04-30 04:57:40,682 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=149202.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 04:57:53,981 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.949e+02 2.990e+02 3.583e+02 4.304e+02 1.223e+03, threshold=7.166e+02, percent-clipped=1.0 2023-04-30 04:58:55,139 INFO [train.py:904] (4/8) Epoch 15, batch 7150, loss[loss=0.1992, simple_loss=0.2868, pruned_loss=0.05584, over 16855.00 frames. ], tot_loss[loss=0.2135, simple_loss=0.2983, pruned_loss=0.06435, over 3105663.50 frames. ], batch size: 83, lr: 4.51e-03, grad_scale: 2.0 2023-04-30 04:59:37,782 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.49 vs. limit=5.0 2023-04-30 04:59:40,840 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9173, 4.1997, 3.9976, 4.0580, 3.7056, 3.7975, 3.8891, 4.1526], device='cuda:4'), covar=tensor([0.1025, 0.0829, 0.0988, 0.0703, 0.0812, 0.1636, 0.0801, 0.0995], device='cuda:4'), in_proj_covar=tensor([0.0592, 0.0729, 0.0608, 0.0528, 0.0459, 0.0475, 0.0609, 0.0552], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:00:02,986 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9155, 4.8569, 4.6634, 4.0102, 4.7655, 1.9279, 4.4909, 4.5276], device='cuda:4'), covar=tensor([0.0071, 0.0067, 0.0180, 0.0393, 0.0086, 0.2566, 0.0125, 0.0190], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0131, 0.0177, 0.0165, 0.0150, 0.0189, 0.0166, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:00:05,782 INFO [train.py:904] (4/8) Epoch 15, batch 7200, loss[loss=0.1787, simple_loss=0.2714, pruned_loss=0.04304, over 16707.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2956, pruned_loss=0.06218, over 3104170.09 frames. ], batch size: 134, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:00:13,267 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=149307.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:00:17,918 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.349e+02 2.693e+02 3.090e+02 3.791e+02 7.265e+02, threshold=6.181e+02, percent-clipped=1.0 2023-04-30 05:01:26,135 INFO [train.py:904] (4/8) Epoch 15, batch 7250, loss[loss=0.2429, simple_loss=0.303, pruned_loss=0.0914, over 11587.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.2932, pruned_loss=0.06108, over 3086529.01 frames. ], batch size: 248, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:01:31,288 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=149355.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:01:57,267 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=149372.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:02:42,286 INFO [train.py:904] (4/8) Epoch 15, batch 7300, loss[loss=0.2607, simple_loss=0.3161, pruned_loss=0.1027, over 11386.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.2928, pruned_loss=0.06128, over 3080373.15 frames. ], batch size: 246, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:02:55,636 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.231e+02 2.888e+02 3.399e+02 4.220e+02 6.338e+02, threshold=6.799e+02, percent-clipped=2.0 2023-04-30 05:03:10,461 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=149420.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:03:49,334 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2797, 2.1579, 2.3383, 3.9656, 2.0845, 2.6039, 2.2528, 2.3696], device='cuda:4'), covar=tensor([0.1140, 0.3283, 0.2543, 0.0484, 0.4006, 0.2228, 0.3172, 0.3363], device='cuda:4'), in_proj_covar=tensor([0.0375, 0.0412, 0.0342, 0.0317, 0.0421, 0.0474, 0.0380, 0.0482], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:03:58,890 INFO [train.py:904] (4/8) Epoch 15, batch 7350, loss[loss=0.2105, simple_loss=0.2963, pruned_loss=0.06236, over 16415.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.294, pruned_loss=0.0623, over 3080307.33 frames. ], batch size: 146, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:05:17,915 INFO [train.py:904] (4/8) Epoch 15, batch 7400, loss[loss=0.2217, simple_loss=0.3089, pruned_loss=0.06721, over 16698.00 frames. ], tot_loss[loss=0.2107, simple_loss=0.2952, pruned_loss=0.06313, over 3101110.98 frames. ], batch size: 89, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:05:31,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0241, 2.3922, 2.5874, 1.9266, 2.6601, 2.7591, 2.4042, 2.4027], device='cuda:4'), covar=tensor([0.0671, 0.0202, 0.0214, 0.0870, 0.0103, 0.0225, 0.0392, 0.0415], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0104, 0.0090, 0.0137, 0.0073, 0.0113, 0.0122, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 05:05:32,175 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.760e+02 2.749e+02 3.166e+02 4.051e+02 9.173e+02, threshold=6.332e+02, percent-clipped=4.0 2023-04-30 05:05:37,366 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=149514.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:06:37,031 INFO [train.py:904] (4/8) Epoch 15, batch 7450, loss[loss=0.1879, simple_loss=0.2881, pruned_loss=0.04388, over 16859.00 frames. ], tot_loss[loss=0.212, simple_loss=0.2962, pruned_loss=0.06391, over 3100844.38 frames. ], batch size: 102, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:07:00,492 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=149565.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:07:16,484 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=149575.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:07:43,226 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.58 vs. limit=5.0 2023-04-30 05:08:00,166 INFO [train.py:904] (4/8) Epoch 15, batch 7500, loss[loss=0.1885, simple_loss=0.281, pruned_loss=0.04805, over 16698.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.2962, pruned_loss=0.06354, over 3101269.46 frames. ], batch size: 89, lr: 4.50e-03, grad_scale: 2.0 2023-04-30 05:08:16,083 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.765e+02 2.987e+02 3.634e+02 4.459e+02 1.121e+03, threshold=7.267e+02, percent-clipped=5.0 2023-04-30 05:08:20,096 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4692, 3.5098, 3.3109, 3.0421, 3.0907, 3.4080, 3.2995, 3.2101], device='cuda:4'), covar=tensor([0.0627, 0.0521, 0.0295, 0.0278, 0.0565, 0.0419, 0.1264, 0.0527], device='cuda:4'), in_proj_covar=tensor([0.0257, 0.0356, 0.0304, 0.0288, 0.0319, 0.0334, 0.0207, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:08:40,022 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=149626.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 05:09:18,804 INFO [train.py:904] (4/8) Epoch 15, batch 7550, loss[loss=0.1868, simple_loss=0.2829, pruned_loss=0.04539, over 16839.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.2954, pruned_loss=0.06375, over 3089026.58 frames. ], batch size: 96, lr: 4.50e-03, grad_scale: 2.0 2023-04-30 05:09:50,000 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1045, 2.0437, 2.1751, 3.6390, 2.0632, 2.4336, 2.2004, 2.2134], device='cuda:4'), covar=tensor([0.1190, 0.3492, 0.2490, 0.0506, 0.3937, 0.2299, 0.3060, 0.3173], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0416, 0.0345, 0.0319, 0.0425, 0.0477, 0.0384, 0.0487], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:10:35,822 INFO [train.py:904] (4/8) Epoch 15, batch 7600, loss[loss=0.2016, simple_loss=0.2856, pruned_loss=0.05874, over 16496.00 frames. ], tot_loss[loss=0.2109, simple_loss=0.2945, pruned_loss=0.06364, over 3093258.77 frames. ], batch size: 68, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:10:50,664 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.987e+02 2.884e+02 3.508e+02 4.101e+02 6.446e+02, threshold=7.017e+02, percent-clipped=0.0 2023-04-30 05:11:33,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7327, 4.7548, 4.5752, 4.2394, 4.2077, 4.6062, 4.5138, 4.3162], device='cuda:4'), covar=tensor([0.0587, 0.0416, 0.0302, 0.0326, 0.1040, 0.0460, 0.0416, 0.0701], device='cuda:4'), in_proj_covar=tensor([0.0258, 0.0357, 0.0305, 0.0289, 0.0321, 0.0336, 0.0209, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:11:45,888 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=149746.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 05:11:55,061 INFO [train.py:904] (4/8) Epoch 15, batch 7650, loss[loss=0.2209, simple_loss=0.3031, pruned_loss=0.06936, over 16898.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.2949, pruned_loss=0.06395, over 3102553.98 frames. ], batch size: 116, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:12:46,902 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-30 05:13:13,691 INFO [train.py:904] (4/8) Epoch 15, batch 7700, loss[loss=0.1859, simple_loss=0.2813, pruned_loss=0.04521, over 16798.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.295, pruned_loss=0.06414, over 3104761.72 frames. ], batch size: 102, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:13:22,700 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=149807.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 05:13:28,958 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-30 05:13:29,264 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.980e+02 3.235e+02 4.085e+02 5.006e+02 1.161e+03, threshold=8.169e+02, percent-clipped=3.0 2023-04-30 05:14:28,445 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.48 vs. limit=5.0 2023-04-30 05:14:33,190 INFO [train.py:904] (4/8) Epoch 15, batch 7750, loss[loss=0.1988, simple_loss=0.2893, pruned_loss=0.05414, over 16780.00 frames. ], tot_loss[loss=0.2126, simple_loss=0.2959, pruned_loss=0.06464, over 3094240.08 frames. ], batch size: 83, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:15:02,328 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=149870.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:15:05,237 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-30 05:15:26,751 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-30 05:15:51,917 INFO [train.py:904] (4/8) Epoch 15, batch 7800, loss[loss=0.199, simple_loss=0.2831, pruned_loss=0.05744, over 16596.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.2976, pruned_loss=0.06608, over 3081750.83 frames. ], batch size: 62, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:16:07,341 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.059e+02 3.237e+02 3.890e+02 4.639e+02 7.802e+02, threshold=7.781e+02, percent-clipped=0.0 2023-04-30 05:16:20,940 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=149921.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 05:17:06,634 INFO [train.py:904] (4/8) Epoch 15, batch 7850, loss[loss=0.195, simple_loss=0.2788, pruned_loss=0.05561, over 16611.00 frames. ], tot_loss[loss=0.2153, simple_loss=0.2983, pruned_loss=0.06614, over 3069110.92 frames. ], batch size: 62, lr: 4.50e-03, grad_scale: 4.0 2023-04-30 05:17:57,450 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-30 05:18:25,132 INFO [train.py:904] (4/8) Epoch 15, batch 7900, loss[loss=0.2524, simple_loss=0.3123, pruned_loss=0.09622, over 11497.00 frames. ], tot_loss[loss=0.2147, simple_loss=0.2977, pruned_loss=0.06585, over 3069990.94 frames. ], batch size: 247, lr: 4.49e-03, grad_scale: 4.0 2023-04-30 05:18:40,305 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.219e+02 2.824e+02 3.475e+02 4.207e+02 7.504e+02, threshold=6.949e+02, percent-clipped=0.0 2023-04-30 05:19:43,693 INFO [train.py:904] (4/8) Epoch 15, batch 7950, loss[loss=0.1916, simple_loss=0.2694, pruned_loss=0.05688, over 17087.00 frames. ], tot_loss[loss=0.215, simple_loss=0.298, pruned_loss=0.066, over 3084674.93 frames. ], batch size: 53, lr: 4.49e-03, grad_scale: 4.0 2023-04-30 05:20:25,869 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150079.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:20:35,087 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8984, 2.0921, 2.3930, 3.1713, 2.1844, 2.2977, 2.3103, 2.1507], device='cuda:4'), covar=tensor([0.1083, 0.3009, 0.2056, 0.0593, 0.3650, 0.2168, 0.2718, 0.3082], device='cuda:4'), in_proj_covar=tensor([0.0374, 0.0414, 0.0343, 0.0318, 0.0423, 0.0476, 0.0380, 0.0483], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:20:58,677 INFO [train.py:904] (4/8) Epoch 15, batch 8000, loss[loss=0.2041, simple_loss=0.296, pruned_loss=0.05608, over 16701.00 frames. ], tot_loss[loss=0.2159, simple_loss=0.2986, pruned_loss=0.06664, over 3080677.96 frames. ], batch size: 76, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:20:59,766 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=150102.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 05:21:14,673 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.969e+02 2.932e+02 3.545e+02 4.036e+02 7.060e+02, threshold=7.089e+02, percent-clipped=1.0 2023-04-30 05:21:55,970 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150140.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:22:14,245 INFO [train.py:904] (4/8) Epoch 15, batch 8050, loss[loss=0.2262, simple_loss=0.296, pruned_loss=0.07822, over 11465.00 frames. ], tot_loss[loss=0.2155, simple_loss=0.2986, pruned_loss=0.06622, over 3081630.64 frames. ], batch size: 246, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:22:23,959 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.33 vs. limit=5.0 2023-04-30 05:22:41,313 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150170.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:23:09,726 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-30 05:23:30,576 INFO [train.py:904] (4/8) Epoch 15, batch 8100, loss[loss=0.239, simple_loss=0.3041, pruned_loss=0.08691, over 11488.00 frames. ], tot_loss[loss=0.2143, simple_loss=0.2979, pruned_loss=0.06531, over 3086507.96 frames. ], batch size: 246, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:23:31,320 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.88 vs. limit=5.0 2023-04-30 05:23:45,512 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.847e+02 2.677e+02 3.415e+02 4.132e+02 1.188e+03, threshold=6.830e+02, percent-clipped=3.0 2023-04-30 05:23:56,038 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=150218.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:23:59,910 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150221.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:24:45,701 INFO [train.py:904] (4/8) Epoch 15, batch 8150, loss[loss=0.2358, simple_loss=0.2969, pruned_loss=0.08732, over 11886.00 frames. ], tot_loss[loss=0.2111, simple_loss=0.2946, pruned_loss=0.06385, over 3096136.68 frames. ], batch size: 248, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:24:50,565 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-30 05:24:56,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3767, 3.4558, 2.0247, 3.7366, 2.6519, 3.7673, 2.1795, 2.7309], device='cuda:4'), covar=tensor([0.0259, 0.0384, 0.1654, 0.0222, 0.0794, 0.0555, 0.1499, 0.0764], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0169, 0.0188, 0.0142, 0.0169, 0.0209, 0.0196, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 05:25:11,110 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=150269.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:25:19,702 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1553, 3.1284, 3.5395, 1.7581, 3.6939, 3.7390, 2.8302, 2.6839], device='cuda:4'), covar=tensor([0.0874, 0.0241, 0.0190, 0.1173, 0.0061, 0.0146, 0.0400, 0.0486], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0104, 0.0090, 0.0137, 0.0072, 0.0114, 0.0122, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 05:25:19,711 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150274.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:25:49,994 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-30 05:26:00,375 INFO [train.py:904] (4/8) Epoch 15, batch 8200, loss[loss=0.2508, simple_loss=0.3135, pruned_loss=0.09406, over 11751.00 frames. ], tot_loss[loss=0.2085, simple_loss=0.2915, pruned_loss=0.06276, over 3100938.09 frames. ], batch size: 247, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:26:16,358 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.058e+02 2.882e+02 3.369e+02 3.879e+02 7.748e+02, threshold=6.737e+02, percent-clipped=3.0 2023-04-30 05:26:43,105 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1565, 1.9448, 2.0885, 3.6747, 1.9649, 2.3965, 2.1147, 2.1257], device='cuda:4'), covar=tensor([0.1090, 0.3939, 0.2882, 0.0494, 0.4455, 0.2568, 0.3571, 0.3544], device='cuda:4'), in_proj_covar=tensor([0.0374, 0.0414, 0.0343, 0.0318, 0.0423, 0.0474, 0.0380, 0.0483], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:26:54,543 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150335.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:27:17,187 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7993, 3.7997, 4.1245, 4.1070, 4.1071, 3.9159, 3.8807, 3.9193], device='cuda:4'), covar=tensor([0.0351, 0.0644, 0.0433, 0.0445, 0.0484, 0.0434, 0.0916, 0.0448], device='cuda:4'), in_proj_covar=tensor([0.0368, 0.0396, 0.0386, 0.0370, 0.0442, 0.0408, 0.0507, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 05:27:22,317 INFO [train.py:904] (4/8) Epoch 15, batch 8250, loss[loss=0.1965, simple_loss=0.285, pruned_loss=0.05397, over 12217.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2909, pruned_loss=0.06089, over 3069153.11 frames. ], batch size: 247, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:28:45,011 INFO [train.py:904] (4/8) Epoch 15, batch 8300, loss[loss=0.189, simple_loss=0.2838, pruned_loss=0.04711, over 16893.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2876, pruned_loss=0.0578, over 3041326.42 frames. ], batch size: 116, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:28:45,736 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150402.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 05:28:51,042 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4414, 3.0445, 2.6987, 2.2300, 2.1746, 2.2151, 2.9802, 2.8857], device='cuda:4'), covar=tensor([0.2355, 0.0740, 0.1536, 0.2592, 0.2904, 0.2176, 0.0453, 0.1223], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0258, 0.0290, 0.0291, 0.0283, 0.0233, 0.0274, 0.0307], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:29:01,272 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.566e+02 2.394e+02 2.913e+02 3.586e+02 6.520e+02, threshold=5.826e+02, percent-clipped=0.0 2023-04-30 05:29:39,971 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=150435.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:29:56,829 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0121, 3.1498, 1.8931, 3.3065, 2.3767, 3.3262, 2.0974, 2.6533], device='cuda:4'), covar=tensor([0.0273, 0.0362, 0.1439, 0.0220, 0.0794, 0.0494, 0.1378, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0168, 0.0187, 0.0140, 0.0168, 0.0206, 0.0196, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 05:30:05,306 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=150450.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 05:30:07,755 INFO [train.py:904] (4/8) Epoch 15, batch 8350, loss[loss=0.2247, simple_loss=0.2986, pruned_loss=0.07535, over 12218.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2872, pruned_loss=0.056, over 3029829.36 frames. ], batch size: 247, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:31:29,420 INFO [train.py:904] (4/8) Epoch 15, batch 8400, loss[loss=0.1887, simple_loss=0.2816, pruned_loss=0.04791, over 16172.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2847, pruned_loss=0.05373, over 3023594.29 frames. ], batch size: 165, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:31:46,287 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.596e+02 2.409e+02 2.912e+02 3.516e+02 5.302e+02, threshold=5.823e+02, percent-clipped=0.0 2023-04-30 05:31:53,287 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7092, 3.9201, 2.3124, 4.3633, 2.9337, 4.3143, 2.6127, 3.1270], device='cuda:4'), covar=tensor([0.0221, 0.0268, 0.1367, 0.0161, 0.0675, 0.0399, 0.1226, 0.0592], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0166, 0.0185, 0.0138, 0.0166, 0.0204, 0.0194, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 05:32:27,364 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150537.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:32:49,402 INFO [train.py:904] (4/8) Epoch 15, batch 8450, loss[loss=0.1786, simple_loss=0.2691, pruned_loss=0.04405, over 17061.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2829, pruned_loss=0.05229, over 3006964.76 frames. ], batch size: 55, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:33:45,065 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0371, 4.1120, 3.9404, 3.6714, 3.6918, 4.0365, 3.7089, 3.8178], device='cuda:4'), covar=tensor([0.0514, 0.0448, 0.0269, 0.0254, 0.0625, 0.0388, 0.1088, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0257, 0.0357, 0.0302, 0.0287, 0.0316, 0.0332, 0.0208, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:34:02,287 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150598.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:34:08,631 INFO [train.py:904] (4/8) Epoch 15, batch 8500, loss[loss=0.1623, simple_loss=0.241, pruned_loss=0.04176, over 12047.00 frames. ], tot_loss[loss=0.189, simple_loss=0.279, pruned_loss=0.04947, over 3020646.32 frames. ], batch size: 248, lr: 4.49e-03, grad_scale: 8.0 2023-04-30 05:34:25,006 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.222e+02 2.200e+02 2.703e+02 3.335e+02 6.908e+02, threshold=5.407e+02, percent-clipped=1.0 2023-04-30 05:34:54,218 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=150630.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:35:30,232 INFO [train.py:904] (4/8) Epoch 15, batch 8550, loss[loss=0.1836, simple_loss=0.2837, pruned_loss=0.04178, over 16687.00 frames. ], tot_loss[loss=0.188, simple_loss=0.2781, pruned_loss=0.04901, over 3029594.85 frames. ], batch size: 89, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:36:15,152 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-30 05:36:21,251 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4349, 3.5884, 3.8155, 1.7480, 3.9268, 4.1481, 3.1873, 3.0711], device='cuda:4'), covar=tensor([0.0882, 0.0165, 0.0168, 0.1238, 0.0055, 0.0102, 0.0323, 0.0454], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0102, 0.0087, 0.0134, 0.0070, 0.0110, 0.0119, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 05:36:43,562 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8462, 3.6948, 3.8919, 4.0033, 4.0907, 3.6443, 4.0417, 4.1018], device='cuda:4'), covar=tensor([0.1337, 0.1086, 0.1229, 0.0640, 0.0498, 0.1974, 0.0644, 0.0573], device='cuda:4'), in_proj_covar=tensor([0.0546, 0.0680, 0.0803, 0.0688, 0.0524, 0.0544, 0.0553, 0.0642], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:36:48,780 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5727, 4.7288, 4.8809, 4.7421, 4.6784, 5.2507, 4.7737, 4.5040], device='cuda:4'), covar=tensor([0.1128, 0.1858, 0.1657, 0.1892, 0.2637, 0.1071, 0.1482, 0.2246], device='cuda:4'), in_proj_covar=tensor([0.0367, 0.0516, 0.0566, 0.0435, 0.0579, 0.0595, 0.0448, 0.0583], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 05:37:12,290 INFO [train.py:904] (4/8) Epoch 15, batch 8600, loss[loss=0.1964, simple_loss=0.2912, pruned_loss=0.05078, over 16764.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2779, pruned_loss=0.04777, over 3039592.80 frames. ], batch size: 124, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:37:32,420 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.969e+02 2.521e+02 3.034e+02 3.543e+02 4.971e+02, threshold=6.069e+02, percent-clipped=0.0 2023-04-30 05:38:14,916 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150732.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:38:20,223 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150735.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:38:51,972 INFO [train.py:904] (4/8) Epoch 15, batch 8650, loss[loss=0.1822, simple_loss=0.2791, pruned_loss=0.04267, over 15250.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.2758, pruned_loss=0.04646, over 3018604.38 frames. ], batch size: 190, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:39:04,154 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3948, 1.9497, 1.6611, 1.7085, 2.2453, 1.8798, 2.0007, 2.3477], device='cuda:4'), covar=tensor([0.0156, 0.0399, 0.0471, 0.0432, 0.0259, 0.0345, 0.0169, 0.0227], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0211, 0.0206, 0.0206, 0.0210, 0.0210, 0.0209, 0.0203], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:40:03,477 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=150783.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:40:19,420 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8963, 2.7988, 2.6813, 2.0919, 2.5181, 2.7812, 2.6812, 1.9051], device='cuda:4'), covar=tensor([0.0392, 0.0062, 0.0052, 0.0306, 0.0096, 0.0075, 0.0079, 0.0386], device='cuda:4'), in_proj_covar=tensor([0.0127, 0.0071, 0.0071, 0.0129, 0.0085, 0.0093, 0.0082, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 05:40:23,688 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150793.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 05:40:39,983 INFO [train.py:904] (4/8) Epoch 15, batch 8700, loss[loss=0.1619, simple_loss=0.2484, pruned_loss=0.03765, over 12436.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2728, pruned_loss=0.04501, over 3012114.45 frames. ], batch size: 247, lr: 4.48e-03, grad_scale: 4.0 2023-04-30 05:41:01,898 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.169e+02 2.751e+02 3.261e+02 6.645e+02, threshold=5.502e+02, percent-clipped=1.0 2023-04-30 05:41:53,967 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7314, 3.2633, 3.4167, 1.9385, 2.9137, 2.1877, 3.1522, 3.3368], device='cuda:4'), covar=tensor([0.0274, 0.0722, 0.0458, 0.1941, 0.0724, 0.0955, 0.0691, 0.0900], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0149, 0.0158, 0.0144, 0.0136, 0.0124, 0.0137, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 05:42:16,603 INFO [train.py:904] (4/8) Epoch 15, batch 8750, loss[loss=0.1898, simple_loss=0.2727, pruned_loss=0.05342, over 12285.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2726, pruned_loss=0.04455, over 3016595.69 frames. ], batch size: 247, lr: 4.48e-03, grad_scale: 4.0 2023-04-30 05:43:18,822 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-30 05:43:52,657 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=150893.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:44:06,369 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150900.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:44:08,758 INFO [train.py:904] (4/8) Epoch 15, batch 8800, loss[loss=0.208, simple_loss=0.2919, pruned_loss=0.06208, over 12798.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2708, pruned_loss=0.0435, over 3011323.19 frames. ], batch size: 247, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:44:28,830 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.691e+02 2.483e+02 3.172e+02 3.684e+02 7.481e+02, threshold=6.344e+02, percent-clipped=2.0 2023-04-30 05:45:06,312 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150930.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:45:50,511 INFO [train.py:904] (4/8) Epoch 15, batch 8850, loss[loss=0.1757, simple_loss=0.281, pruned_loss=0.03525, over 15482.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2735, pruned_loss=0.04294, over 3022735.09 frames. ], batch size: 192, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:46:10,288 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150961.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:46:46,308 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=150978.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:47:03,093 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-30 05:47:20,997 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2762, 3.3904, 3.6089, 3.5928, 3.6011, 3.4370, 3.3227, 3.4803], device='cuda:4'), covar=tensor([0.0530, 0.0867, 0.0630, 0.0624, 0.0667, 0.0651, 0.1210, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0373, 0.0365, 0.0351, 0.0418, 0.0387, 0.0477, 0.0311], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 05:47:36,864 INFO [train.py:904] (4/8) Epoch 15, batch 8900, loss[loss=0.18, simple_loss=0.2663, pruned_loss=0.0469, over 12475.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2735, pruned_loss=0.04229, over 3029493.93 frames. ], batch size: 247, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:47:59,492 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.662e+02 2.270e+02 2.672e+02 3.305e+02 7.174e+02, threshold=5.344e+02, percent-clipped=2.0 2023-04-30 05:48:14,534 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9358, 2.1043, 2.3030, 3.2577, 2.1241, 2.2898, 2.2621, 2.1340], device='cuda:4'), covar=tensor([0.1092, 0.3157, 0.2407, 0.0574, 0.4065, 0.2420, 0.3259, 0.3543], device='cuda:4'), in_proj_covar=tensor([0.0368, 0.0405, 0.0341, 0.0310, 0.0416, 0.0465, 0.0374, 0.0473], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:48:52,934 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-30 05:48:55,843 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-30 05:49:42,110 INFO [train.py:904] (4/8) Epoch 15, batch 8950, loss[loss=0.1644, simple_loss=0.2568, pruned_loss=0.03598, over 16958.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2727, pruned_loss=0.04236, over 3051506.32 frames. ], batch size: 125, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:51:01,604 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151088.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 05:51:30,995 INFO [train.py:904] (4/8) Epoch 15, batch 9000, loss[loss=0.1622, simple_loss=0.253, pruned_loss=0.0357, over 16962.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2695, pruned_loss=0.04104, over 3079514.89 frames. ], batch size: 116, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:51:30,995 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 05:51:40,827 INFO [train.py:938] (4/8) Epoch 15, validation: loss=0.15, simple_loss=0.2539, pruned_loss=0.02307, over 944034.00 frames. 2023-04-30 05:51:40,828 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 05:52:03,718 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.315e+02 2.164e+02 2.577e+02 3.241e+02 6.734e+02, threshold=5.154e+02, percent-clipped=2.0 2023-04-30 05:53:23,275 INFO [train.py:904] (4/8) Epoch 15, batch 9050, loss[loss=0.1759, simple_loss=0.2627, pruned_loss=0.04459, over 16370.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2703, pruned_loss=0.04155, over 3080530.66 frames. ], batch size: 146, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:54:02,965 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-30 05:54:28,393 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151184.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:54:51,863 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151193.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:55:09,971 INFO [train.py:904] (4/8) Epoch 15, batch 9100, loss[loss=0.1963, simple_loss=0.2996, pruned_loss=0.04653, over 16091.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2703, pruned_loss=0.04195, over 3106053.62 frames. ], batch size: 165, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:55:31,073 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.671e+02 2.276e+02 2.684e+02 3.281e+02 5.650e+02, threshold=5.369e+02, percent-clipped=3.0 2023-04-30 05:56:44,480 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=151241.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:56:54,442 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151245.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:57:08,959 INFO [train.py:904] (4/8) Epoch 15, batch 9150, loss[loss=0.1638, simple_loss=0.2563, pruned_loss=0.03561, over 17056.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2703, pruned_loss=0.04154, over 3091091.04 frames. ], batch size: 55, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:57:18,559 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151256.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:57:44,635 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9605, 5.0295, 4.8516, 4.5426, 4.5385, 4.9675, 4.8493, 4.6319], device='cuda:4'), covar=tensor([0.0627, 0.0403, 0.0306, 0.0256, 0.0898, 0.0418, 0.0331, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0348, 0.0299, 0.0282, 0.0308, 0.0326, 0.0205, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:57:46,747 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8241, 2.1782, 1.8353, 1.9723, 2.5050, 2.1680, 2.4392, 2.7092], device='cuda:4'), covar=tensor([0.0114, 0.0378, 0.0432, 0.0428, 0.0241, 0.0378, 0.0141, 0.0216], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0212, 0.0206, 0.0206, 0.0211, 0.0211, 0.0210, 0.0202], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 05:58:12,601 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151282.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:58:14,739 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4953, 1.7918, 2.1571, 2.4379, 2.4488, 2.7558, 1.8722, 2.6749], device='cuda:4'), covar=tensor([0.0178, 0.0449, 0.0284, 0.0286, 0.0269, 0.0173, 0.0463, 0.0166], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0175, 0.0160, 0.0163, 0.0174, 0.0131, 0.0176, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-30 05:58:37,728 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-30 05:58:46,493 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151298.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 05:58:51,676 INFO [train.py:904] (4/8) Epoch 15, batch 9200, loss[loss=0.1879, simple_loss=0.2768, pruned_loss=0.04943, over 16680.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2664, pruned_loss=0.04072, over 3099617.01 frames. ], batch size: 134, lr: 4.48e-03, grad_scale: 8.0 2023-04-30 05:59:12,238 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.606e+02 2.499e+02 3.414e+02 4.138e+02 7.678e+02, threshold=6.827e+02, percent-clipped=7.0 2023-04-30 06:00:09,365 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151343.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:00:24,878 INFO [train.py:904] (4/8) Epoch 15, batch 9250, loss[loss=0.1681, simple_loss=0.2607, pruned_loss=0.03775, over 15297.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2659, pruned_loss=0.04059, over 3072784.35 frames. ], batch size: 190, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:00:41,061 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151359.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:01:43,245 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151388.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:02:14,982 INFO [train.py:904] (4/8) Epoch 15, batch 9300, loss[loss=0.1716, simple_loss=0.2625, pruned_loss=0.04036, over 16143.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2642, pruned_loss=0.03992, over 3075246.94 frames. ], batch size: 165, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:02:23,518 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4055, 4.2795, 4.5041, 4.5864, 4.7497, 4.3224, 4.7448, 4.7771], device='cuda:4'), covar=tensor([0.1526, 0.1150, 0.1170, 0.0617, 0.0445, 0.0884, 0.0457, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0532, 0.0661, 0.0781, 0.0676, 0.0509, 0.0531, 0.0542, 0.0624], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:02:37,910 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.525e+02 2.225e+02 2.604e+02 3.286e+02 5.862e+02, threshold=5.207e+02, percent-clipped=0.0 2023-04-30 06:03:29,231 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=151436.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:03:36,253 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3478, 4.5052, 4.2453, 4.0035, 3.8305, 4.4125, 4.2098, 3.9873], device='cuda:4'), covar=tensor([0.0656, 0.0561, 0.0366, 0.0345, 0.1018, 0.0492, 0.0569, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0252, 0.0347, 0.0299, 0.0282, 0.0308, 0.0327, 0.0205, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:03:59,110 INFO [train.py:904] (4/8) Epoch 15, batch 9350, loss[loss=0.1754, simple_loss=0.26, pruned_loss=0.04538, over 12156.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.264, pruned_loss=0.03979, over 3072195.68 frames. ], batch size: 248, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:04:14,961 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4451, 3.0903, 2.7796, 2.2269, 2.2448, 2.2289, 3.0517, 2.9172], device='cuda:4'), covar=tensor([0.2372, 0.0637, 0.1325, 0.2548, 0.2330, 0.1985, 0.0435, 0.1184], device='cuda:4'), in_proj_covar=tensor([0.0304, 0.0253, 0.0285, 0.0286, 0.0270, 0.0230, 0.0268, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:04:42,714 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-30 06:04:44,200 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3353, 4.3127, 4.7021, 4.6930, 4.6727, 4.3819, 4.3982, 4.2962], device='cuda:4'), covar=tensor([0.0288, 0.0833, 0.0372, 0.0470, 0.0481, 0.0529, 0.0848, 0.0463], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0370, 0.0363, 0.0347, 0.0414, 0.0386, 0.0471, 0.0308], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 06:05:36,966 INFO [train.py:904] (4/8) Epoch 15, batch 9400, loss[loss=0.2087, simple_loss=0.3086, pruned_loss=0.05442, over 16903.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2638, pruned_loss=0.03973, over 3060101.87 frames. ], batch size: 116, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:05:59,179 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.525e+02 2.126e+02 2.557e+02 3.057e+02 4.455e+02, threshold=5.114e+02, percent-clipped=0.0 2023-04-30 06:06:55,073 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151540.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:07:17,456 INFO [train.py:904] (4/8) Epoch 15, batch 9450, loss[loss=0.1763, simple_loss=0.273, pruned_loss=0.03978, over 16203.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2659, pruned_loss=0.03992, over 3068788.99 frames. ], batch size: 165, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:07:24,792 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151556.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:08:18,292 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9455, 2.3287, 2.3231, 3.0491, 1.9377, 3.3220, 1.6951, 2.8054], device='cuda:4'), covar=tensor([0.1236, 0.0605, 0.1014, 0.0144, 0.0080, 0.0375, 0.1435, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0159, 0.0181, 0.0158, 0.0187, 0.0200, 0.0185, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-30 06:08:36,388 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-30 06:08:58,267 INFO [train.py:904] (4/8) Epoch 15, batch 9500, loss[loss=0.1668, simple_loss=0.2605, pruned_loss=0.03654, over 15524.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.266, pruned_loss=0.04006, over 3051325.06 frames. ], batch size: 191, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:09:04,124 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=151604.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:09:22,316 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.398e+02 2.144e+02 2.700e+02 3.388e+02 6.768e+02, threshold=5.400e+02, percent-clipped=4.0 2023-04-30 06:09:49,632 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-30 06:10:11,844 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151638.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:10:44,462 INFO [train.py:904] (4/8) Epoch 15, batch 9550, loss[loss=0.1642, simple_loss=0.2572, pruned_loss=0.03561, over 12438.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2656, pruned_loss=0.04053, over 3045304.71 frames. ], batch size: 246, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:10:49,183 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151654.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:11:40,533 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-30 06:11:58,457 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151688.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:12:12,794 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-04-30 06:12:24,746 INFO [train.py:904] (4/8) Epoch 15, batch 9600, loss[loss=0.18, simple_loss=0.2845, pruned_loss=0.03773, over 16196.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2668, pruned_loss=0.04082, over 3065217.10 frames. ], batch size: 165, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:12:44,296 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.467e+02 2.311e+02 2.682e+02 3.396e+02 6.074e+02, threshold=5.365e+02, percent-clipped=2.0 2023-04-30 06:14:04,290 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151749.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:14:11,685 INFO [train.py:904] (4/8) Epoch 15, batch 9650, loss[loss=0.1909, simple_loss=0.2912, pruned_loss=0.04531, over 16916.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2681, pruned_loss=0.04082, over 3057221.26 frames. ], batch size: 102, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:14:36,026 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-30 06:14:52,101 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-30 06:15:58,106 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151801.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:15:58,768 INFO [train.py:904] (4/8) Epoch 15, batch 9700, loss[loss=0.1737, simple_loss=0.2697, pruned_loss=0.0389, over 15211.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2677, pruned_loss=0.04068, over 3066516.01 frames. ], batch size: 190, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:16:19,905 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.752e+02 2.225e+02 2.684e+02 3.461e+02 6.863e+02, threshold=5.368e+02, percent-clipped=3.0 2023-04-30 06:17:18,824 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151840.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:17:41,644 INFO [train.py:904] (4/8) Epoch 15, batch 9750, loss[loss=0.1629, simple_loss=0.2557, pruned_loss=0.035, over 16718.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2665, pruned_loss=0.04064, over 3072791.05 frames. ], batch size: 83, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:18:01,505 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151862.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:18:56,065 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=151888.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:19:18,821 INFO [train.py:904] (4/8) Epoch 15, batch 9800, loss[loss=0.1708, simple_loss=0.269, pruned_loss=0.03626, over 16560.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.267, pruned_loss=0.04005, over 3075702.91 frames. ], batch size: 68, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:19:40,660 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.482e+02 2.231e+02 2.655e+02 3.314e+02 7.232e+02, threshold=5.310e+02, percent-clipped=1.0 2023-04-30 06:20:30,693 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151938.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:21:03,075 INFO [train.py:904] (4/8) Epoch 15, batch 9850, loss[loss=0.1945, simple_loss=0.2947, pruned_loss=0.0472, over 16331.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2678, pruned_loss=0.0396, over 3076990.60 frames. ], batch size: 146, lr: 4.47e-03, grad_scale: 8.0 2023-04-30 06:21:08,464 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151954.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:22:17,652 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=151986.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:22:57,937 INFO [train.py:904] (4/8) Epoch 15, batch 9900, loss[loss=0.1893, simple_loss=0.2897, pruned_loss=0.04439, over 16424.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2681, pruned_loss=0.0392, over 3086411.58 frames. ], batch size: 146, lr: 4.46e-03, grad_scale: 8.0 2023-04-30 06:22:58,683 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=152002.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:23:08,665 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-30 06:23:24,818 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.484e+02 2.244e+02 2.678e+02 3.249e+02 7.284e+02, threshold=5.355e+02, percent-clipped=2.0 2023-04-30 06:24:37,872 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152044.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:24:55,576 INFO [train.py:904] (4/8) Epoch 15, batch 9950, loss[loss=0.1701, simple_loss=0.2703, pruned_loss=0.03491, over 16280.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2704, pruned_loss=0.03971, over 3091152.18 frames. ], batch size: 146, lr: 4.46e-03, grad_scale: 8.0 2023-04-30 06:24:57,197 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-30 06:25:24,221 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9427, 4.9197, 4.7454, 4.2874, 4.7700, 2.0731, 4.5589, 4.5982], device='cuda:4'), covar=tensor([0.0065, 0.0065, 0.0138, 0.0249, 0.0089, 0.2131, 0.0104, 0.0189], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0127, 0.0171, 0.0153, 0.0147, 0.0187, 0.0160, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:26:56,976 INFO [train.py:904] (4/8) Epoch 15, batch 10000, loss[loss=0.1868, simple_loss=0.286, pruned_loss=0.04383, over 15745.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2688, pruned_loss=0.03926, over 3103828.91 frames. ], batch size: 194, lr: 4.46e-03, grad_scale: 8.0 2023-04-30 06:27:18,766 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.352e+02 2.235e+02 2.836e+02 3.494e+02 9.282e+02, threshold=5.672e+02, percent-clipped=5.0 2023-04-30 06:27:38,879 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-30 06:28:35,912 INFO [train.py:904] (4/8) Epoch 15, batch 10050, loss[loss=0.1865, simple_loss=0.2869, pruned_loss=0.04299, over 15281.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2687, pruned_loss=0.03929, over 3081685.73 frames. ], batch size: 191, lr: 4.46e-03, grad_scale: 4.0 2023-04-30 06:28:45,747 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152157.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:29:04,135 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2966, 3.3540, 2.1275, 3.7146, 2.4180, 3.6524, 2.0782, 2.6809], device='cuda:4'), covar=tensor([0.0275, 0.0394, 0.1443, 0.0162, 0.0898, 0.0555, 0.1509, 0.0707], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0161, 0.0183, 0.0135, 0.0166, 0.0198, 0.0194, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 06:29:04,439 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-30 06:29:27,707 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3695, 2.0296, 2.3331, 4.1316, 2.0332, 2.3049, 2.1882, 2.1665], device='cuda:4'), covar=tensor([0.1159, 0.4266, 0.2597, 0.0470, 0.4815, 0.2813, 0.3566, 0.3808], device='cuda:4'), in_proj_covar=tensor([0.0363, 0.0399, 0.0337, 0.0305, 0.0412, 0.0454, 0.0368, 0.0465], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:30:08,501 INFO [train.py:904] (4/8) Epoch 15, batch 10100, loss[loss=0.18, simple_loss=0.2722, pruned_loss=0.04397, over 12615.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2689, pruned_loss=0.03941, over 3078137.20 frames. ], batch size: 247, lr: 4.46e-03, grad_scale: 4.0 2023-04-30 06:30:28,184 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.524e+02 2.142e+02 2.576e+02 3.219e+02 1.174e+03, threshold=5.152e+02, percent-clipped=3.0 2023-04-30 06:30:55,367 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152225.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:31:53,397 INFO [train.py:904] (4/8) Epoch 16, batch 0, loss[loss=0.2624, simple_loss=0.332, pruned_loss=0.0964, over 15409.00 frames. ], tot_loss[loss=0.2624, simple_loss=0.332, pruned_loss=0.0964, over 15409.00 frames. ], batch size: 191, lr: 4.32e-03, grad_scale: 8.0 2023-04-30 06:31:53,398 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 06:32:00,892 INFO [train.py:938] (4/8) Epoch 16, validation: loss=0.1492, simple_loss=0.2525, pruned_loss=0.02288, over 944034.00 frames. 2023-04-30 06:32:00,892 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 06:32:12,312 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3777, 4.3433, 4.3523, 3.9887, 4.2677, 1.7414, 4.0544, 4.0270], device='cuda:4'), covar=tensor([0.0144, 0.0101, 0.0169, 0.0248, 0.0134, 0.2361, 0.0184, 0.0235], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0127, 0.0171, 0.0153, 0.0147, 0.0187, 0.0161, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:32:29,815 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152273.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:32:30,334 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-30 06:32:48,488 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152286.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:33:09,961 INFO [train.py:904] (4/8) Epoch 16, batch 50, loss[loss=0.2066, simple_loss=0.2786, pruned_loss=0.06732, over 16749.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2815, pruned_loss=0.05679, over 751087.49 frames. ], batch size: 134, lr: 4.32e-03, grad_scale: 1.0 2023-04-30 06:33:12,145 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9542, 4.5806, 4.7331, 5.1101, 5.2635, 4.6628, 5.3196, 5.2732], device='cuda:4'), covar=tensor([0.1685, 0.1547, 0.2423, 0.1053, 0.0782, 0.0989, 0.0716, 0.0963], device='cuda:4'), in_proj_covar=tensor([0.0539, 0.0670, 0.0789, 0.0685, 0.0515, 0.0534, 0.0550, 0.0633], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:33:29,875 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.302e+02 2.475e+02 3.034e+02 3.841e+02 6.460e+02, threshold=6.067e+02, percent-clipped=5.0 2023-04-30 06:33:53,516 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152334.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:34:08,289 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=152344.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:34:18,730 INFO [train.py:904] (4/8) Epoch 16, batch 100, loss[loss=0.1737, simple_loss=0.2705, pruned_loss=0.03838, over 17048.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2755, pruned_loss=0.05389, over 1318748.74 frames. ], batch size: 50, lr: 4.32e-03, grad_scale: 1.0 2023-04-30 06:34:25,648 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-30 06:35:14,870 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=152392.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:35:26,676 INFO [train.py:904] (4/8) Epoch 16, batch 150, loss[loss=0.1883, simple_loss=0.2825, pruned_loss=0.04707, over 16697.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2715, pruned_loss=0.05173, over 1761681.27 frames. ], batch size: 57, lr: 4.32e-03, grad_scale: 1.0 2023-04-30 06:35:48,055 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.712e+02 2.454e+02 2.817e+02 3.397e+02 1.160e+03, threshold=5.634e+02, percent-clipped=3.0 2023-04-30 06:36:24,969 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152445.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:36:35,178 INFO [train.py:904] (4/8) Epoch 16, batch 200, loss[loss=0.1994, simple_loss=0.2838, pruned_loss=0.05751, over 16692.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2708, pruned_loss=0.05165, over 2110381.00 frames. ], batch size: 57, lr: 4.32e-03, grad_scale: 1.0 2023-04-30 06:36:42,880 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=152457.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:37:44,240 INFO [train.py:904] (4/8) Epoch 16, batch 250, loss[loss=0.1879, simple_loss=0.2644, pruned_loss=0.05569, over 16765.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2693, pruned_loss=0.0517, over 2389004.34 frames. ], batch size: 124, lr: 4.31e-03, grad_scale: 1.0 2023-04-30 06:37:48,058 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=152505.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:37:50,052 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152506.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:38:05,708 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.680e+02 2.341e+02 2.836e+02 3.695e+02 5.924e+02, threshold=5.672e+02, percent-clipped=1.0 2023-04-30 06:38:13,084 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 06:38:47,291 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.17 vs. limit=5.0 2023-04-30 06:38:53,026 INFO [train.py:904] (4/8) Epoch 16, batch 300, loss[loss=0.1793, simple_loss=0.2558, pruned_loss=0.05138, over 16778.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2663, pruned_loss=0.04964, over 2607196.78 frames. ], batch size: 124, lr: 4.31e-03, grad_scale: 1.0 2023-04-30 06:39:33,771 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152581.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:39:34,998 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152582.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:39:43,034 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152589.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:39:57,826 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8859, 2.0913, 2.3598, 2.8657, 2.6818, 3.3919, 2.1695, 3.2757], device='cuda:4'), covar=tensor([0.0226, 0.0410, 0.0322, 0.0277, 0.0291, 0.0159, 0.0418, 0.0146], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0181, 0.0166, 0.0169, 0.0179, 0.0136, 0.0182, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:40:01,537 INFO [train.py:904] (4/8) Epoch 16, batch 350, loss[loss=0.1502, simple_loss=0.2342, pruned_loss=0.03315, over 17202.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2626, pruned_loss=0.04732, over 2772331.38 frames. ], batch size: 44, lr: 4.31e-03, grad_scale: 1.0 2023-04-30 06:40:20,730 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.536e+02 2.239e+02 2.608e+02 3.067e+02 7.666e+02, threshold=5.216e+02, percent-clipped=3.0 2023-04-30 06:40:38,376 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152629.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:40:58,713 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152643.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:41:06,972 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152650.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:41:08,822 INFO [train.py:904] (4/8) Epoch 16, batch 400, loss[loss=0.1836, simple_loss=0.2552, pruned_loss=0.05606, over 16352.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2615, pruned_loss=0.04761, over 2892411.70 frames. ], batch size: 165, lr: 4.31e-03, grad_scale: 2.0 2023-04-30 06:41:16,092 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-30 06:42:16,089 INFO [train.py:904] (4/8) Epoch 16, batch 450, loss[loss=0.1574, simple_loss=0.2388, pruned_loss=0.03799, over 16803.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2602, pruned_loss=0.04683, over 2989821.25 frames. ], batch size: 102, lr: 4.31e-03, grad_scale: 2.0 2023-04-30 06:42:36,230 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.633e+02 2.224e+02 2.572e+02 2.956e+02 7.682e+02, threshold=5.144e+02, percent-clipped=1.0 2023-04-30 06:42:50,488 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152726.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:42:54,169 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-30 06:42:57,197 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1502, 4.8250, 5.1313, 5.3244, 5.5555, 4.8967, 5.4971, 5.5317], device='cuda:4'), covar=tensor([0.1675, 0.1430, 0.1682, 0.0833, 0.0549, 0.0741, 0.0580, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0582, 0.0725, 0.0851, 0.0732, 0.0552, 0.0577, 0.0591, 0.0679], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:43:09,062 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-30 06:43:24,043 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8375, 3.9421, 3.0098, 2.3467, 2.7195, 2.5595, 4.1754, 3.5899], device='cuda:4'), covar=tensor([0.2579, 0.0619, 0.1606, 0.2451, 0.2405, 0.1787, 0.0458, 0.1278], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0261, 0.0292, 0.0290, 0.0278, 0.0236, 0.0276, 0.0311], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:43:25,827 INFO [train.py:904] (4/8) Epoch 16, batch 500, loss[loss=0.1856, simple_loss=0.2582, pruned_loss=0.05648, over 16838.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2589, pruned_loss=0.04606, over 3070536.36 frames. ], batch size: 109, lr: 4.31e-03, grad_scale: 2.0 2023-04-30 06:43:41,856 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6612, 1.8664, 2.2098, 2.4923, 2.5545, 2.4636, 1.8063, 2.7239], device='cuda:4'), covar=tensor([0.0143, 0.0390, 0.0264, 0.0229, 0.0243, 0.0287, 0.0437, 0.0133], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0181, 0.0166, 0.0170, 0.0179, 0.0136, 0.0182, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:44:14,753 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152787.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:44:33,457 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152801.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:44:34,216 INFO [train.py:904] (4/8) Epoch 16, batch 550, loss[loss=0.1935, simple_loss=0.2848, pruned_loss=0.05108, over 16666.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2581, pruned_loss=0.0454, over 3126535.52 frames. ], batch size: 57, lr: 4.31e-03, grad_scale: 2.0 2023-04-30 06:44:55,529 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.310e+02 2.218e+02 2.743e+02 3.343e+02 5.376e+02, threshold=5.487e+02, percent-clipped=2.0 2023-04-30 06:45:17,896 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-30 06:45:46,473 INFO [train.py:904] (4/8) Epoch 16, batch 600, loss[loss=0.1709, simple_loss=0.2562, pruned_loss=0.04282, over 17190.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2577, pruned_loss=0.04555, over 3169244.77 frames. ], batch size: 46, lr: 4.31e-03, grad_scale: 2.0 2023-04-30 06:46:15,400 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-30 06:46:25,164 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=152881.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:46:41,333 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-30 06:46:53,580 INFO [train.py:904] (4/8) Epoch 16, batch 650, loss[loss=0.1807, simple_loss=0.2779, pruned_loss=0.0418, over 16756.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2556, pruned_loss=0.04494, over 3199156.79 frames. ], batch size: 57, lr: 4.31e-03, grad_scale: 2.0 2023-04-30 06:47:14,336 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.748e+02 2.133e+02 2.528e+02 3.105e+02 5.746e+02, threshold=5.056e+02, percent-clipped=1.0 2023-04-30 06:47:30,635 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=152929.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:47:30,805 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=152929.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:47:43,509 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152938.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:47:52,080 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152945.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:48:00,896 INFO [train.py:904] (4/8) Epoch 16, batch 700, loss[loss=0.1458, simple_loss=0.2322, pruned_loss=0.02968, over 17234.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2554, pruned_loss=0.04445, over 3227191.56 frames. ], batch size: 45, lr: 4.31e-03, grad_scale: 2.0 2023-04-30 06:48:37,713 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=152977.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:49:10,743 INFO [train.py:904] (4/8) Epoch 16, batch 750, loss[loss=0.1768, simple_loss=0.2691, pruned_loss=0.04225, over 17021.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2564, pruned_loss=0.04465, over 3245370.79 frames. ], batch size: 55, lr: 4.31e-03, grad_scale: 2.0 2023-04-30 06:49:31,098 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.515e+02 2.303e+02 2.598e+02 3.090e+02 5.870e+02, threshold=5.196e+02, percent-clipped=1.0 2023-04-30 06:49:47,340 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-30 06:49:57,723 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6778, 1.7083, 1.5508, 1.5366, 1.9054, 1.5761, 1.6960, 1.9339], device='cuda:4'), covar=tensor([0.0215, 0.0309, 0.0410, 0.0407, 0.0205, 0.0327, 0.0229, 0.0193], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0225, 0.0217, 0.0219, 0.0225, 0.0225, 0.0226, 0.0217], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:50:08,432 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-30 06:50:17,801 INFO [train.py:904] (4/8) Epoch 16, batch 800, loss[loss=0.1799, simple_loss=0.256, pruned_loss=0.05195, over 15479.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.256, pruned_loss=0.04457, over 3264999.89 frames. ], batch size: 190, lr: 4.31e-03, grad_scale: 4.0 2023-04-30 06:50:21,706 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3681, 2.2471, 2.3667, 4.2229, 2.2182, 2.7041, 2.3061, 2.3960], device='cuda:4'), covar=tensor([0.1223, 0.3700, 0.2847, 0.0534, 0.3942, 0.2299, 0.3665, 0.3114], device='cuda:4'), in_proj_covar=tensor([0.0380, 0.0418, 0.0352, 0.0322, 0.0427, 0.0478, 0.0386, 0.0489], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:51:00,654 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=153082.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:51:25,512 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153101.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:51:26,408 INFO [train.py:904] (4/8) Epoch 16, batch 850, loss[loss=0.1652, simple_loss=0.2432, pruned_loss=0.04358, over 16424.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2552, pruned_loss=0.04414, over 3284922.41 frames. ], batch size: 75, lr: 4.31e-03, grad_scale: 4.0 2023-04-30 06:51:46,528 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.323e+02 2.045e+02 2.529e+02 2.957e+02 4.458e+02, threshold=5.058e+02, percent-clipped=0.0 2023-04-30 06:51:47,000 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=153117.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:52:13,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9369, 3.3182, 2.9826, 5.0806, 4.2465, 4.4864, 1.6927, 3.3124], device='cuda:4'), covar=tensor([0.1288, 0.0615, 0.1028, 0.0183, 0.0230, 0.0410, 0.1495, 0.0711], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0165, 0.0187, 0.0167, 0.0197, 0.0212, 0.0191, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 06:52:32,207 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=153149.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:52:35,389 INFO [train.py:904] (4/8) Epoch 16, batch 900, loss[loss=0.1855, simple_loss=0.262, pruned_loss=0.05448, over 16736.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2541, pruned_loss=0.04355, over 3300124.65 frames. ], batch size: 89, lr: 4.31e-03, grad_scale: 4.0 2023-04-30 06:53:10,979 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=153178.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:53:20,122 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0933, 4.7797, 5.0973, 5.2803, 5.4748, 4.8275, 5.4678, 5.4383], device='cuda:4'), covar=tensor([0.1764, 0.1292, 0.1657, 0.0738, 0.0559, 0.0773, 0.0538, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0600, 0.0743, 0.0880, 0.0752, 0.0567, 0.0591, 0.0608, 0.0701], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:53:41,472 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8983, 3.2219, 2.9281, 5.0874, 4.2539, 4.5082, 1.6761, 3.4249], device='cuda:4'), covar=tensor([0.1311, 0.0650, 0.1060, 0.0180, 0.0213, 0.0401, 0.1558, 0.0668], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0165, 0.0187, 0.0168, 0.0197, 0.0212, 0.0191, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 06:53:43,181 INFO [train.py:904] (4/8) Epoch 16, batch 950, loss[loss=0.1683, simple_loss=0.2473, pruned_loss=0.04464, over 16713.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2548, pruned_loss=0.04428, over 3294225.98 frames. ], batch size: 62, lr: 4.30e-03, grad_scale: 4.0 2023-04-30 06:54:04,601 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.182e+02 2.578e+02 3.357e+02 5.954e+02, threshold=5.156e+02, percent-clipped=2.0 2023-04-30 06:54:16,934 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-30 06:54:33,155 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153238.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:54:42,574 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153245.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:54:53,047 INFO [train.py:904] (4/8) Epoch 16, batch 1000, loss[loss=0.1374, simple_loss=0.218, pruned_loss=0.02842, over 17029.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2532, pruned_loss=0.04374, over 3296133.11 frames. ], batch size: 41, lr: 4.30e-03, grad_scale: 4.0 2023-04-30 06:55:29,839 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=153278.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:55:40,234 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=153286.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:55:50,994 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=153293.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:56:02,850 INFO [train.py:904] (4/8) Epoch 16, batch 1050, loss[loss=0.1861, simple_loss=0.2939, pruned_loss=0.03915, over 17112.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2532, pruned_loss=0.04353, over 3305287.66 frames. ], batch size: 47, lr: 4.30e-03, grad_scale: 4.0 2023-04-30 06:56:24,682 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.592e+02 2.258e+02 2.836e+02 3.297e+02 6.244e+02, threshold=5.672e+02, percent-clipped=2.0 2023-04-30 06:56:55,498 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=153339.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:57:12,820 INFO [train.py:904] (4/8) Epoch 16, batch 1100, loss[loss=0.1772, simple_loss=0.2676, pruned_loss=0.0434, over 16640.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2523, pruned_loss=0.04291, over 3305999.32 frames. ], batch size: 57, lr: 4.30e-03, grad_scale: 4.0 2023-04-30 06:57:19,245 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9109, 4.9639, 5.4141, 5.4026, 5.3764, 5.0582, 4.9872, 4.7912], device='cuda:4'), covar=tensor([0.0317, 0.0573, 0.0375, 0.0428, 0.0457, 0.0387, 0.0827, 0.0400], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0407, 0.0395, 0.0380, 0.0447, 0.0420, 0.0511, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 06:57:20,994 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0286, 4.3327, 4.6016, 4.5913, 4.5854, 4.2843, 3.9666, 4.2285], device='cuda:4'), covar=tensor([0.0644, 0.0879, 0.0593, 0.0651, 0.0739, 0.0669, 0.1400, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0375, 0.0407, 0.0395, 0.0380, 0.0447, 0.0420, 0.0511, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 06:57:54,331 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153382.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:58:21,562 INFO [train.py:904] (4/8) Epoch 16, batch 1150, loss[loss=0.1851, simple_loss=0.2564, pruned_loss=0.05686, over 16762.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.252, pruned_loss=0.0424, over 3307188.71 frames. ], batch size: 124, lr: 4.30e-03, grad_scale: 4.0 2023-04-30 06:58:42,013 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 2.113e+02 2.608e+02 3.086e+02 6.684e+02, threshold=5.217e+02, percent-clipped=1.0 2023-04-30 06:58:46,840 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=153421.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:58:58,552 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=153430.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 06:59:27,881 INFO [train.py:904] (4/8) Epoch 16, batch 1200, loss[loss=0.1699, simple_loss=0.2579, pruned_loss=0.04097, over 17059.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2519, pruned_loss=0.04199, over 3309559.51 frames. ], batch size: 55, lr: 4.30e-03, grad_scale: 8.0 2023-04-30 06:59:49,483 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1162, 5.1069, 4.8782, 4.2921, 4.8849, 2.0677, 4.6556, 4.8421], device='cuda:4'), covar=tensor([0.0090, 0.0086, 0.0183, 0.0429, 0.0110, 0.2443, 0.0133, 0.0194], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0138, 0.0185, 0.0169, 0.0158, 0.0198, 0.0173, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 06:59:53,488 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9455, 4.0583, 3.0513, 2.4358, 2.8023, 2.4891, 4.2124, 3.6360], device='cuda:4'), covar=tensor([0.2319, 0.0590, 0.1585, 0.2352, 0.2314, 0.1836, 0.0458, 0.1223], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0263, 0.0293, 0.0290, 0.0282, 0.0237, 0.0278, 0.0315], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 06:59:57,535 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=153473.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:00:10,678 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=153482.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:00:27,333 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.23 vs. limit=5.0 2023-04-30 07:00:37,161 INFO [train.py:904] (4/8) Epoch 16, batch 1250, loss[loss=0.1944, simple_loss=0.2622, pruned_loss=0.06335, over 16776.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2523, pruned_loss=0.04304, over 3298553.19 frames. ], batch size: 124, lr: 4.30e-03, grad_scale: 8.0 2023-04-30 07:00:57,394 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.717e+02 2.233e+02 2.659e+02 3.231e+02 7.143e+02, threshold=5.319e+02, percent-clipped=2.0 2023-04-30 07:01:03,222 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5509, 3.9187, 4.3878, 2.2514, 3.5074, 2.5458, 4.2227, 4.1140], device='cuda:4'), covar=tensor([0.0231, 0.0744, 0.0354, 0.1836, 0.0655, 0.0940, 0.0459, 0.0829], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0155, 0.0161, 0.0149, 0.0140, 0.0127, 0.0142, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 07:01:43,331 INFO [train.py:904] (4/8) Epoch 16, batch 1300, loss[loss=0.157, simple_loss=0.2333, pruned_loss=0.0403, over 15507.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2526, pruned_loss=0.04337, over 3301478.81 frames. ], batch size: 190, lr: 4.30e-03, grad_scale: 8.0 2023-04-30 07:01:57,134 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-30 07:02:52,652 INFO [train.py:904] (4/8) Epoch 16, batch 1350, loss[loss=0.1463, simple_loss=0.2339, pruned_loss=0.02936, over 17003.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2529, pruned_loss=0.04279, over 3312284.02 frames. ], batch size: 41, lr: 4.30e-03, grad_scale: 8.0 2023-04-30 07:03:12,843 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.319e+02 2.697e+02 3.051e+02 7.500e+02, threshold=5.394e+02, percent-clipped=1.0 2023-04-30 07:03:36,988 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=153634.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:03:53,973 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-30 07:04:02,081 INFO [train.py:904] (4/8) Epoch 16, batch 1400, loss[loss=0.175, simple_loss=0.2712, pruned_loss=0.03939, over 17053.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2537, pruned_loss=0.04309, over 3321094.80 frames. ], batch size: 55, lr: 4.30e-03, grad_scale: 8.0 2023-04-30 07:04:37,655 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.51 vs. limit=5.0 2023-04-30 07:05:12,109 INFO [train.py:904] (4/8) Epoch 16, batch 1450, loss[loss=0.1719, simple_loss=0.2688, pruned_loss=0.03752, over 17032.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2524, pruned_loss=0.04256, over 3327169.60 frames. ], batch size: 55, lr: 4.30e-03, grad_scale: 4.0 2023-04-30 07:05:34,100 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 2.274e+02 2.601e+02 3.378e+02 7.170e+02, threshold=5.202e+02, percent-clipped=2.0 2023-04-30 07:06:22,460 INFO [train.py:904] (4/8) Epoch 16, batch 1500, loss[loss=0.1632, simple_loss=0.2405, pruned_loss=0.04291, over 16788.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2523, pruned_loss=0.04255, over 3335764.43 frames. ], batch size: 39, lr: 4.30e-03, grad_scale: 4.0 2023-04-30 07:06:31,709 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-30 07:06:50,946 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153773.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:06:56,029 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=153777.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:07:30,670 INFO [train.py:904] (4/8) Epoch 16, batch 1550, loss[loss=0.1629, simple_loss=0.2549, pruned_loss=0.03546, over 17170.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2535, pruned_loss=0.04402, over 3342880.02 frames. ], batch size: 46, lr: 4.30e-03, grad_scale: 4.0 2023-04-30 07:07:53,743 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.594e+02 2.189e+02 2.739e+02 3.070e+02 6.481e+02, threshold=5.478e+02, percent-clipped=4.0 2023-04-30 07:07:58,215 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=153821.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:08:24,175 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3285, 5.1373, 5.1152, 4.6529, 4.7680, 5.1621, 5.1938, 4.7605], device='cuda:4'), covar=tensor([0.0561, 0.0472, 0.0294, 0.0341, 0.1074, 0.0418, 0.0273, 0.0753], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0388, 0.0331, 0.0318, 0.0347, 0.0366, 0.0226, 0.0396], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 07:08:40,378 INFO [train.py:904] (4/8) Epoch 16, batch 1600, loss[loss=0.2015, simple_loss=0.2871, pruned_loss=0.05794, over 15520.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2563, pruned_loss=0.04522, over 3329135.68 frames. ], batch size: 191, lr: 4.30e-03, grad_scale: 8.0 2023-04-30 07:09:33,533 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.17 vs. limit=5.0 2023-04-30 07:09:47,765 INFO [train.py:904] (4/8) Epoch 16, batch 1650, loss[loss=0.2152, simple_loss=0.2964, pruned_loss=0.06704, over 15425.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2579, pruned_loss=0.04566, over 3330801.30 frames. ], batch size: 190, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:10:09,057 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.581e+02 2.549e+02 2.974e+02 3.703e+02 7.236e+02, threshold=5.949e+02, percent-clipped=5.0 2023-04-30 07:10:16,010 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=153923.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:10:32,424 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153934.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:10:56,056 INFO [train.py:904] (4/8) Epoch 16, batch 1700, loss[loss=0.1867, simple_loss=0.2638, pruned_loss=0.05473, over 16680.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2597, pruned_loss=0.0463, over 3322898.53 frames. ], batch size: 134, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:11:38,410 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=153982.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:11:40,939 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=153984.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:12:09,295 INFO [train.py:904] (4/8) Epoch 16, batch 1750, loss[loss=0.1869, simple_loss=0.2701, pruned_loss=0.05186, over 16444.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2608, pruned_loss=0.04648, over 3323248.90 frames. ], batch size: 75, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:12:33,136 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.607e+02 2.261e+02 2.728e+02 3.333e+02 9.430e+02, threshold=5.456e+02, percent-clipped=2.0 2023-04-30 07:12:46,543 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-30 07:13:15,168 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154049.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 07:13:18,923 INFO [train.py:904] (4/8) Epoch 16, batch 1800, loss[loss=0.199, simple_loss=0.2804, pruned_loss=0.05884, over 16761.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2625, pruned_loss=0.04715, over 3314577.90 frames. ], batch size: 124, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:13:27,771 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8006, 3.7198, 4.2526, 2.0515, 4.3287, 4.4082, 3.1666, 3.3116], device='cuda:4'), covar=tensor([0.0737, 0.0245, 0.0190, 0.1120, 0.0084, 0.0185, 0.0411, 0.0408], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0104, 0.0092, 0.0137, 0.0073, 0.0117, 0.0124, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 07:13:53,179 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154076.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:13:54,191 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154077.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:14:03,034 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154084.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:14:17,334 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 07:14:20,075 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0776, 5.0799, 5.5715, 5.5452, 5.5647, 5.1626, 5.1292, 4.9244], device='cuda:4'), covar=tensor([0.0294, 0.0488, 0.0342, 0.0375, 0.0432, 0.0319, 0.0859, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0384, 0.0414, 0.0403, 0.0384, 0.0455, 0.0430, 0.0524, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 07:14:28,054 INFO [train.py:904] (4/8) Epoch 16, batch 1850, loss[loss=0.1894, simple_loss=0.2649, pruned_loss=0.057, over 16662.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2626, pruned_loss=0.04676, over 3310675.98 frames. ], batch size: 134, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:14:38,647 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154110.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 07:14:50,217 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.501e+02 2.254e+02 2.561e+02 3.083e+02 7.849e+02, threshold=5.121e+02, percent-clipped=3.0 2023-04-30 07:14:59,481 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=154125.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:15:15,718 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154137.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 07:15:25,821 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154145.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:15:33,361 INFO [train.py:904] (4/8) Epoch 16, batch 1900, loss[loss=0.1983, simple_loss=0.2735, pruned_loss=0.06154, over 16505.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.262, pruned_loss=0.0463, over 3304804.26 frames. ], batch size: 68, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:15:36,738 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-30 07:16:44,644 INFO [train.py:904] (4/8) Epoch 16, batch 1950, loss[loss=0.1943, simple_loss=0.2737, pruned_loss=0.0574, over 16736.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2622, pruned_loss=0.04581, over 3304761.05 frames. ], batch size: 134, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:17:04,523 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.551e+02 2.235e+02 2.551e+02 3.026e+02 6.742e+02, threshold=5.103e+02, percent-clipped=2.0 2023-04-30 07:17:10,526 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154221.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:17:28,781 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7058, 4.7851, 4.9731, 4.7887, 4.7963, 5.4112, 4.8779, 4.5300], device='cuda:4'), covar=tensor([0.1584, 0.1917, 0.2284, 0.1947, 0.2692, 0.1017, 0.1590, 0.2543], device='cuda:4'), in_proj_covar=tensor([0.0395, 0.0564, 0.0611, 0.0472, 0.0631, 0.0645, 0.0480, 0.0626], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 07:17:45,989 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6315, 4.6424, 5.0845, 5.0844, 5.1273, 4.7481, 4.7020, 4.5228], device='cuda:4'), covar=tensor([0.0345, 0.0631, 0.0409, 0.0399, 0.0427, 0.0416, 0.0918, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0385, 0.0414, 0.0402, 0.0384, 0.0453, 0.0431, 0.0526, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 07:17:51,475 INFO [train.py:904] (4/8) Epoch 16, batch 2000, loss[loss=0.1592, simple_loss=0.2592, pruned_loss=0.02958, over 17140.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2613, pruned_loss=0.04557, over 3299922.78 frames. ], batch size: 48, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:18:06,118 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7672, 4.5399, 4.7823, 4.9832, 5.1314, 4.5635, 5.1117, 5.1397], device='cuda:4'), covar=tensor([0.1577, 0.1209, 0.1666, 0.0719, 0.0547, 0.0982, 0.0677, 0.0594], device='cuda:4'), in_proj_covar=tensor([0.0617, 0.0765, 0.0907, 0.0773, 0.0580, 0.0606, 0.0618, 0.0716], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:18:27,772 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154279.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:18:33,379 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154282.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 07:18:38,753 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154286.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:18:44,544 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154291.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:18:59,381 INFO [train.py:904] (4/8) Epoch 16, batch 2050, loss[loss=0.1577, simple_loss=0.248, pruned_loss=0.03367, over 17199.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2616, pruned_loss=0.04607, over 3302953.11 frames. ], batch size: 44, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:19:19,611 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.703e+02 2.241e+02 2.674e+02 3.116e+02 4.900e+02, threshold=5.347e+02, percent-clipped=0.0 2023-04-30 07:19:46,291 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154337.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 07:20:01,031 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154347.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:20:06,432 INFO [train.py:904] (4/8) Epoch 16, batch 2100, loss[loss=0.1755, simple_loss=0.2548, pruned_loss=0.04809, over 16786.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2621, pruned_loss=0.04627, over 3314058.58 frames. ], batch size: 39, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:20:06,837 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154352.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:21:09,648 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154398.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 07:21:14,694 INFO [train.py:904] (4/8) Epoch 16, batch 2150, loss[loss=0.1645, simple_loss=0.2609, pruned_loss=0.03401, over 17086.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2628, pruned_loss=0.0467, over 3316765.18 frames. ], batch size: 50, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:21:18,545 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154405.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 07:21:19,119 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-30 07:21:36,517 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.770e+02 2.339e+02 2.706e+02 3.339e+02 5.041e+02, threshold=5.411e+02, percent-clipped=0.0 2023-04-30 07:21:57,838 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154432.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 07:22:07,246 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154440.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:22:25,078 INFO [train.py:904] (4/8) Epoch 16, batch 2200, loss[loss=0.1615, simple_loss=0.2535, pruned_loss=0.03477, over 17186.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2623, pruned_loss=0.04631, over 3323929.45 frames. ], batch size: 46, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:22:55,692 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154474.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:23:32,242 INFO [train.py:904] (4/8) Epoch 16, batch 2250, loss[loss=0.19, simple_loss=0.2676, pruned_loss=0.05618, over 16349.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2616, pruned_loss=0.04603, over 3322940.22 frames. ], batch size: 146, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:23:55,133 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.236e+02 2.597e+02 3.012e+02 8.725e+02, threshold=5.194e+02, percent-clipped=2.0 2023-04-30 07:24:14,244 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9531, 2.7156, 2.7963, 2.1679, 2.6640, 2.2183, 2.6625, 2.8872], device='cuda:4'), covar=tensor([0.0294, 0.0788, 0.0488, 0.1619, 0.0766, 0.0834, 0.0551, 0.0726], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0157, 0.0163, 0.0149, 0.0141, 0.0127, 0.0142, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 07:24:18,875 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154535.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:24:40,019 INFO [train.py:904] (4/8) Epoch 16, batch 2300, loss[loss=0.195, simple_loss=0.2701, pruned_loss=0.05994, over 16815.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.262, pruned_loss=0.0459, over 3320622.53 frames. ], batch size: 102, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:25:16,758 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154577.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 07:25:20,107 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154579.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:25:51,246 INFO [train.py:904] (4/8) Epoch 16, batch 2350, loss[loss=0.1737, simple_loss=0.2697, pruned_loss=0.03881, over 17097.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2631, pruned_loss=0.04637, over 3320608.39 frames. ], batch size: 49, lr: 4.29e-03, grad_scale: 8.0 2023-04-30 07:25:51,694 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2592, 3.2379, 1.9485, 3.4221, 2.5280, 3.4452, 2.1321, 2.6654], device='cuda:4'), covar=tensor([0.0260, 0.0411, 0.1556, 0.0309, 0.0780, 0.0657, 0.1349, 0.0665], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0176, 0.0195, 0.0155, 0.0175, 0.0218, 0.0204, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 07:26:11,866 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.387e+02 2.266e+02 2.563e+02 3.089e+02 4.962e+02, threshold=5.126e+02, percent-clipped=0.0 2023-04-30 07:26:25,330 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=154627.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:26:26,679 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154628.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:26:28,930 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3318, 4.1996, 4.2052, 3.9929, 4.0076, 4.2742, 3.9918, 4.0700], device='cuda:4'), covar=tensor([0.0587, 0.0674, 0.0284, 0.0302, 0.0750, 0.0469, 0.0686, 0.0603], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0393, 0.0335, 0.0323, 0.0352, 0.0371, 0.0230, 0.0402], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 07:26:45,857 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154642.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:26:52,321 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154647.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:26:57,937 INFO [train.py:904] (4/8) Epoch 16, batch 2400, loss[loss=0.1968, simple_loss=0.2795, pruned_loss=0.05706, over 16487.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2648, pruned_loss=0.04761, over 3315769.60 frames. ], batch size: 75, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:27:39,670 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5930, 4.5619, 4.5261, 3.9960, 4.5752, 1.7743, 4.3420, 4.2468], device='cuda:4'), covar=tensor([0.0119, 0.0094, 0.0176, 0.0327, 0.0091, 0.2568, 0.0134, 0.0198], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0142, 0.0190, 0.0174, 0.0162, 0.0201, 0.0178, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:27:51,028 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154689.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:27:55,459 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154693.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 07:28:04,568 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7586, 2.3659, 2.3391, 4.5444, 2.3428, 2.8304, 2.3748, 2.5225], device='cuda:4'), covar=tensor([0.0979, 0.3563, 0.2846, 0.0416, 0.4015, 0.2468, 0.3411, 0.3472], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0423, 0.0355, 0.0326, 0.0427, 0.0488, 0.0393, 0.0497], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:28:06,455 INFO [train.py:904] (4/8) Epoch 16, batch 2450, loss[loss=0.1876, simple_loss=0.2685, pruned_loss=0.05334, over 16747.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2657, pruned_loss=0.04732, over 3323348.31 frames. ], batch size: 102, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:28:12,093 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154705.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 07:28:28,917 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.609e+02 2.204e+02 2.505e+02 3.030e+02 6.585e+02, threshold=5.010e+02, percent-clipped=3.0 2023-04-30 07:28:31,783 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154720.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:28:50,202 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154732.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:28:59,844 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154740.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:29:16,066 INFO [train.py:904] (4/8) Epoch 16, batch 2500, loss[loss=0.1865, simple_loss=0.2685, pruned_loss=0.05224, over 16839.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2649, pruned_loss=0.04674, over 3325236.20 frames. ], batch size: 102, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:29:18,086 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=154753.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 07:29:18,140 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6669, 6.1031, 5.4115, 6.0628, 5.5037, 5.0954, 5.5236, 6.1217], device='cuda:4'), covar=tensor([0.2295, 0.1264, 0.2420, 0.1043, 0.1427, 0.1395, 0.1979, 0.1643], device='cuda:4'), in_proj_covar=tensor([0.0639, 0.0790, 0.0642, 0.0569, 0.0497, 0.0506, 0.0658, 0.0605], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:29:53,783 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=154780.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:29:55,043 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154781.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:30:05,086 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=154788.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:30:24,141 INFO [train.py:904] (4/8) Epoch 16, batch 2550, loss[loss=0.1716, simple_loss=0.2681, pruned_loss=0.03759, over 17128.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2656, pruned_loss=0.0469, over 3311957.90 frames. ], batch size: 48, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:30:47,016 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.496e+02 2.136e+02 2.491e+02 3.165e+02 6.684e+02, threshold=4.981e+02, percent-clipped=5.0 2023-04-30 07:31:02,555 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154830.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:31:18,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2403, 2.0444, 2.2159, 3.8839, 2.1312, 2.4121, 2.0986, 2.2114], device='cuda:4'), covar=tensor([0.1244, 0.3789, 0.2814, 0.0547, 0.3762, 0.2497, 0.3879, 0.3167], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0423, 0.0353, 0.0325, 0.0426, 0.0488, 0.0392, 0.0495], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:31:32,699 INFO [train.py:904] (4/8) Epoch 16, batch 2600, loss[loss=0.174, simple_loss=0.2647, pruned_loss=0.04163, over 16411.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.265, pruned_loss=0.04637, over 3311599.38 frames. ], batch size: 68, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:32:02,194 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3818, 5.3773, 5.2186, 4.7143, 5.3097, 2.2646, 5.0745, 5.1806], device='cuda:4'), covar=tensor([0.0081, 0.0075, 0.0161, 0.0333, 0.0084, 0.2302, 0.0113, 0.0160], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0143, 0.0191, 0.0175, 0.0163, 0.0203, 0.0179, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:32:07,943 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154877.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 07:32:40,408 INFO [train.py:904] (4/8) Epoch 16, batch 2650, loss[loss=0.1831, simple_loss=0.2702, pruned_loss=0.04797, over 16523.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2648, pruned_loss=0.04597, over 3324769.92 frames. ], batch size: 75, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:33:01,394 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.567e+02 2.264e+02 2.671e+02 3.159e+02 6.185e+02, threshold=5.341e+02, percent-clipped=3.0 2023-04-30 07:33:12,796 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=154925.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:33:13,265 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-30 07:33:35,888 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154942.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:33:42,402 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1136, 5.5815, 5.6766, 5.4467, 5.3973, 6.0520, 5.5438, 5.2611], device='cuda:4'), covar=tensor([0.0869, 0.1934, 0.2309, 0.1910, 0.2658, 0.0865, 0.1488, 0.2174], device='cuda:4'), in_proj_covar=tensor([0.0399, 0.0567, 0.0616, 0.0479, 0.0637, 0.0649, 0.0484, 0.0633], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 07:33:43,508 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154947.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:33:49,091 INFO [train.py:904] (4/8) Epoch 16, batch 2700, loss[loss=0.1795, simple_loss=0.2794, pruned_loss=0.03978, over 16775.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2651, pruned_loss=0.04551, over 3319908.63 frames. ], batch size: 62, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:34:21,662 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154976.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:34:32,699 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154984.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:34:41,398 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=154990.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:34:45,815 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154993.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 07:34:47,944 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=154995.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:34:56,796 INFO [train.py:904] (4/8) Epoch 16, batch 2750, loss[loss=0.161, simple_loss=0.2505, pruned_loss=0.0357, over 16800.00 frames. ], tot_loss[loss=0.178, simple_loss=0.265, pruned_loss=0.04546, over 3325205.55 frames. ], batch size: 42, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:35:18,259 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.250e+02 2.149e+02 2.425e+02 2.902e+02 6.172e+02, threshold=4.850e+02, percent-clipped=1.0 2023-04-30 07:35:46,127 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155037.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:35:50,509 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=155041.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 07:36:04,765 INFO [train.py:904] (4/8) Epoch 16, batch 2800, loss[loss=0.1675, simple_loss=0.2613, pruned_loss=0.03691, over 16095.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2649, pruned_loss=0.04526, over 3327923.21 frames. ], batch size: 35, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:36:28,368 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155069.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:36:38,304 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155076.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:37:15,724 INFO [train.py:904] (4/8) Epoch 16, batch 2850, loss[loss=0.1565, simple_loss=0.2513, pruned_loss=0.03084, over 17290.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2644, pruned_loss=0.0452, over 3333474.31 frames. ], batch size: 52, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:37:27,136 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2664, 3.6642, 3.8010, 2.2468, 3.0113, 2.4901, 3.7479, 3.8174], device='cuda:4'), covar=tensor([0.0289, 0.0773, 0.0477, 0.1734, 0.0791, 0.0908, 0.0585, 0.0938], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0155, 0.0161, 0.0147, 0.0139, 0.0125, 0.0140, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 07:37:36,420 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.487e+02 2.238e+02 2.628e+02 3.275e+02 9.061e+02, threshold=5.256e+02, percent-clipped=3.0 2023-04-30 07:37:54,446 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155130.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:37:54,489 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155130.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:38:08,194 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1678, 5.1022, 5.6644, 5.6074, 5.6673, 5.2381, 5.1622, 4.9922], device='cuda:4'), covar=tensor([0.0325, 0.0575, 0.0290, 0.0398, 0.0524, 0.0384, 0.1082, 0.0467], device='cuda:4'), in_proj_covar=tensor([0.0391, 0.0423, 0.0411, 0.0389, 0.0462, 0.0437, 0.0533, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 07:38:23,945 INFO [train.py:904] (4/8) Epoch 16, batch 2900, loss[loss=0.1748, simple_loss=0.2703, pruned_loss=0.03963, over 17296.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2625, pruned_loss=0.04527, over 3328433.86 frames. ], batch size: 52, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:39:01,148 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=155178.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:39:06,992 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1428, 5.6951, 5.8381, 5.5383, 5.6672, 6.1901, 5.7217, 5.3764], device='cuda:4'), covar=tensor([0.0856, 0.1656, 0.2313, 0.1962, 0.2424, 0.0860, 0.1394, 0.2424], device='cuda:4'), in_proj_covar=tensor([0.0400, 0.0572, 0.0621, 0.0484, 0.0640, 0.0649, 0.0487, 0.0638], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 07:39:19,296 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3196, 4.3325, 4.7311, 4.7066, 4.7296, 4.3848, 4.4065, 4.2920], device='cuda:4'), covar=tensor([0.0396, 0.0625, 0.0387, 0.0428, 0.0517, 0.0466, 0.0907, 0.0666], device='cuda:4'), in_proj_covar=tensor([0.0389, 0.0420, 0.0409, 0.0386, 0.0459, 0.0435, 0.0531, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 07:39:33,217 INFO [train.py:904] (4/8) Epoch 16, batch 2950, loss[loss=0.1875, simple_loss=0.2726, pruned_loss=0.05125, over 16687.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2622, pruned_loss=0.04613, over 3321279.36 frames. ], batch size: 62, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:39:34,960 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5720, 2.2034, 2.4378, 4.5242, 2.2296, 2.5571, 2.3453, 2.3837], device='cuda:4'), covar=tensor([0.1326, 0.4181, 0.2754, 0.0489, 0.4675, 0.3115, 0.3658, 0.4212], device='cuda:4'), in_proj_covar=tensor([0.0384, 0.0421, 0.0351, 0.0323, 0.0424, 0.0485, 0.0390, 0.0493], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:39:54,113 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.648e+02 2.372e+02 2.734e+02 3.300e+02 7.430e+02, threshold=5.468e+02, percent-clipped=3.0 2023-04-30 07:40:40,766 INFO [train.py:904] (4/8) Epoch 16, batch 3000, loss[loss=0.1977, simple_loss=0.2888, pruned_loss=0.05328, over 17077.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2619, pruned_loss=0.04612, over 3326118.22 frames. ], batch size: 55, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:40:40,767 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 07:40:49,854 INFO [train.py:938] (4/8) Epoch 16, validation: loss=0.1368, simple_loss=0.2429, pruned_loss=0.01541, over 944034.00 frames. 2023-04-30 07:40:49,855 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 07:41:34,767 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155284.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:41:59,653 INFO [train.py:904] (4/8) Epoch 16, batch 3050, loss[loss=0.1467, simple_loss=0.2379, pruned_loss=0.02776, over 17185.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2614, pruned_loss=0.04581, over 3332382.73 frames. ], batch size: 46, lr: 4.28e-03, grad_scale: 8.0 2023-04-30 07:42:21,041 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 2.286e+02 2.843e+02 3.408e+02 8.038e+02, threshold=5.686e+02, percent-clipped=2.0 2023-04-30 07:42:42,354 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=155332.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:42:42,368 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155332.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:42:48,335 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.95 vs. limit=5.0 2023-04-30 07:43:10,099 INFO [train.py:904] (4/8) Epoch 16, batch 3100, loss[loss=0.1853, simple_loss=0.2748, pruned_loss=0.04787, over 16466.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2617, pruned_loss=0.04611, over 3332134.78 frames. ], batch size: 68, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:43:42,862 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155376.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:44:00,386 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7068, 1.8305, 2.2598, 2.5666, 2.6843, 2.6450, 1.7890, 2.8202], device='cuda:4'), covar=tensor([0.0158, 0.0447, 0.0293, 0.0244, 0.0239, 0.0241, 0.0468, 0.0165], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0186, 0.0172, 0.0175, 0.0187, 0.0143, 0.0185, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:44:16,195 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-30 07:44:17,694 INFO [train.py:904] (4/8) Epoch 16, batch 3150, loss[loss=0.1593, simple_loss=0.2577, pruned_loss=0.03048, over 17040.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2615, pruned_loss=0.04654, over 3324761.62 frames. ], batch size: 50, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:44:39,886 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.743e+02 2.373e+02 2.704e+02 3.278e+02 4.749e+02, threshold=5.408e+02, percent-clipped=0.0 2023-04-30 07:44:47,894 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=155424.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:44:49,093 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155425.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:45:27,253 INFO [train.py:904] (4/8) Epoch 16, batch 3200, loss[loss=0.1602, simple_loss=0.2444, pruned_loss=0.03794, over 16648.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2601, pruned_loss=0.04576, over 3329068.84 frames. ], batch size: 89, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:45:34,212 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155457.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:45:52,695 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-30 07:46:23,740 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-30 07:46:36,135 INFO [train.py:904] (4/8) Epoch 16, batch 3250, loss[loss=0.203, simple_loss=0.2817, pruned_loss=0.0622, over 16336.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2597, pruned_loss=0.04531, over 3334735.40 frames. ], batch size: 165, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:46:57,821 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9756, 2.3863, 1.8130, 2.1005, 2.7735, 2.5771, 3.0103, 2.9709], device='cuda:4'), covar=tensor([0.0212, 0.0480, 0.0665, 0.0573, 0.0304, 0.0452, 0.0258, 0.0339], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0229, 0.0220, 0.0222, 0.0230, 0.0231, 0.0237, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:46:58,461 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.619e+02 2.206e+02 2.546e+02 3.042e+02 5.764e+02, threshold=5.092e+02, percent-clipped=1.0 2023-04-30 07:46:58,939 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155518.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:47:34,935 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-30 07:47:40,182 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.45 vs. limit=5.0 2023-04-30 07:47:45,922 INFO [train.py:904] (4/8) Epoch 16, batch 3300, loss[loss=0.1808, simple_loss=0.2668, pruned_loss=0.04745, over 16837.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2606, pruned_loss=0.04532, over 3339581.34 frames. ], batch size: 102, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:48:40,260 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-30 07:48:56,952 INFO [train.py:904] (4/8) Epoch 16, batch 3350, loss[loss=0.1815, simple_loss=0.2708, pruned_loss=0.04604, over 15529.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2613, pruned_loss=0.04521, over 3341289.15 frames. ], batch size: 190, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:49:01,189 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155605.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:49:19,482 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.838e+02 2.398e+02 2.739e+02 3.162e+02 6.710e+02, threshold=5.477e+02, percent-clipped=3.0 2023-04-30 07:49:39,756 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155632.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:50:08,457 INFO [train.py:904] (4/8) Epoch 16, batch 3400, loss[loss=0.1812, simple_loss=0.2568, pruned_loss=0.05283, over 16788.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2615, pruned_loss=0.04514, over 3339301.31 frames. ], batch size: 124, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:50:16,630 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155657.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:50:27,968 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155666.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:50:49,342 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=155680.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:50:56,071 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-30 07:51:19,270 INFO [train.py:904] (4/8) Epoch 16, batch 3450, loss[loss=0.1946, simple_loss=0.2674, pruned_loss=0.0609, over 16724.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2598, pruned_loss=0.04474, over 3337297.53 frames. ], batch size: 124, lr: 4.27e-03, grad_scale: 16.0 2023-04-30 07:51:41,374 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.454e+02 2.296e+02 2.647e+02 3.295e+02 8.789e+02, threshold=5.295e+02, percent-clipped=1.0 2023-04-30 07:51:41,848 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155718.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:51:46,816 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4036, 2.2643, 1.9033, 2.0355, 2.5671, 2.3271, 2.5487, 2.7246], device='cuda:4'), covar=tensor([0.0203, 0.0329, 0.0427, 0.0402, 0.0190, 0.0288, 0.0201, 0.0221], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0229, 0.0220, 0.0221, 0.0230, 0.0231, 0.0237, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 07:51:52,548 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155725.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:51:54,074 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.20 vs. limit=5.0 2023-04-30 07:51:54,401 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-30 07:52:28,775 INFO [train.py:904] (4/8) Epoch 16, batch 3500, loss[loss=0.1906, simple_loss=0.2647, pruned_loss=0.05824, over 16872.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2591, pruned_loss=0.04439, over 3336128.30 frames. ], batch size: 109, lr: 4.27e-03, grad_scale: 16.0 2023-04-30 07:52:43,885 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-30 07:52:58,254 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=155773.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:52:58,400 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3062, 3.4296, 3.6877, 2.5264, 3.3883, 3.7240, 3.5131, 1.9623], device='cuda:4'), covar=tensor([0.0494, 0.0167, 0.0062, 0.0361, 0.0094, 0.0108, 0.0096, 0.0494], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0077, 0.0076, 0.0131, 0.0089, 0.0099, 0.0088, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 07:53:14,812 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155784.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:53:40,573 INFO [train.py:904] (4/8) Epoch 16, batch 3550, loss[loss=0.1814, simple_loss=0.2731, pruned_loss=0.04481, over 17043.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.258, pruned_loss=0.04413, over 3337789.31 frames. ], batch size: 55, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:53:56,154 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155813.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:54:04,791 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.461e+02 2.310e+02 2.739e+02 3.224e+02 5.856e+02, threshold=5.477e+02, percent-clipped=2.0 2023-04-30 07:54:42,450 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155845.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 07:54:52,030 INFO [train.py:904] (4/8) Epoch 16, batch 3600, loss[loss=0.1772, simple_loss=0.2499, pruned_loss=0.05221, over 16710.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2563, pruned_loss=0.04349, over 3339120.20 frames. ], batch size: 124, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:56:03,880 INFO [train.py:904] (4/8) Epoch 16, batch 3650, loss[loss=0.1863, simple_loss=0.2578, pruned_loss=0.05737, over 16346.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.256, pruned_loss=0.0438, over 3325891.80 frames. ], batch size: 165, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:56:28,564 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.735e+02 2.258e+02 2.651e+02 3.350e+02 6.145e+02, threshold=5.302e+02, percent-clipped=1.0 2023-04-30 07:57:17,739 INFO [train.py:904] (4/8) Epoch 16, batch 3700, loss[loss=0.1887, simple_loss=0.2596, pruned_loss=0.05891, over 16691.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2551, pruned_loss=0.04531, over 3299972.52 frames. ], batch size: 134, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:57:32,537 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155961.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:57:35,159 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5645, 3.6985, 2.2945, 3.8461, 2.8869, 3.8466, 2.2184, 2.9364], device='cuda:4'), covar=tensor([0.0243, 0.0312, 0.1358, 0.0242, 0.0703, 0.0541, 0.1301, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0175, 0.0193, 0.0156, 0.0173, 0.0218, 0.0203, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 07:57:39,656 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6472, 3.7871, 2.3394, 3.9421, 2.9896, 3.8786, 2.3472, 3.0615], device='cuda:4'), covar=tensor([0.0207, 0.0334, 0.1284, 0.0221, 0.0609, 0.0650, 0.1196, 0.0511], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0175, 0.0193, 0.0156, 0.0173, 0.0218, 0.0203, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 07:58:36,927 INFO [train.py:904] (4/8) Epoch 16, batch 3750, loss[loss=0.1918, simple_loss=0.2846, pruned_loss=0.04951, over 17103.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2567, pruned_loss=0.04699, over 3283573.86 frames. ], batch size: 47, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 07:58:53,071 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156013.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 07:59:02,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.696e+02 2.313e+02 2.777e+02 3.166e+02 7.362e+02, threshold=5.555e+02, percent-clipped=3.0 2023-04-30 07:59:47,759 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4453, 3.3393, 3.6400, 1.7231, 3.6762, 3.7758, 3.1516, 2.8023], device='cuda:4'), covar=tensor([0.0782, 0.0231, 0.0166, 0.1283, 0.0118, 0.0164, 0.0324, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0106, 0.0093, 0.0138, 0.0074, 0.0120, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 07:59:51,245 INFO [train.py:904] (4/8) Epoch 16, batch 3800, loss[loss=0.1762, simple_loss=0.2501, pruned_loss=0.05113, over 16845.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2585, pruned_loss=0.04856, over 3269646.07 frames. ], batch size: 102, lr: 4.27e-03, grad_scale: 8.0 2023-04-30 08:00:17,864 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156070.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:00:19,348 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-30 08:01:03,106 INFO [train.py:904] (4/8) Epoch 16, batch 3850, loss[loss=0.1722, simple_loss=0.2498, pruned_loss=0.04728, over 16758.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.258, pruned_loss=0.04871, over 3279939.99 frames. ], batch size: 83, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:01:15,204 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-30 08:01:20,118 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156113.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:01:28,704 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.683e+02 2.283e+02 2.734e+02 3.408e+02 7.535e+02, threshold=5.468e+02, percent-clipped=3.0 2023-04-30 08:01:47,476 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156131.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 08:01:56,306 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-30 08:01:59,099 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156140.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 08:02:16,231 INFO [train.py:904] (4/8) Epoch 16, batch 3900, loss[loss=0.1679, simple_loss=0.2457, pruned_loss=0.04504, over 16392.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2582, pruned_loss=0.04939, over 3288102.94 frames. ], batch size: 75, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:02:29,442 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=156161.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:02:29,622 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6345, 3.4798, 3.8048, 2.0170, 3.9006, 3.9091, 3.1454, 2.9099], device='cuda:4'), covar=tensor([0.0689, 0.0236, 0.0143, 0.1131, 0.0074, 0.0166, 0.0345, 0.0434], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0107, 0.0093, 0.0139, 0.0075, 0.0120, 0.0125, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 08:02:32,158 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4161, 2.2307, 2.4358, 4.2359, 2.2598, 2.6426, 2.2704, 2.5032], device='cuda:4'), covar=tensor([0.1202, 0.3728, 0.2559, 0.0501, 0.3740, 0.2454, 0.3836, 0.2741], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0428, 0.0356, 0.0327, 0.0428, 0.0492, 0.0395, 0.0500], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:03:26,069 INFO [train.py:904] (4/8) Epoch 16, batch 3950, loss[loss=0.1956, simple_loss=0.2713, pruned_loss=0.06002, over 16248.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2576, pruned_loss=0.04999, over 3285612.76 frames. ], batch size: 165, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:03:47,447 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156216.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:03:50,338 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.580e+02 2.349e+02 2.934e+02 3.426e+02 7.028e+02, threshold=5.868e+02, percent-clipped=1.0 2023-04-30 08:04:28,204 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-30 08:04:37,101 INFO [train.py:904] (4/8) Epoch 16, batch 4000, loss[loss=0.1721, simple_loss=0.2475, pruned_loss=0.04833, over 16911.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.257, pruned_loss=0.0501, over 3291023.94 frames. ], batch size: 109, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:04:49,892 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156261.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:05:12,170 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156277.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:05:15,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6276, 1.6620, 2.2315, 2.5790, 2.5398, 2.9145, 1.8200, 2.8695], device='cuda:4'), covar=tensor([0.0208, 0.0471, 0.0299, 0.0283, 0.0301, 0.0165, 0.0492, 0.0107], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0186, 0.0174, 0.0177, 0.0188, 0.0144, 0.0186, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:05:31,048 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156290.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:05:49,243 INFO [train.py:904] (4/8) Epoch 16, batch 4050, loss[loss=0.1749, simple_loss=0.2606, pruned_loss=0.04458, over 16831.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2568, pruned_loss=0.04879, over 3292733.20 frames. ], batch size: 42, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:05:59,640 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=156309.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:06:04,372 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156313.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:06:14,225 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.383e+02 1.898e+02 2.203e+02 2.661e+02 4.268e+02, threshold=4.405e+02, percent-clipped=0.0 2023-04-30 08:07:01,332 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156351.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:07:01,986 INFO [train.py:904] (4/8) Epoch 16, batch 4100, loss[loss=0.1808, simple_loss=0.2728, pruned_loss=0.04443, over 16805.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2588, pruned_loss=0.0484, over 3294228.29 frames. ], batch size: 102, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:07:15,154 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=156361.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:08:15,974 INFO [train.py:904] (4/8) Epoch 16, batch 4150, loss[loss=0.2046, simple_loss=0.299, pruned_loss=0.05509, over 16894.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2659, pruned_loss=0.05084, over 3260619.41 frames. ], batch size: 96, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:08:23,503 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-30 08:08:40,341 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.507e+02 2.118e+02 2.582e+02 3.252e+02 7.284e+02, threshold=5.164e+02, percent-clipped=7.0 2023-04-30 08:08:52,136 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156426.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 08:09:13,333 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156440.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 08:09:30,688 INFO [train.py:904] (4/8) Epoch 16, batch 4200, loss[loss=0.2132, simple_loss=0.3038, pruned_loss=0.06126, over 15417.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.273, pruned_loss=0.05283, over 3234498.40 frames. ], batch size: 191, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:10:24,399 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=156488.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:10:43,947 INFO [train.py:904] (4/8) Epoch 16, batch 4250, loss[loss=0.1842, simple_loss=0.2783, pruned_loss=0.04505, over 16611.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.2767, pruned_loss=0.05286, over 3219650.67 frames. ], batch size: 62, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:10:44,433 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0896, 4.0539, 3.9698, 3.2824, 3.9918, 1.8338, 3.7613, 3.5516], device='cuda:4'), covar=tensor([0.0127, 0.0105, 0.0190, 0.0242, 0.0097, 0.2549, 0.0116, 0.0193], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0139, 0.0186, 0.0172, 0.0159, 0.0197, 0.0175, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:11:09,161 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.668e+02 2.212e+02 2.674e+02 3.089e+02 6.979e+02, threshold=5.347e+02, percent-clipped=2.0 2023-04-30 08:11:56,444 INFO [train.py:904] (4/8) Epoch 16, batch 4300, loss[loss=0.1934, simple_loss=0.2957, pruned_loss=0.04559, over 16895.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2776, pruned_loss=0.05215, over 3203965.12 frames. ], batch size: 96, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:12:10,157 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9468, 2.8823, 2.4181, 2.6667, 3.3189, 2.9448, 3.5619, 3.4560], device='cuda:4'), covar=tensor([0.0053, 0.0300, 0.0382, 0.0324, 0.0189, 0.0289, 0.0160, 0.0177], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0217, 0.0209, 0.0210, 0.0219, 0.0220, 0.0224, 0.0216], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:12:27,168 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156572.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:12:42,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8075, 3.2456, 3.1955, 2.0275, 2.9689, 3.1908, 3.0807, 1.7776], device='cuda:4'), covar=tensor([0.0493, 0.0044, 0.0051, 0.0398, 0.0088, 0.0088, 0.0087, 0.0417], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0076, 0.0077, 0.0130, 0.0089, 0.0099, 0.0088, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 08:13:09,396 INFO [train.py:904] (4/8) Epoch 16, batch 4350, loss[loss=0.2033, simple_loss=0.2899, pruned_loss=0.05834, over 16662.00 frames. ], tot_loss[loss=0.1935, simple_loss=0.2806, pruned_loss=0.05318, over 3197419.61 frames. ], batch size: 62, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:13:17,437 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0853, 4.8728, 5.1241, 5.3046, 5.4423, 4.8284, 5.4229, 5.4515], device='cuda:4'), covar=tensor([0.1455, 0.1148, 0.1226, 0.0547, 0.0411, 0.0689, 0.0467, 0.0518], device='cuda:4'), in_proj_covar=tensor([0.0591, 0.0730, 0.0859, 0.0739, 0.0554, 0.0584, 0.0587, 0.0687], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:13:19,829 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5367, 5.8466, 5.5440, 5.6189, 5.2402, 5.0384, 5.2559, 5.9855], device='cuda:4'), covar=tensor([0.1031, 0.0758, 0.0933, 0.0766, 0.0742, 0.0748, 0.0946, 0.0774], device='cuda:4'), in_proj_covar=tensor([0.0624, 0.0772, 0.0630, 0.0559, 0.0485, 0.0496, 0.0643, 0.0593], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:13:34,596 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.777e+02 2.270e+02 2.623e+02 3.147e+02 7.539e+02, threshold=5.245e+02, percent-clipped=2.0 2023-04-30 08:13:58,856 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7760, 3.2366, 3.1871, 1.9729, 3.0165, 3.1597, 3.0591, 1.8830], device='cuda:4'), covar=tensor([0.0533, 0.0044, 0.0049, 0.0424, 0.0088, 0.0100, 0.0087, 0.0412], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0075, 0.0076, 0.0131, 0.0089, 0.0099, 0.0088, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 08:14:14,560 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156646.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:14:22,117 INFO [train.py:904] (4/8) Epoch 16, batch 4400, loss[loss=0.1787, simple_loss=0.2646, pruned_loss=0.04639, over 17269.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2828, pruned_loss=0.05435, over 3185176.87 frames. ], batch size: 52, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:15:07,950 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6052, 3.6619, 1.8414, 4.3774, 2.8156, 4.2558, 2.1769, 2.9324], device='cuda:4'), covar=tensor([0.0265, 0.0387, 0.2093, 0.0123, 0.0815, 0.0324, 0.1808, 0.0701], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0172, 0.0190, 0.0149, 0.0172, 0.0212, 0.0199, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:15:32,099 INFO [train.py:904] (4/8) Epoch 16, batch 4450, loss[loss=0.2138, simple_loss=0.3092, pruned_loss=0.05923, over 16306.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.286, pruned_loss=0.05559, over 3184461.72 frames. ], batch size: 165, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:15:57,563 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.385e+02 2.034e+02 2.328e+02 2.907e+02 4.699e+02, threshold=4.656e+02, percent-clipped=0.0 2023-04-30 08:16:09,706 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156726.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 08:16:37,917 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5559, 2.4938, 2.5075, 3.7239, 3.0540, 3.9947, 1.3748, 2.9984], device='cuda:4'), covar=tensor([0.1286, 0.0768, 0.1096, 0.0123, 0.0244, 0.0306, 0.1583, 0.0663], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0168, 0.0187, 0.0172, 0.0202, 0.0212, 0.0190, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:16:41,518 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156749.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:16:45,639 INFO [train.py:904] (4/8) Epoch 16, batch 4500, loss[loss=0.1897, simple_loss=0.2814, pruned_loss=0.04905, over 16841.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2861, pruned_loss=0.05618, over 3189405.35 frames. ], batch size: 83, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:17:16,182 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=156774.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:17:43,519 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7269, 2.9108, 2.9550, 4.9871, 4.0072, 4.2923, 1.7385, 3.3258], device='cuda:4'), covar=tensor([0.1306, 0.0761, 0.1034, 0.0109, 0.0337, 0.0344, 0.1535, 0.0768], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0168, 0.0188, 0.0172, 0.0203, 0.0213, 0.0191, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:17:46,999 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-04-30 08:17:56,547 INFO [train.py:904] (4/8) Epoch 16, batch 4550, loss[loss=0.2058, simple_loss=0.294, pruned_loss=0.05883, over 16684.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2872, pruned_loss=0.05699, over 3201856.51 frames. ], batch size: 62, lr: 4.26e-03, grad_scale: 8.0 2023-04-30 08:18:07,896 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156810.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:18:09,174 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2526, 2.1079, 2.6983, 3.1307, 2.9673, 3.6813, 2.0785, 3.6132], device='cuda:4'), covar=tensor([0.0138, 0.0407, 0.0242, 0.0197, 0.0225, 0.0095, 0.0449, 0.0102], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0184, 0.0170, 0.0175, 0.0186, 0.0141, 0.0183, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:18:19,417 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.454e+02 2.005e+02 2.274e+02 2.770e+02 4.941e+02, threshold=4.548e+02, percent-clipped=1.0 2023-04-30 08:19:03,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4774, 2.1544, 2.8986, 3.3181, 3.1409, 3.8768, 2.4319, 3.7929], device='cuda:4'), covar=tensor([0.0128, 0.0417, 0.0233, 0.0187, 0.0224, 0.0103, 0.0405, 0.0105], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0184, 0.0171, 0.0175, 0.0186, 0.0141, 0.0184, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:19:06,328 INFO [train.py:904] (4/8) Epoch 16, batch 4600, loss[loss=0.2016, simple_loss=0.286, pruned_loss=0.0586, over 16631.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.288, pruned_loss=0.0568, over 3216230.31 frames. ], batch size: 62, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:19:36,424 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156872.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:19:55,685 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156885.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:20:19,276 INFO [train.py:904] (4/8) Epoch 16, batch 4650, loss[loss=0.2014, simple_loss=0.2857, pruned_loss=0.05851, over 16421.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2873, pruned_loss=0.05685, over 3218506.05 frames. ], batch size: 75, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:20:36,265 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9004, 1.8544, 2.4915, 2.8103, 2.6979, 3.2630, 1.8659, 3.2379], device='cuda:4'), covar=tensor([0.0154, 0.0447, 0.0254, 0.0252, 0.0258, 0.0127, 0.0510, 0.0108], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0185, 0.0172, 0.0176, 0.0188, 0.0142, 0.0185, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:20:37,490 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9876, 2.2376, 1.6618, 1.9602, 2.6261, 2.2868, 2.8017, 2.9106], device='cuda:4'), covar=tensor([0.0123, 0.0425, 0.0614, 0.0466, 0.0240, 0.0405, 0.0180, 0.0255], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0218, 0.0211, 0.0211, 0.0220, 0.0220, 0.0225, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:20:45,035 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.464e+02 1.893e+02 2.205e+02 2.654e+02 4.694e+02, threshold=4.410e+02, percent-clipped=1.0 2023-04-30 08:20:46,966 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=156920.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:21:25,724 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156946.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:21:25,867 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156946.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:21:33,087 INFO [train.py:904] (4/8) Epoch 16, batch 4700, loss[loss=0.1735, simple_loss=0.2594, pruned_loss=0.04381, over 17023.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2843, pruned_loss=0.05546, over 3225863.90 frames. ], batch size: 41, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:22:33,933 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3093, 5.3089, 5.1728, 4.8071, 4.7631, 5.2228, 5.1266, 4.9582], device='cuda:4'), covar=tensor([0.0505, 0.0411, 0.0245, 0.0247, 0.0953, 0.0422, 0.0253, 0.0604], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0372, 0.0316, 0.0305, 0.0332, 0.0353, 0.0216, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:22:36,836 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=156994.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:22:37,020 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5097, 2.3561, 2.2077, 3.6725, 2.2154, 3.7204, 1.4636, 2.5504], device='cuda:4'), covar=tensor([0.1664, 0.0991, 0.1468, 0.0209, 0.0230, 0.0468, 0.1907, 0.1007], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0168, 0.0187, 0.0172, 0.0202, 0.0212, 0.0190, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:22:47,496 INFO [train.py:904] (4/8) Epoch 16, batch 4750, loss[loss=0.1746, simple_loss=0.2627, pruned_loss=0.0433, over 16307.00 frames. ], tot_loss[loss=0.1938, simple_loss=0.2803, pruned_loss=0.05362, over 3226106.53 frames. ], batch size: 165, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:22:59,840 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157010.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:23:11,419 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.334e+02 1.885e+02 2.255e+02 2.719e+02 5.243e+02, threshold=4.511e+02, percent-clipped=2.0 2023-04-30 08:23:40,698 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3340, 2.0724, 2.3476, 3.8550, 2.0365, 2.3547, 2.2118, 2.1944], device='cuda:4'), covar=tensor([0.1367, 0.4247, 0.2791, 0.0630, 0.4906, 0.2992, 0.3633, 0.3942], device='cuda:4'), in_proj_covar=tensor([0.0382, 0.0424, 0.0351, 0.0322, 0.0427, 0.0489, 0.0390, 0.0494], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:23:42,891 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4033, 4.8201, 4.7093, 3.0948, 3.9031, 4.6129, 4.0403, 2.3447], device='cuda:4'), covar=tensor([0.0551, 0.0023, 0.0026, 0.0347, 0.0081, 0.0061, 0.0073, 0.0450], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0074, 0.0076, 0.0129, 0.0088, 0.0098, 0.0087, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 08:23:48,801 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-30 08:23:59,602 INFO [train.py:904] (4/8) Epoch 16, batch 4800, loss[loss=0.1914, simple_loss=0.285, pruned_loss=0.04892, over 15238.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2776, pruned_loss=0.05219, over 3206998.91 frames. ], batch size: 190, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:24:29,067 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157071.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:24:52,068 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4639, 3.5173, 1.9133, 3.9709, 2.5511, 3.8781, 2.0712, 2.6763], device='cuda:4'), covar=tensor([0.0256, 0.0372, 0.1771, 0.0115, 0.0917, 0.0483, 0.1755, 0.0853], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0170, 0.0190, 0.0147, 0.0170, 0.0209, 0.0198, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:25:14,775 INFO [train.py:904] (4/8) Epoch 16, batch 4850, loss[loss=0.196, simple_loss=0.2885, pruned_loss=0.05176, over 15372.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2782, pruned_loss=0.05136, over 3206480.34 frames. ], batch size: 190, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:25:20,717 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157105.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:25:41,920 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.374e+02 1.885e+02 2.190e+02 2.625e+02 3.825e+02, threshold=4.379e+02, percent-clipped=0.0 2023-04-30 08:26:16,182 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157142.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:26:30,907 INFO [train.py:904] (4/8) Epoch 16, batch 4900, loss[loss=0.1792, simple_loss=0.273, pruned_loss=0.04272, over 16388.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2778, pruned_loss=0.05045, over 3179034.87 frames. ], batch size: 146, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:26:59,081 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-30 08:27:45,016 INFO [train.py:904] (4/8) Epoch 16, batch 4950, loss[loss=0.1909, simple_loss=0.284, pruned_loss=0.04886, over 17106.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2771, pruned_loss=0.0497, over 3193142.47 frames. ], batch size: 47, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:27:46,688 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157203.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:28:08,806 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.535e+02 2.143e+02 2.414e+02 2.690e+02 6.197e+02, threshold=4.827e+02, percent-clipped=2.0 2023-04-30 08:28:33,100 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 08:28:41,076 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157241.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:28:57,707 INFO [train.py:904] (4/8) Epoch 16, batch 5000, loss[loss=0.1796, simple_loss=0.2669, pruned_loss=0.04611, over 17048.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.279, pruned_loss=0.04979, over 3203255.02 frames. ], batch size: 53, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:29:13,704 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5662, 3.0861, 3.0119, 1.8872, 2.6517, 2.1542, 2.9865, 3.2906], device='cuda:4'), covar=tensor([0.0386, 0.0783, 0.0672, 0.1890, 0.0977, 0.0937, 0.0948, 0.0799], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0156, 0.0161, 0.0148, 0.0139, 0.0125, 0.0140, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:29:41,501 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-30 08:30:10,076 INFO [train.py:904] (4/8) Epoch 16, batch 5050, loss[loss=0.189, simple_loss=0.2785, pruned_loss=0.04973, over 16641.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2792, pruned_loss=0.04964, over 3219270.46 frames. ], batch size: 134, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:30:33,952 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.633e+02 2.188e+02 2.518e+02 3.068e+02 6.843e+02, threshold=5.037e+02, percent-clipped=3.0 2023-04-30 08:31:22,540 INFO [train.py:904] (4/8) Epoch 16, batch 5100, loss[loss=0.1706, simple_loss=0.2636, pruned_loss=0.03882, over 15299.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2772, pruned_loss=0.04869, over 3220453.73 frames. ], batch size: 190, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:31:29,171 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157356.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:31:44,276 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157366.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:32:07,538 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157382.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:32:38,192 INFO [train.py:904] (4/8) Epoch 16, batch 5150, loss[loss=0.1814, simple_loss=0.2818, pruned_loss=0.04051, over 16397.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2776, pruned_loss=0.04802, over 3210891.57 frames. ], batch size: 146, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:32:42,248 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157404.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 08:32:43,923 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=157405.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:33:02,260 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157417.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:33:04,659 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 1.981e+02 2.319e+02 2.667e+02 4.266e+02, threshold=4.638e+02, percent-clipped=0.0 2023-04-30 08:33:40,815 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157443.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:33:52,504 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2399, 3.4390, 3.6087, 3.5526, 3.5536, 3.3953, 3.2467, 3.4884], device='cuda:4'), covar=tensor([0.0587, 0.0796, 0.0565, 0.0642, 0.0781, 0.0670, 0.1367, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0374, 0.0400, 0.0390, 0.0370, 0.0443, 0.0416, 0.0512, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 08:33:53,087 INFO [train.py:904] (4/8) Epoch 16, batch 5200, loss[loss=0.2202, simple_loss=0.3082, pruned_loss=0.06606, over 12616.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2758, pruned_loss=0.04725, over 3209566.89 frames. ], batch size: 248, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:33:54,778 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=157453.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:34:12,652 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157465.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 08:35:01,960 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157498.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:35:08,189 INFO [train.py:904] (4/8) Epoch 16, batch 5250, loss[loss=0.2152, simple_loss=0.2928, pruned_loss=0.06882, over 12565.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2737, pruned_loss=0.04743, over 3192185.15 frames. ], batch size: 247, lr: 4.25e-03, grad_scale: 8.0 2023-04-30 08:35:13,345 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157506.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 08:35:32,951 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.377e+02 2.109e+02 2.306e+02 2.689e+02 3.834e+02, threshold=4.611e+02, percent-clipped=0.0 2023-04-30 08:36:06,362 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=157541.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:36:21,503 INFO [train.py:904] (4/8) Epoch 16, batch 5300, loss[loss=0.1701, simple_loss=0.2543, pruned_loss=0.0429, over 16865.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2697, pruned_loss=0.04641, over 3197684.65 frames. ], batch size: 116, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:36:36,696 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8201, 3.3970, 3.3225, 2.1595, 3.0995, 3.3141, 3.1743, 1.8020], device='cuda:4'), covar=tensor([0.0616, 0.0051, 0.0059, 0.0419, 0.0097, 0.0113, 0.0098, 0.0502], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0074, 0.0075, 0.0129, 0.0089, 0.0098, 0.0087, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 08:36:45,022 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157567.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 08:36:46,162 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157568.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:37:17,189 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=157589.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:37:35,558 INFO [train.py:904] (4/8) Epoch 16, batch 5350, loss[loss=0.1884, simple_loss=0.282, pruned_loss=0.04738, over 16807.00 frames. ], tot_loss[loss=0.1794, simple_loss=0.2677, pruned_loss=0.0456, over 3195243.01 frames. ], batch size: 83, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:38:00,208 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.392e+02 1.990e+02 2.338e+02 2.844e+02 4.721e+02, threshold=4.676e+02, percent-clipped=1.0 2023-04-30 08:38:16,648 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157629.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:38:48,259 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7644, 5.1001, 5.2439, 5.0925, 5.0267, 5.6482, 5.1098, 4.8869], device='cuda:4'), covar=tensor([0.0964, 0.1352, 0.1570, 0.1592, 0.2236, 0.0747, 0.1239, 0.2182], device='cuda:4'), in_proj_covar=tensor([0.0384, 0.0539, 0.0582, 0.0452, 0.0609, 0.0619, 0.0460, 0.0610], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 08:38:49,042 INFO [train.py:904] (4/8) Epoch 16, batch 5400, loss[loss=0.1921, simple_loss=0.2887, pruned_loss=0.04776, over 16749.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2703, pruned_loss=0.04648, over 3195030.97 frames. ], batch size: 89, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:39:09,535 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=157666.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:40:04,937 INFO [train.py:904] (4/8) Epoch 16, batch 5450, loss[loss=0.2247, simple_loss=0.2993, pruned_loss=0.07503, over 12323.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2737, pruned_loss=0.04839, over 3172676.96 frames. ], batch size: 248, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:40:20,459 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157712.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:40:23,493 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=157714.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:40:29,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.390e+02 2.375e+02 2.743e+02 3.382e+02 6.991e+02, threshold=5.486e+02, percent-clipped=5.0 2023-04-30 08:41:00,184 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157738.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:41:20,975 INFO [train.py:904] (4/8) Epoch 16, batch 5500, loss[loss=0.2782, simple_loss=0.341, pruned_loss=0.1077, over 11858.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.282, pruned_loss=0.05344, over 3157422.45 frames. ], batch size: 246, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:41:33,656 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157760.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 08:41:46,482 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1476, 5.1337, 4.9009, 4.1557, 5.1097, 1.8584, 4.8016, 4.8671], device='cuda:4'), covar=tensor([0.0074, 0.0070, 0.0174, 0.0435, 0.0068, 0.2496, 0.0121, 0.0182], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0137, 0.0184, 0.0172, 0.0156, 0.0195, 0.0171, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:42:30,919 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=157798.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:42:36,591 INFO [train.py:904] (4/8) Epoch 16, batch 5550, loss[loss=0.2838, simple_loss=0.3366, pruned_loss=0.1155, over 11342.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2894, pruned_loss=0.05868, over 3131391.18 frames. ], batch size: 248, lr: 4.24e-03, grad_scale: 16.0 2023-04-30 08:43:04,446 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.012e+02 3.202e+02 3.935e+02 4.988e+02 8.755e+02, threshold=7.870e+02, percent-clipped=18.0 2023-04-30 08:43:47,866 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=157846.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:43:57,047 INFO [train.py:904] (4/8) Epoch 16, batch 5600, loss[loss=0.2268, simple_loss=0.3098, pruned_loss=0.07192, over 16301.00 frames. ], tot_loss[loss=0.212, simple_loss=0.2955, pruned_loss=0.06422, over 3081768.48 frames. ], batch size: 165, lr: 4.24e-03, grad_scale: 16.0 2023-04-30 08:44:00,073 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157853.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:44:05,636 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9862, 2.1038, 2.3083, 3.5283, 2.0559, 2.3774, 2.2121, 2.2313], device='cuda:4'), covar=tensor([0.1255, 0.3256, 0.2322, 0.0517, 0.3844, 0.2271, 0.3131, 0.3000], device='cuda:4'), in_proj_covar=tensor([0.0377, 0.0417, 0.0347, 0.0317, 0.0420, 0.0482, 0.0385, 0.0486], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:44:14,250 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157862.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 08:44:40,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6381, 3.6001, 4.2696, 1.7450, 4.4184, 4.4220, 3.0370, 3.2041], device='cuda:4'), covar=tensor([0.0777, 0.0252, 0.0147, 0.1268, 0.0054, 0.0109, 0.0432, 0.0430], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0104, 0.0090, 0.0136, 0.0072, 0.0116, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 08:45:21,493 INFO [train.py:904] (4/8) Epoch 16, batch 5650, loss[loss=0.2702, simple_loss=0.3291, pruned_loss=0.1056, over 11119.00 frames. ], tot_loss[loss=0.2192, simple_loss=0.3011, pruned_loss=0.06862, over 3037030.85 frames. ], batch size: 247, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:45:41,602 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157914.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:45:50,669 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.188e+02 3.549e+02 4.142e+02 4.836e+02 1.033e+03, threshold=8.283e+02, percent-clipped=2.0 2023-04-30 08:45:56,909 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157924.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:46:39,301 INFO [train.py:904] (4/8) Epoch 16, batch 5700, loss[loss=0.2095, simple_loss=0.2973, pruned_loss=0.06083, over 17000.00 frames. ], tot_loss[loss=0.2212, simple_loss=0.3022, pruned_loss=0.07007, over 3026775.84 frames. ], batch size: 109, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:48:01,939 INFO [train.py:904] (4/8) Epoch 16, batch 5750, loss[loss=0.2201, simple_loss=0.3009, pruned_loss=0.06966, over 17076.00 frames. ], tot_loss[loss=0.223, simple_loss=0.3043, pruned_loss=0.07086, over 3024279.38 frames. ], batch size: 55, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:48:17,328 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158012.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:48:30,520 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.050e+02 3.002e+02 3.593e+02 4.729e+02 9.320e+02, threshold=7.185e+02, percent-clipped=2.0 2023-04-30 08:49:00,708 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158038.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:49:22,774 INFO [train.py:904] (4/8) Epoch 16, batch 5800, loss[loss=0.2083, simple_loss=0.2971, pruned_loss=0.05981, over 16778.00 frames. ], tot_loss[loss=0.2216, simple_loss=0.3038, pruned_loss=0.06965, over 3026688.35 frames. ], batch size: 124, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:49:33,677 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1515, 2.0494, 1.7819, 1.8294, 2.3062, 2.0287, 2.1192, 2.4755], device='cuda:4'), covar=tensor([0.0219, 0.0353, 0.0436, 0.0422, 0.0201, 0.0328, 0.0202, 0.0229], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0218, 0.0211, 0.0210, 0.0218, 0.0220, 0.0223, 0.0214], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 08:49:35,950 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=158060.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:49:36,094 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158060.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 08:50:16,597 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=158086.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:50:41,299 INFO [train.py:904] (4/8) Epoch 16, batch 5850, loss[loss=0.2079, simple_loss=0.2941, pruned_loss=0.0609, over 16865.00 frames. ], tot_loss[loss=0.2192, simple_loss=0.3018, pruned_loss=0.0683, over 3008592.92 frames. ], batch size: 116, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:50:51,621 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=158108.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 08:51:08,528 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.631e+02 2.851e+02 3.457e+02 4.318e+02 8.019e+02, threshold=6.915e+02, percent-clipped=1.0 2023-04-30 08:51:13,048 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158122.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:52:03,755 INFO [train.py:904] (4/8) Epoch 16, batch 5900, loss[loss=0.1891, simple_loss=0.2748, pruned_loss=0.05165, over 16630.00 frames. ], tot_loss[loss=0.2187, simple_loss=0.3011, pruned_loss=0.06816, over 3014739.89 frames. ], batch size: 62, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:52:07,444 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.09 vs. limit=5.0 2023-04-30 08:52:24,454 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158162.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 08:52:29,769 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9367, 2.7997, 2.7957, 2.1281, 2.5982, 2.1575, 2.7583, 2.9047], device='cuda:4'), covar=tensor([0.0264, 0.0651, 0.0493, 0.1550, 0.0766, 0.0834, 0.0532, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0154, 0.0160, 0.0147, 0.0138, 0.0125, 0.0139, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:52:46,438 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8042, 2.5602, 2.5953, 1.8822, 2.4707, 2.6492, 2.5617, 1.9023], device='cuda:4'), covar=tensor([0.0426, 0.0095, 0.0074, 0.0358, 0.0116, 0.0123, 0.0098, 0.0379], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0074, 0.0075, 0.0129, 0.0089, 0.0098, 0.0086, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 08:52:55,962 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158183.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:53:25,265 INFO [train.py:904] (4/8) Epoch 16, batch 5950, loss[loss=0.2321, simple_loss=0.3158, pruned_loss=0.07414, over 16439.00 frames. ], tot_loss[loss=0.217, simple_loss=0.3014, pruned_loss=0.06626, over 3046637.33 frames. ], batch size: 146, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:53:36,984 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=158209.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:53:38,263 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=158210.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 08:53:55,265 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.638e+02 2.924e+02 3.226e+02 4.367e+02 1.058e+03, threshold=6.453e+02, percent-clipped=4.0 2023-04-30 08:54:01,814 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158224.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:54:24,118 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6027, 3.1329, 3.0755, 1.9193, 2.6748, 2.1191, 3.1594, 3.2391], device='cuda:4'), covar=tensor([0.0277, 0.0655, 0.0615, 0.1947, 0.0875, 0.0992, 0.0674, 0.0850], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0155, 0.0161, 0.0148, 0.0139, 0.0126, 0.0140, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:54:44,845 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6202, 2.6210, 1.8878, 2.7086, 2.1381, 2.7626, 2.0844, 2.3311], device='cuda:4'), covar=tensor([0.0310, 0.0400, 0.1370, 0.0247, 0.0663, 0.0537, 0.1175, 0.0618], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0172, 0.0194, 0.0148, 0.0173, 0.0211, 0.0201, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 08:54:45,453 INFO [train.py:904] (4/8) Epoch 16, batch 6000, loss[loss=0.2352, simple_loss=0.3062, pruned_loss=0.08212, over 11473.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.3007, pruned_loss=0.06596, over 3063765.81 frames. ], batch size: 247, lr: 4.24e-03, grad_scale: 8.0 2023-04-30 08:54:45,453 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 08:54:56,481 INFO [train.py:938] (4/8) Epoch 16, validation: loss=0.1553, simple_loss=0.2682, pruned_loss=0.0212, over 944034.00 frames. 2023-04-30 08:54:56,482 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 08:55:02,060 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7334, 3.9546, 2.9791, 2.3627, 2.8282, 2.5316, 4.2883, 3.6288], device='cuda:4'), covar=tensor([0.2928, 0.0726, 0.1793, 0.2567, 0.2331, 0.1886, 0.0463, 0.1146], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0260, 0.0294, 0.0294, 0.0286, 0.0238, 0.0279, 0.0314], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 08:55:27,051 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=158272.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:56:13,210 INFO [train.py:904] (4/8) Epoch 16, batch 6050, loss[loss=0.198, simple_loss=0.3065, pruned_loss=0.04477, over 16712.00 frames. ], tot_loss[loss=0.2141, simple_loss=0.2986, pruned_loss=0.06484, over 3085240.03 frames. ], batch size: 83, lr: 4.23e-03, grad_scale: 8.0 2023-04-30 08:56:40,245 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.935e+02 2.678e+02 3.324e+02 4.253e+02 8.310e+02, threshold=6.647e+02, percent-clipped=4.0 2023-04-30 08:57:31,985 INFO [train.py:904] (4/8) Epoch 16, batch 6100, loss[loss=0.2298, simple_loss=0.2998, pruned_loss=0.07991, over 11615.00 frames. ], tot_loss[loss=0.2125, simple_loss=0.2977, pruned_loss=0.06365, over 3082845.08 frames. ], batch size: 246, lr: 4.23e-03, grad_scale: 8.0 2023-04-30 08:58:06,890 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-30 08:58:23,970 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158384.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 08:58:49,714 INFO [train.py:904] (4/8) Epoch 16, batch 6150, loss[loss=0.2473, simple_loss=0.3121, pruned_loss=0.09124, over 11641.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.295, pruned_loss=0.06231, over 3108217.66 frames. ], batch size: 248, lr: 4.23e-03, grad_scale: 8.0 2023-04-30 08:59:05,870 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-30 08:59:18,317 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.901e+02 2.793e+02 3.279e+02 3.951e+02 7.811e+02, threshold=6.558e+02, percent-clipped=1.0 2023-04-30 08:59:57,052 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158445.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:00:08,003 INFO [train.py:904] (4/8) Epoch 16, batch 6200, loss[loss=0.1797, simple_loss=0.2752, pruned_loss=0.0421, over 16881.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2934, pruned_loss=0.06189, over 3106674.58 frames. ], batch size: 90, lr: 4.23e-03, grad_scale: 8.0 2023-04-30 09:00:27,492 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0599, 5.3426, 5.0603, 5.0913, 4.9020, 4.8138, 4.7150, 5.4377], device='cuda:4'), covar=tensor([0.1093, 0.0865, 0.0921, 0.0806, 0.0738, 0.0819, 0.1110, 0.0837], device='cuda:4'), in_proj_covar=tensor([0.0597, 0.0739, 0.0604, 0.0540, 0.0462, 0.0472, 0.0613, 0.0565], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:00:38,246 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2154, 4.5861, 3.6387, 2.7506, 3.2008, 3.0033, 4.9384, 4.0631], device='cuda:4'), covar=tensor([0.2352, 0.0534, 0.1364, 0.2289, 0.2394, 0.1674, 0.0345, 0.1015], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0263, 0.0297, 0.0297, 0.0290, 0.0240, 0.0282, 0.0319], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 09:00:44,336 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9567, 4.1916, 3.8908, 3.7258, 3.3235, 4.1312, 3.7615, 3.7354], device='cuda:4'), covar=tensor([0.0880, 0.0685, 0.0458, 0.0404, 0.1543, 0.0536, 0.1176, 0.0755], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0370, 0.0313, 0.0301, 0.0326, 0.0350, 0.0215, 0.0372], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:00:48,543 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=158478.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:01:21,942 INFO [train.py:904] (4/8) Epoch 16, batch 6250, loss[loss=0.1975, simple_loss=0.2942, pruned_loss=0.0504, over 16799.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2928, pruned_loss=0.06168, over 3100496.96 frames. ], batch size: 83, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:01:34,585 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158509.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:01:50,724 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-30 09:01:50,901 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.774e+02 2.618e+02 3.109e+02 4.242e+02 1.066e+03, threshold=6.218e+02, percent-clipped=4.0 2023-04-30 09:01:57,763 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9708, 3.3317, 3.2732, 1.9242, 2.8817, 2.2743, 3.5180, 3.4947], device='cuda:4'), covar=tensor([0.0258, 0.0723, 0.0606, 0.2044, 0.0842, 0.0957, 0.0592, 0.0917], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0155, 0.0161, 0.0148, 0.0139, 0.0125, 0.0140, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:02:38,737 INFO [train.py:904] (4/8) Epoch 16, batch 6300, loss[loss=0.1913, simple_loss=0.279, pruned_loss=0.05174, over 16172.00 frames. ], tot_loss[loss=0.2066, simple_loss=0.2918, pruned_loss=0.06065, over 3114441.90 frames. ], batch size: 165, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:02:45,562 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=158557.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:03:55,407 INFO [train.py:904] (4/8) Epoch 16, batch 6350, loss[loss=0.2149, simple_loss=0.2967, pruned_loss=0.0666, over 16246.00 frames. ], tot_loss[loss=0.209, simple_loss=0.2935, pruned_loss=0.06222, over 3110858.45 frames. ], batch size: 165, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:04:24,044 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.016e+02 2.970e+02 3.577e+02 4.444e+02 9.031e+02, threshold=7.154e+02, percent-clipped=4.0 2023-04-30 09:05:06,638 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5576, 4.4046, 4.3119, 2.9286, 3.8528, 4.3789, 3.9557, 2.4571], device='cuda:4'), covar=tensor([0.0473, 0.0032, 0.0043, 0.0340, 0.0078, 0.0092, 0.0063, 0.0387], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0074, 0.0076, 0.0129, 0.0088, 0.0098, 0.0086, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 09:05:11,913 INFO [train.py:904] (4/8) Epoch 16, batch 6400, loss[loss=0.1987, simple_loss=0.2781, pruned_loss=0.05968, over 16849.00 frames. ], tot_loss[loss=0.2096, simple_loss=0.2934, pruned_loss=0.06294, over 3105226.01 frames. ], batch size: 116, lr: 4.23e-03, grad_scale: 8.0 2023-04-30 09:05:55,475 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1326, 4.0101, 4.2188, 4.3312, 4.4537, 4.0410, 4.4095, 4.4796], device='cuda:4'), covar=tensor([0.1766, 0.1149, 0.1448, 0.0705, 0.0580, 0.1292, 0.0879, 0.0627], device='cuda:4'), in_proj_covar=tensor([0.0585, 0.0720, 0.0847, 0.0727, 0.0545, 0.0572, 0.0583, 0.0679], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:06:27,727 INFO [train.py:904] (4/8) Epoch 16, batch 6450, loss[loss=0.188, simple_loss=0.2756, pruned_loss=0.0502, over 16904.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.2937, pruned_loss=0.06249, over 3099748.91 frames. ], batch size: 116, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:06:56,388 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.032e+02 2.821e+02 3.663e+02 4.597e+02 9.189e+02, threshold=7.326e+02, percent-clipped=6.0 2023-04-30 09:07:26,483 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=158740.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:07:41,177 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-30 09:07:43,536 INFO [train.py:904] (4/8) Epoch 16, batch 6500, loss[loss=0.2217, simple_loss=0.2868, pruned_loss=0.07827, over 11380.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2921, pruned_loss=0.06176, over 3114093.80 frames. ], batch size: 247, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:08:10,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7469, 3.7867, 3.8655, 3.6259, 3.7272, 4.1866, 3.8430, 3.5983], device='cuda:4'), covar=tensor([0.1939, 0.2288, 0.2658, 0.2561, 0.2988, 0.1750, 0.1729, 0.2711], device='cuda:4'), in_proj_covar=tensor([0.0384, 0.0549, 0.0600, 0.0462, 0.0620, 0.0627, 0.0472, 0.0622], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 09:08:21,049 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158777.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:08:22,798 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158778.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:08:31,554 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158783.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:09:01,978 INFO [train.py:904] (4/8) Epoch 16, batch 6550, loss[loss=0.1949, simple_loss=0.2983, pruned_loss=0.04578, over 16675.00 frames. ], tot_loss[loss=0.2107, simple_loss=0.2954, pruned_loss=0.06296, over 3100689.61 frames. ], batch size: 68, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:09:28,746 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3889, 3.2875, 3.6481, 1.8616, 3.8365, 3.8788, 2.9510, 2.8448], device='cuda:4'), covar=tensor([0.0779, 0.0241, 0.0178, 0.1195, 0.0062, 0.0150, 0.0400, 0.0461], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0105, 0.0091, 0.0138, 0.0073, 0.0117, 0.0124, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 09:09:33,175 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.010e+02 2.836e+02 3.301e+02 3.944e+02 7.483e+02, threshold=6.603e+02, percent-clipped=1.0 2023-04-30 09:09:39,593 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=158826.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:09:53,261 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158835.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 09:09:58,224 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158838.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:10:06,293 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158844.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:10:18,010 INFO [train.py:904] (4/8) Epoch 16, batch 6600, loss[loss=0.2361, simple_loss=0.3056, pruned_loss=0.08332, over 11413.00 frames. ], tot_loss[loss=0.2129, simple_loss=0.2976, pruned_loss=0.06405, over 3071198.27 frames. ], batch size: 248, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:11:25,976 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158896.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 09:11:34,212 INFO [train.py:904] (4/8) Epoch 16, batch 6650, loss[loss=0.1928, simple_loss=0.2764, pruned_loss=0.05462, over 16744.00 frames. ], tot_loss[loss=0.215, simple_loss=0.2987, pruned_loss=0.06568, over 3047632.66 frames. ], batch size: 124, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:11:43,536 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8541, 3.6758, 4.2153, 2.0803, 4.4263, 4.4054, 3.2078, 3.1525], device='cuda:4'), covar=tensor([0.0701, 0.0242, 0.0144, 0.1156, 0.0048, 0.0104, 0.0357, 0.0454], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0106, 0.0091, 0.0138, 0.0073, 0.0117, 0.0125, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 09:12:04,649 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.962e+02 3.031e+02 3.577e+02 4.481e+02 9.334e+02, threshold=7.154e+02, percent-clipped=3.0 2023-04-30 09:12:14,608 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3518, 3.3842, 1.6561, 3.8255, 2.4978, 3.7843, 1.8938, 2.6400], device='cuda:4'), covar=tensor([0.0283, 0.0402, 0.2055, 0.0195, 0.0892, 0.0471, 0.1933, 0.0812], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0170, 0.0192, 0.0147, 0.0173, 0.0209, 0.0200, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:12:32,337 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 09:12:50,515 INFO [train.py:904] (4/8) Epoch 16, batch 6700, loss[loss=0.2136, simple_loss=0.2971, pruned_loss=0.06506, over 16898.00 frames. ], tot_loss[loss=0.2136, simple_loss=0.2969, pruned_loss=0.06517, over 3072987.23 frames. ], batch size: 109, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:12:54,012 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8746, 5.0639, 4.7197, 4.5288, 3.9840, 5.0187, 4.9121, 4.5822], device='cuda:4'), covar=tensor([0.0853, 0.0454, 0.0483, 0.0392, 0.1939, 0.0425, 0.0348, 0.0768], device='cuda:4'), in_proj_covar=tensor([0.0264, 0.0369, 0.0311, 0.0298, 0.0325, 0.0348, 0.0214, 0.0371], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:14:07,727 INFO [train.py:904] (4/8) Epoch 16, batch 6750, loss[loss=0.1979, simple_loss=0.2818, pruned_loss=0.05698, over 16200.00 frames. ], tot_loss[loss=0.212, simple_loss=0.295, pruned_loss=0.06447, over 3094135.26 frames. ], batch size: 165, lr: 4.23e-03, grad_scale: 4.0 2023-04-30 09:14:37,800 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.101e+02 2.962e+02 3.496e+02 4.058e+02 1.383e+03, threshold=6.992e+02, percent-clipped=2.0 2023-04-30 09:14:51,734 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-30 09:15:05,830 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159040.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:15:20,875 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4208, 2.1565, 1.7587, 1.8856, 2.4438, 2.1135, 2.3304, 2.5861], device='cuda:4'), covar=tensor([0.0178, 0.0376, 0.0483, 0.0447, 0.0219, 0.0384, 0.0207, 0.0250], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0219, 0.0213, 0.0213, 0.0219, 0.0219, 0.0223, 0.0215], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:15:23,303 INFO [train.py:904] (4/8) Epoch 16, batch 6800, loss[loss=0.2316, simple_loss=0.3081, pruned_loss=0.07754, over 11452.00 frames. ], tot_loss[loss=0.2118, simple_loss=0.2952, pruned_loss=0.06418, over 3097822.03 frames. ], batch size: 248, lr: 4.22e-03, grad_scale: 8.0 2023-04-30 09:15:25,660 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6229, 2.5690, 1.8740, 2.6635, 2.1475, 2.7310, 2.0760, 2.3612], device='cuda:4'), covar=tensor([0.0288, 0.0373, 0.1317, 0.0226, 0.0688, 0.0497, 0.1192, 0.0633], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0168, 0.0191, 0.0145, 0.0171, 0.0208, 0.0198, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:15:50,652 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159069.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:16:20,371 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=159088.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:16:40,956 INFO [train.py:904] (4/8) Epoch 16, batch 6850, loss[loss=0.2317, simple_loss=0.335, pruned_loss=0.06419, over 16696.00 frames. ], tot_loss[loss=0.214, simple_loss=0.2973, pruned_loss=0.06533, over 3082965.51 frames. ], batch size: 62, lr: 4.22e-03, grad_scale: 2.0 2023-04-30 09:17:12,961 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.918e+02 2.851e+02 3.389e+02 4.147e+02 9.414e+02, threshold=6.778e+02, percent-clipped=4.0 2023-04-30 09:17:22,470 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159130.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:17:26,682 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159133.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:17:35,362 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159139.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:17:54,571 INFO [train.py:904] (4/8) Epoch 16, batch 6900, loss[loss=0.2124, simple_loss=0.3042, pruned_loss=0.06025, over 16630.00 frames. ], tot_loss[loss=0.2144, simple_loss=0.2995, pruned_loss=0.06465, over 3097663.59 frames. ], batch size: 68, lr: 4.22e-03, grad_scale: 2.0 2023-04-30 09:18:07,168 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159159.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:18:13,882 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4523, 3.0375, 3.0269, 1.9446, 2.7624, 2.1933, 3.0107, 3.1557], device='cuda:4'), covar=tensor([0.0301, 0.0659, 0.0566, 0.1870, 0.0787, 0.0895, 0.0669, 0.0813], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0156, 0.0162, 0.0149, 0.0140, 0.0126, 0.0141, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:18:14,136 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-30 09:18:56,517 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159191.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 09:19:13,870 INFO [train.py:904] (4/8) Epoch 16, batch 6950, loss[loss=0.2092, simple_loss=0.3014, pruned_loss=0.05851, over 16898.00 frames. ], tot_loss[loss=0.216, simple_loss=0.3005, pruned_loss=0.06581, over 3102712.52 frames. ], batch size: 96, lr: 4.22e-03, grad_scale: 2.0 2023-04-30 09:19:42,695 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159220.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:19:48,719 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.671e+02 3.353e+02 4.093e+02 5.038e+02 1.287e+03, threshold=8.186e+02, percent-clipped=11.0 2023-04-30 09:20:23,349 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-30 09:20:29,895 INFO [train.py:904] (4/8) Epoch 16, batch 7000, loss[loss=0.1958, simple_loss=0.2966, pruned_loss=0.04751, over 16660.00 frames. ], tot_loss[loss=0.2164, simple_loss=0.3015, pruned_loss=0.06567, over 3092430.84 frames. ], batch size: 89, lr: 4.22e-03, grad_scale: 2.0 2023-04-30 09:21:27,981 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159290.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:21:42,800 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1590, 1.8547, 2.5553, 3.0111, 2.7774, 3.4742, 2.0495, 3.4775], device='cuda:4'), covar=tensor([0.0171, 0.0489, 0.0317, 0.0260, 0.0286, 0.0144, 0.0523, 0.0128], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0187, 0.0171, 0.0175, 0.0185, 0.0143, 0.0185, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:21:45,218 INFO [train.py:904] (4/8) Epoch 16, batch 7050, loss[loss=0.211, simple_loss=0.2963, pruned_loss=0.06287, over 16636.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.302, pruned_loss=0.06526, over 3098271.69 frames. ], batch size: 134, lr: 4.22e-03, grad_scale: 2.0 2023-04-30 09:22:18,896 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.833e+02 2.988e+02 3.653e+02 4.362e+02 9.002e+02, threshold=7.306e+02, percent-clipped=3.0 2023-04-30 09:23:00,608 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159351.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:23:01,353 INFO [train.py:904] (4/8) Epoch 16, batch 7100, loss[loss=0.2001, simple_loss=0.2864, pruned_loss=0.05688, over 16425.00 frames. ], tot_loss[loss=0.2156, simple_loss=0.3006, pruned_loss=0.06526, over 3079839.81 frames. ], batch size: 146, lr: 4.22e-03, grad_scale: 2.0 2023-04-30 09:23:24,931 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8536, 3.7577, 4.3670, 2.0782, 4.5273, 4.4888, 3.0538, 3.2274], device='cuda:4'), covar=tensor([0.0743, 0.0220, 0.0146, 0.1184, 0.0046, 0.0125, 0.0425, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0106, 0.0092, 0.0138, 0.0073, 0.0117, 0.0126, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 09:24:17,480 INFO [train.py:904] (4/8) Epoch 16, batch 7150, loss[loss=0.2105, simple_loss=0.2974, pruned_loss=0.06177, over 16712.00 frames. ], tot_loss[loss=0.2142, simple_loss=0.2986, pruned_loss=0.06484, over 3081192.15 frames. ], batch size: 89, lr: 4.22e-03, grad_scale: 2.0 2023-04-30 09:24:49,379 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.558e+02 3.206e+02 4.048e+02 4.680e+02 7.501e+02, threshold=8.096e+02, percent-clipped=1.0 2023-04-30 09:24:51,546 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159425.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:25:02,933 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159433.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:25:11,743 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159439.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:25:30,654 INFO [train.py:904] (4/8) Epoch 16, batch 7200, loss[loss=0.2005, simple_loss=0.2868, pruned_loss=0.05715, over 15395.00 frames. ], tot_loss[loss=0.2121, simple_loss=0.2971, pruned_loss=0.06357, over 3066807.75 frames. ], batch size: 191, lr: 4.22e-03, grad_scale: 4.0 2023-04-30 09:26:06,561 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.02 vs. limit=5.0 2023-04-30 09:26:15,843 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=159481.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:26:26,897 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=159487.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:26:27,356 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-30 09:26:33,340 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159491.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 09:26:50,665 INFO [train.py:904] (4/8) Epoch 16, batch 7250, loss[loss=0.2373, simple_loss=0.3013, pruned_loss=0.08664, over 11423.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2943, pruned_loss=0.06228, over 3067373.04 frames. ], batch size: 248, lr: 4.22e-03, grad_scale: 4.0 2023-04-30 09:26:59,797 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5540, 3.2422, 2.7124, 2.2065, 2.2863, 2.2868, 3.3476, 3.1211], device='cuda:4'), covar=tensor([0.2458, 0.0686, 0.1612, 0.2439, 0.2246, 0.1874, 0.0508, 0.1078], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0263, 0.0296, 0.0298, 0.0289, 0.0240, 0.0281, 0.0318], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 09:27:11,555 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159515.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:27:23,390 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.928e+02 2.845e+02 3.560e+02 4.449e+02 7.164e+02, threshold=7.119e+02, percent-clipped=0.0 2023-04-30 09:27:27,586 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9606, 3.1364, 3.1911, 2.1265, 2.9803, 3.1552, 3.0445, 1.9760], device='cuda:4'), covar=tensor([0.0496, 0.0064, 0.0059, 0.0384, 0.0104, 0.0130, 0.0095, 0.0398], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0075, 0.0076, 0.0131, 0.0089, 0.0099, 0.0088, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 09:27:47,290 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=159539.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 09:28:06,272 INFO [train.py:904] (4/8) Epoch 16, batch 7300, loss[loss=0.243, simple_loss=0.3109, pruned_loss=0.08759, over 11433.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2937, pruned_loss=0.06179, over 3074597.29 frames. ], batch size: 246, lr: 4.22e-03, grad_scale: 4.0 2023-04-30 09:28:35,553 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-30 09:28:36,749 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-30 09:29:22,396 INFO [train.py:904] (4/8) Epoch 16, batch 7350, loss[loss=0.217, simple_loss=0.3096, pruned_loss=0.06221, over 16862.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2948, pruned_loss=0.06307, over 3050521.39 frames. ], batch size: 102, lr: 4.22e-03, grad_scale: 4.0 2023-04-30 09:29:34,552 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-30 09:29:56,734 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.741e+02 3.169e+02 3.768e+02 4.450e+02 9.358e+02, threshold=7.536e+02, percent-clipped=3.0 2023-04-30 09:30:07,673 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159631.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:30:31,263 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159646.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:30:39,217 INFO [train.py:904] (4/8) Epoch 16, batch 7400, loss[loss=0.2232, simple_loss=0.3043, pruned_loss=0.07105, over 16950.00 frames. ], tot_loss[loss=0.2113, simple_loss=0.2959, pruned_loss=0.06336, over 3068433.04 frames. ], batch size: 109, lr: 4.22e-03, grad_scale: 4.0 2023-04-30 09:30:49,091 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4490, 3.4022, 1.8750, 3.8569, 2.5095, 3.8221, 2.0575, 2.6990], device='cuda:4'), covar=tensor([0.0282, 0.0458, 0.1998, 0.0249, 0.0933, 0.0612, 0.1699, 0.0823], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0167, 0.0190, 0.0145, 0.0171, 0.0207, 0.0198, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:31:27,305 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3271, 4.3354, 4.1914, 3.9220, 3.9000, 4.2838, 3.9705, 4.0049], device='cuda:4'), covar=tensor([0.0523, 0.0524, 0.0287, 0.0288, 0.0761, 0.0424, 0.0716, 0.0605], device='cuda:4'), in_proj_covar=tensor([0.0260, 0.0364, 0.0304, 0.0293, 0.0318, 0.0342, 0.0210, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:31:41,327 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159692.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:31:57,477 INFO [train.py:904] (4/8) Epoch 16, batch 7450, loss[loss=0.216, simple_loss=0.3059, pruned_loss=0.06303, over 16890.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.2965, pruned_loss=0.06404, over 3081428.33 frames. ], batch size: 116, lr: 4.22e-03, grad_scale: 4.0 2023-04-30 09:32:33,422 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.461e+02 3.124e+02 3.664e+02 4.558e+02 8.463e+02, threshold=7.327e+02, percent-clipped=4.0 2023-04-30 09:32:36,579 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159725.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:33:17,743 INFO [train.py:904] (4/8) Epoch 16, batch 7500, loss[loss=0.2575, simple_loss=0.3259, pruned_loss=0.09451, over 11349.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.2963, pruned_loss=0.06345, over 3070316.32 frames. ], batch size: 248, lr: 4.22e-03, grad_scale: 4.0 2023-04-30 09:33:50,299 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=159773.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:34:37,012 INFO [train.py:904] (4/8) Epoch 16, batch 7550, loss[loss=0.1917, simple_loss=0.2813, pruned_loss=0.05106, over 16211.00 frames. ], tot_loss[loss=0.2112, simple_loss=0.2953, pruned_loss=0.06353, over 3069053.97 frames. ], batch size: 165, lr: 4.21e-03, grad_scale: 4.0 2023-04-30 09:34:44,469 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1821, 4.0487, 2.5756, 4.9787, 3.2540, 4.8323, 2.9393, 3.4841], device='cuda:4'), covar=tensor([0.0234, 0.0360, 0.1574, 0.0149, 0.0703, 0.0455, 0.1265, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0168, 0.0190, 0.0145, 0.0171, 0.0208, 0.0199, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:34:58,353 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159815.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:35:09,541 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5396, 3.5046, 2.5591, 2.1271, 2.4475, 2.2406, 3.6065, 3.1824], device='cuda:4'), covar=tensor([0.3125, 0.0797, 0.2114, 0.2683, 0.2741, 0.2208, 0.0571, 0.1333], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0264, 0.0297, 0.0298, 0.0290, 0.0241, 0.0282, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 09:35:10,140 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.970e+02 2.719e+02 3.178e+02 3.820e+02 7.109e+02, threshold=6.356e+02, percent-clipped=0.0 2023-04-30 09:35:53,836 INFO [train.py:904] (4/8) Epoch 16, batch 7600, loss[loss=0.2132, simple_loss=0.3063, pruned_loss=0.06005, over 16958.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2942, pruned_loss=0.06342, over 3072795.55 frames. ], batch size: 109, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:35:59,114 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159855.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:36:07,886 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5805, 4.6396, 5.0315, 4.9727, 5.0072, 4.7055, 4.5999, 4.4853], device='cuda:4'), covar=tensor([0.0469, 0.0580, 0.0411, 0.0562, 0.0569, 0.0452, 0.1129, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0372, 0.0400, 0.0389, 0.0370, 0.0441, 0.0415, 0.0511, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 09:36:10,414 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=159863.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:37:09,859 INFO [train.py:904] (4/8) Epoch 16, batch 7650, loss[loss=0.2645, simple_loss=0.3224, pruned_loss=0.1034, over 11383.00 frames. ], tot_loss[loss=0.2108, simple_loss=0.2947, pruned_loss=0.06348, over 3083645.09 frames. ], batch size: 248, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:37:21,682 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159910.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:37:30,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159916.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:37:43,258 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.780e+02 3.207e+02 4.068e+02 5.310e+02 2.281e+03, threshold=8.135e+02, percent-clipped=15.0 2023-04-30 09:38:13,537 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159946.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:38:22,085 INFO [train.py:904] (4/8) Epoch 16, batch 7700, loss[loss=0.2639, simple_loss=0.3233, pruned_loss=0.1022, over 11731.00 frames. ], tot_loss[loss=0.212, simple_loss=0.2953, pruned_loss=0.06437, over 3071337.19 frames. ], batch size: 248, lr: 4.21e-03, grad_scale: 4.0 2023-04-30 09:38:51,198 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159971.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:38:56,715 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8185, 5.0144, 4.7209, 4.5304, 4.0138, 4.9672, 4.8837, 4.4921], device='cuda:4'), covar=tensor([0.0957, 0.0699, 0.0490, 0.0423, 0.1951, 0.0512, 0.0404, 0.0862], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0368, 0.0306, 0.0295, 0.0320, 0.0345, 0.0212, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:39:15,433 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159987.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:39:26,410 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=159994.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:39:41,052 INFO [train.py:904] (4/8) Epoch 16, batch 7750, loss[loss=0.2104, simple_loss=0.2991, pruned_loss=0.06083, over 16416.00 frames. ], tot_loss[loss=0.212, simple_loss=0.2958, pruned_loss=0.0641, over 3088592.00 frames. ], batch size: 146, lr: 4.21e-03, grad_scale: 4.0 2023-04-30 09:39:56,929 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6076, 2.6339, 2.2730, 3.5881, 2.5912, 3.8274, 1.4226, 2.7448], device='cuda:4'), covar=tensor([0.1389, 0.0735, 0.1308, 0.0155, 0.0221, 0.0397, 0.1708, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0166, 0.0188, 0.0169, 0.0202, 0.0211, 0.0192, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:40:13,104 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7486, 3.8355, 2.9407, 2.2777, 2.6033, 2.4886, 4.0567, 3.4251], device='cuda:4'), covar=tensor([0.2711, 0.0708, 0.1788, 0.2728, 0.2626, 0.1926, 0.0453, 0.1287], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0266, 0.0299, 0.0300, 0.0292, 0.0242, 0.0283, 0.0322], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 09:40:13,567 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.949e+02 3.102e+02 3.551e+02 4.185e+02 8.391e+02, threshold=7.102e+02, percent-clipped=1.0 2023-04-30 09:40:53,103 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-30 09:40:53,467 INFO [train.py:904] (4/8) Epoch 16, batch 7800, loss[loss=0.2945, simple_loss=0.3555, pruned_loss=0.1168, over 11354.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.2967, pruned_loss=0.06491, over 3073481.40 frames. ], batch size: 246, lr: 4.21e-03, grad_scale: 4.0 2023-04-30 09:41:17,901 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1307, 2.9196, 3.0399, 1.6474, 3.1120, 3.3018, 2.7273, 2.5733], device='cuda:4'), covar=tensor([0.0863, 0.0247, 0.0250, 0.1325, 0.0118, 0.0221, 0.0419, 0.0495], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0105, 0.0092, 0.0137, 0.0073, 0.0117, 0.0125, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 09:41:40,771 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-30 09:42:08,841 INFO [train.py:904] (4/8) Epoch 16, batch 7850, loss[loss=0.2391, simple_loss=0.3094, pruned_loss=0.08442, over 11508.00 frames. ], tot_loss[loss=0.2137, simple_loss=0.2974, pruned_loss=0.06496, over 3069915.09 frames. ], batch size: 247, lr: 4.21e-03, grad_scale: 4.0 2023-04-30 09:42:43,408 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.920e+02 2.758e+02 3.305e+02 4.051e+02 1.069e+03, threshold=6.609e+02, percent-clipped=3.0 2023-04-30 09:43:25,067 INFO [train.py:904] (4/8) Epoch 16, batch 7900, loss[loss=0.2066, simple_loss=0.2871, pruned_loss=0.06298, over 16642.00 frames. ], tot_loss[loss=0.2126, simple_loss=0.2964, pruned_loss=0.06443, over 3058179.40 frames. ], batch size: 62, lr: 4.21e-03, grad_scale: 4.0 2023-04-30 09:43:45,856 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-04-30 09:44:43,661 INFO [train.py:904] (4/8) Epoch 16, batch 7950, loss[loss=0.2099, simple_loss=0.2951, pruned_loss=0.06238, over 16679.00 frames. ], tot_loss[loss=0.2134, simple_loss=0.2967, pruned_loss=0.06502, over 3058513.36 frames. ], batch size: 134, lr: 4.21e-03, grad_scale: 4.0 2023-04-30 09:44:56,865 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=160211.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:45:16,535 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.952e+02 2.700e+02 3.245e+02 4.105e+02 1.062e+03, threshold=6.489e+02, percent-clipped=6.0 2023-04-30 09:45:44,795 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7962, 2.9631, 2.5908, 5.0386, 3.5178, 4.2377, 1.7532, 2.7524], device='cuda:4'), covar=tensor([0.1365, 0.0843, 0.1337, 0.0197, 0.0505, 0.0461, 0.1667, 0.1077], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0166, 0.0188, 0.0170, 0.0203, 0.0210, 0.0192, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:45:56,353 INFO [train.py:904] (4/8) Epoch 16, batch 8000, loss[loss=0.2221, simple_loss=0.3074, pruned_loss=0.0684, over 16845.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.2975, pruned_loss=0.06613, over 3028748.41 frames. ], batch size: 116, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:46:17,932 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=160266.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:46:20,133 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-30 09:46:31,022 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1350, 3.1793, 1.8159, 3.4324, 2.3780, 3.4731, 1.9932, 2.5447], device='cuda:4'), covar=tensor([0.0297, 0.0397, 0.1723, 0.0233, 0.0884, 0.0602, 0.1599, 0.0791], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0170, 0.0193, 0.0147, 0.0173, 0.0211, 0.0201, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:46:50,290 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=160287.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:47:12,589 INFO [train.py:904] (4/8) Epoch 16, batch 8050, loss[loss=0.1985, simple_loss=0.2954, pruned_loss=0.05079, over 16736.00 frames. ], tot_loss[loss=0.2138, simple_loss=0.2971, pruned_loss=0.06526, over 3042402.31 frames. ], batch size: 89, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:47:20,249 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=160307.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 09:47:47,811 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.763e+02 2.953e+02 3.727e+02 4.663e+02 8.321e+02, threshold=7.453e+02, percent-clipped=3.0 2023-04-30 09:47:57,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6685, 1.7113, 2.2735, 2.5884, 2.5153, 3.0096, 1.9189, 2.9215], device='cuda:4'), covar=tensor([0.0197, 0.0436, 0.0279, 0.0255, 0.0283, 0.0131, 0.0427, 0.0122], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0184, 0.0168, 0.0172, 0.0183, 0.0140, 0.0183, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:48:04,048 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=160335.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:48:29,618 INFO [train.py:904] (4/8) Epoch 16, batch 8100, loss[loss=0.2299, simple_loss=0.3091, pruned_loss=0.0754, over 16382.00 frames. ], tot_loss[loss=0.2127, simple_loss=0.2967, pruned_loss=0.06438, over 3058903.95 frames. ], batch size: 146, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:48:54,849 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=160368.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 09:49:46,014 INFO [train.py:904] (4/8) Epoch 16, batch 8150, loss[loss=0.2361, simple_loss=0.3029, pruned_loss=0.08463, over 11766.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2941, pruned_loss=0.06322, over 3065893.85 frames. ], batch size: 248, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:50:21,735 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.821e+02 3.128e+02 3.748e+02 4.647e+02 9.176e+02, threshold=7.497e+02, percent-clipped=1.0 2023-04-30 09:51:00,122 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-30 09:51:05,065 INFO [train.py:904] (4/8) Epoch 16, batch 8200, loss[loss=0.2016, simple_loss=0.2852, pruned_loss=0.05904, over 15349.00 frames. ], tot_loss[loss=0.2083, simple_loss=0.2917, pruned_loss=0.06249, over 3063155.41 frames. ], batch size: 190, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:51:05,530 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4989, 4.5449, 4.9594, 4.8927, 4.9229, 4.5907, 4.5850, 4.4179], device='cuda:4'), covar=tensor([0.0349, 0.0611, 0.0365, 0.0442, 0.0491, 0.0407, 0.1060, 0.0486], device='cuda:4'), in_proj_covar=tensor([0.0378, 0.0409, 0.0399, 0.0377, 0.0450, 0.0422, 0.0517, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 09:51:18,377 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2991, 2.0333, 1.6972, 1.7039, 2.2743, 1.9665, 1.9388, 2.3459], device='cuda:4'), covar=tensor([0.0211, 0.0348, 0.0457, 0.0407, 0.0226, 0.0315, 0.0209, 0.0241], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0215, 0.0210, 0.0210, 0.0215, 0.0215, 0.0220, 0.0212], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:52:27,310 INFO [train.py:904] (4/8) Epoch 16, batch 8250, loss[loss=0.1762, simple_loss=0.2676, pruned_loss=0.04236, over 12323.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.2898, pruned_loss=0.06027, over 3021790.92 frames. ], batch size: 248, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:52:42,382 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=160511.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:53:02,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7853, 1.3541, 1.6541, 1.6701, 1.8165, 1.8860, 1.6148, 1.7642], device='cuda:4'), covar=tensor([0.0221, 0.0360, 0.0208, 0.0251, 0.0268, 0.0171, 0.0368, 0.0129], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0184, 0.0169, 0.0173, 0.0183, 0.0141, 0.0184, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:53:04,605 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.122e+02 2.811e+02 3.275e+02 4.165e+02 1.278e+03, threshold=6.551e+02, percent-clipped=3.0 2023-04-30 09:53:18,917 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.20 vs. limit=5.0 2023-04-30 09:53:49,474 INFO [train.py:904] (4/8) Epoch 16, batch 8300, loss[loss=0.1693, simple_loss=0.2711, pruned_loss=0.03374, over 16460.00 frames. ], tot_loss[loss=0.2003, simple_loss=0.2866, pruned_loss=0.05698, over 3017206.21 frames. ], batch size: 75, lr: 4.21e-03, grad_scale: 8.0 2023-04-30 09:54:01,474 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=160559.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:54:10,191 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-30 09:54:13,054 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=160566.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:54:27,762 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5064, 2.5474, 2.3750, 3.7671, 2.3441, 4.0383, 1.4148, 2.9086], device='cuda:4'), covar=tensor([0.1513, 0.0791, 0.1172, 0.0166, 0.0137, 0.0311, 0.1702, 0.0718], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0165, 0.0187, 0.0168, 0.0201, 0.0210, 0.0191, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:55:10,567 INFO [train.py:904] (4/8) Epoch 16, batch 8350, loss[loss=0.1771, simple_loss=0.2691, pruned_loss=0.04258, over 16435.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.286, pruned_loss=0.05518, over 3018052.16 frames. ], batch size: 75, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 09:55:30,914 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=160614.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 09:55:48,785 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 2.217e+02 2.608e+02 3.326e+02 6.898e+02, threshold=5.216e+02, percent-clipped=1.0 2023-04-30 09:56:10,634 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0443, 4.0182, 3.9385, 3.2634, 3.9603, 1.8825, 3.7747, 3.5963], device='cuda:4'), covar=tensor([0.0108, 0.0100, 0.0178, 0.0263, 0.0107, 0.2469, 0.0129, 0.0215], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0133, 0.0179, 0.0163, 0.0151, 0.0191, 0.0165, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 09:56:33,054 INFO [train.py:904] (4/8) Epoch 16, batch 8400, loss[loss=0.1681, simple_loss=0.264, pruned_loss=0.03608, over 15499.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2829, pruned_loss=0.05252, over 3025550.53 frames. ], batch size: 191, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 09:56:43,275 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2703, 4.3360, 4.4897, 4.2727, 4.2863, 4.8189, 4.3803, 4.0816], device='cuda:4'), covar=tensor([0.1516, 0.1909, 0.2094, 0.2289, 0.2943, 0.1146, 0.1573, 0.2716], device='cuda:4'), in_proj_covar=tensor([0.0372, 0.0537, 0.0583, 0.0449, 0.0599, 0.0615, 0.0460, 0.0603], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 09:56:51,249 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=160663.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 09:57:54,560 INFO [train.py:904] (4/8) Epoch 16, batch 8450, loss[loss=0.181, simple_loss=0.2741, pruned_loss=0.04398, over 15234.00 frames. ], tot_loss[loss=0.191, simple_loss=0.281, pruned_loss=0.05054, over 3033726.33 frames. ], batch size: 191, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 09:58:31,816 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.268e+02 2.257e+02 2.696e+02 3.215e+02 7.413e+02, threshold=5.391e+02, percent-clipped=3.0 2023-04-30 09:58:45,893 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9627, 2.3541, 2.4541, 3.0168, 2.0242, 3.3249, 1.7011, 2.8350], device='cuda:4'), covar=tensor([0.1198, 0.0593, 0.0849, 0.0151, 0.0095, 0.0315, 0.1390, 0.0575], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0163, 0.0185, 0.0166, 0.0199, 0.0208, 0.0190, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 09:59:11,464 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.17 vs. limit=5.0 2023-04-30 09:59:15,544 INFO [train.py:904] (4/8) Epoch 16, batch 8500, loss[loss=0.1914, simple_loss=0.2897, pruned_loss=0.04653, over 16605.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2771, pruned_loss=0.04839, over 3019475.04 frames. ], batch size: 62, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:00:42,444 INFO [train.py:904] (4/8) Epoch 16, batch 8550, loss[loss=0.1872, simple_loss=0.2857, pruned_loss=0.04438, over 16895.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2749, pruned_loss=0.0473, over 3013069.27 frames. ], batch size: 96, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:01:27,141 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.280e+02 2.208e+02 2.559e+02 3.052e+02 5.214e+02, threshold=5.118e+02, percent-clipped=0.0 2023-04-30 10:01:32,831 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0253, 3.0682, 1.9408, 3.2533, 2.3029, 3.2950, 2.1273, 2.5791], device='cuda:4'), covar=tensor([0.0320, 0.0363, 0.1497, 0.0266, 0.0833, 0.0494, 0.1433, 0.0746], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0166, 0.0187, 0.0143, 0.0168, 0.0203, 0.0196, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 10:02:22,609 INFO [train.py:904] (4/8) Epoch 16, batch 8600, loss[loss=0.1837, simple_loss=0.2766, pruned_loss=0.04537, over 16524.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2752, pruned_loss=0.04667, over 2995087.02 frames. ], batch size: 75, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:02:27,427 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9270, 4.2431, 4.0747, 4.0702, 3.7531, 3.8253, 3.8808, 4.2279], device='cuda:4'), covar=tensor([0.1085, 0.0838, 0.0877, 0.0784, 0.0810, 0.1677, 0.0964, 0.1019], device='cuda:4'), in_proj_covar=tensor([0.0591, 0.0717, 0.0587, 0.0530, 0.0452, 0.0470, 0.0600, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:02:30,188 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6300, 3.7114, 3.4875, 3.2289, 3.3055, 3.6061, 3.3612, 3.4386], device='cuda:4'), covar=tensor([0.0569, 0.0595, 0.0276, 0.0231, 0.0543, 0.0498, 0.1487, 0.0506], device='cuda:4'), in_proj_covar=tensor([0.0261, 0.0366, 0.0304, 0.0294, 0.0316, 0.0344, 0.0211, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:04:02,532 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=160901.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:04:03,213 INFO [train.py:904] (4/8) Epoch 16, batch 8650, loss[loss=0.1778, simple_loss=0.2793, pruned_loss=0.03815, over 16903.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2734, pruned_loss=0.04504, over 2994045.47 frames. ], batch size: 116, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:04:27,070 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.45 vs. limit=5.0 2023-04-30 10:04:56,266 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 2.325e+02 2.764e+02 3.397e+02 5.575e+02, threshold=5.528e+02, percent-clipped=1.0 2023-04-30 10:05:52,474 INFO [train.py:904] (4/8) Epoch 16, batch 8700, loss[loss=0.1738, simple_loss=0.258, pruned_loss=0.04477, over 12365.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2706, pruned_loss=0.04373, over 3013299.80 frames. ], batch size: 250, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:06:13,612 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=160962.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:06:15,195 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=160963.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 10:06:18,487 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=160965.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:06:31,570 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-30 10:07:29,335 INFO [train.py:904] (4/8) Epoch 16, batch 8750, loss[loss=0.1532, simple_loss=0.2477, pruned_loss=0.02934, over 12245.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2697, pruned_loss=0.04266, over 3017782.30 frames. ], batch size: 248, lr: 4.20e-03, grad_scale: 4.0 2023-04-30 10:07:53,443 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=161011.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 10:08:12,882 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.37 vs. limit=5.0 2023-04-30 10:08:27,329 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.33 vs. limit=5.0 2023-04-30 10:08:27,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.454e+02 2.172e+02 2.627e+02 3.227e+02 5.767e+02, threshold=5.254e+02, percent-clipped=1.0 2023-04-30 10:08:28,253 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161026.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:09:20,775 INFO [train.py:904] (4/8) Epoch 16, batch 8800, loss[loss=0.188, simple_loss=0.2795, pruned_loss=0.0482, over 16715.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2678, pruned_loss=0.04134, over 3028804.58 frames. ], batch size: 134, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:11:05,641 INFO [train.py:904] (4/8) Epoch 16, batch 8850, loss[loss=0.176, simple_loss=0.2605, pruned_loss=0.0458, over 12304.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2707, pruned_loss=0.04133, over 3034821.73 frames. ], batch size: 250, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:11:55,943 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.456e+02 2.310e+02 2.876e+02 3.561e+02 5.593e+02, threshold=5.752e+02, percent-clipped=2.0 2023-04-30 10:12:53,421 INFO [train.py:904] (4/8) Epoch 16, batch 8900, loss[loss=0.1825, simple_loss=0.2785, pruned_loss=0.04328, over 16618.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2718, pruned_loss=0.04091, over 3058073.08 frames. ], batch size: 62, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:12:54,936 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9335, 2.0523, 2.2903, 3.2702, 2.0615, 2.2730, 2.2502, 2.1366], device='cuda:4'), covar=tensor([0.1176, 0.3394, 0.2569, 0.0563, 0.4342, 0.2643, 0.3390, 0.3702], device='cuda:4'), in_proj_covar=tensor([0.0372, 0.0413, 0.0345, 0.0310, 0.0419, 0.0473, 0.0382, 0.0479], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:14:54,865 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7437, 3.2990, 2.8951, 5.1280, 3.9103, 4.6068, 1.8279, 3.3484], device='cuda:4'), covar=tensor([0.1278, 0.0609, 0.1038, 0.0101, 0.0173, 0.0284, 0.1488, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0162, 0.0184, 0.0164, 0.0194, 0.0205, 0.0189, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 10:14:59,662 INFO [train.py:904] (4/8) Epoch 16, batch 8950, loss[loss=0.1633, simple_loss=0.2563, pruned_loss=0.03516, over 16653.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2715, pruned_loss=0.04123, over 3057820.47 frames. ], batch size: 134, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:15:04,531 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5692, 2.3069, 2.3674, 4.3472, 2.2741, 2.7290, 2.3243, 2.4807], device='cuda:4'), covar=tensor([0.1043, 0.3560, 0.2772, 0.0354, 0.4035, 0.2565, 0.3651, 0.3231], device='cuda:4'), in_proj_covar=tensor([0.0370, 0.0411, 0.0343, 0.0310, 0.0417, 0.0471, 0.0381, 0.0477], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:15:29,188 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161216.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:15:47,454 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.408e+02 2.276e+02 2.813e+02 3.338e+02 5.711e+02, threshold=5.626e+02, percent-clipped=0.0 2023-04-30 10:16:01,868 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161231.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:16:46,879 INFO [train.py:904] (4/8) Epoch 16, batch 9000, loss[loss=0.1661, simple_loss=0.2583, pruned_loss=0.03695, over 16428.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2685, pruned_loss=0.04013, over 3056301.96 frames. ], batch size: 146, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:16:46,880 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 10:16:56,909 INFO [train.py:938] (4/8) Epoch 16, validation: loss=0.1491, simple_loss=0.2531, pruned_loss=0.02259, over 944034.00 frames. 2023-04-30 10:16:56,910 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 10:17:08,071 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161257.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:17:49,250 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161277.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:18:20,878 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161292.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:18:40,780 INFO [train.py:904] (4/8) Epoch 16, batch 9050, loss[loss=0.1729, simple_loss=0.2615, pruned_loss=0.04214, over 16697.00 frames. ], tot_loss[loss=0.175, simple_loss=0.269, pruned_loss=0.04052, over 3051415.55 frames. ], batch size: 134, lr: 4.20e-03, grad_scale: 8.0 2023-04-30 10:19:19,365 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161321.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:19:27,691 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.612e+02 2.317e+02 2.661e+02 3.212e+02 6.846e+02, threshold=5.322e+02, percent-clipped=5.0 2023-04-30 10:20:02,808 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5355, 4.3647, 4.6049, 4.7479, 4.9150, 4.4253, 4.9203, 4.9105], device='cuda:4'), covar=tensor([0.1618, 0.1111, 0.1470, 0.0652, 0.0495, 0.0940, 0.0489, 0.0671], device='cuda:4'), in_proj_covar=tensor([0.0555, 0.0684, 0.0803, 0.0694, 0.0521, 0.0546, 0.0555, 0.0653], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:20:22,817 INFO [train.py:904] (4/8) Epoch 16, batch 9100, loss[loss=0.1707, simple_loss=0.2699, pruned_loss=0.03574, over 16899.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.268, pruned_loss=0.04051, over 3059258.72 frames. ], batch size: 102, lr: 4.19e-03, grad_scale: 4.0 2023-04-30 10:21:13,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3544, 3.6632, 3.9237, 2.1391, 3.1956, 2.4959, 3.7056, 3.8255], device='cuda:4'), covar=tensor([0.0226, 0.0717, 0.0475, 0.1895, 0.0761, 0.0933, 0.0640, 0.0903], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0148, 0.0158, 0.0145, 0.0137, 0.0123, 0.0137, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 10:22:19,259 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161401.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:22:19,973 INFO [train.py:904] (4/8) Epoch 16, batch 9150, loss[loss=0.1769, simple_loss=0.2683, pruned_loss=0.04277, over 15485.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2691, pruned_loss=0.04065, over 3049129.23 frames. ], batch size: 191, lr: 4.19e-03, grad_scale: 4.0 2023-04-30 10:23:13,970 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.610e+02 2.279e+02 2.699e+02 3.193e+02 4.519e+02, threshold=5.398e+02, percent-clipped=0.0 2023-04-30 10:23:58,854 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-30 10:24:04,517 INFO [train.py:904] (4/8) Epoch 16, batch 9200, loss[loss=0.1703, simple_loss=0.2598, pruned_loss=0.04045, over 16831.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2646, pruned_loss=0.03979, over 3043505.28 frames. ], batch size: 124, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:24:20,232 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9182, 3.9858, 4.3083, 4.2698, 4.2984, 4.0538, 4.0540, 4.0265], device='cuda:4'), covar=tensor([0.0364, 0.0760, 0.0439, 0.0439, 0.0454, 0.0460, 0.0844, 0.0457], device='cuda:4'), in_proj_covar=tensor([0.0356, 0.0385, 0.0379, 0.0355, 0.0424, 0.0398, 0.0485, 0.0316], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 10:24:20,268 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161460.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:24:24,499 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161462.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:25:10,075 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6159, 3.0130, 3.2442, 1.9770, 2.8043, 2.0785, 3.0931, 3.1896], device='cuda:4'), covar=tensor([0.0298, 0.0829, 0.0494, 0.1936, 0.0819, 0.1044, 0.0725, 0.0914], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0148, 0.0158, 0.0145, 0.0137, 0.0123, 0.0137, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 10:25:42,685 INFO [train.py:904] (4/8) Epoch 16, batch 9250, loss[loss=0.1578, simple_loss=0.2411, pruned_loss=0.0373, over 12521.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2645, pruned_loss=0.0396, over 3062714.15 frames. ], batch size: 248, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:25:56,356 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8786, 1.2934, 1.6916, 1.6152, 1.8567, 1.9187, 1.6164, 1.8421], device='cuda:4'), covar=tensor([0.0256, 0.0390, 0.0202, 0.0286, 0.0263, 0.0187, 0.0375, 0.0121], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0179, 0.0166, 0.0167, 0.0179, 0.0136, 0.0181, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:26:03,685 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7809, 3.8116, 4.1309, 4.0962, 4.0882, 3.8877, 3.8882, 3.9133], device='cuda:4'), covar=tensor([0.0378, 0.0652, 0.0443, 0.0514, 0.0523, 0.0491, 0.0849, 0.0421], device='cuda:4'), in_proj_covar=tensor([0.0357, 0.0386, 0.0381, 0.0356, 0.0425, 0.0399, 0.0487, 0.0317], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 10:26:23,360 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161521.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:26:36,795 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.651e+02 2.444e+02 2.938e+02 3.576e+02 6.685e+02, threshold=5.875e+02, percent-clipped=5.0 2023-04-30 10:27:16,076 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9662, 4.9416, 4.6934, 4.2517, 4.8097, 2.0318, 4.5814, 4.6580], device='cuda:4'), covar=tensor([0.0080, 0.0088, 0.0190, 0.0294, 0.0097, 0.2168, 0.0128, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0128, 0.0173, 0.0156, 0.0148, 0.0188, 0.0161, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:27:23,337 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5448, 3.5389, 3.5128, 2.8344, 3.4092, 2.0660, 3.2458, 2.8240], device='cuda:4'), covar=tensor([0.0121, 0.0104, 0.0147, 0.0179, 0.0086, 0.2166, 0.0121, 0.0185], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0128, 0.0173, 0.0156, 0.0148, 0.0188, 0.0161, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:27:34,801 INFO [train.py:904] (4/8) Epoch 16, batch 9300, loss[loss=0.1465, simple_loss=0.239, pruned_loss=0.027, over 16694.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.263, pruned_loss=0.03926, over 3034457.06 frames. ], batch size: 134, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:27:45,404 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=161557.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:28:23,212 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161572.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:28:41,692 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7681, 3.7958, 4.0294, 1.8401, 4.1958, 4.2479, 3.2373, 3.0546], device='cuda:4'), covar=tensor([0.0653, 0.0191, 0.0180, 0.1212, 0.0054, 0.0105, 0.0332, 0.0429], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0102, 0.0087, 0.0133, 0.0071, 0.0111, 0.0121, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 10:28:54,331 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161587.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:28:56,659 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8279, 4.8041, 4.6221, 4.0906, 4.6904, 1.7213, 4.4491, 4.5757], device='cuda:4'), covar=tensor([0.0103, 0.0110, 0.0215, 0.0411, 0.0128, 0.2557, 0.0165, 0.0220], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0128, 0.0173, 0.0156, 0.0147, 0.0187, 0.0161, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:29:21,682 INFO [train.py:904] (4/8) Epoch 16, batch 9350, loss[loss=0.1667, simple_loss=0.2472, pruned_loss=0.04316, over 12818.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2626, pruned_loss=0.03921, over 3031492.63 frames. ], batch size: 250, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:29:28,430 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=161605.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:30:02,106 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=161621.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:30:12,559 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.368e+02 2.307e+02 2.790e+02 3.265e+02 7.040e+02, threshold=5.579e+02, percent-clipped=1.0 2023-04-30 10:31:03,992 INFO [train.py:904] (4/8) Epoch 16, batch 9400, loss[loss=0.1488, simple_loss=0.232, pruned_loss=0.03279, over 12279.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2624, pruned_loss=0.03891, over 3035701.76 frames. ], batch size: 246, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:31:20,407 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-30 10:31:39,351 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=161669.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:32:44,155 INFO [train.py:904] (4/8) Epoch 16, batch 9450, loss[loss=0.1711, simple_loss=0.265, pruned_loss=0.03862, over 16622.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.264, pruned_loss=0.03894, over 3038707.96 frames. ], batch size: 62, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:32:48,086 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4779, 4.1738, 4.2099, 4.6039, 4.7904, 4.3871, 4.7658, 4.7996], device='cuda:4'), covar=tensor([0.1673, 0.1490, 0.2576, 0.1011, 0.0785, 0.1216, 0.0946, 0.0943], device='cuda:4'), in_proj_covar=tensor([0.0553, 0.0683, 0.0799, 0.0690, 0.0521, 0.0546, 0.0555, 0.0648], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:33:33,849 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.598e+02 2.284e+02 2.576e+02 3.195e+02 6.155e+02, threshold=5.152e+02, percent-clipped=3.0 2023-04-30 10:34:23,901 INFO [train.py:904] (4/8) Epoch 16, batch 9500, loss[loss=0.1725, simple_loss=0.2657, pruned_loss=0.03963, over 16659.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2632, pruned_loss=0.0386, over 3038234.70 frames. ], batch size: 134, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:34:37,541 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161757.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:35:57,326 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-30 10:36:08,222 INFO [train.py:904] (4/8) Epoch 16, batch 9550, loss[loss=0.1937, simple_loss=0.292, pruned_loss=0.04771, over 16151.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2631, pruned_loss=0.03856, over 3046630.39 frames. ], batch size: 165, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:36:38,167 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161815.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:36:40,703 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161816.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:36:48,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8525, 4.9108, 5.3194, 5.3068, 5.2960, 5.0382, 4.9864, 4.7484], device='cuda:4'), covar=tensor([0.0332, 0.0599, 0.0412, 0.0343, 0.0440, 0.0313, 0.0815, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0355, 0.0382, 0.0378, 0.0351, 0.0420, 0.0394, 0.0482, 0.0313], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 10:37:00,497 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.628e+02 2.146e+02 2.559e+02 3.093e+02 5.125e+02, threshold=5.118e+02, percent-clipped=0.0 2023-04-30 10:37:20,921 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3473, 3.5612, 3.5987, 2.4826, 3.2330, 3.5816, 3.4322, 1.9914], device='cuda:4'), covar=tensor([0.0450, 0.0035, 0.0040, 0.0343, 0.0092, 0.0075, 0.0071, 0.0475], device='cuda:4'), in_proj_covar=tensor([0.0130, 0.0073, 0.0074, 0.0129, 0.0088, 0.0096, 0.0085, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 10:37:29,520 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6878, 1.8581, 2.2210, 2.6942, 2.6105, 3.0362, 1.9787, 2.9875], device='cuda:4'), covar=tensor([0.0178, 0.0434, 0.0338, 0.0234, 0.0277, 0.0150, 0.0436, 0.0135], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0177, 0.0164, 0.0165, 0.0178, 0.0134, 0.0179, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:37:51,434 INFO [train.py:904] (4/8) Epoch 16, batch 9600, loss[loss=0.1598, simple_loss=0.2473, pruned_loss=0.03614, over 12393.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2647, pruned_loss=0.03954, over 3024171.68 frames. ], batch size: 248, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:38:25,694 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0248, 3.9969, 3.9052, 3.3233, 3.9566, 1.8231, 3.7995, 3.5746], device='cuda:4'), covar=tensor([0.0092, 0.0093, 0.0168, 0.0224, 0.0098, 0.2449, 0.0122, 0.0228], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0128, 0.0172, 0.0156, 0.0147, 0.0186, 0.0160, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:38:29,893 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=161872.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:38:31,896 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2033, 1.5161, 1.9189, 2.2198, 2.2835, 2.5083, 1.6652, 2.4280], device='cuda:4'), covar=tensor([0.0240, 0.0466, 0.0300, 0.0279, 0.0310, 0.0174, 0.0459, 0.0128], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0176, 0.0164, 0.0165, 0.0178, 0.0134, 0.0179, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:38:38,422 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161876.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:39:02,394 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=161887.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:39:37,499 INFO [train.py:904] (4/8) Epoch 16, batch 9650, loss[loss=0.1566, simple_loss=0.2448, pruned_loss=0.03427, over 17197.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2662, pruned_loss=0.03989, over 3020478.86 frames. ], batch size: 44, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:40:11,745 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6711, 2.6772, 1.8365, 2.8592, 2.0385, 2.8453, 2.0507, 2.4310], device='cuda:4'), covar=tensor([0.0283, 0.0350, 0.1332, 0.0236, 0.0733, 0.0439, 0.1249, 0.0685], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0184, 0.0140, 0.0166, 0.0199, 0.0194, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 10:40:22,071 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=161920.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:40:36,068 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.427e+02 2.311e+02 2.824e+02 3.460e+02 5.730e+02, threshold=5.647e+02, percent-clipped=1.0 2023-04-30 10:40:49,470 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=161935.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:41:23,666 INFO [train.py:904] (4/8) Epoch 16, batch 9700, loss[loss=0.1799, simple_loss=0.2713, pruned_loss=0.04424, over 16962.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2652, pruned_loss=0.03959, over 3027581.88 frames. ], batch size: 109, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:42:01,848 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6506, 2.0650, 1.7752, 1.8396, 2.4132, 2.0675, 2.2335, 2.5408], device='cuda:4'), covar=tensor([0.0132, 0.0376, 0.0466, 0.0476, 0.0238, 0.0378, 0.0180, 0.0238], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0215, 0.0209, 0.0208, 0.0214, 0.0215, 0.0213, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:42:48,080 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-30 10:43:08,497 INFO [train.py:904] (4/8) Epoch 16, batch 9750, loss[loss=0.175, simple_loss=0.2761, pruned_loss=0.037, over 16291.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2646, pruned_loss=0.0397, over 3041813.84 frames. ], batch size: 146, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:43:44,078 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-30 10:43:58,494 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.156e+02 2.615e+02 3.105e+02 5.589e+02, threshold=5.230e+02, percent-clipped=0.0 2023-04-30 10:44:44,851 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0193, 1.8098, 1.6569, 1.5045, 1.9388, 1.6035, 1.6271, 1.9737], device='cuda:4'), covar=tensor([0.0150, 0.0279, 0.0374, 0.0342, 0.0199, 0.0269, 0.0164, 0.0215], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0215, 0.0208, 0.0207, 0.0213, 0.0213, 0.0212, 0.0205], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:44:46,280 INFO [train.py:904] (4/8) Epoch 16, batch 9800, loss[loss=0.1688, simple_loss=0.2614, pruned_loss=0.03813, over 12188.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2646, pruned_loss=0.03871, over 3061289.61 frames. ], batch size: 248, lr: 4.19e-03, grad_scale: 8.0 2023-04-30 10:44:56,983 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162057.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:45:34,286 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=162077.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:46:31,255 INFO [train.py:904] (4/8) Epoch 16, batch 9850, loss[loss=0.1749, simple_loss=0.271, pruned_loss=0.03943, over 16166.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2657, pruned_loss=0.03837, over 3051801.15 frames. ], batch size: 165, lr: 4.18e-03, grad_scale: 8.0 2023-04-30 10:46:38,614 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=162105.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:46:59,780 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162116.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:47:23,357 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.414e+02 2.292e+02 2.812e+02 3.288e+02 4.754e+02, threshold=5.623e+02, percent-clipped=0.0 2023-04-30 10:47:51,495 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=162138.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:48:22,258 INFO [train.py:904] (4/8) Epoch 16, batch 9900, loss[loss=0.1627, simple_loss=0.2666, pruned_loss=0.0294, over 17156.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2659, pruned_loss=0.03824, over 3037583.79 frames. ], batch size: 48, lr: 4.18e-03, grad_scale: 8.0 2023-04-30 10:48:33,213 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-30 10:48:52,696 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=162164.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:49:08,581 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=162171.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:49:18,953 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-30 10:49:54,941 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5242, 4.6209, 4.7856, 4.6084, 4.6780, 5.1604, 4.7240, 4.4884], device='cuda:4'), covar=tensor([0.1150, 0.1927, 0.1992, 0.1976, 0.2377, 0.0956, 0.1463, 0.2295], device='cuda:4'), in_proj_covar=tensor([0.0356, 0.0514, 0.0562, 0.0432, 0.0573, 0.0598, 0.0447, 0.0578], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 10:49:57,193 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6787, 4.6958, 4.5508, 4.1716, 4.2196, 4.6045, 4.4571, 4.3091], device='cuda:4'), covar=tensor([0.0583, 0.0619, 0.0309, 0.0301, 0.0948, 0.0555, 0.0404, 0.0696], device='cuda:4'), in_proj_covar=tensor([0.0254, 0.0351, 0.0296, 0.0284, 0.0303, 0.0329, 0.0203, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-30 10:50:22,156 INFO [train.py:904] (4/8) Epoch 16, batch 9950, loss[loss=0.1579, simple_loss=0.2559, pruned_loss=0.02994, over 16624.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2686, pruned_loss=0.03861, over 3053551.75 frames. ], batch size: 57, lr: 4.18e-03, grad_scale: 8.0 2023-04-30 10:51:26,523 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 2.161e+02 2.630e+02 3.300e+02 8.357e+02, threshold=5.260e+02, percent-clipped=1.0 2023-04-30 10:51:36,430 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=162231.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:52:23,901 INFO [train.py:904] (4/8) Epoch 16, batch 10000, loss[loss=0.1514, simple_loss=0.2448, pruned_loss=0.02902, over 17026.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2669, pruned_loss=0.03826, over 3071275.12 frames. ], batch size: 55, lr: 4.18e-03, grad_scale: 8.0 2023-04-30 10:52:50,895 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3183, 4.3671, 4.2304, 3.8566, 3.9368, 4.3016, 3.9785, 4.0350], device='cuda:4'), covar=tensor([0.0540, 0.0418, 0.0268, 0.0280, 0.0710, 0.0389, 0.0589, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0254, 0.0350, 0.0295, 0.0284, 0.0303, 0.0329, 0.0202, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-30 10:53:46,611 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=162292.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:53:55,222 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8029, 3.7542, 3.9510, 3.7838, 3.8771, 4.3006, 3.9781, 3.7137], device='cuda:4'), covar=tensor([0.2034, 0.2352, 0.2184, 0.2373, 0.2609, 0.1633, 0.1465, 0.2685], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0512, 0.0561, 0.0431, 0.0573, 0.0596, 0.0446, 0.0577], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 10:54:04,077 INFO [train.py:904] (4/8) Epoch 16, batch 10050, loss[loss=0.1848, simple_loss=0.2826, pruned_loss=0.04355, over 16192.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2671, pruned_loss=0.0382, over 3072035.51 frames. ], batch size: 165, lr: 4.18e-03, grad_scale: 8.0 2023-04-30 10:54:54,659 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.562e+02 2.225e+02 2.711e+02 3.158e+02 7.314e+02, threshold=5.421e+02, percent-clipped=6.0 2023-04-30 10:55:03,391 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3894, 3.4732, 2.0897, 3.8337, 2.5337, 3.7860, 2.1232, 2.8659], device='cuda:4'), covar=tensor([0.0278, 0.0375, 0.1514, 0.0184, 0.0830, 0.0548, 0.1477, 0.0711], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0161, 0.0183, 0.0138, 0.0165, 0.0197, 0.0193, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 10:55:38,456 INFO [train.py:904] (4/8) Epoch 16, batch 10100, loss[loss=0.1573, simple_loss=0.2463, pruned_loss=0.03412, over 15422.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2674, pruned_loss=0.03829, over 3067115.86 frames. ], batch size: 191, lr: 4.18e-03, grad_scale: 8.0 2023-04-30 10:56:29,750 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-30 10:57:23,037 INFO [train.py:904] (4/8) Epoch 17, batch 0, loss[loss=0.1978, simple_loss=0.2864, pruned_loss=0.05457, over 16448.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2864, pruned_loss=0.05457, over 16448.00 frames. ], batch size: 68, lr: 4.05e-03, grad_scale: 8.0 2023-04-30 10:57:23,037 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 10:57:30,745 INFO [train.py:938] (4/8) Epoch 17, validation: loss=0.1481, simple_loss=0.2518, pruned_loss=0.02217, over 944034.00 frames. 2023-04-30 10:57:30,746 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 10:57:39,296 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2693, 5.6194, 5.3477, 5.4226, 5.0732, 4.9869, 5.1376, 5.6617], device='cuda:4'), covar=tensor([0.1379, 0.0959, 0.1186, 0.0781, 0.0951, 0.0844, 0.1075, 0.1003], device='cuda:4'), in_proj_covar=tensor([0.0582, 0.0715, 0.0578, 0.0526, 0.0452, 0.0464, 0.0595, 0.0553], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 10:58:09,504 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.615e+02 2.232e+02 2.721e+02 3.359e+02 8.921e+02, threshold=5.442e+02, percent-clipped=3.0 2023-04-30 10:58:14,192 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=162433.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:58:38,651 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-30 10:58:39,851 INFO [train.py:904] (4/8) Epoch 17, batch 50, loss[loss=0.1844, simple_loss=0.2755, pruned_loss=0.04668, over 16535.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2802, pruned_loss=0.05575, over 735029.79 frames. ], batch size: 68, lr: 4.05e-03, grad_scale: 2.0 2023-04-30 10:59:06,486 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162471.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 10:59:41,377 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6702, 2.4624, 2.3192, 3.7183, 2.9117, 3.7801, 1.3202, 2.7345], device='cuda:4'), covar=tensor([0.1469, 0.0722, 0.1191, 0.0164, 0.0157, 0.0402, 0.1714, 0.0842], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0162, 0.0185, 0.0164, 0.0190, 0.0206, 0.0191, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 10:59:47,951 INFO [train.py:904] (4/8) Epoch 17, batch 100, loss[loss=0.1655, simple_loss=0.2582, pruned_loss=0.0364, over 17119.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.273, pruned_loss=0.05181, over 1315304.13 frames. ], batch size: 48, lr: 4.05e-03, grad_scale: 2.0 2023-04-30 10:59:58,046 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-30 11:00:00,634 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=162511.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:00:12,329 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=162519.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:00:22,584 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9637, 2.9807, 3.1343, 1.6593, 3.2449, 3.2819, 2.6552, 2.4799], device='cuda:4'), covar=tensor([0.0965, 0.0218, 0.0198, 0.1364, 0.0102, 0.0203, 0.0477, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0105, 0.0090, 0.0138, 0.0072, 0.0115, 0.0125, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 11:00:26,385 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.493e+02 2.355e+02 2.682e+02 3.198e+02 6.473e+02, threshold=5.364e+02, percent-clipped=1.0 2023-04-30 11:00:56,545 INFO [train.py:904] (4/8) Epoch 17, batch 150, loss[loss=0.1829, simple_loss=0.2674, pruned_loss=0.04922, over 16436.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2699, pruned_loss=0.05021, over 1761417.23 frames. ], batch size: 75, lr: 4.05e-03, grad_scale: 2.0 2023-04-30 11:01:23,614 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=162572.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 11:01:44,469 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=162587.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:02:05,937 INFO [train.py:904] (4/8) Epoch 17, batch 200, loss[loss=0.1878, simple_loss=0.2671, pruned_loss=0.05425, over 16233.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2694, pruned_loss=0.04994, over 2099744.65 frames. ], batch size: 165, lr: 4.05e-03, grad_scale: 2.0 2023-04-30 11:02:25,611 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4833, 3.5251, 3.8645, 2.6606, 3.4465, 3.9138, 3.6507, 2.2384], device='cuda:4'), covar=tensor([0.0481, 0.0294, 0.0045, 0.0346, 0.0109, 0.0088, 0.0083, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0077, 0.0076, 0.0132, 0.0089, 0.0099, 0.0087, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:02:43,589 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.632e+02 2.407e+02 2.768e+02 3.167e+02 5.394e+02, threshold=5.535e+02, percent-clipped=0.0 2023-04-30 11:03:05,790 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1023, 5.6621, 5.8004, 5.5518, 5.7043, 6.2122, 5.6919, 5.4318], device='cuda:4'), covar=tensor([0.0980, 0.1922, 0.2373, 0.2275, 0.2740, 0.1058, 0.1425, 0.2262], device='cuda:4'), in_proj_covar=tensor([0.0373, 0.0542, 0.0596, 0.0455, 0.0605, 0.0627, 0.0469, 0.0605], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:03:12,317 INFO [train.py:904] (4/8) Epoch 17, batch 250, loss[loss=0.169, simple_loss=0.2527, pruned_loss=0.04266, over 16804.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2675, pruned_loss=0.04989, over 2369357.30 frames. ], batch size: 102, lr: 4.05e-03, grad_scale: 2.0 2023-04-30 11:03:22,253 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3049, 5.7319, 5.4807, 5.5683, 5.1782, 5.1543, 5.1722, 5.8335], device='cuda:4'), covar=tensor([0.1503, 0.1053, 0.1078, 0.0836, 0.1049, 0.0693, 0.1114, 0.0990], device='cuda:4'), in_proj_covar=tensor([0.0608, 0.0750, 0.0609, 0.0551, 0.0473, 0.0482, 0.0626, 0.0577], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:04:20,330 INFO [train.py:904] (4/8) Epoch 17, batch 300, loss[loss=0.177, simple_loss=0.2524, pruned_loss=0.05079, over 16787.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2646, pruned_loss=0.04808, over 2589658.72 frames. ], batch size: 102, lr: 4.05e-03, grad_scale: 2.0 2023-04-30 11:04:55,530 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9847, 2.1775, 2.3542, 2.7340, 2.1157, 3.1823, 1.6630, 2.7481], device='cuda:4'), covar=tensor([0.1126, 0.0728, 0.1033, 0.0163, 0.0137, 0.0372, 0.1463, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0164, 0.0187, 0.0168, 0.0194, 0.0209, 0.0193, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 11:04:59,732 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.569e+02 2.237e+02 2.646e+02 3.133e+02 7.879e+02, threshold=5.293e+02, percent-clipped=2.0 2023-04-30 11:05:03,720 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162733.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:05:29,583 INFO [train.py:904] (4/8) Epoch 17, batch 350, loss[loss=0.1838, simple_loss=0.2525, pruned_loss=0.05757, over 16771.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2624, pruned_loss=0.04679, over 2747247.75 frames. ], batch size: 89, lr: 4.05e-03, grad_scale: 2.0 2023-04-30 11:05:45,429 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5224, 2.3614, 2.4652, 4.2371, 2.3212, 2.7520, 2.4423, 2.5675], device='cuda:4'), covar=tensor([0.1180, 0.3525, 0.2714, 0.0542, 0.3805, 0.2321, 0.3404, 0.3096], device='cuda:4'), in_proj_covar=tensor([0.0379, 0.0419, 0.0353, 0.0319, 0.0426, 0.0480, 0.0390, 0.0488], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:06:07,010 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3492, 5.7403, 5.2885, 5.7149, 5.2539, 4.9960, 5.3250, 5.8255], device='cuda:4'), covar=tensor([0.2472, 0.1842, 0.2886, 0.1413, 0.1581, 0.1357, 0.2203, 0.2104], device='cuda:4'), in_proj_covar=tensor([0.0617, 0.0762, 0.0619, 0.0559, 0.0480, 0.0489, 0.0635, 0.0586], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:06:07,975 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=162781.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:06:34,848 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1157, 5.0576, 4.9857, 4.4329, 4.5530, 4.9853, 4.9219, 4.6243], device='cuda:4'), covar=tensor([0.0530, 0.0524, 0.0291, 0.0353, 0.1120, 0.0456, 0.0322, 0.0742], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0374, 0.0313, 0.0302, 0.0323, 0.0350, 0.0216, 0.0375], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:06:36,763 INFO [train.py:904] (4/8) Epoch 17, batch 400, loss[loss=0.1841, simple_loss=0.2628, pruned_loss=0.05273, over 16535.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2613, pruned_loss=0.04664, over 2878871.89 frames. ], batch size: 68, lr: 4.05e-03, grad_scale: 4.0 2023-04-30 11:06:37,427 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-30 11:07:10,943 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=162827.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:07:15,898 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.410e+02 2.145e+02 2.685e+02 3.328e+02 1.079e+03, threshold=5.370e+02, percent-clipped=6.0 2023-04-30 11:07:34,451 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-30 11:07:35,483 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6503, 4.4983, 4.6312, 4.8422, 4.9968, 4.5176, 4.8969, 4.9828], device='cuda:4'), covar=tensor([0.1808, 0.1556, 0.1733, 0.1087, 0.0747, 0.1011, 0.1579, 0.1744], device='cuda:4'), in_proj_covar=tensor([0.0589, 0.0728, 0.0859, 0.0729, 0.0553, 0.0579, 0.0595, 0.0688], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:07:42,831 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6297, 3.8029, 4.1117, 2.1704, 3.3934, 2.7001, 3.9716, 4.0040], device='cuda:4'), covar=tensor([0.0252, 0.0806, 0.0441, 0.1903, 0.0685, 0.0895, 0.0570, 0.1001], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0151, 0.0160, 0.0149, 0.0139, 0.0126, 0.0139, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 11:07:46,582 INFO [train.py:904] (4/8) Epoch 17, batch 450, loss[loss=0.1693, simple_loss=0.2599, pruned_loss=0.03935, over 16596.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2595, pruned_loss=0.04584, over 2980584.54 frames. ], batch size: 62, lr: 4.05e-03, grad_scale: 4.0 2023-04-30 11:08:06,775 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=162867.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 11:08:34,357 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162887.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:08:35,581 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=162888.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:08:55,330 INFO [train.py:904] (4/8) Epoch 17, batch 500, loss[loss=0.1471, simple_loss=0.2291, pruned_loss=0.03255, over 16793.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2581, pruned_loss=0.04513, over 3058673.85 frames. ], batch size: 39, lr: 4.05e-03, grad_scale: 4.0 2023-04-30 11:09:32,593 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.536e+02 2.133e+02 2.499e+02 3.161e+02 5.064e+02, threshold=4.998e+02, percent-clipped=0.0 2023-04-30 11:09:39,866 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=162935.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:10:01,801 INFO [train.py:904] (4/8) Epoch 17, batch 550, loss[loss=0.1753, simple_loss=0.265, pruned_loss=0.04279, over 17181.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2569, pruned_loss=0.04436, over 3107971.27 frames. ], batch size: 46, lr: 4.05e-03, grad_scale: 4.0 2023-04-30 11:11:10,218 INFO [train.py:904] (4/8) Epoch 17, batch 600, loss[loss=0.1859, simple_loss=0.2536, pruned_loss=0.05904, over 16858.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2567, pruned_loss=0.04501, over 3157778.60 frames. ], batch size: 109, lr: 4.05e-03, grad_scale: 4.0 2023-04-30 11:11:47,292 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.692e+02 2.256e+02 2.602e+02 3.226e+02 5.329e+02, threshold=5.204e+02, percent-clipped=2.0 2023-04-30 11:11:49,676 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-04-30 11:12:16,979 INFO [train.py:904] (4/8) Epoch 17, batch 650, loss[loss=0.1656, simple_loss=0.2438, pruned_loss=0.04366, over 16578.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2548, pruned_loss=0.04499, over 3183487.24 frames. ], batch size: 68, lr: 4.05e-03, grad_scale: 4.0 2023-04-30 11:13:09,349 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=163090.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:13:18,939 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9996, 2.0637, 2.5068, 2.8936, 2.7855, 2.8896, 2.0590, 3.0985], device='cuda:4'), covar=tensor([0.0156, 0.0396, 0.0264, 0.0213, 0.0258, 0.0237, 0.0439, 0.0131], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0185, 0.0172, 0.0176, 0.0186, 0.0142, 0.0186, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:13:25,534 INFO [train.py:904] (4/8) Epoch 17, batch 700, loss[loss=0.1767, simple_loss=0.2641, pruned_loss=0.04469, over 16792.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.254, pruned_loss=0.04426, over 3208126.66 frames. ], batch size: 57, lr: 4.05e-03, grad_scale: 4.0 2023-04-30 11:13:47,796 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-30 11:14:04,489 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.287e+02 2.752e+02 3.270e+02 5.446e+02, threshold=5.503e+02, percent-clipped=2.0 2023-04-30 11:14:34,382 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=163151.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:14:35,142 INFO [train.py:904] (4/8) Epoch 17, batch 750, loss[loss=0.1771, simple_loss=0.2538, pruned_loss=0.05013, over 16796.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2542, pruned_loss=0.04422, over 3238249.40 frames. ], batch size: 116, lr: 4.05e-03, grad_scale: 4.0 2023-04-30 11:14:57,182 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=163167.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 11:15:01,183 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7726, 5.1099, 4.8781, 4.8921, 4.6289, 4.6240, 4.5931, 5.1639], device='cuda:4'), covar=tensor([0.1206, 0.0951, 0.1034, 0.0783, 0.0788, 0.1058, 0.1120, 0.0888], device='cuda:4'), in_proj_covar=tensor([0.0632, 0.0781, 0.0634, 0.0573, 0.0491, 0.0497, 0.0649, 0.0602], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:15:17,982 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=163183.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:15:44,388 INFO [train.py:904] (4/8) Epoch 17, batch 800, loss[loss=0.191, simple_loss=0.2703, pruned_loss=0.05587, over 16776.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2541, pruned_loss=0.0436, over 3250270.54 frames. ], batch size: 124, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:15:47,051 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-30 11:16:03,279 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=163215.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:16:23,806 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.244e+02 2.209e+02 2.648e+02 3.284e+02 7.557e+02, threshold=5.297e+02, percent-clipped=2.0 2023-04-30 11:16:36,599 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=163240.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:16:53,792 INFO [train.py:904] (4/8) Epoch 17, batch 850, loss[loss=0.2013, simple_loss=0.2729, pruned_loss=0.06481, over 12118.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2534, pruned_loss=0.04307, over 3257685.94 frames. ], batch size: 246, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:18:01,087 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=163301.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:18:01,854 INFO [train.py:904] (4/8) Epoch 17, batch 900, loss[loss=0.1517, simple_loss=0.2299, pruned_loss=0.03674, over 16460.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2533, pruned_loss=0.04297, over 3271499.50 frames. ], batch size: 146, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:18:26,592 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3422, 4.4195, 4.7560, 4.7327, 4.7592, 4.4347, 4.4643, 4.3179], device='cuda:4'), covar=tensor([0.0369, 0.0631, 0.0388, 0.0394, 0.0436, 0.0413, 0.0768, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0387, 0.0418, 0.0408, 0.0383, 0.0456, 0.0431, 0.0525, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 11:18:40,393 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.550e+02 2.201e+02 2.666e+02 3.227e+02 4.456e+02, threshold=5.332e+02, percent-clipped=0.0 2023-04-30 11:18:42,635 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0831, 2.4931, 2.0250, 2.3016, 2.9042, 2.7069, 2.9959, 3.0280], device='cuda:4'), covar=tensor([0.0238, 0.0403, 0.0561, 0.0463, 0.0278, 0.0322, 0.0279, 0.0291], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0230, 0.0221, 0.0220, 0.0228, 0.0230, 0.0234, 0.0222], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:18:52,680 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.11 vs. limit=2.0 2023-04-30 11:19:09,573 INFO [train.py:904] (4/8) Epoch 17, batch 950, loss[loss=0.1508, simple_loss=0.2501, pruned_loss=0.02571, over 17120.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2539, pruned_loss=0.04302, over 3287226.06 frames. ], batch size: 49, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:19:33,948 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0420, 5.0746, 5.5796, 5.5087, 5.5316, 5.1893, 5.1348, 4.9477], device='cuda:4'), covar=tensor([0.0352, 0.0534, 0.0317, 0.0448, 0.0516, 0.0381, 0.0952, 0.0399], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0419, 0.0411, 0.0385, 0.0459, 0.0432, 0.0528, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 11:20:00,132 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6537, 3.7087, 2.2082, 3.9335, 2.8978, 3.9392, 2.3890, 3.0000], device='cuda:4'), covar=tensor([0.0209, 0.0366, 0.1414, 0.0281, 0.0631, 0.0614, 0.1164, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0171, 0.0192, 0.0152, 0.0173, 0.0212, 0.0201, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 11:20:13,017 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-30 11:20:17,969 INFO [train.py:904] (4/8) Epoch 17, batch 1000, loss[loss=0.161, simple_loss=0.2421, pruned_loss=0.03989, over 16688.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2523, pruned_loss=0.04316, over 3299137.89 frames. ], batch size: 89, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:20:54,826 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.618e+02 2.252e+02 2.741e+02 3.139e+02 5.661e+02, threshold=5.483e+02, percent-clipped=2.0 2023-04-30 11:20:59,913 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6046, 3.4422, 2.7641, 2.1804, 2.2681, 2.2333, 3.5691, 3.0925], device='cuda:4'), covar=tensor([0.2533, 0.0660, 0.1554, 0.2645, 0.2723, 0.2044, 0.0491, 0.1551], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0262, 0.0295, 0.0298, 0.0286, 0.0241, 0.0281, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:21:18,941 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=163446.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:21:26,450 INFO [train.py:904] (4/8) Epoch 17, batch 1050, loss[loss=0.156, simple_loss=0.2326, pruned_loss=0.03965, over 16728.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2518, pruned_loss=0.04271, over 3303106.66 frames. ], batch size: 83, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:22:10,622 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=163483.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:22:36,965 INFO [train.py:904] (4/8) Epoch 17, batch 1100, loss[loss=0.1548, simple_loss=0.2521, pruned_loss=0.02875, over 17112.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2514, pruned_loss=0.04231, over 3309405.18 frames. ], batch size: 48, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:22:40,157 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.85 vs. limit=5.0 2023-04-30 11:23:08,666 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8658, 4.6360, 4.9209, 5.1353, 5.3214, 4.6459, 5.2984, 5.3034], device='cuda:4'), covar=tensor([0.1950, 0.1327, 0.1819, 0.0819, 0.0580, 0.1070, 0.0571, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0617, 0.0765, 0.0903, 0.0769, 0.0579, 0.0610, 0.0622, 0.0721], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:23:14,822 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.362e+02 2.061e+02 2.430e+02 2.825e+02 4.071e+02, threshold=4.861e+02, percent-clipped=0.0 2023-04-30 11:23:16,186 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=163531.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:23:43,925 INFO [train.py:904] (4/8) Epoch 17, batch 1150, loss[loss=0.1558, simple_loss=0.2432, pruned_loss=0.0342, over 16843.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.251, pruned_loss=0.04135, over 3318727.49 frames. ], batch size: 42, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:23:46,877 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0776, 4.1709, 2.8620, 4.8098, 3.4634, 4.8054, 3.1531, 3.6184], device='cuda:4'), covar=tensor([0.0272, 0.0355, 0.1274, 0.0233, 0.0667, 0.0426, 0.1118, 0.0523], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0173, 0.0193, 0.0154, 0.0174, 0.0214, 0.0201, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 11:24:34,585 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9226, 2.0248, 2.5103, 2.9272, 2.7602, 3.3164, 2.2121, 3.2216], device='cuda:4'), covar=tensor([0.0199, 0.0421, 0.0286, 0.0246, 0.0258, 0.0158, 0.0406, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0186, 0.0174, 0.0176, 0.0187, 0.0144, 0.0187, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:24:43,075 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=163596.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:24:52,260 INFO [train.py:904] (4/8) Epoch 17, batch 1200, loss[loss=0.1539, simple_loss=0.2506, pruned_loss=0.02857, over 17198.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2506, pruned_loss=0.04077, over 3322081.09 frames. ], batch size: 44, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:25:03,572 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5546, 3.5277, 3.8607, 2.8328, 3.4059, 3.8648, 3.6722, 2.3159], device='cuda:4'), covar=tensor([0.0429, 0.0194, 0.0043, 0.0308, 0.0101, 0.0076, 0.0071, 0.0417], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0078, 0.0077, 0.0132, 0.0091, 0.0101, 0.0090, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:25:12,719 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.78 vs. limit=5.0 2023-04-30 11:25:29,957 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 2.305e+02 2.671e+02 3.522e+02 8.114e+02, threshold=5.342e+02, percent-clipped=7.0 2023-04-30 11:25:58,589 INFO [train.py:904] (4/8) Epoch 17, batch 1250, loss[loss=0.1889, simple_loss=0.257, pruned_loss=0.06037, over 16910.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2504, pruned_loss=0.0419, over 3298254.49 frames. ], batch size: 116, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:27:06,160 INFO [train.py:904] (4/8) Epoch 17, batch 1300, loss[loss=0.1589, simple_loss=0.2379, pruned_loss=0.03993, over 15637.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2507, pruned_loss=0.04197, over 3302675.28 frames. ], batch size: 191, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:27:44,997 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.402e+02 2.213e+02 2.483e+02 3.083e+02 7.276e+02, threshold=4.967e+02, percent-clipped=1.0 2023-04-30 11:27:48,630 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-30 11:28:08,492 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=163746.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:28:16,451 INFO [train.py:904] (4/8) Epoch 17, batch 1350, loss[loss=0.1307, simple_loss=0.2201, pruned_loss=0.02064, over 16945.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2506, pruned_loss=0.04149, over 3291258.54 frames. ], batch size: 41, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:28:38,645 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5377, 3.6317, 1.8469, 3.8633, 2.7893, 3.7731, 1.8936, 2.7424], device='cuda:4'), covar=tensor([0.0279, 0.0359, 0.1787, 0.0282, 0.0706, 0.0740, 0.1805, 0.0736], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0196, 0.0156, 0.0176, 0.0217, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 11:29:15,148 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=163794.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:29:25,094 INFO [train.py:904] (4/8) Epoch 17, batch 1400, loss[loss=0.1856, simple_loss=0.262, pruned_loss=0.05461, over 16792.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2509, pruned_loss=0.04176, over 3303004.04 frames. ], batch size: 102, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:30:05,130 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.424e+02 2.111e+02 2.569e+02 3.362e+02 6.310e+02, threshold=5.138e+02, percent-clipped=5.0 2023-04-30 11:30:36,590 INFO [train.py:904] (4/8) Epoch 17, batch 1450, loss[loss=0.1758, simple_loss=0.2444, pruned_loss=0.05359, over 16749.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2497, pruned_loss=0.0417, over 3310533.50 frames. ], batch size: 124, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:30:42,238 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.52 vs. limit=5.0 2023-04-30 11:31:02,191 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.43 vs. limit=5.0 2023-04-30 11:31:38,011 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=163896.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:31:45,953 INFO [train.py:904] (4/8) Epoch 17, batch 1500, loss[loss=0.182, simple_loss=0.2514, pruned_loss=0.05634, over 16319.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2502, pruned_loss=0.04196, over 3310631.47 frames. ], batch size: 145, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:31:53,070 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-30 11:32:16,971 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4326, 2.3431, 2.3866, 4.3032, 2.3075, 2.7666, 2.3656, 2.5315], device='cuda:4'), covar=tensor([0.1201, 0.3368, 0.2703, 0.0556, 0.3867, 0.2315, 0.3391, 0.2991], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0426, 0.0359, 0.0328, 0.0432, 0.0493, 0.0397, 0.0501], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:32:24,618 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.269e+02 2.188e+02 2.579e+02 3.200e+02 4.820e+02, threshold=5.157e+02, percent-clipped=0.0 2023-04-30 11:32:40,189 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 11:32:45,102 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=163944.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:32:56,202 INFO [train.py:904] (4/8) Epoch 17, batch 1550, loss[loss=0.1657, simple_loss=0.2469, pruned_loss=0.04227, over 16847.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2527, pruned_loss=0.04255, over 3315106.88 frames. ], batch size: 96, lr: 4.04e-03, grad_scale: 8.0 2023-04-30 11:33:00,256 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1686, 2.1123, 2.2351, 3.9229, 2.1779, 2.5069, 2.1810, 2.2940], device='cuda:4'), covar=tensor([0.1306, 0.3545, 0.2648, 0.0568, 0.3651, 0.2370, 0.3523, 0.2885], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0426, 0.0360, 0.0328, 0.0432, 0.0493, 0.0397, 0.0501], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:33:14,964 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-30 11:34:07,075 INFO [train.py:904] (4/8) Epoch 17, batch 1600, loss[loss=0.1834, simple_loss=0.257, pruned_loss=0.05488, over 16495.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2544, pruned_loss=0.04318, over 3319679.74 frames. ], batch size: 75, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:34:45,102 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.680e+02 2.274e+02 2.815e+02 3.289e+02 5.501e+02, threshold=5.629e+02, percent-clipped=1.0 2023-04-30 11:35:10,886 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0223, 1.9985, 2.5624, 2.9865, 2.7954, 3.4371, 2.3098, 3.4280], device='cuda:4'), covar=tensor([0.0210, 0.0470, 0.0321, 0.0271, 0.0300, 0.0154, 0.0428, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0187, 0.0175, 0.0178, 0.0187, 0.0145, 0.0188, 0.0137], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:35:15,602 INFO [train.py:904] (4/8) Epoch 17, batch 1650, loss[loss=0.1854, simple_loss=0.2735, pruned_loss=0.0487, over 17070.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2552, pruned_loss=0.04349, over 3328534.70 frames. ], batch size: 55, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:35:39,289 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8651, 4.4558, 4.5852, 3.2328, 3.8412, 4.4960, 4.1023, 2.5712], device='cuda:4'), covar=tensor([0.0455, 0.0067, 0.0030, 0.0325, 0.0101, 0.0072, 0.0072, 0.0447], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0079, 0.0078, 0.0133, 0.0091, 0.0102, 0.0090, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:36:08,294 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164090.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:36:24,908 INFO [train.py:904] (4/8) Epoch 17, batch 1700, loss[loss=0.1782, simple_loss=0.2587, pruned_loss=0.04883, over 16814.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2566, pruned_loss=0.04405, over 3335941.24 frames. ], batch size: 96, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:36:40,702 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1195, 5.1331, 5.6260, 5.5196, 5.6054, 5.2136, 5.1753, 4.9162], device='cuda:4'), covar=tensor([0.0316, 0.0517, 0.0333, 0.0463, 0.0456, 0.0383, 0.0970, 0.0429], device='cuda:4'), in_proj_covar=tensor([0.0393, 0.0425, 0.0415, 0.0390, 0.0462, 0.0438, 0.0535, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 11:36:50,128 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1610, 5.0436, 5.0228, 4.4716, 4.5897, 5.0229, 5.0228, 4.6499], device='cuda:4'), covar=tensor([0.0604, 0.0560, 0.0310, 0.0402, 0.1235, 0.0472, 0.0331, 0.0867], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0400, 0.0335, 0.0328, 0.0350, 0.0378, 0.0230, 0.0404], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:37:01,949 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.646e+02 2.209e+02 2.796e+02 3.472e+02 6.333e+02, threshold=5.591e+02, percent-clipped=2.0 2023-04-30 11:37:05,848 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-30 11:37:32,257 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164151.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:37:32,930 INFO [train.py:904] (4/8) Epoch 17, batch 1750, loss[loss=0.1476, simple_loss=0.2437, pruned_loss=0.02579, over 17118.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2579, pruned_loss=0.04344, over 3340221.19 frames. ], batch size: 47, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:37:36,733 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164155.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:38:41,937 INFO [train.py:904] (4/8) Epoch 17, batch 1800, loss[loss=0.1789, simple_loss=0.2618, pruned_loss=0.04802, over 16234.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2595, pruned_loss=0.04423, over 3326112.27 frames. ], batch size: 165, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:38:58,676 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.86 vs. limit=5.0 2023-04-30 11:39:01,653 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164216.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:39:17,731 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8672, 5.2489, 5.0003, 5.0231, 4.7595, 4.6951, 4.7337, 5.3242], device='cuda:4'), covar=tensor([0.1291, 0.0861, 0.0958, 0.0747, 0.0775, 0.0987, 0.1132, 0.0852], device='cuda:4'), in_proj_covar=tensor([0.0648, 0.0794, 0.0649, 0.0590, 0.0505, 0.0507, 0.0661, 0.0615], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:39:19,764 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.509e+02 2.179e+02 2.567e+02 3.142e+02 4.810e+02, threshold=5.135e+02, percent-clipped=0.0 2023-04-30 11:39:21,321 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7462, 4.5418, 4.7686, 4.9598, 5.1313, 4.5348, 5.1037, 5.0928], device='cuda:4'), covar=tensor([0.1684, 0.1279, 0.1668, 0.0710, 0.0559, 0.0912, 0.0598, 0.0679], device='cuda:4'), in_proj_covar=tensor([0.0622, 0.0771, 0.0916, 0.0774, 0.0583, 0.0621, 0.0627, 0.0731], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:39:49,954 INFO [train.py:904] (4/8) Epoch 17, batch 1850, loss[loss=0.1779, simple_loss=0.2698, pruned_loss=0.04296, over 16647.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2598, pruned_loss=0.04393, over 3331698.17 frames. ], batch size: 62, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:41:01,367 INFO [train.py:904] (4/8) Epoch 17, batch 1900, loss[loss=0.169, simple_loss=0.2615, pruned_loss=0.0383, over 17142.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2591, pruned_loss=0.04326, over 3325757.41 frames. ], batch size: 48, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:41:06,093 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-30 11:41:26,283 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7159, 2.3772, 1.8359, 2.1326, 2.8299, 2.5662, 2.7929, 2.8618], device='cuda:4'), covar=tensor([0.0184, 0.0344, 0.0513, 0.0427, 0.0195, 0.0318, 0.0225, 0.0267], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0231, 0.0222, 0.0221, 0.0230, 0.0232, 0.0236, 0.0225], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:41:40,535 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7000, 2.3185, 1.7261, 2.1311, 2.7767, 2.5210, 2.8234, 2.8504], device='cuda:4'), covar=tensor([0.0182, 0.0397, 0.0548, 0.0411, 0.0201, 0.0305, 0.0210, 0.0251], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0230, 0.0221, 0.0221, 0.0230, 0.0231, 0.0236, 0.0225], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:41:41,204 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.647e+02 2.257e+02 2.647e+02 3.150e+02 4.785e+02, threshold=5.294e+02, percent-clipped=0.0 2023-04-30 11:42:09,131 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6936, 2.3724, 1.9304, 2.1841, 2.8637, 2.6005, 2.8567, 2.8869], device='cuda:4'), covar=tensor([0.0204, 0.0387, 0.0502, 0.0417, 0.0188, 0.0310, 0.0188, 0.0249], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0230, 0.0221, 0.0221, 0.0230, 0.0231, 0.0236, 0.0225], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:42:12,315 INFO [train.py:904] (4/8) Epoch 17, batch 1950, loss[loss=0.1416, simple_loss=0.2272, pruned_loss=0.02804, over 16767.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2587, pruned_loss=0.04304, over 3325355.59 frames. ], batch size: 39, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:42:24,731 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164360.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:42:47,931 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-30 11:42:56,785 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8286, 3.9505, 2.1398, 4.6720, 3.0843, 4.5085, 2.3070, 3.2411], device='cuda:4'), covar=tensor([0.0311, 0.0373, 0.1907, 0.0191, 0.0788, 0.0509, 0.1795, 0.0688], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0195, 0.0158, 0.0175, 0.0218, 0.0204, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 11:43:23,961 INFO [train.py:904] (4/8) Epoch 17, batch 2000, loss[loss=0.172, simple_loss=0.2607, pruned_loss=0.04159, over 15351.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2589, pruned_loss=0.04315, over 3314008.24 frames. ], batch size: 190, lr: 4.03e-03, grad_scale: 8.0 2023-04-30 11:43:51,020 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164421.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:44:02,065 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.637e+02 2.498e+02 2.950e+02 3.427e+02 6.648e+02, threshold=5.900e+02, percent-clipped=3.0 2023-04-30 11:44:25,523 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164446.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:44:32,478 INFO [train.py:904] (4/8) Epoch 17, batch 2050, loss[loss=0.1877, simple_loss=0.2587, pruned_loss=0.05837, over 16672.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2591, pruned_loss=0.04402, over 3310045.37 frames. ], batch size: 89, lr: 4.03e-03, grad_scale: 16.0 2023-04-30 11:44:47,247 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6164, 3.7451, 2.4139, 4.0040, 2.9778, 3.9203, 2.3044, 2.9690], device='cuda:4'), covar=tensor([0.0248, 0.0361, 0.1321, 0.0219, 0.0632, 0.0713, 0.1277, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0195, 0.0157, 0.0176, 0.0218, 0.0204, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 11:44:53,394 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1551, 2.1319, 2.2627, 3.8598, 2.1321, 2.4632, 2.2194, 2.2873], device='cuda:4'), covar=tensor([0.1286, 0.3597, 0.2698, 0.0572, 0.3726, 0.2413, 0.3468, 0.2980], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0428, 0.0358, 0.0326, 0.0431, 0.0493, 0.0397, 0.0500], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:45:07,662 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.55 vs. limit=5.0 2023-04-30 11:45:41,559 INFO [train.py:904] (4/8) Epoch 17, batch 2100, loss[loss=0.177, simple_loss=0.2566, pruned_loss=0.04872, over 16884.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2598, pruned_loss=0.04503, over 3306715.49 frames. ], batch size: 116, lr: 4.03e-03, grad_scale: 16.0 2023-04-30 11:45:54,974 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164511.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:46:13,724 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7249, 4.8308, 4.9748, 4.8362, 4.7557, 5.4278, 4.9589, 4.6806], device='cuda:4'), covar=tensor([0.1437, 0.1887, 0.2378, 0.2073, 0.3096, 0.1136, 0.1634, 0.2710], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0570, 0.0630, 0.0483, 0.0651, 0.0665, 0.0497, 0.0646], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:46:20,646 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.600e+02 2.320e+02 2.783e+02 3.356e+02 6.919e+02, threshold=5.567e+02, percent-clipped=1.0 2023-04-30 11:46:50,964 INFO [train.py:904] (4/8) Epoch 17, batch 2150, loss[loss=0.1971, simple_loss=0.2759, pruned_loss=0.05912, over 16573.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2605, pruned_loss=0.04564, over 3312646.02 frames. ], batch size: 75, lr: 4.03e-03, grad_scale: 16.0 2023-04-30 11:47:13,572 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164568.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:47:27,027 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164578.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:47:58,403 INFO [train.py:904] (4/8) Epoch 17, batch 2200, loss[loss=0.2131, simple_loss=0.2936, pruned_loss=0.06628, over 16727.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2615, pruned_loss=0.04597, over 3312866.82 frames. ], batch size: 76, lr: 4.03e-03, grad_scale: 16.0 2023-04-30 11:48:36,632 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164629.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:48:37,253 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.246e+02 2.692e+02 3.358e+02 7.856e+02, threshold=5.383e+02, percent-clipped=4.0 2023-04-30 11:48:44,215 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4360, 3.5846, 3.8868, 2.7848, 3.4550, 3.9167, 3.6490, 2.1271], device='cuda:4'), covar=tensor([0.0485, 0.0199, 0.0050, 0.0335, 0.0101, 0.0090, 0.0083, 0.0481], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0077, 0.0077, 0.0132, 0.0091, 0.0101, 0.0089, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:48:44,450 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.01 vs. limit=5.0 2023-04-30 11:48:50,067 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164639.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 11:48:51,349 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.51 vs. limit=2.0 2023-04-30 11:48:58,199 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9865, 3.9511, 4.3661, 2.1481, 4.6059, 4.5867, 3.2390, 3.5446], device='cuda:4'), covar=tensor([0.0685, 0.0232, 0.0210, 0.1159, 0.0060, 0.0185, 0.0457, 0.0364], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0095, 0.0139, 0.0075, 0.0122, 0.0127, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 11:49:02,793 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-30 11:49:06,804 INFO [train.py:904] (4/8) Epoch 17, batch 2250, loss[loss=0.1456, simple_loss=0.2344, pruned_loss=0.02837, over 16825.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2617, pruned_loss=0.04589, over 3317551.42 frames. ], batch size: 42, lr: 4.03e-03, grad_scale: 16.0 2023-04-30 11:49:15,027 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164658.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:50:15,835 INFO [train.py:904] (4/8) Epoch 17, batch 2300, loss[loss=0.1607, simple_loss=0.2417, pruned_loss=0.03979, over 16830.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.262, pruned_loss=0.04579, over 3312057.92 frames. ], batch size: 102, lr: 4.03e-03, grad_scale: 16.0 2023-04-30 11:50:35,158 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164716.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:50:38,858 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164719.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:50:53,238 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.267e+02 2.732e+02 3.248e+02 6.270e+02, threshold=5.465e+02, percent-clipped=1.0 2023-04-30 11:51:16,805 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=164746.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:51:16,824 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164746.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:51:24,583 INFO [train.py:904] (4/8) Epoch 17, batch 2350, loss[loss=0.2034, simple_loss=0.2864, pruned_loss=0.0602, over 16701.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2631, pruned_loss=0.04651, over 3297610.87 frames. ], batch size: 76, lr: 4.03e-03, grad_scale: 16.0 2023-04-30 11:52:01,897 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5905, 2.5094, 1.8512, 2.2238, 2.8106, 2.6139, 3.2237, 3.2245], device='cuda:4'), covar=tensor([0.0154, 0.0431, 0.0621, 0.0535, 0.0318, 0.0415, 0.0278, 0.0250], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0232, 0.0223, 0.0223, 0.0233, 0.0233, 0.0239, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:52:23,522 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=164794.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:52:34,376 INFO [train.py:904] (4/8) Epoch 17, batch 2400, loss[loss=0.1703, simple_loss=0.257, pruned_loss=0.0418, over 16877.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2643, pruned_loss=0.04652, over 3297217.70 frames. ], batch size: 96, lr: 4.03e-03, grad_scale: 16.0 2023-04-30 11:52:40,491 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3446, 2.3318, 2.4039, 4.2087, 2.2307, 2.6625, 2.3706, 2.5202], device='cuda:4'), covar=tensor([0.1312, 0.3655, 0.2773, 0.0553, 0.4030, 0.2550, 0.3647, 0.2935], device='cuda:4'), in_proj_covar=tensor([0.0390, 0.0430, 0.0360, 0.0328, 0.0432, 0.0497, 0.0399, 0.0502], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 11:52:41,630 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164807.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:52:46,724 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=164811.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:53:12,672 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.558e+02 2.216e+02 2.538e+02 3.112e+02 6.074e+02, threshold=5.077e+02, percent-clipped=1.0 2023-04-30 11:53:18,391 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8684, 3.9571, 2.1738, 4.6475, 3.1733, 4.5809, 2.4492, 3.2771], device='cuda:4'), covar=tensor([0.0268, 0.0333, 0.1672, 0.0192, 0.0681, 0.0425, 0.1476, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0174, 0.0193, 0.0157, 0.0173, 0.0217, 0.0202, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 11:53:41,596 INFO [train.py:904] (4/8) Epoch 17, batch 2450, loss[loss=0.1856, simple_loss=0.2788, pruned_loss=0.04617, over 16714.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2647, pruned_loss=0.04634, over 3304476.32 frames. ], batch size: 57, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 11:53:51,211 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=164859.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:54:17,469 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-30 11:54:46,718 INFO [train.py:904] (4/8) Epoch 17, batch 2500, loss[loss=0.148, simple_loss=0.2376, pruned_loss=0.02926, over 16839.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2637, pruned_loss=0.0454, over 3307650.73 frames. ], batch size: 42, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 11:55:17,515 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164924.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:55:26,643 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.613e+02 2.192e+02 2.472e+02 2.956e+02 6.595e+02, threshold=4.945e+02, percent-clipped=3.0 2023-04-30 11:55:30,971 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164934.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 11:55:35,458 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-30 11:55:55,231 INFO [train.py:904] (4/8) Epoch 17, batch 2550, loss[loss=0.1521, simple_loss=0.2398, pruned_loss=0.03225, over 16838.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2642, pruned_loss=0.04555, over 3298072.18 frames. ], batch size: 42, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 11:56:00,826 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5000, 4.5809, 4.6794, 4.4628, 4.4618, 5.1012, 4.5982, 4.2786], device='cuda:4'), covar=tensor([0.1579, 0.2182, 0.2229, 0.2335, 0.2974, 0.1226, 0.1770, 0.2528], device='cuda:4'), in_proj_covar=tensor([0.0394, 0.0569, 0.0627, 0.0479, 0.0645, 0.0662, 0.0495, 0.0643], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:56:16,737 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-30 11:56:18,768 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8367, 3.9289, 2.9871, 2.3484, 2.5242, 2.3709, 3.9874, 3.4110], device='cuda:4'), covar=tensor([0.2400, 0.0492, 0.1507, 0.2923, 0.2949, 0.2061, 0.0429, 0.1367], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0263, 0.0297, 0.0298, 0.0290, 0.0243, 0.0282, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:57:02,097 INFO [train.py:904] (4/8) Epoch 17, batch 2600, loss[loss=0.1689, simple_loss=0.2526, pruned_loss=0.04259, over 16877.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2635, pruned_loss=0.04548, over 3305343.91 frames. ], batch size: 116, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 11:57:17,929 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=165014.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:57:20,666 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165016.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:57:41,409 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.570e+02 2.187e+02 2.619e+02 3.229e+02 7.768e+02, threshold=5.237e+02, percent-clipped=3.0 2023-04-30 11:58:08,461 INFO [train.py:904] (4/8) Epoch 17, batch 2650, loss[loss=0.1879, simple_loss=0.2721, pruned_loss=0.05185, over 16384.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2642, pruned_loss=0.04552, over 3315712.68 frames. ], batch size: 145, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 11:58:26,259 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=165064.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:58:27,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7941, 3.8580, 2.8858, 2.2596, 2.5655, 2.3588, 3.9477, 3.4208], device='cuda:4'), covar=tensor([0.2500, 0.0587, 0.1662, 0.3065, 0.2731, 0.2053, 0.0500, 0.1353], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0264, 0.0298, 0.0300, 0.0290, 0.0244, 0.0283, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:58:57,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1038, 3.0221, 3.1001, 2.3016, 2.8993, 3.2197, 3.0468, 1.9102], device='cuda:4'), covar=tensor([0.0440, 0.0114, 0.0069, 0.0357, 0.0122, 0.0093, 0.0098, 0.0466], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0076, 0.0076, 0.0129, 0.0090, 0.0099, 0.0088, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 11:59:18,027 INFO [train.py:904] (4/8) Epoch 17, batch 2700, loss[loss=0.1814, simple_loss=0.2673, pruned_loss=0.04775, over 12621.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2645, pruned_loss=0.04525, over 3307615.05 frames. ], batch size: 246, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 11:59:18,300 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=165102.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:59:20,982 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-30 11:59:48,042 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=165123.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 11:59:55,438 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2884, 3.5537, 3.8817, 1.8567, 4.0461, 4.1442, 3.0583, 3.0501], device='cuda:4'), covar=tensor([0.1127, 0.0226, 0.0207, 0.1306, 0.0103, 0.0196, 0.0449, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0095, 0.0139, 0.0075, 0.0122, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 11:59:59,034 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.378e+02 2.045e+02 2.516e+02 2.952e+02 6.082e+02, threshold=5.032e+02, percent-clipped=3.0 2023-04-30 12:00:28,481 INFO [train.py:904] (4/8) Epoch 17, batch 2750, loss[loss=0.1837, simple_loss=0.2632, pruned_loss=0.0521, over 16708.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2651, pruned_loss=0.04529, over 3303244.48 frames. ], batch size: 124, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 12:01:13,265 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=165184.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:01:38,385 INFO [train.py:904] (4/8) Epoch 17, batch 2800, loss[loss=0.1842, simple_loss=0.2701, pruned_loss=0.04911, over 16250.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2645, pruned_loss=0.04457, over 3312076.28 frames. ], batch size: 165, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 12:02:10,325 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165224.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:02:19,320 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.573e+02 2.251e+02 2.601e+02 3.544e+02 6.990e+02, threshold=5.203e+02, percent-clipped=6.0 2023-04-30 12:02:24,068 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165234.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 12:02:48,648 INFO [train.py:904] (4/8) Epoch 17, batch 2850, loss[loss=0.1585, simple_loss=0.2537, pruned_loss=0.03166, over 17158.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2633, pruned_loss=0.04424, over 3308378.56 frames. ], batch size: 46, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 12:03:18,271 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=165272.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:03:18,401 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8112, 4.5760, 4.8878, 5.0462, 5.2129, 4.5852, 5.1945, 5.2124], device='cuda:4'), covar=tensor([0.1715, 0.1310, 0.1527, 0.0728, 0.0594, 0.0992, 0.0636, 0.0635], device='cuda:4'), in_proj_covar=tensor([0.0633, 0.0779, 0.0930, 0.0789, 0.0593, 0.0629, 0.0633, 0.0738], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:03:31,430 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=165282.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:03:33,892 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2351, 5.7883, 5.9064, 5.5299, 5.6642, 6.2335, 5.6334, 5.3934], device='cuda:4'), covar=tensor([0.0842, 0.1816, 0.2029, 0.2195, 0.2524, 0.0991, 0.1500, 0.2435], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0572, 0.0631, 0.0485, 0.0650, 0.0666, 0.0500, 0.0650], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 12:03:47,547 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7330, 4.7888, 4.9198, 4.7442, 4.7834, 5.3755, 4.8624, 4.6087], device='cuda:4'), covar=tensor([0.1279, 0.2218, 0.2394, 0.2152, 0.2914, 0.1156, 0.1665, 0.2641], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0571, 0.0631, 0.0484, 0.0649, 0.0665, 0.0499, 0.0649], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 12:03:58,714 INFO [train.py:904] (4/8) Epoch 17, batch 2900, loss[loss=0.2646, simple_loss=0.3111, pruned_loss=0.1091, over 11961.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.263, pruned_loss=0.04514, over 3302899.40 frames. ], batch size: 246, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 12:04:16,984 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165314.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:04:39,533 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.654e+02 2.311e+02 2.666e+02 3.214e+02 5.469e+02, threshold=5.331e+02, percent-clipped=1.0 2023-04-30 12:05:09,243 INFO [train.py:904] (4/8) Epoch 17, batch 2950, loss[loss=0.1741, simple_loss=0.2704, pruned_loss=0.03886, over 17092.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2629, pruned_loss=0.04568, over 3302596.12 frames. ], batch size: 47, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 12:05:19,681 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1941, 5.7159, 5.8575, 5.5205, 5.6470, 6.2137, 5.7633, 5.4929], device='cuda:4'), covar=tensor([0.0858, 0.2023, 0.2357, 0.2426, 0.2748, 0.1098, 0.1425, 0.2601], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0575, 0.0636, 0.0489, 0.0654, 0.0671, 0.0502, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 12:05:23,903 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=165362.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:06:20,614 INFO [train.py:904] (4/8) Epoch 17, batch 3000, loss[loss=0.1834, simple_loss=0.2621, pruned_loss=0.05236, over 16670.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2636, pruned_loss=0.04638, over 3293903.82 frames. ], batch size: 134, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 12:06:20,614 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 12:06:29,127 INFO [train.py:938] (4/8) Epoch 17, validation: loss=0.1364, simple_loss=0.2422, pruned_loss=0.0153, over 944034.00 frames. 2023-04-30 12:06:29,127 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 12:06:29,503 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165402.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:07:09,775 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.715e+02 2.395e+02 2.828e+02 3.269e+02 8.319e+02, threshold=5.656e+02, percent-clipped=3.0 2023-04-30 12:07:23,749 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8806, 3.0899, 3.1131, 2.0553, 2.7437, 2.1753, 3.3895, 3.3717], device='cuda:4'), covar=tensor([0.0267, 0.0896, 0.0625, 0.1860, 0.0866, 0.1046, 0.0572, 0.0897], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0158, 0.0163, 0.0149, 0.0140, 0.0126, 0.0140, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 12:07:36,653 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=165450.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:07:38,650 INFO [train.py:904] (4/8) Epoch 17, batch 3050, loss[loss=0.1719, simple_loss=0.2614, pruned_loss=0.04117, over 16485.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2623, pruned_loss=0.04567, over 3306284.20 frames. ], batch size: 68, lr: 4.02e-03, grad_scale: 4.0 2023-04-30 12:08:12,764 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.53 vs. limit=2.0 2023-04-30 12:08:15,873 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=165479.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:08:36,269 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-04-30 12:08:45,827 INFO [train.py:904] (4/8) Epoch 17, batch 3100, loss[loss=0.1906, simple_loss=0.2672, pruned_loss=0.05706, over 16321.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2624, pruned_loss=0.04573, over 3317887.30 frames. ], batch size: 145, lr: 4.02e-03, grad_scale: 4.0 2023-04-30 12:09:10,285 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6986, 3.1006, 2.7565, 5.0064, 4.0967, 4.3984, 1.9086, 2.9131], device='cuda:4'), covar=tensor([0.1436, 0.0747, 0.1160, 0.0231, 0.0233, 0.0445, 0.1544, 0.0904], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0165, 0.0186, 0.0175, 0.0200, 0.0212, 0.0192, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 12:09:27,660 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 2.216e+02 2.589e+02 3.162e+02 6.656e+02, threshold=5.179e+02, percent-clipped=4.0 2023-04-30 12:09:55,014 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=165551.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:09:55,706 INFO [train.py:904] (4/8) Epoch 17, batch 3150, loss[loss=0.1711, simple_loss=0.2512, pruned_loss=0.04547, over 16757.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2611, pruned_loss=0.04533, over 3320370.25 frames. ], batch size: 83, lr: 4.02e-03, grad_scale: 4.0 2023-04-30 12:11:06,275 INFO [train.py:904] (4/8) Epoch 17, batch 3200, loss[loss=0.1996, simple_loss=0.2809, pruned_loss=0.05914, over 16289.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2599, pruned_loss=0.04445, over 3322435.10 frames. ], batch size: 165, lr: 4.02e-03, grad_scale: 8.0 2023-04-30 12:11:07,600 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.28 vs. limit=5.0 2023-04-30 12:11:21,068 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=165612.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:11:49,641 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.479e+02 2.164e+02 2.494e+02 3.096e+02 4.677e+02, threshold=4.988e+02, percent-clipped=0.0 2023-04-30 12:11:58,716 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6619, 2.2670, 2.4165, 4.5300, 2.2996, 2.7377, 2.3666, 2.4974], device='cuda:4'), covar=tensor([0.1115, 0.3780, 0.2887, 0.0426, 0.4010, 0.2665, 0.3480, 0.3620], device='cuda:4'), in_proj_covar=tensor([0.0394, 0.0431, 0.0359, 0.0329, 0.0432, 0.0498, 0.0401, 0.0505], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:12:15,438 INFO [train.py:904] (4/8) Epoch 17, batch 3250, loss[loss=0.2015, simple_loss=0.2823, pruned_loss=0.06036, over 16524.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2604, pruned_loss=0.04428, over 3326263.89 frames. ], batch size: 68, lr: 4.01e-03, grad_scale: 4.0 2023-04-30 12:13:23,339 INFO [train.py:904] (4/8) Epoch 17, batch 3300, loss[loss=0.1925, simple_loss=0.27, pruned_loss=0.0575, over 16916.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.262, pruned_loss=0.04577, over 3323583.06 frames. ], batch size: 116, lr: 4.01e-03, grad_scale: 4.0 2023-04-30 12:13:37,602 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-30 12:14:00,813 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-30 12:14:06,777 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.766e+02 2.306e+02 2.673e+02 3.210e+02 6.560e+02, threshold=5.346e+02, percent-clipped=3.0 2023-04-30 12:14:30,279 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=165750.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:14:32,748 INFO [train.py:904] (4/8) Epoch 17, batch 3350, loss[loss=0.151, simple_loss=0.2374, pruned_loss=0.03234, over 16996.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.262, pruned_loss=0.04516, over 3325733.26 frames. ], batch size: 41, lr: 4.01e-03, grad_scale: 4.0 2023-04-30 12:15:11,142 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165779.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:15:42,358 INFO [train.py:904] (4/8) Epoch 17, batch 3400, loss[loss=0.1559, simple_loss=0.2548, pruned_loss=0.02847, over 17270.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2626, pruned_loss=0.0453, over 3321046.46 frames. ], batch size: 52, lr: 4.01e-03, grad_scale: 4.0 2023-04-30 12:15:44,583 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=165803.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:15:56,245 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=165811.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:16:18,380 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=165827.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:16:27,736 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.598e+02 2.228e+02 2.640e+02 3.185e+02 6.119e+02, threshold=5.280e+02, percent-clipped=1.0 2023-04-30 12:16:39,562 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-30 12:16:54,367 INFO [train.py:904] (4/8) Epoch 17, batch 3450, loss[loss=0.1614, simple_loss=0.2564, pruned_loss=0.0332, over 17079.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2605, pruned_loss=0.04486, over 3327667.03 frames. ], batch size: 55, lr: 4.01e-03, grad_scale: 4.0 2023-04-30 12:17:00,953 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 12:17:11,499 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=165864.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:17:14,449 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2406, 4.1985, 4.3941, 4.1442, 4.1902, 4.7969, 4.3496, 4.0292], device='cuda:4'), covar=tensor([0.1886, 0.2255, 0.1948, 0.2339, 0.3059, 0.1167, 0.1542, 0.2682], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0572, 0.0629, 0.0484, 0.0651, 0.0663, 0.0497, 0.0648], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 12:18:05,609 INFO [train.py:904] (4/8) Epoch 17, batch 3500, loss[loss=0.1842, simple_loss=0.2734, pruned_loss=0.0475, over 17076.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2589, pruned_loss=0.04434, over 3320530.04 frames. ], batch size: 53, lr: 4.01e-03, grad_scale: 4.0 2023-04-30 12:18:10,388 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2456, 4.9240, 5.2028, 5.3854, 5.6143, 4.8193, 5.6090, 5.5781], device='cuda:4'), covar=tensor([0.1707, 0.1124, 0.1735, 0.0784, 0.0511, 0.0864, 0.0428, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0640, 0.0790, 0.0945, 0.0803, 0.0602, 0.0639, 0.0640, 0.0748], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:18:13,073 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=165907.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:18:25,137 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.83 vs. limit=5.0 2023-04-30 12:18:49,909 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.598e+02 2.135e+02 2.523e+02 3.067e+02 7.185e+02, threshold=5.046e+02, percent-clipped=1.0 2023-04-30 12:19:02,461 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-30 12:19:05,087 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-30 12:19:16,033 INFO [train.py:904] (4/8) Epoch 17, batch 3550, loss[loss=0.1655, simple_loss=0.2618, pruned_loss=0.03456, over 17159.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2581, pruned_loss=0.04407, over 3322334.89 frames. ], batch size: 46, lr: 4.01e-03, grad_scale: 4.0 2023-04-30 12:20:27,926 INFO [train.py:904] (4/8) Epoch 17, batch 3600, loss[loss=0.2128, simple_loss=0.2778, pruned_loss=0.07388, over 11398.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2571, pruned_loss=0.04344, over 3314114.12 frames. ], batch size: 246, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:21:12,107 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.506e+02 2.087e+02 2.490e+02 2.982e+02 5.501e+02, threshold=4.979e+02, percent-clipped=1.0 2023-04-30 12:21:29,507 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7233, 3.7697, 2.8049, 2.2506, 2.5856, 2.2459, 3.8574, 3.4052], device='cuda:4'), covar=tensor([0.2659, 0.0630, 0.1673, 0.2641, 0.2547, 0.2065, 0.0525, 0.1309], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0265, 0.0297, 0.0300, 0.0291, 0.0244, 0.0283, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 12:21:40,963 INFO [train.py:904] (4/8) Epoch 17, batch 3650, loss[loss=0.1752, simple_loss=0.2458, pruned_loss=0.05232, over 16900.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2549, pruned_loss=0.04334, over 3310032.88 frames. ], batch size: 116, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:21:44,588 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166054.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 12:21:55,116 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-30 12:22:21,828 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3758, 5.3481, 5.1578, 4.6281, 5.2001, 2.0831, 4.9578, 5.1110], device='cuda:4'), covar=tensor([0.0088, 0.0074, 0.0169, 0.0351, 0.0081, 0.2419, 0.0115, 0.0149], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0142, 0.0190, 0.0173, 0.0163, 0.0198, 0.0179, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:22:55,839 INFO [train.py:904] (4/8) Epoch 17, batch 3700, loss[loss=0.1628, simple_loss=0.2397, pruned_loss=0.04293, over 16854.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.254, pruned_loss=0.04484, over 3297903.67 frames. ], batch size: 90, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:23:02,845 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166106.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:23:16,025 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166115.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 12:23:29,629 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1446, 2.0664, 1.6959, 1.8026, 2.3617, 2.0664, 2.2223, 2.4589], device='cuda:4'), covar=tensor([0.0240, 0.0348, 0.0471, 0.0441, 0.0221, 0.0311, 0.0199, 0.0246], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0231, 0.0221, 0.0223, 0.0233, 0.0232, 0.0237, 0.0228], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:23:42,388 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.852e+02 2.227e+02 2.592e+02 3.005e+02 5.986e+02, threshold=5.184e+02, percent-clipped=3.0 2023-04-30 12:24:03,613 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166147.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:24:10,338 INFO [train.py:904] (4/8) Epoch 17, batch 3750, loss[loss=0.18, simple_loss=0.2577, pruned_loss=0.05115, over 16843.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.255, pruned_loss=0.04623, over 3284463.68 frames. ], batch size: 96, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:24:21,424 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166159.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:24:48,238 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.11 vs. limit=5.0 2023-04-30 12:25:24,422 INFO [train.py:904] (4/8) Epoch 17, batch 3800, loss[loss=0.1812, simple_loss=0.2545, pruned_loss=0.05394, over 16785.00 frames. ], tot_loss[loss=0.1757, simple_loss=0.256, pruned_loss=0.04769, over 3289282.43 frames. ], batch size: 124, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:25:32,315 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166207.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:25:33,757 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166208.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:25:50,901 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1206, 3.2025, 3.2298, 2.1070, 2.9254, 2.4305, 3.6530, 3.5715], device='cuda:4'), covar=tensor([0.0255, 0.0886, 0.0667, 0.1887, 0.0786, 0.0981, 0.0480, 0.0770], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0160, 0.0164, 0.0150, 0.0141, 0.0127, 0.0140, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 12:26:10,618 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.840e+02 2.250e+02 2.656e+02 3.137e+02 5.830e+02, threshold=5.312e+02, percent-clipped=1.0 2023-04-30 12:26:38,503 INFO [train.py:904] (4/8) Epoch 17, batch 3850, loss[loss=0.1816, simple_loss=0.2526, pruned_loss=0.05528, over 16878.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2572, pruned_loss=0.04874, over 3278584.22 frames. ], batch size: 109, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:26:44,003 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=166255.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:27:52,934 INFO [train.py:904] (4/8) Epoch 17, batch 3900, loss[loss=0.1684, simple_loss=0.243, pruned_loss=0.04692, over 16508.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2565, pruned_loss=0.04905, over 3279468.80 frames. ], batch size: 75, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:28:00,967 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166307.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:28:01,184 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-30 12:28:37,861 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.433e+02 2.279e+02 2.722e+02 3.230e+02 5.998e+02, threshold=5.443e+02, percent-clipped=1.0 2023-04-30 12:29:07,065 INFO [train.py:904] (4/8) Epoch 17, batch 3950, loss[loss=0.1765, simple_loss=0.2461, pruned_loss=0.05349, over 16753.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2563, pruned_loss=0.05005, over 3263944.80 frames. ], batch size: 124, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:29:30,588 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166368.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:29:37,905 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 12:29:58,075 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7884, 3.1596, 2.8491, 5.1068, 4.0099, 4.1983, 1.8291, 3.2602], device='cuda:4'), covar=tensor([0.1316, 0.0685, 0.1187, 0.0101, 0.0379, 0.0390, 0.1588, 0.0827], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0167, 0.0189, 0.0178, 0.0204, 0.0214, 0.0194, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 12:30:18,532 INFO [train.py:904] (4/8) Epoch 17, batch 4000, loss[loss=0.1863, simple_loss=0.2593, pruned_loss=0.05667, over 16453.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2564, pruned_loss=0.05052, over 3278781.70 frames. ], batch size: 146, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:30:25,219 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166406.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:30:31,136 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166410.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 12:31:03,720 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.351e+02 2.185e+02 2.579e+02 2.981e+02 5.548e+02, threshold=5.157e+02, percent-clipped=1.0 2023-04-30 12:31:31,403 INFO [train.py:904] (4/8) Epoch 17, batch 4050, loss[loss=0.1696, simple_loss=0.2583, pruned_loss=0.0404, over 15535.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2564, pruned_loss=0.04904, over 3280092.22 frames. ], batch size: 191, lr: 4.01e-03, grad_scale: 8.0 2023-04-30 12:31:34,056 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=166454.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:31:42,524 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166459.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:31:46,065 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7897, 3.9177, 4.2963, 1.9448, 4.5958, 4.6220, 3.1371, 3.2890], device='cuda:4'), covar=tensor([0.0787, 0.0236, 0.0144, 0.1299, 0.0047, 0.0078, 0.0377, 0.0442], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0108, 0.0095, 0.0139, 0.0077, 0.0123, 0.0128, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 12:32:14,536 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166481.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:32:44,116 INFO [train.py:904] (4/8) Epoch 17, batch 4100, loss[loss=0.1958, simple_loss=0.2839, pruned_loss=0.05382, over 16609.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2576, pruned_loss=0.04809, over 3282980.91 frames. ], batch size: 68, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:32:46,306 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166503.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:32:51,044 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=166507.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:32:54,595 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166509.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:33:29,237 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166531.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:33:31,669 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.352e+02 2.012e+02 2.257e+02 2.846e+02 6.984e+02, threshold=4.513e+02, percent-clipped=1.0 2023-04-30 12:33:45,774 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166542.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:33:59,704 INFO [train.py:904] (4/8) Epoch 17, batch 4150, loss[loss=0.1875, simple_loss=0.2856, pruned_loss=0.04467, over 16907.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2646, pruned_loss=0.05063, over 3239316.22 frames. ], batch size: 96, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:34:28,434 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166570.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:35:01,100 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166592.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:35:04,691 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9085, 5.4588, 5.6373, 5.3372, 5.3087, 5.9972, 5.4230, 5.1672], device='cuda:4'), covar=tensor([0.0898, 0.1562, 0.1408, 0.1611, 0.2525, 0.0783, 0.1295, 0.2266], device='cuda:4'), in_proj_covar=tensor([0.0391, 0.0562, 0.0615, 0.0474, 0.0633, 0.0653, 0.0488, 0.0634], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 12:35:14,003 INFO [train.py:904] (4/8) Epoch 17, batch 4200, loss[loss=0.2102, simple_loss=0.3019, pruned_loss=0.05927, over 16423.00 frames. ], tot_loss[loss=0.188, simple_loss=0.2715, pruned_loss=0.05228, over 3215574.64 frames. ], batch size: 146, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:36:00,016 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.639e+02 2.403e+02 2.795e+02 3.348e+02 5.640e+02, threshold=5.591e+02, percent-clipped=5.0 2023-04-30 12:36:13,721 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-30 12:36:27,721 INFO [train.py:904] (4/8) Epoch 17, batch 4250, loss[loss=0.1747, simple_loss=0.2683, pruned_loss=0.04051, over 17244.00 frames. ], tot_loss[loss=0.1906, simple_loss=0.2757, pruned_loss=0.05275, over 3195021.83 frames. ], batch size: 52, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:36:43,547 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166663.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:36:49,759 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1116, 3.9234, 3.8720, 4.2886, 4.4250, 4.0753, 4.3095, 4.5089], device='cuda:4'), covar=tensor([0.1803, 0.1346, 0.2390, 0.1111, 0.0872, 0.1870, 0.1552, 0.0850], device='cuda:4'), in_proj_covar=tensor([0.0612, 0.0751, 0.0894, 0.0764, 0.0574, 0.0612, 0.0608, 0.0711], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:37:33,377 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0404, 2.9492, 2.3286, 2.6642, 3.3783, 2.9064, 3.5417, 3.5798], device='cuda:4'), covar=tensor([0.0056, 0.0378, 0.0511, 0.0445, 0.0207, 0.0375, 0.0245, 0.0195], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0225, 0.0216, 0.0218, 0.0227, 0.0226, 0.0230, 0.0222], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:37:39,002 INFO [train.py:904] (4/8) Epoch 17, batch 4300, loss[loss=0.2125, simple_loss=0.3172, pruned_loss=0.05385, over 16673.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2769, pruned_loss=0.05151, over 3210011.20 frames. ], batch size: 76, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:37:51,463 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166710.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 12:38:24,692 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.536e+02 2.116e+02 2.589e+02 3.113e+02 5.603e+02, threshold=5.178e+02, percent-clipped=2.0 2023-04-30 12:38:28,866 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166735.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:38:52,935 INFO [train.py:904] (4/8) Epoch 17, batch 4350, loss[loss=0.2141, simple_loss=0.2951, pruned_loss=0.06656, over 16637.00 frames. ], tot_loss[loss=0.1919, simple_loss=0.2794, pruned_loss=0.05225, over 3204767.89 frames. ], batch size: 57, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:39:01,857 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=166758.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 12:39:10,325 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166764.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:39:57,510 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166796.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:40:05,343 INFO [train.py:904] (4/8) Epoch 17, batch 4400, loss[loss=0.2069, simple_loss=0.2879, pruned_loss=0.06291, over 16430.00 frames. ], tot_loss[loss=0.1945, simple_loss=0.2818, pruned_loss=0.0536, over 3188445.78 frames. ], batch size: 68, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:40:06,936 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166803.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:40:26,500 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-30 12:40:37,959 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166825.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:40:49,146 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.478e+02 2.232e+02 2.626e+02 3.012e+02 5.313e+02, threshold=5.252e+02, percent-clipped=1.0 2023-04-30 12:40:55,203 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166837.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:41:15,723 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=166851.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:41:16,588 INFO [train.py:904] (4/8) Epoch 17, batch 4450, loss[loss=0.2136, simple_loss=0.2997, pruned_loss=0.06371, over 15403.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.2851, pruned_loss=0.05463, over 3199828.46 frames. ], batch size: 190, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:41:36,425 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166865.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:42:02,676 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1805, 1.9261, 2.6542, 3.0177, 2.8432, 3.3102, 2.1083, 3.3211], device='cuda:4'), covar=tensor([0.0159, 0.0445, 0.0258, 0.0227, 0.0239, 0.0143, 0.0446, 0.0111], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0187, 0.0173, 0.0177, 0.0186, 0.0143, 0.0187, 0.0137], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:42:06,782 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-30 12:42:07,449 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166887.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:42:17,774 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.95 vs. limit=5.0 2023-04-30 12:42:28,857 INFO [train.py:904] (4/8) Epoch 17, batch 4500, loss[loss=0.1859, simple_loss=0.2716, pruned_loss=0.05014, over 16675.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2851, pruned_loss=0.05492, over 3225214.35 frames. ], batch size: 57, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:43:07,243 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166929.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:43:12,783 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.487e+02 1.901e+02 2.204e+02 2.572e+02 5.344e+02, threshold=4.409e+02, percent-clipped=1.0 2023-04-30 12:43:40,951 INFO [train.py:904] (4/8) Epoch 17, batch 4550, loss[loss=0.1962, simple_loss=0.2823, pruned_loss=0.05511, over 17120.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2858, pruned_loss=0.0557, over 3242518.75 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:43:57,100 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166963.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:44:09,655 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.03 vs. limit=5.0 2023-04-30 12:44:35,641 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166990.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:44:53,178 INFO [train.py:904] (4/8) Epoch 17, batch 4600, loss[loss=0.1959, simple_loss=0.2839, pruned_loss=0.054, over 16575.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2869, pruned_loss=0.05602, over 3238989.09 frames. ], batch size: 68, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:44:58,245 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8517, 2.0913, 2.3817, 3.1437, 2.1683, 2.3003, 2.2591, 2.1977], device='cuda:4'), covar=tensor([0.1274, 0.3083, 0.2138, 0.0628, 0.3678, 0.2331, 0.2921, 0.3245], device='cuda:4'), in_proj_covar=tensor([0.0389, 0.0429, 0.0354, 0.0324, 0.0429, 0.0496, 0.0397, 0.0501], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:45:07,098 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167011.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:45:38,256 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.244e+02 1.923e+02 2.253e+02 2.642e+02 5.543e+02, threshold=4.506e+02, percent-clipped=2.0 2023-04-30 12:46:05,391 INFO [train.py:904] (4/8) Epoch 17, batch 4650, loss[loss=0.1903, simple_loss=0.2711, pruned_loss=0.0547, over 16565.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2868, pruned_loss=0.05666, over 3224432.78 frames. ], batch size: 62, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:46:49,844 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2312, 3.2498, 1.9463, 3.5987, 2.4961, 3.6013, 2.1017, 2.6011], device='cuda:4'), covar=tensor([0.0289, 0.0358, 0.1601, 0.0125, 0.0837, 0.0467, 0.1454, 0.0817], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0169, 0.0189, 0.0149, 0.0169, 0.0209, 0.0196, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 12:47:00,699 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167091.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:47:10,833 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 12:47:13,808 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167100.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:47:16,413 INFO [train.py:904] (4/8) Epoch 17, batch 4700, loss[loss=0.2007, simple_loss=0.2845, pruned_loss=0.05848, over 16709.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2844, pruned_loss=0.05559, over 3219863.82 frames. ], batch size: 124, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:47:18,967 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167104.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:47:42,084 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167120.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:48:00,930 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.373e+02 1.982e+02 2.244e+02 2.685e+02 3.929e+02, threshold=4.489e+02, percent-clipped=0.0 2023-04-30 12:48:06,819 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167137.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:48:28,392 INFO [train.py:904] (4/8) Epoch 17, batch 4750, loss[loss=0.1805, simple_loss=0.2616, pruned_loss=0.04968, over 16634.00 frames. ], tot_loss[loss=0.1935, simple_loss=0.2801, pruned_loss=0.0535, over 3229143.16 frames. ], batch size: 62, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:48:42,008 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167161.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:48:48,031 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167165.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:48:48,180 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167165.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:49:16,407 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167185.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:49:18,967 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167187.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:49:40,730 INFO [train.py:904] (4/8) Epoch 17, batch 4800, loss[loss=0.1821, simple_loss=0.2784, pruned_loss=0.0429, over 16694.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2767, pruned_loss=0.05163, over 3217377.73 frames. ], batch size: 134, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:49:58,979 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167213.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:50:27,972 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.365e+02 2.008e+02 2.170e+02 2.594e+02 4.937e+02, threshold=4.340e+02, percent-clipped=1.0 2023-04-30 12:50:32,466 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167235.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:50:58,072 INFO [train.py:904] (4/8) Epoch 17, batch 4850, loss[loss=0.1754, simple_loss=0.2712, pruned_loss=0.03983, over 16634.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.2766, pruned_loss=0.05041, over 3213045.49 frames. ], batch size: 134, lr: 4.00e-03, grad_scale: 8.0 2023-04-30 12:51:46,659 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167285.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:52:12,188 INFO [train.py:904] (4/8) Epoch 17, batch 4900, loss[loss=0.1762, simple_loss=0.2619, pruned_loss=0.04523, over 16522.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2757, pruned_loss=0.04944, over 3184149.75 frames. ], batch size: 68, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 12:52:55,965 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.440e+02 1.938e+02 2.305e+02 2.710e+02 4.454e+02, threshold=4.611e+02, percent-clipped=1.0 2023-04-30 12:53:24,978 INFO [train.py:904] (4/8) Epoch 17, batch 4950, loss[loss=0.2069, simple_loss=0.2972, pruned_loss=0.05826, over 16894.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2758, pruned_loss=0.04918, over 3181510.72 frames. ], batch size: 109, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 12:54:22,575 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167391.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:54:37,753 INFO [train.py:904] (4/8) Epoch 17, batch 5000, loss[loss=0.1703, simple_loss=0.2629, pruned_loss=0.03886, over 16540.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2769, pruned_loss=0.04897, over 3201767.35 frames. ], batch size: 68, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 12:54:38,639 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-04-30 12:54:50,040 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2613, 1.9925, 2.7989, 3.1688, 2.9931, 3.8237, 2.1651, 3.6062], device='cuda:4'), covar=tensor([0.0173, 0.0468, 0.0276, 0.0263, 0.0263, 0.0103, 0.0488, 0.0105], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0188, 0.0173, 0.0178, 0.0187, 0.0144, 0.0188, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 12:55:02,762 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167420.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:55:19,934 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167432.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:55:20,678 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.629e+02 2.024e+02 2.500e+02 2.980e+02 7.048e+02, threshold=5.000e+02, percent-clipped=1.0 2023-04-30 12:55:29,659 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167439.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:55:46,692 INFO [train.py:904] (4/8) Epoch 17, batch 5050, loss[loss=0.1907, simple_loss=0.2853, pruned_loss=0.04808, over 16970.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2773, pruned_loss=0.04868, over 3215068.79 frames. ], batch size: 109, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 12:55:49,557 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5477, 2.9484, 3.0283, 1.8857, 2.6909, 2.0714, 3.1482, 3.1191], device='cuda:4'), covar=tensor([0.0290, 0.0793, 0.0674, 0.1989, 0.0872, 0.0982, 0.0641, 0.0870], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0160, 0.0165, 0.0150, 0.0142, 0.0128, 0.0142, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 12:55:51,605 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167456.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:55:57,173 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167460.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:56:08,484 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167468.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:56:26,386 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7141, 2.7541, 2.6106, 4.5907, 3.4822, 4.1474, 1.5599, 3.0710], device='cuda:4'), covar=tensor([0.1291, 0.0736, 0.1188, 0.0115, 0.0269, 0.0366, 0.1560, 0.0792], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0168, 0.0190, 0.0176, 0.0203, 0.0211, 0.0194, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 12:56:43,508 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167493.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:56:56,777 INFO [train.py:904] (4/8) Epoch 17, batch 5100, loss[loss=0.1808, simple_loss=0.2684, pruned_loss=0.04659, over 17054.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2758, pruned_loss=0.04807, over 3229878.13 frames. ], batch size: 53, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 12:57:39,789 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.439e+02 2.010e+02 2.231e+02 2.542e+02 5.876e+02, threshold=4.463e+02, percent-clipped=1.0 2023-04-30 12:57:40,485 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.06 vs. limit=2.0 2023-04-30 12:58:08,925 INFO [train.py:904] (4/8) Epoch 17, batch 5150, loss[loss=0.1912, simple_loss=0.2876, pruned_loss=0.0474, over 15303.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2751, pruned_loss=0.04729, over 3218740.78 frames. ], batch size: 190, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 12:58:56,855 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167585.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 12:59:21,203 INFO [train.py:904] (4/8) Epoch 17, batch 5200, loss[loss=0.1919, simple_loss=0.2754, pruned_loss=0.05417, over 16941.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2738, pruned_loss=0.04687, over 3226796.79 frames. ], batch size: 116, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 12:59:38,311 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1373, 4.5105, 3.9668, 4.3795, 4.1115, 4.0685, 4.0068, 4.4825], device='cuda:4'), covar=tensor([0.2461, 0.1595, 0.2555, 0.1374, 0.1608, 0.2422, 0.2306, 0.1969], device='cuda:4'), in_proj_covar=tensor([0.0615, 0.0756, 0.0617, 0.0557, 0.0477, 0.0484, 0.0626, 0.0584], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:00:07,259 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.425e+02 2.038e+02 2.260e+02 2.754e+02 5.460e+02, threshold=4.520e+02, percent-clipped=3.0 2023-04-30 13:00:08,173 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167633.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:00:36,319 INFO [train.py:904] (4/8) Epoch 17, batch 5250, loss[loss=0.2025, simple_loss=0.2944, pruned_loss=0.05532, over 15330.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2725, pruned_loss=0.04682, over 3212476.62 frames. ], batch size: 190, lr: 3.99e-03, grad_scale: 16.0 2023-04-30 13:00:55,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5474, 5.8689, 5.5545, 5.6666, 5.3216, 5.2110, 5.3776, 5.9932], device='cuda:4'), covar=tensor([0.1134, 0.0766, 0.0931, 0.0727, 0.0795, 0.0611, 0.1030, 0.0777], device='cuda:4'), in_proj_covar=tensor([0.0619, 0.0759, 0.0621, 0.0560, 0.0479, 0.0486, 0.0629, 0.0588], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:01:14,769 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-30 13:01:48,350 INFO [train.py:904] (4/8) Epoch 17, batch 5300, loss[loss=0.1837, simple_loss=0.2697, pruned_loss=0.04891, over 16883.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2689, pruned_loss=0.04537, over 3226836.20 frames. ], batch size: 109, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 13:02:12,751 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167719.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:02:35,320 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.320e+02 2.142e+02 2.405e+02 3.044e+02 5.207e+02, threshold=4.809e+02, percent-clipped=2.0 2023-04-30 13:03:01,793 INFO [train.py:904] (4/8) Epoch 17, batch 5350, loss[loss=0.1992, simple_loss=0.2848, pruned_loss=0.05679, over 15391.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2675, pruned_loss=0.04506, over 3220416.00 frames. ], batch size: 190, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 13:03:08,226 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167756.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:03:13,810 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167760.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:03:23,334 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167767.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:03:41,683 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167780.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:03:53,661 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167788.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:04:14,426 INFO [train.py:904] (4/8) Epoch 17, batch 5400, loss[loss=0.1961, simple_loss=0.2858, pruned_loss=0.05322, over 16658.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2705, pruned_loss=0.04586, over 3220418.67 frames. ], batch size: 134, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 13:04:17,677 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167804.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:04:24,145 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=167808.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:04:54,088 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167828.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:05:02,680 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.524e+02 2.037e+02 2.360e+02 2.749e+02 5.264e+02, threshold=4.720e+02, percent-clipped=1.0 2023-04-30 13:05:11,159 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9811, 3.2202, 3.2195, 2.1557, 2.9773, 3.2059, 3.0268, 1.8483], device='cuda:4'), covar=tensor([0.0475, 0.0050, 0.0054, 0.0373, 0.0092, 0.0098, 0.0096, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0075, 0.0076, 0.0129, 0.0090, 0.0099, 0.0088, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:05:31,690 INFO [train.py:904] (4/8) Epoch 17, batch 5450, loss[loss=0.1964, simple_loss=0.2926, pruned_loss=0.05011, over 16838.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2737, pruned_loss=0.04746, over 3200917.73 frames. ], batch size: 96, lr: 3.99e-03, grad_scale: 4.0 2023-04-30 13:06:13,365 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2118, 3.4244, 3.5576, 3.5623, 3.5512, 3.3800, 3.4102, 3.4532], device='cuda:4'), covar=tensor([0.0416, 0.0643, 0.0548, 0.0444, 0.0534, 0.0600, 0.0864, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0375, 0.0409, 0.0401, 0.0375, 0.0446, 0.0420, 0.0519, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 13:06:48,592 INFO [train.py:904] (4/8) Epoch 17, batch 5500, loss[loss=0.2738, simple_loss=0.3403, pruned_loss=0.1037, over 11904.00 frames. ], tot_loss[loss=0.1928, simple_loss=0.281, pruned_loss=0.05232, over 3161812.50 frames. ], batch size: 248, lr: 3.99e-03, grad_scale: 4.0 2023-04-30 13:07:39,672 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.878e+02 2.929e+02 3.465e+02 4.184e+02 8.083e+02, threshold=6.929e+02, percent-clipped=17.0 2023-04-30 13:08:05,266 INFO [train.py:904] (4/8) Epoch 17, batch 5550, loss[loss=0.2072, simple_loss=0.2897, pruned_loss=0.06234, over 16301.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2872, pruned_loss=0.05638, over 3154740.99 frames. ], batch size: 146, lr: 3.99e-03, grad_scale: 4.0 2023-04-30 13:09:31,455 INFO [train.py:904] (4/8) Epoch 17, batch 5600, loss[loss=0.2729, simple_loss=0.3364, pruned_loss=0.1047, over 11263.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.292, pruned_loss=0.06032, over 3137683.98 frames. ], batch size: 246, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 13:10:27,345 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.908e+02 3.389e+02 4.058e+02 4.967e+02 1.149e+03, threshold=8.116e+02, percent-clipped=3.0 2023-04-30 13:10:27,870 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=168035.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:10:40,920 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-30 13:10:54,415 INFO [train.py:904] (4/8) Epoch 17, batch 5650, loss[loss=0.2163, simple_loss=0.3, pruned_loss=0.06626, over 15307.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.2962, pruned_loss=0.06346, over 3127457.09 frames. ], batch size: 191, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 13:11:30,930 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=168075.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:11:50,819 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168088.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:12:04,650 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=168096.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 13:12:07,945 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0102, 4.0718, 3.8714, 3.6265, 3.6296, 4.0087, 3.6778, 3.7902], device='cuda:4'), covar=tensor([0.0618, 0.0483, 0.0272, 0.0279, 0.0702, 0.0438, 0.0906, 0.0607], device='cuda:4'), in_proj_covar=tensor([0.0274, 0.0386, 0.0326, 0.0314, 0.0335, 0.0368, 0.0222, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:12:12,308 INFO [train.py:904] (4/8) Epoch 17, batch 5700, loss[loss=0.2818, simple_loss=0.335, pruned_loss=0.1144, over 11015.00 frames. ], tot_loss[loss=0.2151, simple_loss=0.2985, pruned_loss=0.06588, over 3099985.80 frames. ], batch size: 248, lr: 3.99e-03, grad_scale: 8.0 2023-04-30 13:12:45,743 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=168123.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:13:04,294 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.185e+02 3.066e+02 3.948e+02 4.929e+02 1.585e+03, threshold=7.895e+02, percent-clipped=5.0 2023-04-30 13:13:06,079 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=168136.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:13:31,682 INFO [train.py:904] (4/8) Epoch 17, batch 5750, loss[loss=0.2199, simple_loss=0.3024, pruned_loss=0.06867, over 16465.00 frames. ], tot_loss[loss=0.218, simple_loss=0.3011, pruned_loss=0.06743, over 3084656.46 frames. ], batch size: 146, lr: 3.98e-03, grad_scale: 8.0 2023-04-30 13:14:50,309 INFO [train.py:904] (4/8) Epoch 17, batch 5800, loss[loss=0.2467, simple_loss=0.3119, pruned_loss=0.09077, over 12009.00 frames. ], tot_loss[loss=0.2176, simple_loss=0.3014, pruned_loss=0.06694, over 3076523.64 frames. ], batch size: 248, lr: 3.98e-03, grad_scale: 8.0 2023-04-30 13:15:41,128 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=168233.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:15:43,217 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.651e+02 2.888e+02 3.447e+02 4.230e+02 9.306e+02, threshold=6.893e+02, percent-clipped=1.0 2023-04-30 13:16:09,620 INFO [train.py:904] (4/8) Epoch 17, batch 5850, loss[loss=0.2242, simple_loss=0.296, pruned_loss=0.07615, over 11781.00 frames. ], tot_loss[loss=0.2158, simple_loss=0.2997, pruned_loss=0.06593, over 3048926.28 frames. ], batch size: 246, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:16:39,653 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0895, 2.4635, 2.6223, 1.8261, 2.6870, 2.8092, 2.4773, 2.3992], device='cuda:4'), covar=tensor([0.0700, 0.0226, 0.0235, 0.1013, 0.0106, 0.0243, 0.0410, 0.0429], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0103, 0.0092, 0.0135, 0.0074, 0.0117, 0.0123, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 13:17:19,339 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=168294.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:17:32,943 INFO [train.py:904] (4/8) Epoch 17, batch 5900, loss[loss=0.2426, simple_loss=0.3094, pruned_loss=0.08794, over 11416.00 frames. ], tot_loss[loss=0.216, simple_loss=0.2998, pruned_loss=0.0661, over 3049902.50 frames. ], batch size: 248, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:18:01,400 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5404, 3.7391, 2.8012, 2.1256, 2.4733, 2.2878, 3.9203, 3.3675], device='cuda:4'), covar=tensor([0.3029, 0.0595, 0.1786, 0.2979, 0.2603, 0.2001, 0.0483, 0.1178], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0261, 0.0297, 0.0300, 0.0289, 0.0242, 0.0285, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:18:22,475 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3519, 3.8682, 3.9663, 2.7012, 3.6414, 3.9948, 3.7633, 2.1707], device='cuda:4'), covar=tensor([0.0519, 0.0058, 0.0055, 0.0385, 0.0086, 0.0102, 0.0076, 0.0482], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0077, 0.0078, 0.0132, 0.0091, 0.0102, 0.0090, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:18:28,194 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.895e+02 2.786e+02 3.302e+02 4.334e+02 9.819e+02, threshold=6.604e+02, percent-clipped=5.0 2023-04-30 13:18:52,630 INFO [train.py:904] (4/8) Epoch 17, batch 5950, loss[loss=0.2046, simple_loss=0.2946, pruned_loss=0.05724, over 17038.00 frames. ], tot_loss[loss=0.2135, simple_loss=0.2993, pruned_loss=0.0639, over 3077842.12 frames. ], batch size: 50, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:18:55,194 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-30 13:19:29,609 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168375.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:19:52,886 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=168391.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 13:20:10,354 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.70 vs. limit=5.0 2023-04-30 13:20:12,026 INFO [train.py:904] (4/8) Epoch 17, batch 6000, loss[loss=0.243, simple_loss=0.3048, pruned_loss=0.09057, over 11354.00 frames. ], tot_loss[loss=0.2131, simple_loss=0.2987, pruned_loss=0.06371, over 3089435.88 frames. ], batch size: 246, lr: 3.98e-03, grad_scale: 8.0 2023-04-30 13:20:12,026 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 13:20:21,980 INFO [train.py:938] (4/8) Epoch 17, validation: loss=0.1535, simple_loss=0.2667, pruned_loss=0.0202, over 944034.00 frames. 2023-04-30 13:20:21,980 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 13:20:53,933 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=168423.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:20:53,976 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168423.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:21:15,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.147e+02 2.768e+02 3.388e+02 4.185e+02 7.743e+02, threshold=6.777e+02, percent-clipped=1.0 2023-04-30 13:21:39,917 INFO [train.py:904] (4/8) Epoch 17, batch 6050, loss[loss=0.192, simple_loss=0.2901, pruned_loss=0.04697, over 16823.00 frames. ], tot_loss[loss=0.2115, simple_loss=0.2974, pruned_loss=0.0628, over 3103454.36 frames. ], batch size: 102, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:21:56,938 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8299, 2.7881, 2.8098, 2.1774, 2.6940, 2.1714, 2.7316, 2.9393], device='cuda:4'), covar=tensor([0.0234, 0.0710, 0.0462, 0.1575, 0.0711, 0.0817, 0.0536, 0.0641], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0157, 0.0162, 0.0148, 0.0141, 0.0126, 0.0140, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 13:22:09,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=168471.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:22:59,844 INFO [train.py:904] (4/8) Epoch 17, batch 6100, loss[loss=0.1911, simple_loss=0.2814, pruned_loss=0.05035, over 15417.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2967, pruned_loss=0.0619, over 3112262.94 frames. ], batch size: 190, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:23:55,308 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.024e+02 2.750e+02 3.352e+02 4.017e+02 8.288e+02, threshold=6.704e+02, percent-clipped=2.0 2023-04-30 13:24:02,635 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5645, 2.2647, 1.8606, 2.1048, 2.5829, 2.2504, 2.3896, 2.7301], device='cuda:4'), covar=tensor([0.0168, 0.0355, 0.0467, 0.0402, 0.0226, 0.0324, 0.0195, 0.0235], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0222, 0.0215, 0.0216, 0.0224, 0.0223, 0.0224, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:24:18,366 INFO [train.py:904] (4/8) Epoch 17, batch 6150, loss[loss=0.2095, simple_loss=0.2944, pruned_loss=0.06233, over 16710.00 frames. ], tot_loss[loss=0.2091, simple_loss=0.2949, pruned_loss=0.06163, over 3100744.95 frames. ], batch size: 83, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:24:49,846 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4528, 5.7675, 5.4667, 5.5480, 5.1630, 5.1140, 5.2250, 5.9246], device='cuda:4'), covar=tensor([0.1256, 0.0824, 0.1098, 0.0829, 0.0828, 0.0746, 0.1018, 0.0840], device='cuda:4'), in_proj_covar=tensor([0.0623, 0.0765, 0.0625, 0.0567, 0.0480, 0.0493, 0.0637, 0.0590], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:25:17,314 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=168589.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:25:36,300 INFO [train.py:904] (4/8) Epoch 17, batch 6200, loss[loss=0.1801, simple_loss=0.2686, pruned_loss=0.04577, over 16892.00 frames. ], tot_loss[loss=0.207, simple_loss=0.2925, pruned_loss=0.06077, over 3093724.63 frames. ], batch size: 96, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:26:11,256 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-30 13:26:31,025 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.813e+02 2.748e+02 3.467e+02 4.001e+02 6.501e+02, threshold=6.934e+02, percent-clipped=0.0 2023-04-30 13:26:42,693 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-30 13:26:53,706 INFO [train.py:904] (4/8) Epoch 17, batch 6250, loss[loss=0.1941, simple_loss=0.2921, pruned_loss=0.04807, over 17100.00 frames. ], tot_loss[loss=0.2066, simple_loss=0.292, pruned_loss=0.06062, over 3084940.68 frames. ], batch size: 49, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:27:36,878 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5210, 3.4780, 3.4395, 2.5875, 3.3613, 2.0503, 3.1384, 2.7334], device='cuda:4'), covar=tensor([0.0176, 0.0143, 0.0197, 0.0270, 0.0117, 0.2314, 0.0157, 0.0248], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0138, 0.0185, 0.0170, 0.0159, 0.0195, 0.0172, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:27:46,014 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8264, 2.7265, 2.7459, 2.0395, 2.6286, 2.1690, 2.7020, 2.9167], device='cuda:4'), covar=tensor([0.0257, 0.0724, 0.0545, 0.1768, 0.0808, 0.0903, 0.0563, 0.0697], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0157, 0.0163, 0.0149, 0.0142, 0.0127, 0.0141, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 13:27:53,795 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168691.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:28:10,636 INFO [train.py:904] (4/8) Epoch 17, batch 6300, loss[loss=0.2037, simple_loss=0.2843, pruned_loss=0.06158, over 16686.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2919, pruned_loss=0.06014, over 3095058.48 frames. ], batch size: 57, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:29:06,429 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.898e+02 2.739e+02 3.263e+02 4.043e+02 7.710e+02, threshold=6.525e+02, percent-clipped=1.0 2023-04-30 13:29:09,982 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=168739.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:29:29,581 INFO [train.py:904] (4/8) Epoch 17, batch 6350, loss[loss=0.1794, simple_loss=0.271, pruned_loss=0.04393, over 16864.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2925, pruned_loss=0.06139, over 3086082.87 frames. ], batch size: 96, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:30:08,870 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2615, 5.3355, 5.7094, 5.7167, 5.7294, 5.4140, 5.3106, 5.0748], device='cuda:4'), covar=tensor([0.0305, 0.0461, 0.0320, 0.0369, 0.0447, 0.0313, 0.0848, 0.0406], device='cuda:4'), in_proj_covar=tensor([0.0381, 0.0415, 0.0406, 0.0381, 0.0452, 0.0428, 0.0525, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 13:30:46,693 INFO [train.py:904] (4/8) Epoch 17, batch 6400, loss[loss=0.1913, simple_loss=0.2776, pruned_loss=0.0525, over 16888.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2921, pruned_loss=0.06172, over 3096455.88 frames. ], batch size: 96, lr: 3.98e-03, grad_scale: 8.0 2023-04-30 13:31:42,103 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.043e+02 2.957e+02 3.621e+02 4.188e+02 7.946e+02, threshold=7.242e+02, percent-clipped=4.0 2023-04-30 13:32:03,554 INFO [train.py:904] (4/8) Epoch 17, batch 6450, loss[loss=0.2114, simple_loss=0.3081, pruned_loss=0.05732, over 16852.00 frames. ], tot_loss[loss=0.207, simple_loss=0.2922, pruned_loss=0.06089, over 3089246.82 frames. ], batch size: 90, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:32:06,443 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.57 vs. limit=5.0 2023-04-30 13:32:16,532 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.70 vs. limit=5.0 2023-04-30 13:33:02,283 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168889.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:33:20,845 INFO [train.py:904] (4/8) Epoch 17, batch 6500, loss[loss=0.1692, simple_loss=0.2523, pruned_loss=0.04304, over 17208.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2903, pruned_loss=0.06024, over 3103223.31 frames. ], batch size: 44, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:33:34,980 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5879, 1.7800, 2.1653, 2.5047, 2.5995, 2.9161, 1.7528, 2.7569], device='cuda:4'), covar=tensor([0.0178, 0.0430, 0.0285, 0.0281, 0.0252, 0.0145, 0.0481, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0186, 0.0172, 0.0175, 0.0185, 0.0143, 0.0188, 0.0137], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:33:43,162 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5254, 3.4919, 2.7196, 2.2011, 2.3135, 2.2590, 3.6136, 3.2219], device='cuda:4'), covar=tensor([0.2775, 0.0680, 0.1675, 0.2607, 0.2653, 0.2064, 0.0456, 0.1244], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0259, 0.0295, 0.0298, 0.0288, 0.0241, 0.0282, 0.0318], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:33:47,134 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.61 vs. limit=5.0 2023-04-30 13:34:15,420 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=168937.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:34:15,486 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5699, 4.8537, 4.6202, 4.6187, 4.3625, 4.3461, 4.2662, 4.9012], device='cuda:4'), covar=tensor([0.1131, 0.0817, 0.0892, 0.0821, 0.0828, 0.1389, 0.1127, 0.0894], device='cuda:4'), in_proj_covar=tensor([0.0631, 0.0772, 0.0632, 0.0574, 0.0486, 0.0497, 0.0642, 0.0597], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:34:15,611 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1514, 3.2314, 1.8928, 3.4047, 2.3633, 3.4498, 2.0685, 2.6431], device='cuda:4'), covar=tensor([0.0280, 0.0374, 0.1646, 0.0259, 0.0922, 0.0692, 0.1528, 0.0705], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0170, 0.0191, 0.0150, 0.0171, 0.0210, 0.0198, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 13:34:16,252 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.781e+02 2.744e+02 3.257e+02 4.082e+02 7.063e+02, threshold=6.513e+02, percent-clipped=0.0 2023-04-30 13:34:22,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2058, 4.2372, 4.3989, 4.1719, 4.3003, 4.8184, 4.3504, 4.0526], device='cuda:4'), covar=tensor([0.1712, 0.2391, 0.2433, 0.2538, 0.2824, 0.1151, 0.1643, 0.2650], device='cuda:4'), in_proj_covar=tensor([0.0387, 0.0558, 0.0609, 0.0468, 0.0629, 0.0643, 0.0482, 0.0628], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:34:40,482 INFO [train.py:904] (4/8) Epoch 17, batch 6550, loss[loss=0.1966, simple_loss=0.2987, pruned_loss=0.04727, over 16784.00 frames. ], tot_loss[loss=0.2085, simple_loss=0.2939, pruned_loss=0.06156, over 3096887.42 frames. ], batch size: 83, lr: 3.98e-03, grad_scale: 4.0 2023-04-30 13:35:03,781 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=168967.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:35:06,295 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=168969.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 13:35:10,329 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-30 13:35:58,162 INFO [train.py:904] (4/8) Epoch 17, batch 6600, loss[loss=0.1908, simple_loss=0.2882, pruned_loss=0.04668, over 16923.00 frames. ], tot_loss[loss=0.2101, simple_loss=0.2962, pruned_loss=0.06202, over 3099264.26 frames. ], batch size: 96, lr: 3.97e-03, grad_scale: 4.0 2023-04-30 13:36:17,568 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2012, 4.2373, 4.5716, 4.5376, 4.5716, 4.2920, 4.2633, 4.1980], device='cuda:4'), covar=tensor([0.0329, 0.0549, 0.0415, 0.0437, 0.0459, 0.0416, 0.0897, 0.0530], device='cuda:4'), in_proj_covar=tensor([0.0382, 0.0416, 0.0407, 0.0383, 0.0453, 0.0429, 0.0527, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 13:36:37,604 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169028.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:36:39,983 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169030.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 13:36:50,873 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.775e+02 2.872e+02 3.571e+02 4.497e+02 9.620e+02, threshold=7.142e+02, percent-clipped=5.0 2023-04-30 13:36:52,655 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8115, 4.6818, 4.8512, 5.0258, 5.2204, 4.6617, 5.2094, 5.2187], device='cuda:4'), covar=tensor([0.1643, 0.1263, 0.1590, 0.0702, 0.0498, 0.0800, 0.0521, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0586, 0.0721, 0.0859, 0.0740, 0.0555, 0.0588, 0.0592, 0.0690], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:37:13,179 INFO [train.py:904] (4/8) Epoch 17, batch 6650, loss[loss=0.1834, simple_loss=0.2698, pruned_loss=0.04849, over 16389.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2962, pruned_loss=0.06234, over 3125693.26 frames. ], batch size: 35, lr: 3.97e-03, grad_scale: 4.0 2023-04-30 13:37:59,851 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169082.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:38:30,528 INFO [train.py:904] (4/8) Epoch 17, batch 6700, loss[loss=0.2621, simple_loss=0.3214, pruned_loss=0.1014, over 11721.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2948, pruned_loss=0.06259, over 3115301.26 frames. ], batch size: 247, lr: 3.97e-03, grad_scale: 4.0 2023-04-30 13:38:53,622 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169116.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:39:26,519 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.893e+02 2.876e+02 3.506e+02 4.097e+02 7.829e+02, threshold=7.011e+02, percent-clipped=1.0 2023-04-30 13:39:34,910 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169143.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:39:47,624 INFO [train.py:904] (4/8) Epoch 17, batch 6750, loss[loss=0.1937, simple_loss=0.2797, pruned_loss=0.0539, over 16529.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2925, pruned_loss=0.06176, over 3126877.88 frames. ], batch size: 68, lr: 3.97e-03, grad_scale: 4.0 2023-04-30 13:39:58,662 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169159.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:39:58,697 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4060, 2.1734, 1.6919, 1.9584, 2.5082, 2.1742, 2.3650, 2.6554], device='cuda:4'), covar=tensor([0.0176, 0.0386, 0.0504, 0.0410, 0.0212, 0.0337, 0.0193, 0.0230], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0224, 0.0217, 0.0217, 0.0225, 0.0223, 0.0226, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:40:14,531 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1904, 5.8360, 5.9976, 5.6371, 5.8462, 6.2725, 5.7979, 5.5892], device='cuda:4'), covar=tensor([0.0821, 0.1702, 0.1714, 0.1773, 0.1833, 0.0881, 0.1444, 0.2114], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0559, 0.0611, 0.0469, 0.0631, 0.0646, 0.0488, 0.0631], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:40:26,155 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169177.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:40:38,894 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9618, 5.4881, 5.6842, 5.3612, 5.4381, 6.0030, 5.4500, 5.2404], device='cuda:4'), covar=tensor([0.0985, 0.1750, 0.1841, 0.1816, 0.2341, 0.0889, 0.1506, 0.2272], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0559, 0.0611, 0.0468, 0.0632, 0.0646, 0.0487, 0.0631], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:40:59,502 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0993, 5.0692, 4.8851, 4.2036, 4.9588, 1.8215, 4.6897, 4.6902], device='cuda:4'), covar=tensor([0.0076, 0.0073, 0.0167, 0.0383, 0.0084, 0.2655, 0.0134, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0139, 0.0186, 0.0170, 0.0159, 0.0196, 0.0172, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:41:04,381 INFO [train.py:904] (4/8) Epoch 17, batch 6800, loss[loss=0.2195, simple_loss=0.3061, pruned_loss=0.06644, over 16743.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2928, pruned_loss=0.06185, over 3114419.91 frames. ], batch size: 124, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:41:33,643 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169220.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:41:44,481 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-30 13:41:58,900 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3638, 3.5097, 3.5637, 2.2528, 2.9969, 2.4021, 3.7029, 3.8311], device='cuda:4'), covar=tensor([0.0302, 0.0919, 0.0654, 0.2116, 0.0993, 0.0999, 0.0723, 0.1104], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0159, 0.0165, 0.0151, 0.0143, 0.0127, 0.0142, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 13:42:02,687 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.174e+02 2.920e+02 3.483e+02 4.404e+02 7.991e+02, threshold=6.967e+02, percent-clipped=1.0 2023-04-30 13:42:08,565 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8092, 4.0370, 3.0856, 2.3547, 2.7935, 2.6459, 4.2972, 3.6823], device='cuda:4'), covar=tensor([0.2695, 0.0610, 0.1663, 0.2658, 0.2634, 0.1788, 0.0484, 0.1099], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0261, 0.0297, 0.0299, 0.0290, 0.0243, 0.0284, 0.0321], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:42:23,105 INFO [train.py:904] (4/8) Epoch 17, batch 6850, loss[loss=0.1839, simple_loss=0.2879, pruned_loss=0.03992, over 16823.00 frames. ], tot_loss[loss=0.2089, simple_loss=0.294, pruned_loss=0.06192, over 3122188.57 frames. ], batch size: 102, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:43:37,898 INFO [train.py:904] (4/8) Epoch 17, batch 6900, loss[loss=0.2583, simple_loss=0.3195, pruned_loss=0.09856, over 11612.00 frames. ], tot_loss[loss=0.2097, simple_loss=0.296, pruned_loss=0.06169, over 3117899.43 frames. ], batch size: 246, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:44:02,250 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-30 13:44:03,426 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-30 13:44:06,504 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-04-30 13:44:10,300 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169323.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:44:13,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3435, 3.3381, 3.3752, 3.4817, 3.5095, 3.2854, 3.4839, 3.5559], device='cuda:4'), covar=tensor([0.1246, 0.0956, 0.1069, 0.0629, 0.0711, 0.2468, 0.1058, 0.0839], device='cuda:4'), in_proj_covar=tensor([0.0586, 0.0724, 0.0861, 0.0740, 0.0558, 0.0587, 0.0594, 0.0690], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:44:14,314 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169325.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 13:44:21,554 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8357, 3.2700, 3.2624, 1.8641, 2.8788, 2.2124, 3.3036, 3.5253], device='cuda:4'), covar=tensor([0.0350, 0.0825, 0.0687, 0.2170, 0.0896, 0.1051, 0.0748, 0.0867], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0158, 0.0164, 0.0150, 0.0142, 0.0127, 0.0142, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 13:44:33,264 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.645e+02 2.634e+02 3.268e+02 4.133e+02 7.362e+02, threshold=6.535e+02, percent-clipped=1.0 2023-04-30 13:44:55,623 INFO [train.py:904] (4/8) Epoch 17, batch 6950, loss[loss=0.2135, simple_loss=0.2987, pruned_loss=0.0641, over 17057.00 frames. ], tot_loss[loss=0.2137, simple_loss=0.2988, pruned_loss=0.06433, over 3089368.85 frames. ], batch size: 53, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:45:58,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1920, 5.4715, 5.2046, 5.2177, 4.9390, 4.9643, 4.8666, 5.5467], device='cuda:4'), covar=tensor([0.1051, 0.0779, 0.0968, 0.0801, 0.0851, 0.0744, 0.1187, 0.0798], device='cuda:4'), in_proj_covar=tensor([0.0629, 0.0766, 0.0625, 0.0569, 0.0481, 0.0493, 0.0634, 0.0593], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:46:11,848 INFO [train.py:904] (4/8) Epoch 17, batch 7000, loss[loss=0.1937, simple_loss=0.2962, pruned_loss=0.04561, over 16448.00 frames. ], tot_loss[loss=0.2138, simple_loss=0.2993, pruned_loss=0.06411, over 3086217.22 frames. ], batch size: 68, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:46:41,088 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.70 vs. limit=5.0 2023-04-30 13:47:07,099 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.910e+02 3.013e+02 3.558e+02 4.321e+02 1.050e+03, threshold=7.116e+02, percent-clipped=3.0 2023-04-30 13:47:07,440 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169438.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:47:28,661 INFO [train.py:904] (4/8) Epoch 17, batch 7050, loss[loss=0.1858, simple_loss=0.2828, pruned_loss=0.04438, over 16876.00 frames. ], tot_loss[loss=0.2135, simple_loss=0.2998, pruned_loss=0.06364, over 3089166.76 frames. ], batch size: 96, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:47:47,705 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-30 13:48:00,525 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169472.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:48:47,182 INFO [train.py:904] (4/8) Epoch 17, batch 7100, loss[loss=0.2083, simple_loss=0.2896, pruned_loss=0.06356, over 16374.00 frames. ], tot_loss[loss=0.2135, simple_loss=0.2988, pruned_loss=0.06417, over 3059839.72 frames. ], batch size: 146, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:49:08,054 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169515.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:49:22,031 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-30 13:49:27,128 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3227, 2.9555, 2.9243, 1.9209, 2.6542, 2.0407, 2.9939, 3.1482], device='cuda:4'), covar=tensor([0.0308, 0.0712, 0.0625, 0.2054, 0.0861, 0.1011, 0.0712, 0.0840], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0158, 0.0164, 0.0150, 0.0142, 0.0127, 0.0142, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 13:49:42,141 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.978e+02 2.876e+02 3.484e+02 4.198e+02 1.044e+03, threshold=6.968e+02, percent-clipped=3.0 2023-04-30 13:50:02,970 INFO [train.py:904] (4/8) Epoch 17, batch 7150, loss[loss=0.2621, simple_loss=0.3136, pruned_loss=0.1053, over 11544.00 frames. ], tot_loss[loss=0.212, simple_loss=0.2964, pruned_loss=0.06382, over 3067123.26 frames. ], batch size: 248, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:50:35,879 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8842, 1.9527, 2.3430, 3.1192, 2.1813, 2.1911, 2.2012, 2.0561], device='cuda:4'), covar=tensor([0.1185, 0.3510, 0.2170, 0.0605, 0.3861, 0.2372, 0.2978, 0.3450], device='cuda:4'), in_proj_covar=tensor([0.0380, 0.0420, 0.0349, 0.0316, 0.0424, 0.0486, 0.0391, 0.0491], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:50:49,485 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-30 13:51:19,559 INFO [train.py:904] (4/8) Epoch 17, batch 7200, loss[loss=0.2017, simple_loss=0.2793, pruned_loss=0.06204, over 11655.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2935, pruned_loss=0.06147, over 3070567.97 frames. ], batch size: 246, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:51:48,691 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169621.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:51:51,988 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169623.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:51:55,185 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169625.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 13:52:16,448 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.648e+02 2.818e+02 3.396e+02 4.262e+02 6.944e+02, threshold=6.791e+02, percent-clipped=0.0 2023-04-30 13:52:38,958 INFO [train.py:904] (4/8) Epoch 17, batch 7250, loss[loss=0.1654, simple_loss=0.2555, pruned_loss=0.03765, over 16703.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.2911, pruned_loss=0.06023, over 3060898.03 frames. ], batch size: 89, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:53:08,592 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=169671.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:53:12,520 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=169673.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 13:53:14,884 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7946, 4.2819, 4.4322, 1.9814, 4.6772, 4.9037, 3.4122, 3.4284], device='cuda:4'), covar=tensor([0.1137, 0.0141, 0.0183, 0.1438, 0.0076, 0.0100, 0.0429, 0.0543], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0105, 0.0093, 0.0138, 0.0075, 0.0119, 0.0126, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 13:53:25,490 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169682.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:53:37,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0772, 3.5366, 3.4949, 2.2187, 3.2321, 3.5009, 3.3356, 1.9977], device='cuda:4'), covar=tensor([0.0499, 0.0044, 0.0049, 0.0400, 0.0085, 0.0108, 0.0088, 0.0414], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0076, 0.0077, 0.0131, 0.0090, 0.0102, 0.0089, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 13:53:57,113 INFO [train.py:904] (4/8) Epoch 17, batch 7300, loss[loss=0.22, simple_loss=0.2916, pruned_loss=0.07415, over 11369.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.291, pruned_loss=0.0603, over 3068750.02 frames. ], batch size: 247, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:54:41,436 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169729.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:54:55,061 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.978e+02 2.965e+02 3.498e+02 4.419e+02 8.107e+02, threshold=6.996e+02, percent-clipped=3.0 2023-04-30 13:54:55,567 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169738.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:55:10,620 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169747.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:55:16,956 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9364, 4.7718, 4.9588, 5.1334, 5.3017, 4.7502, 5.2972, 5.3264], device='cuda:4'), covar=tensor([0.1651, 0.1195, 0.1558, 0.0661, 0.0549, 0.0807, 0.0527, 0.0549], device='cuda:4'), in_proj_covar=tensor([0.0577, 0.0714, 0.0850, 0.0729, 0.0552, 0.0580, 0.0590, 0.0681], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:55:17,626 INFO [train.py:904] (4/8) Epoch 17, batch 7350, loss[loss=0.219, simple_loss=0.2969, pruned_loss=0.07057, over 17040.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2917, pruned_loss=0.06068, over 3082658.81 frames. ], batch size: 53, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:55:49,801 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169772.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:56:11,061 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9546, 3.2746, 3.4481, 1.9945, 2.9688, 2.2931, 3.4899, 3.5367], device='cuda:4'), covar=tensor([0.0268, 0.0797, 0.0586, 0.2058, 0.0838, 0.0952, 0.0676, 0.0989], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0160, 0.0165, 0.0151, 0.0144, 0.0128, 0.0143, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 13:56:12,136 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=169786.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:56:12,341 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169786.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:56:19,812 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169790.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:56:37,950 INFO [train.py:904] (4/8) Epoch 17, batch 7400, loss[loss=0.2214, simple_loss=0.3047, pruned_loss=0.06904, over 16428.00 frames. ], tot_loss[loss=0.2075, simple_loss=0.2926, pruned_loss=0.0612, over 3073021.36 frames. ], batch size: 75, lr: 3.97e-03, grad_scale: 8.0 2023-04-30 13:56:48,610 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169808.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:56:59,824 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169815.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:57:07,983 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=169820.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:57:36,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.712e+02 2.775e+02 3.366e+02 3.998e+02 8.087e+02, threshold=6.732e+02, percent-clipped=3.0 2023-04-30 13:57:52,295 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169847.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:57:59,465 INFO [train.py:904] (4/8) Epoch 17, batch 7450, loss[loss=0.1942, simple_loss=0.2936, pruned_loss=0.04743, over 16879.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2936, pruned_loss=0.062, over 3075967.03 frames. ], batch size: 83, lr: 3.96e-03, grad_scale: 8.0 2023-04-30 13:58:18,999 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=169863.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 13:58:25,454 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.62 vs. limit=5.0 2023-04-30 13:58:50,184 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8414, 2.7825, 2.2614, 2.6500, 3.2598, 2.8942, 3.5289, 3.4561], device='cuda:4'), covar=tensor([0.0082, 0.0386, 0.0494, 0.0379, 0.0199, 0.0330, 0.0172, 0.0224], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0220, 0.0213, 0.0214, 0.0220, 0.0219, 0.0221, 0.0214], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 13:59:20,328 INFO [train.py:904] (4/8) Epoch 17, batch 7500, loss[loss=0.2026, simple_loss=0.2806, pruned_loss=0.06226, over 16976.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2934, pruned_loss=0.06146, over 3089068.88 frames. ], batch size: 55, lr: 3.96e-03, grad_scale: 8.0 2023-04-30 14:00:17,627 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.828e+02 2.669e+02 3.210e+02 3.850e+02 6.683e+02, threshold=6.420e+02, percent-clipped=0.0 2023-04-30 14:00:23,259 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 14:00:39,248 INFO [train.py:904] (4/8) Epoch 17, batch 7550, loss[loss=0.1906, simple_loss=0.2782, pruned_loss=0.05147, over 16355.00 frames. ], tot_loss[loss=0.2074, simple_loss=0.2921, pruned_loss=0.06132, over 3090607.21 frames. ], batch size: 146, lr: 3.96e-03, grad_scale: 8.0 2023-04-30 14:01:19,594 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169977.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:02:00,832 INFO [train.py:904] (4/8) Epoch 17, batch 7600, loss[loss=0.2093, simple_loss=0.2929, pruned_loss=0.06284, over 15350.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2919, pruned_loss=0.0619, over 3087171.58 frames. ], batch size: 190, lr: 3.96e-03, grad_scale: 8.0 2023-04-30 14:02:40,027 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4845, 3.5475, 3.3109, 2.8916, 3.1576, 3.4482, 3.2963, 3.2538], device='cuda:4'), covar=tensor([0.0594, 0.0538, 0.0266, 0.0271, 0.0534, 0.0423, 0.1134, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0377, 0.0316, 0.0302, 0.0324, 0.0353, 0.0216, 0.0376], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:02:58,189 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.022e+02 2.819e+02 3.511e+02 4.461e+02 7.625e+02, threshold=7.022e+02, percent-clipped=5.0 2023-04-30 14:03:08,083 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=170045.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:03:17,364 INFO [train.py:904] (4/8) Epoch 17, batch 7650, loss[loss=0.2059, simple_loss=0.2925, pruned_loss=0.05959, over 16291.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2935, pruned_loss=0.06316, over 3076357.95 frames. ], batch size: 165, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:04:07,611 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170085.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:04:24,703 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7551, 4.7462, 4.5940, 3.8919, 4.6961, 1.6680, 4.4406, 4.3747], device='cuda:4'), covar=tensor([0.0091, 0.0076, 0.0171, 0.0366, 0.0079, 0.2731, 0.0120, 0.0208], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0136, 0.0183, 0.0167, 0.0155, 0.0193, 0.0169, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:04:33,127 INFO [train.py:904] (4/8) Epoch 17, batch 7700, loss[loss=0.209, simple_loss=0.2951, pruned_loss=0.06146, over 16358.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.2938, pruned_loss=0.0637, over 3081984.64 frames. ], batch size: 146, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:04:35,306 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170103.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:04:39,599 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=170106.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:05:29,557 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.892e+02 2.959e+02 3.624e+02 4.617e+02 9.999e+02, threshold=7.248e+02, percent-clipped=4.0 2023-04-30 14:05:34,959 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170142.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:05:50,026 INFO [train.py:904] (4/8) Epoch 17, batch 7750, loss[loss=0.2191, simple_loss=0.3049, pruned_loss=0.06663, over 16258.00 frames. ], tot_loss[loss=0.2102, simple_loss=0.2936, pruned_loss=0.06346, over 3075438.88 frames. ], batch size: 165, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:06:57,183 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6247, 2.7019, 2.1249, 2.3134, 3.0782, 2.7042, 3.2670, 3.3350], device='cuda:4'), covar=tensor([0.0102, 0.0372, 0.0515, 0.0438, 0.0212, 0.0343, 0.0229, 0.0198], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0222, 0.0215, 0.0216, 0.0221, 0.0220, 0.0222, 0.0215], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:07:00,130 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8515, 1.2804, 1.7116, 1.6207, 1.8030, 1.9087, 1.5319, 1.7691], device='cuda:4'), covar=tensor([0.0221, 0.0363, 0.0185, 0.0277, 0.0234, 0.0156, 0.0378, 0.0120], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0183, 0.0170, 0.0173, 0.0182, 0.0141, 0.0186, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:07:07,218 INFO [train.py:904] (4/8) Epoch 17, batch 7800, loss[loss=0.2534, simple_loss=0.3128, pruned_loss=0.09699, over 11374.00 frames. ], tot_loss[loss=0.2108, simple_loss=0.2945, pruned_loss=0.06353, over 3089988.23 frames. ], batch size: 246, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:08:03,351 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.635e+02 2.884e+02 3.642e+02 4.850e+02 8.293e+02, threshold=7.285e+02, percent-clipped=4.0 2023-04-30 14:08:09,959 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-30 14:08:23,347 INFO [train.py:904] (4/8) Epoch 17, batch 7850, loss[loss=0.2233, simple_loss=0.3055, pruned_loss=0.07053, over 16649.00 frames. ], tot_loss[loss=0.2102, simple_loss=0.2947, pruned_loss=0.06282, over 3079793.37 frames. ], batch size: 134, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:08:29,322 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7767, 3.8008, 3.9304, 3.7485, 3.8548, 4.2786, 3.8750, 3.6742], device='cuda:4'), covar=tensor([0.2430, 0.2210, 0.2741, 0.2516, 0.2874, 0.1793, 0.1867, 0.2684], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0560, 0.0616, 0.0469, 0.0632, 0.0648, 0.0488, 0.0630], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 14:09:01,008 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170277.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:09:38,288 INFO [train.py:904] (4/8) Epoch 17, batch 7900, loss[loss=0.2484, simple_loss=0.3114, pruned_loss=0.09273, over 11589.00 frames. ], tot_loss[loss=0.2089, simple_loss=0.2937, pruned_loss=0.06209, over 3094271.80 frames. ], batch size: 247, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:10:13,071 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=170325.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:10:20,361 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6765, 3.1708, 3.1860, 1.8976, 2.7878, 2.1405, 3.2962, 3.3691], device='cuda:4'), covar=tensor([0.0273, 0.0773, 0.0600, 0.2006, 0.0854, 0.1024, 0.0614, 0.0816], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0157, 0.0163, 0.0149, 0.0142, 0.0126, 0.0141, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 14:10:26,873 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-30 14:10:30,088 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8446, 3.1114, 3.0596, 1.9217, 2.9556, 3.1226, 3.0189, 1.7455], device='cuda:4'), covar=tensor([0.0588, 0.0089, 0.0089, 0.0487, 0.0123, 0.0159, 0.0114, 0.0567], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0077, 0.0078, 0.0132, 0.0091, 0.0103, 0.0090, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 14:10:36,113 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.088e+02 2.800e+02 3.465e+02 4.175e+02 8.586e+02, threshold=6.929e+02, percent-clipped=4.0 2023-04-30 14:10:55,565 INFO [train.py:904] (4/8) Epoch 17, batch 7950, loss[loss=0.253, simple_loss=0.3201, pruned_loss=0.09294, over 11741.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2942, pruned_loss=0.06282, over 3084552.44 frames. ], batch size: 246, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:11:49,407 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170385.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:11:53,819 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3713, 3.4315, 1.9508, 3.7371, 2.5343, 3.7808, 2.1973, 2.7602], device='cuda:4'), covar=tensor([0.0269, 0.0353, 0.1708, 0.0207, 0.0800, 0.0586, 0.1505, 0.0711], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0170, 0.0192, 0.0150, 0.0172, 0.0211, 0.0200, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 14:12:13,321 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170401.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:12:14,140 INFO [train.py:904] (4/8) Epoch 17, batch 8000, loss[loss=0.1888, simple_loss=0.2781, pruned_loss=0.0497, over 17002.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.2951, pruned_loss=0.06388, over 3067251.59 frames. ], batch size: 41, lr: 3.96e-03, grad_scale: 8.0 2023-04-30 14:12:15,876 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170403.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:13:00,978 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=170433.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:13:11,711 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.959e+02 2.905e+02 3.477e+02 4.096e+02 7.787e+02, threshold=6.954e+02, percent-clipped=1.0 2023-04-30 14:13:15,193 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170442.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:13:29,061 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=170451.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:13:29,848 INFO [train.py:904] (4/8) Epoch 17, batch 8050, loss[loss=0.195, simple_loss=0.2818, pruned_loss=0.05412, over 16365.00 frames. ], tot_loss[loss=0.211, simple_loss=0.2952, pruned_loss=0.06343, over 3093920.09 frames. ], batch size: 146, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:13:44,249 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=170461.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:14:28,130 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=170490.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:14:28,396 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4226, 2.9380, 2.6929, 2.2493, 2.2395, 2.2090, 2.9404, 2.8599], device='cuda:4'), covar=tensor([0.2330, 0.0810, 0.1503, 0.2394, 0.2440, 0.2039, 0.0509, 0.1234], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0261, 0.0300, 0.0301, 0.0291, 0.0244, 0.0287, 0.0322], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 14:14:37,502 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3586, 3.4126, 2.0146, 3.7251, 2.5387, 3.7964, 2.1931, 2.7834], device='cuda:4'), covar=tensor([0.0247, 0.0348, 0.1631, 0.0201, 0.0811, 0.0516, 0.1525, 0.0715], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0169, 0.0191, 0.0149, 0.0171, 0.0209, 0.0198, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 14:14:46,136 INFO [train.py:904] (4/8) Epoch 17, batch 8100, loss[loss=0.196, simple_loss=0.2944, pruned_loss=0.04881, over 16528.00 frames. ], tot_loss[loss=0.2097, simple_loss=0.2948, pruned_loss=0.06224, over 3110096.16 frames. ], batch size: 68, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:14:50,583 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2556, 3.4177, 3.6056, 3.5817, 3.5836, 3.3862, 3.4333, 3.4819], device='cuda:4'), covar=tensor([0.0405, 0.0666, 0.0408, 0.0399, 0.0508, 0.0530, 0.0791, 0.0505], device='cuda:4'), in_proj_covar=tensor([0.0376, 0.0408, 0.0402, 0.0379, 0.0452, 0.0423, 0.0519, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 14:14:52,539 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-30 14:15:16,582 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=170522.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 14:15:42,385 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.803e+02 2.570e+02 3.195e+02 3.934e+02 7.205e+02, threshold=6.390e+02, percent-clipped=3.0 2023-04-30 14:15:58,978 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2121, 4.3318, 4.4858, 4.2938, 4.3669, 4.8368, 4.3667, 4.1376], device='cuda:4'), covar=tensor([0.1735, 0.1985, 0.2156, 0.1939, 0.2512, 0.1095, 0.1623, 0.2447], device='cuda:4'), in_proj_covar=tensor([0.0384, 0.0556, 0.0611, 0.0466, 0.0626, 0.0647, 0.0483, 0.0627], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 14:16:00,978 INFO [train.py:904] (4/8) Epoch 17, batch 8150, loss[loss=0.2237, simple_loss=0.2989, pruned_loss=0.07424, over 15342.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2928, pruned_loss=0.06161, over 3096562.84 frames. ], batch size: 191, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:16:27,580 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=170569.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:17:11,447 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4999, 3.2722, 2.7590, 2.0865, 2.1872, 2.2870, 3.3611, 3.0494], device='cuda:4'), covar=tensor([0.2771, 0.0707, 0.1689, 0.2766, 0.2614, 0.2029, 0.0519, 0.1254], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0262, 0.0300, 0.0302, 0.0292, 0.0244, 0.0286, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 14:17:18,489 INFO [train.py:904] (4/8) Epoch 17, batch 8200, loss[loss=0.2192, simple_loss=0.2949, pruned_loss=0.07176, over 11479.00 frames. ], tot_loss[loss=0.206, simple_loss=0.29, pruned_loss=0.06095, over 3089336.86 frames. ], batch size: 246, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:18:05,930 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=170630.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:18:21,306 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.868e+02 2.599e+02 3.232e+02 3.801e+02 6.958e+02, threshold=6.463e+02, percent-clipped=1.0 2023-04-30 14:18:41,069 INFO [train.py:904] (4/8) Epoch 17, batch 8250, loss[loss=0.1673, simple_loss=0.2614, pruned_loss=0.03664, over 12588.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2887, pruned_loss=0.05842, over 3083782.90 frames. ], batch size: 248, lr: 3.96e-03, grad_scale: 4.0 2023-04-30 14:20:04,044 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170701.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:20:05,297 INFO [train.py:904] (4/8) Epoch 17, batch 8300, loss[loss=0.162, simple_loss=0.2563, pruned_loss=0.03381, over 16856.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2859, pruned_loss=0.05583, over 3069832.38 frames. ], batch size: 42, lr: 3.95e-03, grad_scale: 4.0 2023-04-30 14:21:09,571 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 2.327e+02 2.774e+02 3.132e+02 7.452e+02, threshold=5.547e+02, percent-clipped=1.0 2023-04-30 14:21:26,073 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=170749.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:21:29,854 INFO [train.py:904] (4/8) Epoch 17, batch 8350, loss[loss=0.164, simple_loss=0.2676, pruned_loss=0.03014, over 16893.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.2852, pruned_loss=0.05385, over 3080535.48 frames. ], batch size: 96, lr: 3.95e-03, grad_scale: 4.0 2023-04-30 14:21:36,771 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9609, 3.8550, 4.0461, 4.1587, 4.2766, 3.8704, 4.2010, 4.2748], device='cuda:4'), covar=tensor([0.1658, 0.1089, 0.1244, 0.0657, 0.0569, 0.1526, 0.0762, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0578, 0.0715, 0.0845, 0.0726, 0.0552, 0.0575, 0.0588, 0.0684], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:22:51,858 INFO [train.py:904] (4/8) Epoch 17, batch 8400, loss[loss=0.1795, simple_loss=0.2738, pruned_loss=0.04262, over 16698.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2823, pruned_loss=0.05183, over 3074599.67 frames. ], batch size: 76, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:23:16,980 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170817.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 14:23:52,183 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.513e+02 2.262e+02 2.769e+02 3.288e+02 8.219e+02, threshold=5.539e+02, percent-clipped=2.0 2023-04-30 14:24:10,749 INFO [train.py:904] (4/8) Epoch 17, batch 8450, loss[loss=0.1929, simple_loss=0.2742, pruned_loss=0.0558, over 12638.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.2801, pruned_loss=0.0501, over 3063190.13 frames. ], batch size: 247, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:25:32,473 INFO [train.py:904] (4/8) Epoch 17, batch 8500, loss[loss=0.1632, simple_loss=0.2554, pruned_loss=0.0355, over 15212.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2766, pruned_loss=0.04763, over 3071727.15 frames. ], batch size: 191, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:26:10,255 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170925.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:26:14,681 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=170928.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:26:34,259 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.530e+02 2.118e+02 2.662e+02 3.144e+02 4.893e+02, threshold=5.323e+02, percent-clipped=0.0 2023-04-30 14:26:56,006 INFO [train.py:904] (4/8) Epoch 17, batch 8550, loss[loss=0.1834, simple_loss=0.2636, pruned_loss=0.05157, over 12151.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2742, pruned_loss=0.04649, over 3070398.70 frames. ], batch size: 247, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:28:08,373 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=170989.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:28:15,915 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0159, 1.8024, 1.6042, 1.5551, 1.9674, 1.6260, 1.5768, 1.9892], device='cuda:4'), covar=tensor([0.0175, 0.0302, 0.0428, 0.0387, 0.0228, 0.0301, 0.0198, 0.0238], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0218, 0.0212, 0.0212, 0.0218, 0.0217, 0.0217, 0.0210], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:28:33,701 INFO [train.py:904] (4/8) Epoch 17, batch 8600, loss[loss=0.177, simple_loss=0.2718, pruned_loss=0.04111, over 15367.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.2742, pruned_loss=0.04554, over 3049168.75 frames. ], batch size: 191, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:29:52,152 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.585e+02 2.454e+02 2.845e+02 3.449e+02 5.171e+02, threshold=5.691e+02, percent-clipped=0.0 2023-04-30 14:30:15,669 INFO [train.py:904] (4/8) Epoch 17, batch 8650, loss[loss=0.1768, simple_loss=0.2754, pruned_loss=0.03905, over 16779.00 frames. ], tot_loss[loss=0.1805, simple_loss=0.2725, pruned_loss=0.04418, over 3052159.16 frames. ], batch size: 124, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:31:00,423 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 14:31:43,988 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171092.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:32:02,416 INFO [train.py:904] (4/8) Epoch 17, batch 8700, loss[loss=0.1771, simple_loss=0.2683, pruned_loss=0.043, over 16723.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2694, pruned_loss=0.04289, over 3048976.22 frames. ], batch size: 134, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:32:31,766 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171117.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 14:33:13,274 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.576e+02 2.121e+02 2.631e+02 3.199e+02 5.449e+02, threshold=5.263e+02, percent-clipped=0.0 2023-04-30 14:33:38,783 INFO [train.py:904] (4/8) Epoch 17, batch 8750, loss[loss=0.1617, simple_loss=0.2521, pruned_loss=0.03567, over 12282.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2691, pruned_loss=0.04239, over 3043053.55 frames. ], batch size: 246, lr: 3.95e-03, grad_scale: 4.0 2023-04-30 14:33:43,329 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171153.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:34:00,455 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-30 14:34:07,867 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5764, 3.6428, 3.4129, 3.1405, 3.2531, 3.5339, 3.3119, 3.3914], device='cuda:4'), covar=tensor([0.0591, 0.0501, 0.0257, 0.0230, 0.0479, 0.0397, 0.1388, 0.0438], device='cuda:4'), in_proj_covar=tensor([0.0264, 0.0371, 0.0312, 0.0298, 0.0318, 0.0348, 0.0215, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:34:13,581 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=171165.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:34:43,258 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8352, 1.7761, 2.3149, 2.7066, 2.5668, 3.1954, 2.1892, 3.0654], device='cuda:4'), covar=tensor([0.0193, 0.0537, 0.0356, 0.0287, 0.0326, 0.0159, 0.0427, 0.0139], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0183, 0.0170, 0.0172, 0.0183, 0.0139, 0.0185, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:35:21,395 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171196.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:35:33,195 INFO [train.py:904] (4/8) Epoch 17, batch 8800, loss[loss=0.1756, simple_loss=0.2655, pruned_loss=0.04286, over 12871.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2671, pruned_loss=0.04097, over 3043112.86 frames. ], batch size: 248, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:36:05,718 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1892, 2.9706, 3.3069, 1.5488, 3.4181, 3.5353, 2.7035, 2.5619], device='cuda:4'), covar=tensor([0.0809, 0.0279, 0.0217, 0.1271, 0.0083, 0.0153, 0.0469, 0.0484], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0100, 0.0087, 0.0131, 0.0070, 0.0111, 0.0118, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 14:36:22,000 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171225.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:36:54,975 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.623e+02 2.375e+02 2.916e+02 3.471e+02 5.213e+02, threshold=5.833e+02, percent-clipped=0.0 2023-04-30 14:37:18,413 INFO [train.py:904] (4/8) Epoch 17, batch 8850, loss[loss=0.1942, simple_loss=0.301, pruned_loss=0.04376, over 16307.00 frames. ], tot_loss[loss=0.1757, simple_loss=0.2701, pruned_loss=0.04069, over 3033217.86 frames. ], batch size: 146, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:37:28,386 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171257.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:38:03,152 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=171273.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:38:27,564 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171284.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:39:01,950 INFO [train.py:904] (4/8) Epoch 17, batch 8900, loss[loss=0.1677, simple_loss=0.2576, pruned_loss=0.03893, over 12600.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2707, pruned_loss=0.04005, over 3036422.37 frames. ], batch size: 250, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:40:38,364 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.477e+02 2.247e+02 2.656e+02 3.091e+02 5.659e+02, threshold=5.312e+02, percent-clipped=0.0 2023-04-30 14:40:51,332 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171345.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:41:04,760 INFO [train.py:904] (4/8) Epoch 17, batch 8950, loss[loss=0.1568, simple_loss=0.2491, pruned_loss=0.03227, over 15559.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.27, pruned_loss=0.04036, over 3037192.64 frames. ], batch size: 192, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:41:20,060 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7527, 3.0090, 3.3806, 1.9895, 2.9150, 2.1779, 3.2405, 3.2206], device='cuda:4'), covar=tensor([0.0247, 0.0811, 0.0521, 0.1981, 0.0727, 0.0941, 0.0753, 0.0921], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0150, 0.0158, 0.0145, 0.0137, 0.0123, 0.0136, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 14:42:53,003 INFO [train.py:904] (4/8) Epoch 17, batch 9000, loss[loss=0.1573, simple_loss=0.2488, pruned_loss=0.03291, over 16934.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2667, pruned_loss=0.03924, over 3034262.63 frames. ], batch size: 116, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:42:53,003 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 14:43:02,948 INFO [train.py:938] (4/8) Epoch 17, validation: loss=0.1478, simple_loss=0.2519, pruned_loss=0.02187, over 944034.00 frames. 2023-04-30 14:43:02,948 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 14:43:10,971 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171406.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:44:23,103 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-30 14:44:23,316 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.259e+02 2.012e+02 2.627e+02 3.134e+02 6.797e+02, threshold=5.255e+02, percent-clipped=3.0 2023-04-30 14:44:41,564 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171448.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:44:47,386 INFO [train.py:904] (4/8) Epoch 17, batch 9050, loss[loss=0.1634, simple_loss=0.2524, pruned_loss=0.03725, over 16347.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2685, pruned_loss=0.03995, over 3059176.95 frames. ], batch size: 146, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:45:07,607 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8661, 1.4439, 1.7440, 1.7642, 1.9066, 1.8778, 1.6435, 1.8419], device='cuda:4'), covar=tensor([0.0208, 0.0375, 0.0193, 0.0272, 0.0244, 0.0190, 0.0375, 0.0136], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0182, 0.0169, 0.0171, 0.0182, 0.0139, 0.0184, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:46:24,613 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171498.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:46:30,501 INFO [train.py:904] (4/8) Epoch 17, batch 9100, loss[loss=0.1759, simple_loss=0.2813, pruned_loss=0.03526, over 16276.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2689, pruned_loss=0.04079, over 3068138.50 frames. ], batch size: 146, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:46:46,632 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-30 14:47:02,102 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8462, 2.7125, 2.5144, 4.5376, 2.7801, 4.0180, 1.6616, 2.8131], device='cuda:4'), covar=tensor([0.1418, 0.0858, 0.1331, 0.0168, 0.0207, 0.0407, 0.1721, 0.0928], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0163, 0.0185, 0.0168, 0.0194, 0.0206, 0.0190, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-30 14:48:02,852 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.545e+02 2.178e+02 2.662e+02 3.383e+02 6.301e+02, threshold=5.324e+02, percent-clipped=5.0 2023-04-30 14:48:28,389 INFO [train.py:904] (4/8) Epoch 17, batch 9150, loss[loss=0.1587, simple_loss=0.2548, pruned_loss=0.03127, over 16913.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2685, pruned_loss=0.04022, over 3055795.64 frames. ], batch size: 96, lr: 3.95e-03, grad_scale: 8.0 2023-04-30 14:48:28,878 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171552.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:48:44,313 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171559.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:49:36,114 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171584.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:50:00,680 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8530, 3.8730, 3.9876, 3.7505, 3.8681, 4.3190, 3.9379, 3.6158], device='cuda:4'), covar=tensor([0.1863, 0.1949, 0.1982, 0.2426, 0.2770, 0.1439, 0.1560, 0.2656], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0534, 0.0585, 0.0447, 0.0598, 0.0621, 0.0466, 0.0600], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 14:50:13,333 INFO [train.py:904] (4/8) Epoch 17, batch 9200, loss[loss=0.1775, simple_loss=0.2725, pruned_loss=0.04129, over 16610.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2645, pruned_loss=0.03952, over 3038082.19 frames. ], batch size: 134, lr: 3.94e-03, grad_scale: 8.0 2023-04-30 14:50:13,933 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9543, 4.5885, 4.5283, 3.1340, 3.8777, 4.5337, 4.0990, 2.7445], device='cuda:4'), covar=tensor([0.0453, 0.0027, 0.0041, 0.0373, 0.0108, 0.0079, 0.0062, 0.0419], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0075, 0.0076, 0.0130, 0.0089, 0.0100, 0.0087, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 14:51:10,342 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=171632.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:51:27,839 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.614e+02 2.277e+02 2.610e+02 3.229e+02 7.994e+02, threshold=5.220e+02, percent-clipped=2.0 2023-04-30 14:51:50,813 INFO [train.py:904] (4/8) Epoch 17, batch 9250, loss[loss=0.1794, simple_loss=0.2705, pruned_loss=0.04416, over 16506.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2643, pruned_loss=0.03937, over 3047022.53 frames. ], batch size: 75, lr: 3.94e-03, grad_scale: 8.0 2023-04-30 14:53:42,073 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171701.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:53:42,880 INFO [train.py:904] (4/8) Epoch 17, batch 9300, loss[loss=0.1596, simple_loss=0.2492, pruned_loss=0.03506, over 15341.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.263, pruned_loss=0.03892, over 3047568.62 frames. ], batch size: 191, lr: 3.94e-03, grad_scale: 8.0 2023-04-30 14:54:17,850 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-30 14:54:21,810 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.22 vs. limit=5.0 2023-04-30 14:55:09,819 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 2.061e+02 2.530e+02 2.950e+02 5.223e+02, threshold=5.060e+02, percent-clipped=1.0 2023-04-30 14:55:21,199 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171748.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:55:27,769 INFO [train.py:904] (4/8) Epoch 17, batch 9350, loss[loss=0.1696, simple_loss=0.2611, pruned_loss=0.03902, over 15421.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2623, pruned_loss=0.03865, over 3058158.87 frames. ], batch size: 191, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 14:55:37,375 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-30 14:55:48,817 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1542, 3.4476, 3.3829, 2.3031, 3.1561, 3.4327, 3.2869, 1.9675], device='cuda:4'), covar=tensor([0.0502, 0.0038, 0.0052, 0.0414, 0.0091, 0.0080, 0.0080, 0.0496], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0074, 0.0075, 0.0130, 0.0089, 0.0099, 0.0087, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 14:55:49,189 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.61 vs. limit=5.0 2023-04-30 14:56:00,519 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-30 14:56:58,070 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=171796.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:57:07,774 INFO [train.py:904] (4/8) Epoch 17, batch 9400, loss[loss=0.1754, simple_loss=0.2759, pruned_loss=0.03748, over 16743.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2625, pruned_loss=0.03877, over 3039735.03 frames. ], batch size: 83, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 14:57:28,346 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4302, 3.0249, 2.7373, 2.2641, 2.1961, 2.2828, 2.9822, 2.8129], device='cuda:4'), covar=tensor([0.2443, 0.0683, 0.1395, 0.2713, 0.2689, 0.1990, 0.0424, 0.1348], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0252, 0.0289, 0.0292, 0.0275, 0.0236, 0.0275, 0.0310], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 14:58:29,671 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.619e+02 2.242e+02 2.701e+02 3.369e+02 7.708e+02, threshold=5.403e+02, percent-clipped=2.0 2023-04-30 14:58:48,377 INFO [train.py:904] (4/8) Epoch 17, batch 9450, loss[loss=0.1763, simple_loss=0.2665, pruned_loss=0.04303, over 16954.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2645, pruned_loss=0.03931, over 3027653.11 frames. ], batch size: 109, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 14:58:49,579 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171852.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 14:58:52,140 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171854.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:00:27,218 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=171900.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:00:31,018 INFO [train.py:904] (4/8) Epoch 17, batch 9500, loss[loss=0.1526, simple_loss=0.2527, pruned_loss=0.02621, over 16852.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2642, pruned_loss=0.03875, over 3049229.63 frames. ], batch size: 102, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:01:41,184 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171937.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:01:51,140 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.446e+02 2.125e+02 2.584e+02 3.095e+02 6.778e+02, threshold=5.168e+02, percent-clipped=1.0 2023-04-30 15:02:14,948 INFO [train.py:904] (4/8) Epoch 17, batch 9550, loss[loss=0.1781, simple_loss=0.2684, pruned_loss=0.04396, over 12612.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2644, pruned_loss=0.03886, over 3068268.10 frames. ], batch size: 247, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:03:49,270 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171998.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:03:57,317 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=172001.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:03:58,206 INFO [train.py:904] (4/8) Epoch 17, batch 9600, loss[loss=0.195, simple_loss=0.2913, pruned_loss=0.04935, over 15438.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2659, pruned_loss=0.03964, over 3068735.19 frames. ], batch size: 191, lr: 3.94e-03, grad_scale: 8.0 2023-04-30 15:05:20,542 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.576e+02 2.253e+02 2.956e+02 3.661e+02 6.195e+02, threshold=5.912e+02, percent-clipped=3.0 2023-04-30 15:05:38,178 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=172049.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:05:44,197 INFO [train.py:904] (4/8) Epoch 17, batch 9650, loss[loss=0.1579, simple_loss=0.2529, pruned_loss=0.03144, over 17000.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2676, pruned_loss=0.04002, over 3055139.67 frames. ], batch size: 109, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:06:20,990 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3676, 2.1055, 2.1559, 3.9339, 2.0856, 2.3835, 2.2713, 2.2414], device='cuda:4'), covar=tensor([0.1123, 0.3879, 0.2973, 0.0501, 0.4442, 0.2852, 0.3577, 0.3897], device='cuda:4'), in_proj_covar=tensor([0.0370, 0.0413, 0.0345, 0.0308, 0.0416, 0.0473, 0.0384, 0.0478], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:07:32,479 INFO [train.py:904] (4/8) Epoch 17, batch 9700, loss[loss=0.1814, simple_loss=0.2803, pruned_loss=0.04126, over 15196.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2668, pruned_loss=0.03968, over 3056522.62 frames. ], batch size: 191, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:08:05,524 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=172118.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:08:57,398 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.657e+02 2.330e+02 2.721e+02 3.165e+02 5.976e+02, threshold=5.442e+02, percent-clipped=1.0 2023-04-30 15:09:14,185 INFO [train.py:904] (4/8) Epoch 17, batch 9750, loss[loss=0.1433, simple_loss=0.2479, pruned_loss=0.01929, over 16722.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2652, pruned_loss=0.03961, over 3064064.03 frames. ], batch size: 89, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:09:18,971 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=172154.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:10:10,091 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=172179.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:10:19,101 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1673, 4.5201, 4.5003, 3.4586, 3.9288, 4.4517, 3.9851, 3.0209], device='cuda:4'), covar=tensor([0.0351, 0.0024, 0.0026, 0.0240, 0.0080, 0.0063, 0.0057, 0.0314], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0075, 0.0075, 0.0131, 0.0089, 0.0099, 0.0087, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 15:10:54,091 INFO [train.py:904] (4/8) Epoch 17, batch 9800, loss[loss=0.1899, simple_loss=0.2911, pruned_loss=0.04428, over 15400.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2654, pruned_loss=0.03852, over 3073647.86 frames. ], batch size: 190, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:10:54,779 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=172202.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:10:56,998 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2032, 4.2898, 4.4597, 4.2358, 4.3946, 4.8348, 4.4193, 4.0916], device='cuda:4'), covar=tensor([0.1560, 0.2033, 0.1869, 0.2183, 0.2430, 0.1109, 0.1631, 0.2786], device='cuda:4'), in_proj_covar=tensor([0.0360, 0.0531, 0.0584, 0.0444, 0.0597, 0.0616, 0.0462, 0.0590], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 15:11:14,244 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8432, 5.1535, 4.9538, 4.9917, 4.7122, 4.6251, 4.6249, 5.2357], device='cuda:4'), covar=tensor([0.1063, 0.0808, 0.0852, 0.0651, 0.0686, 0.0885, 0.0978, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0601, 0.0734, 0.0593, 0.0541, 0.0462, 0.0470, 0.0607, 0.0567], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:12:17,139 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.406e+02 2.047e+02 2.504e+02 3.250e+02 5.993e+02, threshold=5.009e+02, percent-clipped=1.0 2023-04-30 15:12:38,992 INFO [train.py:904] (4/8) Epoch 17, batch 9850, loss[loss=0.1631, simple_loss=0.2494, pruned_loss=0.03834, over 12315.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2661, pruned_loss=0.0382, over 3061314.13 frames. ], batch size: 247, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:13:57,817 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-04-30 15:14:13,936 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=172293.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:14:29,923 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8224, 3.8715, 4.1485, 4.1185, 4.1203, 3.9369, 3.9372, 3.9061], device='cuda:4'), covar=tensor([0.0338, 0.0727, 0.0407, 0.0398, 0.0417, 0.0408, 0.0686, 0.0418], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0384, 0.0378, 0.0358, 0.0422, 0.0395, 0.0481, 0.0315], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 15:14:32,002 INFO [train.py:904] (4/8) Epoch 17, batch 9900, loss[loss=0.1702, simple_loss=0.2684, pruned_loss=0.03597, over 16581.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2663, pruned_loss=0.03797, over 3057032.49 frames. ], batch size: 62, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:14:43,369 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-30 15:16:10,689 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.496e+02 2.206e+02 2.589e+02 2.968e+02 8.308e+02, threshold=5.178e+02, percent-clipped=1.0 2023-04-30 15:16:31,647 INFO [train.py:904] (4/8) Epoch 17, batch 9950, loss[loss=0.1614, simple_loss=0.26, pruned_loss=0.03142, over 17171.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2686, pruned_loss=0.03831, over 3064765.88 frames. ], batch size: 44, lr: 3.94e-03, grad_scale: 4.0 2023-04-30 15:16:35,483 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-30 15:18:33,055 INFO [train.py:904] (4/8) Epoch 17, batch 10000, loss[loss=0.1841, simple_loss=0.2694, pruned_loss=0.04936, over 12552.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2673, pruned_loss=0.03796, over 3073508.17 frames. ], batch size: 250, lr: 3.94e-03, grad_scale: 8.0 2023-04-30 15:19:52,321 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-30 15:19:56,011 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.275e+02 2.112e+02 2.435e+02 3.073e+02 5.359e+02, threshold=4.869e+02, percent-clipped=1.0 2023-04-30 15:20:13,519 INFO [train.py:904] (4/8) Epoch 17, batch 10050, loss[loss=0.1893, simple_loss=0.2814, pruned_loss=0.04861, over 12343.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2674, pruned_loss=0.03822, over 3055760.35 frames. ], batch size: 250, lr: 3.93e-03, grad_scale: 8.0 2023-04-30 15:20:58,574 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=172474.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:21:17,477 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-30 15:21:32,806 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6852, 3.0458, 3.3446, 2.0166, 2.8019, 2.1668, 3.2133, 3.2603], device='cuda:4'), covar=tensor([0.0287, 0.0783, 0.0483, 0.1859, 0.0777, 0.0976, 0.0680, 0.0949], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0147, 0.0158, 0.0145, 0.0136, 0.0123, 0.0136, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 15:21:47,146 INFO [train.py:904] (4/8) Epoch 17, batch 10100, loss[loss=0.1714, simple_loss=0.2609, pruned_loss=0.04098, over 16333.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2675, pruned_loss=0.03832, over 3059073.45 frames. ], batch size: 146, lr: 3.93e-03, grad_scale: 8.0 2023-04-30 15:22:57,875 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.467e+02 2.137e+02 2.474e+02 3.171e+02 5.916e+02, threshold=4.949e+02, percent-clipped=1.0 2023-04-30 15:23:33,082 INFO [train.py:904] (4/8) Epoch 18, batch 0, loss[loss=0.2152, simple_loss=0.2808, pruned_loss=0.07486, over 16774.00 frames. ], tot_loss[loss=0.2152, simple_loss=0.2808, pruned_loss=0.07486, over 16774.00 frames. ], batch size: 83, lr: 3.82e-03, grad_scale: 8.0 2023-04-30 15:23:33,082 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 15:23:40,344 INFO [train.py:938] (4/8) Epoch 18, validation: loss=0.148, simple_loss=0.2515, pruned_loss=0.02223, over 944034.00 frames. 2023-04-30 15:23:40,345 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 15:24:36,440 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=172593.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:24:44,233 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=172598.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:24:50,175 INFO [train.py:904] (4/8) Epoch 18, batch 50, loss[loss=0.1596, simple_loss=0.2449, pruned_loss=0.03716, over 15918.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2738, pruned_loss=0.05052, over 752383.91 frames. ], batch size: 35, lr: 3.82e-03, grad_scale: 1.0 2023-04-30 15:25:21,750 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=172625.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:25:42,961 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=172641.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:25:48,587 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.874e+02 2.462e+02 2.938e+02 3.654e+02 9.185e+02, threshold=5.877e+02, percent-clipped=7.0 2023-04-30 15:25:50,892 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8630, 4.5878, 4.9139, 5.0714, 5.2986, 4.6597, 5.2823, 5.2725], device='cuda:4'), covar=tensor([0.1870, 0.1282, 0.1604, 0.0793, 0.0562, 0.0885, 0.0565, 0.0581], device='cuda:4'), in_proj_covar=tensor([0.0561, 0.0694, 0.0813, 0.0708, 0.0530, 0.0558, 0.0573, 0.0665], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:25:56,033 INFO [train.py:904] (4/8) Epoch 18, batch 100, loss[loss=0.1969, simple_loss=0.2709, pruned_loss=0.0614, over 16886.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.271, pruned_loss=0.05065, over 1327587.90 frames. ], batch size: 96, lr: 3.82e-03, grad_scale: 1.0 2023-04-30 15:25:59,571 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3422, 2.1802, 2.4047, 4.1330, 2.2668, 2.5325, 2.2718, 2.3709], device='cuda:4'), covar=tensor([0.1261, 0.3799, 0.2725, 0.0547, 0.3839, 0.2509, 0.3628, 0.2860], device='cuda:4'), in_proj_covar=tensor([0.0374, 0.0415, 0.0347, 0.0311, 0.0418, 0.0474, 0.0385, 0.0481], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:26:05,927 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=172659.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:26:35,519 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4117, 3.3437, 3.4503, 3.5324, 3.5781, 3.2981, 3.4839, 3.6194], device='cuda:4'), covar=tensor([0.1265, 0.0913, 0.0960, 0.0614, 0.0586, 0.2317, 0.1141, 0.0692], device='cuda:4'), in_proj_covar=tensor([0.0567, 0.0700, 0.0821, 0.0714, 0.0535, 0.0563, 0.0578, 0.0671], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:26:44,037 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=172686.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:27:02,718 INFO [train.py:904] (4/8) Epoch 18, batch 150, loss[loss=0.1896, simple_loss=0.2642, pruned_loss=0.05744, over 16809.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2684, pruned_loss=0.04945, over 1766916.36 frames. ], batch size: 83, lr: 3.82e-03, grad_scale: 1.0 2023-04-30 15:27:46,306 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=172734.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:28:01,449 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.661e+02 2.271e+02 2.721e+02 3.424e+02 7.505e+02, threshold=5.441e+02, percent-clipped=3.0 2023-04-30 15:28:09,983 INFO [train.py:904] (4/8) Epoch 18, batch 200, loss[loss=0.1773, simple_loss=0.2735, pruned_loss=0.04058, over 17120.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2681, pruned_loss=0.04899, over 2118467.45 frames. ], batch size: 49, lr: 3.82e-03, grad_scale: 1.0 2023-04-30 15:28:41,385 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=172774.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:29:10,054 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=172795.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:29:18,648 INFO [train.py:904] (4/8) Epoch 18, batch 250, loss[loss=0.1743, simple_loss=0.249, pruned_loss=0.04983, over 16878.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2665, pruned_loss=0.04902, over 2377472.94 frames. ], batch size: 96, lr: 3.82e-03, grad_scale: 1.0 2023-04-30 15:29:30,270 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-30 15:29:47,727 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=172822.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:30:03,430 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4874, 2.7730, 2.9783, 2.0805, 2.6946, 2.0745, 3.1170, 3.0845], device='cuda:4'), covar=tensor([0.0284, 0.0875, 0.0623, 0.1851, 0.0863, 0.0979, 0.0585, 0.0919], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0153, 0.0163, 0.0150, 0.0140, 0.0126, 0.0140, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 15:30:18,796 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=172845.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:30:19,548 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.374e+02 2.240e+02 2.726e+02 3.152e+02 7.045e+02, threshold=5.451e+02, percent-clipped=2.0 2023-04-30 15:30:24,712 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0060, 1.9757, 2.4724, 2.9247, 2.6757, 3.3486, 2.2950, 3.2799], device='cuda:4'), covar=tensor([0.0214, 0.0450, 0.0319, 0.0273, 0.0308, 0.0169, 0.0422, 0.0171], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0184, 0.0169, 0.0172, 0.0182, 0.0140, 0.0185, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:30:28,996 INFO [train.py:904] (4/8) Epoch 18, batch 300, loss[loss=0.1638, simple_loss=0.2508, pruned_loss=0.03839, over 16560.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2635, pruned_loss=0.0471, over 2595041.13 frames. ], batch size: 75, lr: 3.82e-03, grad_scale: 1.0 2023-04-30 15:31:32,726 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5788, 1.7608, 2.2271, 2.4407, 2.6020, 2.4828, 1.9136, 2.6649], device='cuda:4'), covar=tensor([0.0178, 0.0448, 0.0296, 0.0276, 0.0268, 0.0286, 0.0447, 0.0166], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0185, 0.0170, 0.0172, 0.0182, 0.0141, 0.0186, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:31:39,798 INFO [train.py:904] (4/8) Epoch 18, batch 350, loss[loss=0.1388, simple_loss=0.2218, pruned_loss=0.02794, over 16780.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2601, pruned_loss=0.04532, over 2762256.68 frames. ], batch size: 39, lr: 3.82e-03, grad_scale: 1.0 2023-04-30 15:31:45,366 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=172906.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:31:50,142 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6386, 2.6391, 2.2505, 2.4768, 3.0282, 2.7083, 3.3509, 3.1733], device='cuda:4'), covar=tensor([0.0149, 0.0400, 0.0524, 0.0444, 0.0260, 0.0360, 0.0261, 0.0292], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0229, 0.0220, 0.0222, 0.0228, 0.0227, 0.0228, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:31:54,870 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1067, 2.0368, 2.6258, 3.0249, 2.8461, 3.5208, 2.3670, 3.3881], device='cuda:4'), covar=tensor([0.0228, 0.0492, 0.0326, 0.0294, 0.0319, 0.0156, 0.0470, 0.0173], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0185, 0.0170, 0.0173, 0.0183, 0.0141, 0.0186, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:32:40,483 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.138e+02 2.555e+02 2.961e+02 4.551e+02, threshold=5.109e+02, percent-clipped=0.0 2023-04-30 15:32:47,898 INFO [train.py:904] (4/8) Epoch 18, batch 400, loss[loss=0.1635, simple_loss=0.2593, pruned_loss=0.03382, over 17171.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2583, pruned_loss=0.04503, over 2891461.43 frames. ], batch size: 46, lr: 3.82e-03, grad_scale: 2.0 2023-04-30 15:32:51,726 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=172954.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:33:10,276 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9827, 4.8706, 4.6590, 3.4400, 4.7382, 1.5688, 4.2986, 4.3978], device='cuda:4'), covar=tensor([0.0132, 0.0132, 0.0265, 0.0776, 0.0157, 0.3655, 0.0246, 0.0400], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0139, 0.0183, 0.0166, 0.0159, 0.0198, 0.0172, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:33:30,367 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=172981.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:33:44,208 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.51 vs. limit=5.0 2023-04-30 15:33:59,531 INFO [train.py:904] (4/8) Epoch 18, batch 450, loss[loss=0.1922, simple_loss=0.2651, pruned_loss=0.05959, over 16175.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2575, pruned_loss=0.04453, over 2990674.50 frames. ], batch size: 164, lr: 3.82e-03, grad_scale: 2.0 2023-04-30 15:35:00,082 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.590e+02 2.073e+02 2.463e+02 2.995e+02 6.633e+02, threshold=4.925e+02, percent-clipped=1.0 2023-04-30 15:35:08,319 INFO [train.py:904] (4/8) Epoch 18, batch 500, loss[loss=0.1689, simple_loss=0.245, pruned_loss=0.04639, over 16831.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2561, pruned_loss=0.04386, over 3067181.00 frames. ], batch size: 102, lr: 3.82e-03, grad_scale: 2.0 2023-04-30 15:35:22,612 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6100, 2.5730, 1.6297, 2.6854, 2.1062, 2.8059, 1.9261, 2.3269], device='cuda:4'), covar=tensor([0.0317, 0.0430, 0.1808, 0.0373, 0.0823, 0.0492, 0.1672, 0.0747], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0173, 0.0197, 0.0154, 0.0175, 0.0214, 0.0205, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 15:35:28,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0537, 3.0731, 3.2943, 2.0819, 2.9097, 2.2535, 3.5731, 3.5153], device='cuda:4'), covar=tensor([0.0224, 0.0913, 0.0642, 0.1874, 0.0823, 0.0960, 0.0457, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0155, 0.0163, 0.0150, 0.0141, 0.0126, 0.0140, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 15:35:59,655 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=173090.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:36:14,708 INFO [train.py:904] (4/8) Epoch 18, batch 550, loss[loss=0.1614, simple_loss=0.239, pruned_loss=0.04193, over 16754.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2559, pruned_loss=0.04352, over 3122880.21 frames. ], batch size: 83, lr: 3.82e-03, grad_scale: 2.0 2023-04-30 15:36:34,057 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.68 vs. limit=5.0 2023-04-30 15:37:14,571 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.114e+02 2.403e+02 3.161e+02 6.470e+02, threshold=4.805e+02, percent-clipped=1.0 2023-04-30 15:37:20,868 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=173150.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:37:22,687 INFO [train.py:904] (4/8) Epoch 18, batch 600, loss[loss=0.1807, simple_loss=0.2523, pruned_loss=0.05458, over 16473.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.255, pruned_loss=0.04409, over 3163583.38 frames. ], batch size: 75, lr: 3.81e-03, grad_scale: 2.0 2023-04-30 15:37:45,920 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2834, 2.0826, 2.4069, 4.0041, 2.1648, 2.3414, 2.1857, 2.2341], device='cuda:4'), covar=tensor([0.1542, 0.4266, 0.2933, 0.0740, 0.4750, 0.3176, 0.4009, 0.3808], device='cuda:4'), in_proj_covar=tensor([0.0383, 0.0425, 0.0354, 0.0320, 0.0425, 0.0488, 0.0394, 0.0494], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:37:48,713 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1161, 5.7065, 5.7605, 5.5388, 5.5499, 6.1541, 5.6557, 5.3300], device='cuda:4'), covar=tensor([0.0952, 0.2064, 0.2301, 0.2200, 0.3067, 0.1037, 0.1609, 0.2472], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0564, 0.0619, 0.0468, 0.0634, 0.0653, 0.0492, 0.0631], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 15:38:08,414 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.95 vs. limit=5.0 2023-04-30 15:38:30,577 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=173201.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:38:31,377 INFO [train.py:904] (4/8) Epoch 18, batch 650, loss[loss=0.1812, simple_loss=0.2702, pruned_loss=0.04609, over 17266.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2538, pruned_loss=0.04363, over 3201228.49 frames. ], batch size: 52, lr: 3.81e-03, grad_scale: 2.0 2023-04-30 15:38:45,565 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=173211.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:39:32,444 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.711e+02 2.266e+02 2.757e+02 3.739e+02 8.562e+02, threshold=5.513e+02, percent-clipped=7.0 2023-04-30 15:39:32,828 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2345, 5.1781, 5.0638, 4.5337, 4.6534, 5.1173, 5.0187, 4.6706], device='cuda:4'), covar=tensor([0.0580, 0.0447, 0.0278, 0.0328, 0.1094, 0.0448, 0.0355, 0.0776], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0396, 0.0329, 0.0317, 0.0339, 0.0368, 0.0225, 0.0393], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:39:40,833 INFO [train.py:904] (4/8) Epoch 18, batch 700, loss[loss=0.189, simple_loss=0.2612, pruned_loss=0.05836, over 16444.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2542, pruned_loss=0.04331, over 3235055.15 frames. ], batch size: 146, lr: 3.81e-03, grad_scale: 2.0 2023-04-30 15:39:44,116 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=173254.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:39:58,144 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1738, 5.2370, 5.6561, 5.6393, 5.6815, 5.3104, 5.2435, 5.1013], device='cuda:4'), covar=tensor([0.0327, 0.0521, 0.0453, 0.0517, 0.0440, 0.0354, 0.0903, 0.0453], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0422, 0.0413, 0.0390, 0.0456, 0.0434, 0.0528, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 15:40:22,682 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=173281.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:40:50,347 INFO [train.py:904] (4/8) Epoch 18, batch 750, loss[loss=0.1449, simple_loss=0.231, pruned_loss=0.02937, over 16769.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2553, pruned_loss=0.04376, over 3240699.35 frames. ], batch size: 39, lr: 3.81e-03, grad_scale: 2.0 2023-04-30 15:40:51,630 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=173302.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:41:00,044 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=173309.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:41:26,654 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=173329.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:41:50,586 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.634e+02 2.162e+02 2.529e+02 2.992e+02 5.194e+02, threshold=5.057e+02, percent-clipped=0.0 2023-04-30 15:41:58,280 INFO [train.py:904] (4/8) Epoch 18, batch 800, loss[loss=0.1593, simple_loss=0.2375, pruned_loss=0.0406, over 16742.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2541, pruned_loss=0.04331, over 3257628.11 frames. ], batch size: 89, lr: 3.81e-03, grad_scale: 4.0 2023-04-30 15:42:22,702 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=173370.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:42:49,208 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=173390.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:43:05,169 INFO [train.py:904] (4/8) Epoch 18, batch 850, loss[loss=0.1536, simple_loss=0.239, pruned_loss=0.03409, over 16785.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2539, pruned_loss=0.04301, over 3271577.94 frames. ], batch size: 39, lr: 3.81e-03, grad_scale: 4.0 2023-04-30 15:43:55,000 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=173438.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:44:07,463 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.560e+02 2.076e+02 2.484e+02 3.047e+02 7.232e+02, threshold=4.969e+02, percent-clipped=3.0 2023-04-30 15:44:15,663 INFO [train.py:904] (4/8) Epoch 18, batch 900, loss[loss=0.1629, simple_loss=0.2398, pruned_loss=0.04306, over 16894.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2535, pruned_loss=0.04264, over 3282319.26 frames. ], batch size: 96, lr: 3.81e-03, grad_scale: 4.0 2023-04-30 15:44:30,292 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-30 15:44:32,865 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9719, 4.7052, 4.9756, 5.1913, 5.4614, 4.7137, 5.4044, 5.4128], device='cuda:4'), covar=tensor([0.1898, 0.1313, 0.1774, 0.0837, 0.0503, 0.0947, 0.0498, 0.0560], device='cuda:4'), in_proj_covar=tensor([0.0608, 0.0748, 0.0883, 0.0760, 0.0567, 0.0602, 0.0617, 0.0720], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:45:11,326 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=173491.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 15:45:24,066 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=173501.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:45:25,649 INFO [train.py:904] (4/8) Epoch 18, batch 950, loss[loss=0.1724, simple_loss=0.2601, pruned_loss=0.04229, over 15910.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2542, pruned_loss=0.04249, over 3296032.86 frames. ], batch size: 35, lr: 3.81e-03, grad_scale: 4.0 2023-04-30 15:45:31,802 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=173506.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:45:50,920 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-30 15:46:00,706 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-30 15:46:04,255 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0171, 5.3692, 5.1001, 5.1711, 4.8731, 4.8242, 4.8569, 5.4779], device='cuda:4'), covar=tensor([0.1306, 0.0920, 0.1160, 0.0882, 0.0845, 0.0981, 0.1226, 0.0879], device='cuda:4'), in_proj_covar=tensor([0.0652, 0.0799, 0.0646, 0.0592, 0.0503, 0.0506, 0.0663, 0.0615], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:46:09,860 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8509, 1.9361, 2.4088, 2.6759, 2.7754, 2.7983, 2.0724, 2.9367], device='cuda:4'), covar=tensor([0.0177, 0.0439, 0.0287, 0.0284, 0.0244, 0.0230, 0.0431, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0187, 0.0173, 0.0176, 0.0185, 0.0144, 0.0190, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:46:26,059 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.594e+02 2.297e+02 2.589e+02 3.111e+02 6.149e+02, threshold=5.178e+02, percent-clipped=2.0 2023-04-30 15:46:31,947 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=173549.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:46:35,237 INFO [train.py:904] (4/8) Epoch 18, batch 1000, loss[loss=0.1628, simple_loss=0.2417, pruned_loss=0.04195, over 16877.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2534, pruned_loss=0.04272, over 3300836.58 frames. ], batch size: 90, lr: 3.81e-03, grad_scale: 4.0 2023-04-30 15:46:35,670 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=173552.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 15:47:14,874 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0707, 2.0467, 2.6191, 3.0288, 2.8152, 3.4851, 2.3020, 3.3811], device='cuda:4'), covar=tensor([0.0237, 0.0485, 0.0313, 0.0275, 0.0307, 0.0169, 0.0491, 0.0179], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0187, 0.0173, 0.0176, 0.0184, 0.0144, 0.0190, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:47:44,517 INFO [train.py:904] (4/8) Epoch 18, batch 1050, loss[loss=0.1616, simple_loss=0.2386, pruned_loss=0.04226, over 15597.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2526, pruned_loss=0.04254, over 3302003.93 frames. ], batch size: 191, lr: 3.81e-03, grad_scale: 4.0 2023-04-30 15:48:02,482 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=173614.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:48:46,277 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.302e+02 2.654e+02 3.076e+02 5.985e+02, threshold=5.308e+02, percent-clipped=1.0 2023-04-30 15:48:54,961 INFO [train.py:904] (4/8) Epoch 18, batch 1100, loss[loss=0.1525, simple_loss=0.2316, pruned_loss=0.0367, over 16739.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2512, pruned_loss=0.04192, over 3303069.34 frames. ], batch size: 83, lr: 3.81e-03, grad_scale: 4.0 2023-04-30 15:49:12,458 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=173665.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:49:26,690 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=173675.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 15:50:01,540 INFO [train.py:904] (4/8) Epoch 18, batch 1150, loss[loss=0.1537, simple_loss=0.2395, pruned_loss=0.03398, over 16796.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2519, pruned_loss=0.04159, over 3306813.14 frames. ], batch size: 39, lr: 3.81e-03, grad_scale: 4.0 2023-04-30 15:51:02,212 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.425e+02 2.041e+02 2.391e+02 2.834e+02 4.820e+02, threshold=4.781e+02, percent-clipped=0.0 2023-04-30 15:51:11,442 INFO [train.py:904] (4/8) Epoch 18, batch 1200, loss[loss=0.1791, simple_loss=0.2484, pruned_loss=0.05488, over 16513.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2505, pruned_loss=0.04117, over 3306715.04 frames. ], batch size: 146, lr: 3.81e-03, grad_scale: 8.0 2023-04-30 15:51:33,425 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-30 15:52:19,913 INFO [train.py:904] (4/8) Epoch 18, batch 1250, loss[loss=0.161, simple_loss=0.2421, pruned_loss=0.03993, over 16816.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2506, pruned_loss=0.04183, over 3299032.82 frames. ], batch size: 39, lr: 3.81e-03, grad_scale: 8.0 2023-04-30 15:52:24,681 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=173806.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:53:21,075 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.313e+02 2.602e+02 3.076e+02 5.284e+02, threshold=5.204e+02, percent-clipped=3.0 2023-04-30 15:53:22,575 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=173847.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 15:53:28,943 INFO [train.py:904] (4/8) Epoch 18, batch 1300, loss[loss=0.1508, simple_loss=0.2409, pruned_loss=0.03034, over 17213.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2505, pruned_loss=0.04191, over 3301429.47 frames. ], batch size: 44, lr: 3.81e-03, grad_scale: 8.0 2023-04-30 15:53:32,737 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=173854.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:53:42,985 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1326, 4.1105, 4.4757, 4.4702, 4.5046, 4.2067, 4.2257, 4.1236], device='cuda:4'), covar=tensor([0.0407, 0.0809, 0.0489, 0.0457, 0.0490, 0.0447, 0.0851, 0.0671], device='cuda:4'), in_proj_covar=tensor([0.0396, 0.0430, 0.0419, 0.0394, 0.0464, 0.0441, 0.0535, 0.0349], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 15:54:09,918 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6149, 4.9309, 4.7437, 4.7557, 4.4790, 4.4228, 4.4579, 5.0088], device='cuda:4'), covar=tensor([0.1148, 0.0898, 0.1011, 0.0873, 0.0798, 0.1267, 0.1089, 0.0911], device='cuda:4'), in_proj_covar=tensor([0.0651, 0.0799, 0.0644, 0.0593, 0.0503, 0.0505, 0.0663, 0.0617], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:54:37,707 INFO [train.py:904] (4/8) Epoch 18, batch 1350, loss[loss=0.1869, simple_loss=0.2618, pruned_loss=0.05596, over 16479.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2509, pruned_loss=0.04237, over 3304001.04 frames. ], batch size: 146, lr: 3.81e-03, grad_scale: 8.0 2023-04-30 15:55:06,029 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3756, 3.6483, 4.1012, 2.2315, 3.3128, 2.4634, 3.8601, 3.9150], device='cuda:4'), covar=tensor([0.0304, 0.0843, 0.0465, 0.1945, 0.0728, 0.0953, 0.0610, 0.0985], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0158, 0.0164, 0.0151, 0.0142, 0.0127, 0.0142, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 15:55:37,173 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.363e+02 2.229e+02 2.739e+02 3.130e+02 5.090e+02, threshold=5.477e+02, percent-clipped=0.0 2023-04-30 15:55:44,245 INFO [train.py:904] (4/8) Epoch 18, batch 1400, loss[loss=0.1848, simple_loss=0.2793, pruned_loss=0.04512, over 17034.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2507, pruned_loss=0.04172, over 3313882.06 frames. ], batch size: 55, lr: 3.81e-03, grad_scale: 8.0 2023-04-30 15:56:04,277 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=173965.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:56:09,764 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=173970.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 15:56:26,741 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-30 15:56:56,559 INFO [train.py:904] (4/8) Epoch 18, batch 1450, loss[loss=0.1716, simple_loss=0.2489, pruned_loss=0.04709, over 12141.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2508, pruned_loss=0.04209, over 3310394.98 frames. ], batch size: 246, lr: 3.81e-03, grad_scale: 8.0 2023-04-30 15:57:11,462 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=174013.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:57:14,621 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=174015.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:57:16,896 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=174017.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:57:41,579 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1896, 5.6418, 5.0979, 5.6022, 5.1284, 4.9979, 5.1648, 5.7024], device='cuda:4'), covar=tensor([0.2306, 0.1591, 0.2721, 0.1315, 0.1776, 0.1287, 0.2129, 0.2028], device='cuda:4'), in_proj_covar=tensor([0.0653, 0.0799, 0.0647, 0.0595, 0.0505, 0.0507, 0.0667, 0.0620], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 15:57:54,752 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.387e+02 2.047e+02 2.425e+02 3.116e+02 5.904e+02, threshold=4.850e+02, percent-clipped=1.0 2023-04-30 15:58:03,183 INFO [train.py:904] (4/8) Epoch 18, batch 1500, loss[loss=0.1391, simple_loss=0.231, pruned_loss=0.02355, over 16857.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2511, pruned_loss=0.04192, over 3315960.47 frames. ], batch size: 42, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 15:58:36,090 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=174076.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 15:58:38,866 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=174078.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 15:59:10,941 INFO [train.py:904] (4/8) Epoch 18, batch 1550, loss[loss=0.1979, simple_loss=0.2744, pruned_loss=0.06073, over 16196.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2527, pruned_loss=0.04216, over 3321404.88 frames. ], batch size: 165, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:00:00,536 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=174139.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:00:09,836 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.550e+02 2.219e+02 2.746e+02 3.375e+02 7.585e+02, threshold=5.492e+02, percent-clipped=5.0 2023-04-30 16:00:11,874 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=174147.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 16:00:17,419 INFO [train.py:904] (4/8) Epoch 18, batch 1600, loss[loss=0.1392, simple_loss=0.227, pruned_loss=0.02573, over 16757.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2545, pruned_loss=0.04305, over 3324293.67 frames. ], batch size: 39, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:01:00,645 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-30 16:01:16,803 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=174195.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 16:01:23,702 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=174200.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:01:25,621 INFO [train.py:904] (4/8) Epoch 18, batch 1650, loss[loss=0.168, simple_loss=0.2444, pruned_loss=0.04576, over 16861.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2559, pruned_loss=0.04382, over 3329276.99 frames. ], batch size: 96, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:02:23,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.798e+02 2.550e+02 2.934e+02 3.695e+02 6.371e+02, threshold=5.868e+02, percent-clipped=1.0 2023-04-30 16:02:32,758 INFO [train.py:904] (4/8) Epoch 18, batch 1700, loss[loss=0.181, simple_loss=0.2683, pruned_loss=0.04682, over 16808.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2582, pruned_loss=0.04478, over 3336103.19 frames. ], batch size: 83, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:02:57,241 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=174270.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 16:03:15,588 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=174283.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:03:25,131 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=174291.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:03:40,505 INFO [train.py:904] (4/8) Epoch 18, batch 1750, loss[loss=0.1664, simple_loss=0.266, pruned_loss=0.03339, over 17041.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2593, pruned_loss=0.04507, over 3327666.54 frames. ], batch size: 55, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:04:01,306 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=174318.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:04:36,729 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=174344.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:04:38,536 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.659e+02 2.402e+02 2.710e+02 3.242e+02 7.254e+02, threshold=5.419e+02, percent-clipped=1.0 2023-04-30 16:04:45,342 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0735, 4.1952, 4.5047, 4.4990, 4.5139, 4.2258, 4.2674, 4.1329], device='cuda:4'), covar=tensor([0.0379, 0.0599, 0.0382, 0.0372, 0.0448, 0.0423, 0.0760, 0.0622], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0438, 0.0425, 0.0398, 0.0473, 0.0449, 0.0543, 0.0355], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 16:04:46,157 INFO [train.py:904] (4/8) Epoch 18, batch 1800, loss[loss=0.1763, simple_loss=0.261, pruned_loss=0.04579, over 16664.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.26, pruned_loss=0.04443, over 3329761.35 frames. ], batch size: 134, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:04:46,571 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=174352.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 16:05:10,902 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=174371.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 16:05:13,946 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=174373.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:05:33,449 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=174388.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:05:51,520 INFO [train.py:904] (4/8) Epoch 18, batch 1850, loss[loss=0.1762, simple_loss=0.256, pruned_loss=0.04818, over 16242.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2595, pruned_loss=0.04394, over 3336940.64 frames. ], batch size: 165, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:06:28,714 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9439, 4.4952, 3.3152, 2.3520, 2.8840, 2.6783, 4.7601, 3.7700], device='cuda:4'), covar=tensor([0.2721, 0.0566, 0.1688, 0.3080, 0.2984, 0.1985, 0.0388, 0.1462], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0264, 0.0301, 0.0303, 0.0290, 0.0247, 0.0286, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 16:06:34,328 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-30 16:06:51,028 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.281e+02 2.184e+02 2.475e+02 2.961e+02 4.707e+02, threshold=4.951e+02, percent-clipped=0.0 2023-04-30 16:06:55,531 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=174449.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:06:58,376 INFO [train.py:904] (4/8) Epoch 18, batch 1900, loss[loss=0.1449, simple_loss=0.2358, pruned_loss=0.02702, over 17208.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2595, pruned_loss=0.04343, over 3323149.00 frames. ], batch size: 45, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:07:06,784 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8842, 5.2793, 5.4264, 5.1889, 5.1565, 5.7962, 5.3113, 4.9771], device='cuda:4'), covar=tensor([0.1253, 0.1837, 0.1954, 0.1875, 0.2636, 0.1045, 0.1510, 0.2600], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0581, 0.0638, 0.0481, 0.0654, 0.0666, 0.0502, 0.0652], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 16:07:15,449 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1621, 3.2522, 3.3082, 2.2574, 3.1803, 3.4341, 3.1992, 1.7922], device='cuda:4'), covar=tensor([0.0530, 0.0125, 0.0087, 0.0433, 0.0136, 0.0131, 0.0123, 0.0575], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0080, 0.0080, 0.0134, 0.0093, 0.0104, 0.0091, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 16:07:39,392 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-30 16:07:57,453 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=174495.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:08:06,701 INFO [train.py:904] (4/8) Epoch 18, batch 1950, loss[loss=0.1536, simple_loss=0.2351, pruned_loss=0.03606, over 16987.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.259, pruned_loss=0.04309, over 3326294.28 frames. ], batch size: 41, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:08:47,444 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.56 vs. limit=5.0 2023-04-30 16:09:05,465 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.435e+02 2.251e+02 2.513e+02 3.165e+02 5.916e+02, threshold=5.027e+02, percent-clipped=3.0 2023-04-30 16:09:14,567 INFO [train.py:904] (4/8) Epoch 18, batch 2000, loss[loss=0.1902, simple_loss=0.2875, pruned_loss=0.04644, over 17071.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2587, pruned_loss=0.04319, over 3322996.91 frames. ], batch size: 55, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:09:49,193 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-30 16:10:23,407 INFO [train.py:904] (4/8) Epoch 18, batch 2050, loss[loss=0.1852, simple_loss=0.2718, pruned_loss=0.04929, over 16658.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2591, pruned_loss=0.04399, over 3324935.50 frames. ], batch size: 76, lr: 3.80e-03, grad_scale: 16.0 2023-04-30 16:11:02,127 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2273, 1.5029, 1.9619, 2.1052, 2.2107, 2.2789, 1.7047, 2.3209], device='cuda:4'), covar=tensor([0.0204, 0.0482, 0.0269, 0.0299, 0.0288, 0.0258, 0.0480, 0.0168], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0190, 0.0176, 0.0181, 0.0188, 0.0148, 0.0193, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:11:17,516 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=174639.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:11:26,649 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.708e+02 2.224e+02 2.527e+02 3.035e+02 6.825e+02, threshold=5.053e+02, percent-clipped=2.0 2023-04-30 16:11:28,734 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=174647.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 16:11:35,057 INFO [train.py:904] (4/8) Epoch 18, batch 2100, loss[loss=0.1819, simple_loss=0.2622, pruned_loss=0.05081, over 16860.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2603, pruned_loss=0.04432, over 3326712.89 frames. ], batch size: 102, lr: 3.80e-03, grad_scale: 16.0 2023-04-30 16:12:02,033 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=174671.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 16:12:06,169 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=174673.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:12:45,803 INFO [train.py:904] (4/8) Epoch 18, batch 2150, loss[loss=0.1747, simple_loss=0.2548, pruned_loss=0.04732, over 16716.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2612, pruned_loss=0.04454, over 3323365.00 frames. ], batch size: 134, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:12:50,313 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=174705.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:13:10,638 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=174719.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:13:13,522 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=174721.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:13:34,677 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8735, 5.0336, 5.4496, 5.3993, 5.4123, 5.0523, 4.9801, 4.7622], device='cuda:4'), covar=tensor([0.0344, 0.0527, 0.0365, 0.0403, 0.0448, 0.0380, 0.0962, 0.0479], device='cuda:4'), in_proj_covar=tensor([0.0400, 0.0434, 0.0422, 0.0396, 0.0470, 0.0445, 0.0539, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 16:13:45,381 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=174744.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:13:49,507 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 2.235e+02 2.673e+02 3.216e+02 6.695e+02, threshold=5.347e+02, percent-clipped=2.0 2023-04-30 16:13:56,940 INFO [train.py:904] (4/8) Epoch 18, batch 2200, loss[loss=0.1684, simple_loss=0.2459, pruned_loss=0.04549, over 16301.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2608, pruned_loss=0.04439, over 3327876.52 frames. ], batch size: 36, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:14:16,333 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=174766.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:14:51,039 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1048, 1.9025, 2.5286, 3.0736, 2.8294, 3.4976, 2.0516, 3.5335], device='cuda:4'), covar=tensor([0.0173, 0.0493, 0.0296, 0.0220, 0.0255, 0.0144, 0.0567, 0.0115], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0189, 0.0175, 0.0180, 0.0188, 0.0148, 0.0192, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:14:56,015 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=174795.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:15:06,365 INFO [train.py:904] (4/8) Epoch 18, batch 2250, loss[loss=0.1866, simple_loss=0.2624, pruned_loss=0.05541, over 16722.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2622, pruned_loss=0.04489, over 3325982.66 frames. ], batch size: 124, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:16:02,772 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=174843.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:16:07,686 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.616e+02 2.164e+02 2.481e+02 3.118e+02 6.907e+02, threshold=4.963e+02, percent-clipped=2.0 2023-04-30 16:16:14,117 INFO [train.py:904] (4/8) Epoch 18, batch 2300, loss[loss=0.1541, simple_loss=0.2365, pruned_loss=0.03584, over 16954.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2617, pruned_loss=0.04466, over 3317117.25 frames. ], batch size: 41, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:17:22,134 INFO [train.py:904] (4/8) Epoch 18, batch 2350, loss[loss=0.1572, simple_loss=0.243, pruned_loss=0.03568, over 16840.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.262, pruned_loss=0.04559, over 3317879.11 frames. ], batch size: 42, lr: 3.80e-03, grad_scale: 8.0 2023-04-30 16:18:09,353 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6894, 5.0856, 4.5221, 4.9844, 4.6767, 4.5025, 4.6448, 5.1479], device='cuda:4'), covar=tensor([0.2454, 0.1775, 0.3032, 0.1466, 0.1748, 0.2179, 0.2441, 0.2154], device='cuda:4'), in_proj_covar=tensor([0.0659, 0.0815, 0.0654, 0.0602, 0.0511, 0.0515, 0.0674, 0.0626], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:18:14,701 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=174939.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:18:25,283 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.403e+02 2.298e+02 2.659e+02 3.331e+02 6.396e+02, threshold=5.319e+02, percent-clipped=2.0 2023-04-30 16:18:25,678 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=174947.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 16:18:32,764 INFO [train.py:904] (4/8) Epoch 18, batch 2400, loss[loss=0.163, simple_loss=0.2493, pruned_loss=0.03832, over 16798.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.263, pruned_loss=0.04563, over 3316330.08 frames. ], batch size: 39, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:19:21,465 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=174987.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:19:33,333 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=174995.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:19:42,204 INFO [train.py:904] (4/8) Epoch 18, batch 2450, loss[loss=0.1638, simple_loss=0.2645, pruned_loss=0.03155, over 17074.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.263, pruned_loss=0.04516, over 3312444.35 frames. ], batch size: 50, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:20:37,052 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0511, 5.5666, 5.7522, 5.4091, 5.4811, 6.1159, 5.6028, 5.2928], device='cuda:4'), covar=tensor([0.0895, 0.1972, 0.2049, 0.1960, 0.2621, 0.0863, 0.1416, 0.2344], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0582, 0.0639, 0.0481, 0.0651, 0.0666, 0.0502, 0.0652], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 16:20:40,478 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=175044.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:20:43,743 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.516e+02 2.297e+02 2.808e+02 3.204e+02 6.791e+02, threshold=5.617e+02, percent-clipped=2.0 2023-04-30 16:20:51,502 INFO [train.py:904] (4/8) Epoch 18, batch 2500, loss[loss=0.153, simple_loss=0.2542, pruned_loss=0.02592, over 17139.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.263, pruned_loss=0.04495, over 3311999.83 frames. ], batch size: 47, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:21:04,688 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=175061.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:21:46,586 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=175092.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:22:00,356 INFO [train.py:904] (4/8) Epoch 18, batch 2550, loss[loss=0.163, simple_loss=0.2561, pruned_loss=0.03495, over 17202.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2621, pruned_loss=0.04448, over 3321512.13 frames. ], batch size: 44, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:22:30,167 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2845, 2.1459, 2.3644, 4.0651, 2.1869, 2.5680, 2.2414, 2.3840], device='cuda:4'), covar=tensor([0.1351, 0.3749, 0.2749, 0.0564, 0.3746, 0.2481, 0.3623, 0.3084], device='cuda:4'), in_proj_covar=tensor([0.0393, 0.0432, 0.0361, 0.0327, 0.0432, 0.0500, 0.0401, 0.0505], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:22:34,697 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-30 16:22:51,973 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=175139.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:23:03,049 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.483e+02 2.226e+02 2.530e+02 3.220e+02 6.959e+02, threshold=5.059e+02, percent-clipped=1.0 2023-04-30 16:23:09,303 INFO [train.py:904] (4/8) Epoch 18, batch 2600, loss[loss=0.1721, simple_loss=0.2699, pruned_loss=0.03719, over 17060.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.262, pruned_loss=0.0441, over 3323388.50 frames. ], batch size: 50, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:24:17,737 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=175200.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:24:19,525 INFO [train.py:904] (4/8) Epoch 18, batch 2650, loss[loss=0.1837, simple_loss=0.2649, pruned_loss=0.05132, over 12442.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.262, pruned_loss=0.04386, over 3316427.67 frames. ], batch size: 246, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:24:27,569 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=175207.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:25:22,161 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.596e+02 2.085e+02 2.405e+02 2.864e+02 4.837e+02, threshold=4.811e+02, percent-clipped=0.0 2023-04-30 16:25:27,538 INFO [train.py:904] (4/8) Epoch 18, batch 2700, loss[loss=0.1731, simple_loss=0.26, pruned_loss=0.04311, over 16759.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2625, pruned_loss=0.04354, over 3320987.71 frames. ], batch size: 89, lr: 3.79e-03, grad_scale: 4.0 2023-04-30 16:25:50,112 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=175268.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:26:36,861 INFO [train.py:904] (4/8) Epoch 18, batch 2750, loss[loss=0.1718, simple_loss=0.2578, pruned_loss=0.04289, over 16699.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2622, pruned_loss=0.04304, over 3323750.67 frames. ], batch size: 124, lr: 3.79e-03, grad_scale: 4.0 2023-04-30 16:27:40,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.507e+02 2.063e+02 2.386e+02 2.912e+02 4.096e+02, threshold=4.773e+02, percent-clipped=0.0 2023-04-30 16:27:45,205 INFO [train.py:904] (4/8) Epoch 18, batch 2800, loss[loss=0.1811, simple_loss=0.2628, pruned_loss=0.04968, over 16672.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.262, pruned_loss=0.04293, over 3329526.63 frames. ], batch size: 134, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:27:58,759 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=175361.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:28:11,815 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=175370.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:28:44,369 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9298, 4.8301, 4.7756, 4.4171, 4.4314, 4.8371, 4.7011, 4.5617], device='cuda:4'), covar=tensor([0.0603, 0.0787, 0.0298, 0.0309, 0.0950, 0.0515, 0.0406, 0.0635], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0416, 0.0344, 0.0336, 0.0359, 0.0389, 0.0238, 0.0415], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 16:28:45,654 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=175394.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:28:56,406 INFO [train.py:904] (4/8) Epoch 18, batch 2850, loss[loss=0.1592, simple_loss=0.2635, pruned_loss=0.02747, over 17092.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2615, pruned_loss=0.04318, over 3338122.79 frames. ], batch size: 47, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:29:05,423 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=175409.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:29:05,651 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9513, 2.0406, 2.5189, 2.8766, 2.6721, 3.4028, 2.2813, 3.3558], device='cuda:4'), covar=tensor([0.0242, 0.0480, 0.0320, 0.0314, 0.0323, 0.0178, 0.0443, 0.0147], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0191, 0.0177, 0.0182, 0.0190, 0.0149, 0.0194, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:29:35,300 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=175431.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:29:58,776 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.560e+02 2.293e+02 2.737e+02 3.416e+02 1.541e+03, threshold=5.475e+02, percent-clipped=7.0 2023-04-30 16:30:04,917 INFO [train.py:904] (4/8) Epoch 18, batch 2900, loss[loss=0.16, simple_loss=0.2445, pruned_loss=0.03773, over 16841.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2603, pruned_loss=0.04332, over 3330514.71 frames. ], batch size: 96, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:30:08,973 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=175455.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:30:45,717 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=175481.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:30:50,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9271, 2.9681, 2.5608, 2.7558, 3.1957, 2.9723, 3.6037, 3.4292], device='cuda:4'), covar=tensor([0.0114, 0.0339, 0.0452, 0.0402, 0.0269, 0.0361, 0.0235, 0.0255], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0235, 0.0225, 0.0225, 0.0236, 0.0234, 0.0240, 0.0228], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:31:04,477 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=175495.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:31:13,720 INFO [train.py:904] (4/8) Epoch 18, batch 2950, loss[loss=0.1679, simple_loss=0.2458, pruned_loss=0.04501, over 16790.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2599, pruned_loss=0.04383, over 3328788.65 frames. ], batch size: 102, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:31:50,269 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9565, 2.0933, 2.5614, 2.9476, 2.8145, 3.3957, 2.2406, 3.3786], device='cuda:4'), covar=tensor([0.0233, 0.0462, 0.0318, 0.0296, 0.0302, 0.0181, 0.0474, 0.0176], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0192, 0.0179, 0.0185, 0.0192, 0.0150, 0.0195, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:32:09,640 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=175542.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:32:17,355 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.593e+02 2.280e+02 2.587e+02 3.034e+02 5.964e+02, threshold=5.174e+02, percent-clipped=1.0 2023-04-30 16:32:23,314 INFO [train.py:904] (4/8) Epoch 18, batch 3000, loss[loss=0.1684, simple_loss=0.2525, pruned_loss=0.04211, over 16493.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2605, pruned_loss=0.04406, over 3328048.70 frames. ], batch size: 75, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:32:23,314 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 16:32:32,126 INFO [train.py:938] (4/8) Epoch 18, validation: loss=0.1359, simple_loss=0.2415, pruned_loss=0.01516, over 944034.00 frames. 2023-04-30 16:32:32,126 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 16:32:48,740 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=175563.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:33:12,064 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-30 16:33:42,215 INFO [train.py:904] (4/8) Epoch 18, batch 3050, loss[loss=0.1783, simple_loss=0.2615, pruned_loss=0.04757, over 15517.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2601, pruned_loss=0.04378, over 3330810.06 frames. ], batch size: 190, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:34:38,016 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1777, 2.5677, 1.9725, 2.3369, 2.9130, 2.6222, 3.0701, 3.0532], device='cuda:4'), covar=tensor([0.0182, 0.0388, 0.0557, 0.0450, 0.0247, 0.0365, 0.0226, 0.0263], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0234, 0.0224, 0.0223, 0.0235, 0.0233, 0.0238, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:34:46,185 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.517e+02 2.265e+02 2.697e+02 3.394e+02 4.892e+02, threshold=5.395e+02, percent-clipped=0.0 2023-04-30 16:34:51,893 INFO [train.py:904] (4/8) Epoch 18, batch 3100, loss[loss=0.1693, simple_loss=0.2505, pruned_loss=0.04411, over 16503.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2588, pruned_loss=0.04328, over 3331990.69 frames. ], batch size: 75, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:36:01,368 INFO [train.py:904] (4/8) Epoch 18, batch 3150, loss[loss=0.1996, simple_loss=0.2682, pruned_loss=0.06547, over 16850.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2577, pruned_loss=0.04307, over 3339911.95 frames. ], batch size: 116, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:36:13,026 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=175710.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:36:36,342 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=175726.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:37:03,627 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8394, 2.9555, 2.6912, 5.0556, 4.0968, 4.5430, 1.7831, 3.2466], device='cuda:4'), covar=tensor([0.1409, 0.0760, 0.1209, 0.0212, 0.0234, 0.0367, 0.1560, 0.0780], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0169, 0.0189, 0.0181, 0.0203, 0.0216, 0.0194, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 16:37:05,952 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.459e+02 2.210e+02 2.531e+02 3.035e+02 7.463e+02, threshold=5.061e+02, percent-clipped=4.0 2023-04-30 16:37:09,223 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=175750.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:37:11,226 INFO [train.py:904] (4/8) Epoch 18, batch 3200, loss[loss=0.1467, simple_loss=0.2389, pruned_loss=0.02722, over 16754.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2572, pruned_loss=0.04271, over 3336728.02 frames. ], batch size: 39, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:37:37,896 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=175771.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:37:43,631 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7161, 4.1662, 4.2210, 3.0285, 3.6596, 4.2204, 3.8112, 2.2794], device='cuda:4'), covar=tensor([0.0484, 0.0084, 0.0053, 0.0344, 0.0118, 0.0091, 0.0093, 0.0470], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0080, 0.0080, 0.0132, 0.0094, 0.0104, 0.0091, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 16:38:08,889 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8317, 1.9947, 2.3860, 2.7324, 2.7617, 2.9464, 2.0463, 3.0772], device='cuda:4'), covar=tensor([0.0187, 0.0445, 0.0304, 0.0284, 0.0262, 0.0197, 0.0464, 0.0136], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0191, 0.0178, 0.0183, 0.0190, 0.0149, 0.0194, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:38:11,015 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=175795.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:38:20,671 INFO [train.py:904] (4/8) Epoch 18, batch 3250, loss[loss=0.1576, simple_loss=0.2331, pruned_loss=0.04103, over 16949.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2569, pruned_loss=0.04316, over 3329081.37 frames. ], batch size: 41, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:39:10,170 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=175837.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:39:17,730 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=175843.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:39:23,828 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.575e+02 2.233e+02 2.626e+02 2.962e+02 5.772e+02, threshold=5.253e+02, percent-clipped=1.0 2023-04-30 16:39:29,873 INFO [train.py:904] (4/8) Epoch 18, batch 3300, loss[loss=0.1939, simple_loss=0.2905, pruned_loss=0.04868, over 17073.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2582, pruned_loss=0.04359, over 3319785.13 frames. ], batch size: 53, lr: 3.79e-03, grad_scale: 8.0 2023-04-30 16:39:45,089 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=175863.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:40:10,990 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=175881.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:40:38,820 INFO [train.py:904] (4/8) Epoch 18, batch 3350, loss[loss=0.1585, simple_loss=0.2546, pruned_loss=0.0312, over 17169.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2595, pruned_loss=0.04424, over 3311892.73 frames. ], batch size: 46, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:40:50,862 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=175911.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:41:34,493 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=175942.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:41:38,071 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=175945.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:41:41,869 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.365e+02 2.233e+02 2.642e+02 3.041e+02 9.151e+02, threshold=5.284e+02, percent-clipped=6.0 2023-04-30 16:41:47,202 INFO [train.py:904] (4/8) Epoch 18, batch 3400, loss[loss=0.1923, simple_loss=0.279, pruned_loss=0.05284, over 12180.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2596, pruned_loss=0.0445, over 3308666.73 frames. ], batch size: 246, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:42:51,934 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-30 16:43:01,858 INFO [train.py:904] (4/8) Epoch 18, batch 3450, loss[loss=0.1646, simple_loss=0.2402, pruned_loss=0.04452, over 16873.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2574, pruned_loss=0.04368, over 3307127.55 frames. ], batch size: 96, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:43:08,038 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=176006.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:43:12,504 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-30 16:43:35,396 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=176026.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:44:06,111 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 2.158e+02 2.512e+02 2.983e+02 6.813e+02, threshold=5.023e+02, percent-clipped=2.0 2023-04-30 16:44:09,501 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=176050.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:44:11,307 INFO [train.py:904] (4/8) Epoch 18, batch 3500, loss[loss=0.1723, simple_loss=0.2511, pruned_loss=0.04675, over 16832.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2563, pruned_loss=0.04298, over 3313528.95 frames. ], batch size: 42, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:44:31,791 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=176066.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:44:41,808 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=176074.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:45:14,938 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=176098.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:45:19,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6049, 3.6617, 3.9569, 2.8341, 3.5819, 3.9845, 3.6585, 2.3538], device='cuda:4'), covar=tensor([0.0430, 0.0206, 0.0049, 0.0331, 0.0091, 0.0091, 0.0095, 0.0420], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0080, 0.0080, 0.0133, 0.0094, 0.0104, 0.0091, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 16:45:20,530 INFO [train.py:904] (4/8) Epoch 18, batch 3550, loss[loss=0.1519, simple_loss=0.2335, pruned_loss=0.03517, over 16750.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2552, pruned_loss=0.04225, over 3307730.81 frames. ], batch size: 89, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:46:10,539 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=176137.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:46:26,481 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.470e+02 2.026e+02 2.447e+02 2.938e+02 5.668e+02, threshold=4.895e+02, percent-clipped=1.0 2023-04-30 16:46:32,322 INFO [train.py:904] (4/8) Epoch 18, batch 3600, loss[loss=0.1786, simple_loss=0.254, pruned_loss=0.05157, over 16451.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2548, pruned_loss=0.04213, over 3300412.27 frames. ], batch size: 146, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:47:18,726 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=176185.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:47:43,325 INFO [train.py:904] (4/8) Epoch 18, batch 3650, loss[loss=0.1748, simple_loss=0.2483, pruned_loss=0.05062, over 16535.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2539, pruned_loss=0.04263, over 3296777.47 frames. ], batch size: 146, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:48:30,686 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7179, 2.6112, 2.5544, 4.0203, 3.3599, 4.0510, 1.5110, 2.8953], device='cuda:4'), covar=tensor([0.1369, 0.0729, 0.1138, 0.0190, 0.0171, 0.0380, 0.1617, 0.0800], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0169, 0.0189, 0.0182, 0.0203, 0.0215, 0.0194, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 16:48:35,860 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=176237.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:48:43,994 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=176242.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:48:48,001 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8889, 3.0109, 3.1681, 2.1209, 2.7489, 2.1822, 3.4850, 3.3847], device='cuda:4'), covar=tensor([0.0249, 0.0925, 0.0616, 0.1856, 0.0854, 0.1006, 0.0509, 0.0857], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0162, 0.0166, 0.0152, 0.0143, 0.0128, 0.0144, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 16:48:53,036 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.679e+02 2.298e+02 2.711e+02 3.292e+02 6.434e+02, threshold=5.422e+02, percent-clipped=3.0 2023-04-30 16:48:58,528 INFO [train.py:904] (4/8) Epoch 18, batch 3700, loss[loss=0.1747, simple_loss=0.2467, pruned_loss=0.05136, over 16865.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2532, pruned_loss=0.04421, over 3280742.31 frames. ], batch size: 109, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:49:07,010 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=176257.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:50:10,821 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=176301.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:50:11,691 INFO [train.py:904] (4/8) Epoch 18, batch 3750, loss[loss=0.1617, simple_loss=0.237, pruned_loss=0.04317, over 16455.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2548, pruned_loss=0.04593, over 3264729.51 frames. ], batch size: 146, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:50:13,740 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=176303.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:50:35,167 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=176318.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:51:17,334 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.625e+02 2.173e+02 2.603e+02 3.284e+02 5.784e+02, threshold=5.206e+02, percent-clipped=1.0 2023-04-30 16:51:23,241 INFO [train.py:904] (4/8) Epoch 18, batch 3800, loss[loss=0.1562, simple_loss=0.2365, pruned_loss=0.03796, over 16834.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2564, pruned_loss=0.04765, over 3263734.85 frames. ], batch size: 102, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:51:44,234 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=176366.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:52:31,405 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7767, 3.7748, 2.3283, 4.1170, 3.0525, 4.0896, 2.5776, 3.0574], device='cuda:4'), covar=tensor([0.0216, 0.0346, 0.1379, 0.0248, 0.0625, 0.0652, 0.1179, 0.0620], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0176, 0.0193, 0.0160, 0.0174, 0.0219, 0.0202, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 16:52:35,499 INFO [train.py:904] (4/8) Epoch 18, batch 3850, loss[loss=0.1411, simple_loss=0.2206, pruned_loss=0.03079, over 16902.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2562, pruned_loss=0.04796, over 3265344.02 frames. ], batch size: 96, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:52:38,003 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.60 vs. limit=2.0 2023-04-30 16:52:53,510 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=176414.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:52:54,716 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0453, 5.5963, 5.7400, 5.4509, 5.5461, 6.1296, 5.6546, 5.3483], device='cuda:4'), covar=tensor([0.0745, 0.1579, 0.1398, 0.1675, 0.2169, 0.0784, 0.1206, 0.2073], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0582, 0.0637, 0.0490, 0.0657, 0.0666, 0.0505, 0.0655], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 16:53:33,626 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-30 16:53:42,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.371e+02 2.342e+02 2.782e+02 3.184e+02 4.336e+02, threshold=5.564e+02, percent-clipped=0.0 2023-04-30 16:53:49,440 INFO [train.py:904] (4/8) Epoch 18, batch 3900, loss[loss=0.1729, simple_loss=0.2592, pruned_loss=0.04328, over 17014.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2559, pruned_loss=0.04849, over 3272323.37 frames. ], batch size: 50, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:54:11,154 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=176466.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:54:37,509 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8416, 3.7245, 4.0340, 2.3149, 4.2635, 4.2161, 3.3627, 3.0925], device='cuda:4'), covar=tensor([0.0644, 0.0188, 0.0139, 0.0983, 0.0061, 0.0146, 0.0294, 0.0383], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0106, 0.0094, 0.0138, 0.0076, 0.0122, 0.0125, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 16:55:00,904 INFO [train.py:904] (4/8) Epoch 18, batch 3950, loss[loss=0.1864, simple_loss=0.2654, pruned_loss=0.05367, over 12595.00 frames. ], tot_loss[loss=0.1764, simple_loss=0.2549, pruned_loss=0.04893, over 3269681.56 frames. ], batch size: 246, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:55:10,786 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9263, 3.0633, 3.2595, 2.0036, 2.7662, 2.2136, 3.3654, 3.4482], device='cuda:4'), covar=tensor([0.0241, 0.0846, 0.0608, 0.1872, 0.0874, 0.0993, 0.0546, 0.0748], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0161, 0.0165, 0.0151, 0.0142, 0.0127, 0.0142, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 16:55:35,323 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=176527.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:55:50,459 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=176537.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:56:05,762 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.227e+02 2.521e+02 3.093e+02 7.043e+02, threshold=5.041e+02, percent-clipped=2.0 2023-04-30 16:56:12,320 INFO [train.py:904] (4/8) Epoch 18, batch 4000, loss[loss=0.1722, simple_loss=0.2474, pruned_loss=0.04846, over 16654.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2548, pruned_loss=0.04939, over 3271320.20 frames. ], batch size: 57, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:57:00,440 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=176585.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:57:07,647 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9827, 2.0681, 2.6131, 2.9372, 2.8779, 3.3539, 2.1560, 3.2809], device='cuda:4'), covar=tensor([0.0187, 0.0469, 0.0277, 0.0300, 0.0262, 0.0145, 0.0446, 0.0100], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0188, 0.0177, 0.0180, 0.0187, 0.0148, 0.0191, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:57:18,434 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=176598.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:57:23,563 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=176601.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:57:24,371 INFO [train.py:904] (4/8) Epoch 18, batch 4050, loss[loss=0.1577, simple_loss=0.2484, pruned_loss=0.03346, over 16437.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2552, pruned_loss=0.04846, over 3262013.14 frames. ], batch size: 146, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:57:39,762 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=176613.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:57:58,531 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=176625.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:58:13,167 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0513, 3.1801, 3.5628, 2.0271, 2.9380, 2.2168, 3.4796, 3.5288], device='cuda:4'), covar=tensor([0.0224, 0.0800, 0.0521, 0.2035, 0.0850, 0.1004, 0.0572, 0.0836], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0161, 0.0165, 0.0151, 0.0142, 0.0127, 0.0142, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 16:58:30,484 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.366e+02 1.835e+02 2.096e+02 2.511e+02 3.496e+02, threshold=4.192e+02, percent-clipped=0.0 2023-04-30 16:58:33,596 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=176649.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:58:36,778 INFO [train.py:904] (4/8) Epoch 18, batch 4100, loss[loss=0.1838, simple_loss=0.2611, pruned_loss=0.05324, over 16829.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2573, pruned_loss=0.04822, over 3258557.64 frames. ], batch size: 42, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 16:58:42,927 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6423, 2.1058, 1.5988, 1.8172, 2.4298, 2.1193, 2.4430, 2.6894], device='cuda:4'), covar=tensor([0.0153, 0.0413, 0.0620, 0.0492, 0.0265, 0.0390, 0.0189, 0.0235], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0234, 0.0224, 0.0223, 0.0234, 0.0231, 0.0238, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 16:58:57,272 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=176665.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:59:06,986 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.42 vs. limit=5.0 2023-04-30 16:59:28,858 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=176686.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 16:59:51,167 INFO [train.py:904] (4/8) Epoch 18, batch 4150, loss[loss=0.1887, simple_loss=0.2787, pruned_loss=0.0494, over 16978.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2643, pruned_loss=0.05032, over 3223177.71 frames. ], batch size: 90, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 17:00:05,475 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5418, 4.5063, 4.3314, 3.5907, 4.4422, 1.6481, 4.1872, 3.9672], device='cuda:4'), covar=tensor([0.0081, 0.0075, 0.0170, 0.0318, 0.0076, 0.2956, 0.0116, 0.0247], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0148, 0.0195, 0.0178, 0.0169, 0.0204, 0.0186, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:00:27,841 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=176726.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:00:51,101 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6062, 4.6957, 4.8649, 4.6561, 4.6332, 5.2282, 4.7648, 4.4271], device='cuda:4'), covar=tensor([0.1036, 0.1723, 0.1585, 0.1973, 0.2533, 0.1053, 0.1432, 0.2460], device='cuda:4'), in_proj_covar=tensor([0.0396, 0.0579, 0.0631, 0.0485, 0.0649, 0.0661, 0.0502, 0.0649], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 17:01:00,609 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.553e+02 2.334e+02 2.630e+02 3.373e+02 5.770e+02, threshold=5.259e+02, percent-clipped=9.0 2023-04-30 17:01:06,244 INFO [train.py:904] (4/8) Epoch 18, batch 4200, loss[loss=0.2311, simple_loss=0.3166, pruned_loss=0.07277, over 16734.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2718, pruned_loss=0.05229, over 3187370.08 frames. ], batch size: 57, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 17:02:08,510 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.57 vs. limit=5.0 2023-04-30 17:02:21,464 INFO [train.py:904] (4/8) Epoch 18, batch 4250, loss[loss=0.1882, simple_loss=0.2774, pruned_loss=0.04944, over 17190.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2754, pruned_loss=0.05234, over 3168099.84 frames. ], batch size: 44, lr: 3.78e-03, grad_scale: 8.0 2023-04-30 17:02:51,593 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=176822.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:03:30,835 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.576e+02 2.197e+02 2.694e+02 3.150e+02 7.622e+02, threshold=5.388e+02, percent-clipped=1.0 2023-04-30 17:03:35,707 INFO [train.py:904] (4/8) Epoch 18, batch 4300, loss[loss=0.1813, simple_loss=0.2767, pruned_loss=0.043, over 16653.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2761, pruned_loss=0.05148, over 3160539.28 frames. ], batch size: 76, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:03:37,909 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=176853.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:04:44,757 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=176898.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:04:49,705 INFO [train.py:904] (4/8) Epoch 18, batch 4350, loss[loss=0.1995, simple_loss=0.2918, pruned_loss=0.05361, over 16854.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2794, pruned_loss=0.05248, over 3165263.90 frames. ], batch size: 90, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:05:06,434 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=176913.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:05:07,716 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=176914.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 17:05:21,586 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3633, 4.3333, 4.2047, 3.5788, 4.2997, 1.7457, 4.0527, 3.8539], device='cuda:4'), covar=tensor([0.0081, 0.0069, 0.0164, 0.0308, 0.0077, 0.2792, 0.0118, 0.0228], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0148, 0.0195, 0.0177, 0.0169, 0.0204, 0.0185, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:05:51,770 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-30 17:05:55,206 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=176946.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:05:57,198 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.674e+02 2.251e+02 2.629e+02 3.304e+02 5.605e+02, threshold=5.258e+02, percent-clipped=2.0 2023-04-30 17:06:03,297 INFO [train.py:904] (4/8) Epoch 18, batch 4400, loss[loss=0.2001, simple_loss=0.2892, pruned_loss=0.05544, over 16787.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2818, pruned_loss=0.05332, over 3176382.57 frames. ], batch size: 83, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:06:07,471 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0327, 4.9374, 5.0836, 5.2608, 5.4100, 4.8205, 5.4009, 5.4046], device='cuda:4'), covar=tensor([0.1633, 0.1067, 0.1384, 0.0570, 0.0432, 0.0715, 0.0453, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0607, 0.0753, 0.0883, 0.0767, 0.0565, 0.0606, 0.0618, 0.0717], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:06:17,358 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=176961.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:06:28,881 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-04-30 17:06:45,616 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=176981.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:07:01,853 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8208, 4.0913, 4.1526, 2.2860, 3.4132, 2.6961, 3.9703, 4.2543], device='cuda:4'), covar=tensor([0.0158, 0.0588, 0.0452, 0.1857, 0.0721, 0.0847, 0.0486, 0.0695], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0160, 0.0164, 0.0149, 0.0141, 0.0126, 0.0141, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 17:07:16,203 INFO [train.py:904] (4/8) Epoch 18, batch 4450, loss[loss=0.2033, simple_loss=0.2968, pruned_loss=0.05489, over 16917.00 frames. ], tot_loss[loss=0.1963, simple_loss=0.2845, pruned_loss=0.05403, over 3182165.22 frames. ], batch size: 109, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:07:29,065 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0347, 2.1411, 2.1304, 3.6429, 2.0088, 2.5019, 2.2305, 2.3145], device='cuda:4'), covar=tensor([0.1295, 0.3307, 0.2755, 0.0540, 0.4073, 0.2221, 0.3167, 0.3228], device='cuda:4'), in_proj_covar=tensor([0.0389, 0.0432, 0.0355, 0.0324, 0.0427, 0.0500, 0.0401, 0.0504], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:07:43,886 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=177021.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:08:24,607 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.309e+02 1.969e+02 2.287e+02 2.678e+02 4.361e+02, threshold=4.575e+02, percent-clipped=0.0 2023-04-30 17:08:30,152 INFO [train.py:904] (4/8) Epoch 18, batch 4500, loss[loss=0.196, simple_loss=0.2838, pruned_loss=0.0541, over 16407.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.285, pruned_loss=0.05437, over 3201574.24 frames. ], batch size: 146, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:09:43,614 INFO [train.py:904] (4/8) Epoch 18, batch 4550, loss[loss=0.2115, simple_loss=0.2914, pruned_loss=0.06575, over 17116.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.286, pruned_loss=0.05533, over 3225109.40 frames. ], batch size: 49, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:09:52,992 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=177109.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:09:55,611 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=177111.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:10:11,673 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=177122.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:10:47,514 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.38 vs. limit=5.0 2023-04-30 17:10:48,535 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.278e+02 1.819e+02 2.100e+02 2.524e+02 4.531e+02, threshold=4.200e+02, percent-clipped=0.0 2023-04-30 17:10:53,841 INFO [train.py:904] (4/8) Epoch 18, batch 4600, loss[loss=0.1876, simple_loss=0.2801, pruned_loss=0.04756, over 16670.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2868, pruned_loss=0.0554, over 3234128.63 frames. ], batch size: 134, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:11:19,982 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=177170.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:11:20,172 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=177170.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:11:24,037 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=177172.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:11:34,709 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4934, 2.6509, 2.1218, 2.4060, 3.0519, 2.6494, 3.1452, 3.2104], device='cuda:4'), covar=tensor([0.0088, 0.0321, 0.0467, 0.0346, 0.0186, 0.0294, 0.0176, 0.0209], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0229, 0.0220, 0.0219, 0.0229, 0.0228, 0.0231, 0.0224], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:12:05,777 INFO [train.py:904] (4/8) Epoch 18, batch 4650, loss[loss=0.2232, simple_loss=0.293, pruned_loss=0.07667, over 11915.00 frames. ], tot_loss[loss=0.1981, simple_loss=0.2855, pruned_loss=0.05534, over 3229587.26 frames. ], batch size: 247, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:12:16,074 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=177209.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 17:13:10,932 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.284e+02 1.973e+02 2.204e+02 2.634e+02 3.745e+02, threshold=4.408e+02, percent-clipped=0.0 2023-04-30 17:13:16,214 INFO [train.py:904] (4/8) Epoch 18, batch 4700, loss[loss=0.1732, simple_loss=0.2632, pruned_loss=0.04163, over 16952.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2826, pruned_loss=0.05455, over 3224433.09 frames. ], batch size: 90, lr: 3.77e-03, grad_scale: 16.0 2023-04-30 17:13:33,364 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-30 17:13:59,447 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=177281.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:14:28,060 INFO [train.py:904] (4/8) Epoch 18, batch 4750, loss[loss=0.176, simple_loss=0.2683, pruned_loss=0.0419, over 15560.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.2784, pruned_loss=0.0523, over 3210940.62 frames. ], batch size: 190, lr: 3.77e-03, grad_scale: 16.0 2023-04-30 17:14:41,395 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7303, 4.7226, 4.6000, 3.1891, 3.9301, 4.5339, 3.9585, 2.4538], device='cuda:4'), covar=tensor([0.0488, 0.0020, 0.0027, 0.0329, 0.0087, 0.0061, 0.0078, 0.0428], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0079, 0.0080, 0.0132, 0.0095, 0.0105, 0.0092, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 17:14:55,990 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=177321.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:15:07,309 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=177329.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:15:32,644 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3446, 4.4027, 4.7225, 4.6981, 4.6881, 4.4018, 4.3951, 4.2585], device='cuda:4'), covar=tensor([0.0282, 0.0481, 0.0368, 0.0373, 0.0378, 0.0353, 0.0799, 0.0437], device='cuda:4'), in_proj_covar=tensor([0.0379, 0.0413, 0.0406, 0.0380, 0.0450, 0.0426, 0.0520, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 17:15:34,632 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.255e+02 1.837e+02 2.097e+02 2.577e+02 4.933e+02, threshold=4.195e+02, percent-clipped=1.0 2023-04-30 17:15:40,560 INFO [train.py:904] (4/8) Epoch 18, batch 4800, loss[loss=0.1759, simple_loss=0.2645, pruned_loss=0.04361, over 16455.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2751, pruned_loss=0.05074, over 3195449.97 frames. ], batch size: 75, lr: 3.77e-03, grad_scale: 16.0 2023-04-30 17:16:07,488 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=177369.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:16:57,032 INFO [train.py:904] (4/8) Epoch 18, batch 4850, loss[loss=0.1768, simple_loss=0.2725, pruned_loss=0.04057, over 15312.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2756, pruned_loss=0.05016, over 3165336.86 frames. ], batch size: 190, lr: 3.77e-03, grad_scale: 16.0 2023-04-30 17:17:19,828 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7927, 3.9145, 4.1230, 4.0930, 4.1063, 3.8585, 3.8744, 3.8407], device='cuda:4'), covar=tensor([0.0306, 0.0531, 0.0384, 0.0387, 0.0388, 0.0399, 0.0776, 0.0469], device='cuda:4'), in_proj_covar=tensor([0.0379, 0.0412, 0.0406, 0.0380, 0.0451, 0.0425, 0.0520, 0.0338], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 17:17:28,969 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4902, 3.4651, 2.0956, 3.9568, 2.6065, 3.9181, 2.1390, 2.7939], device='cuda:4'), covar=tensor([0.0268, 0.0377, 0.1634, 0.0113, 0.0895, 0.0526, 0.1598, 0.0764], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0188, 0.0150, 0.0171, 0.0210, 0.0197, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 17:17:36,043 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1913, 4.1510, 4.1157, 3.3428, 4.1083, 1.7752, 3.8428, 3.7153], device='cuda:4'), covar=tensor([0.0099, 0.0109, 0.0145, 0.0356, 0.0090, 0.2673, 0.0147, 0.0237], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0144, 0.0188, 0.0173, 0.0163, 0.0199, 0.0179, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:18:05,879 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.349e+02 1.952e+02 2.254e+02 2.679e+02 4.935e+02, threshold=4.508e+02, percent-clipped=3.0 2023-04-30 17:18:11,220 INFO [train.py:904] (4/8) Epoch 18, batch 4900, loss[loss=0.1686, simple_loss=0.2601, pruned_loss=0.03853, over 16625.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2748, pruned_loss=0.04873, over 3170573.30 frames. ], batch size: 57, lr: 3.77e-03, grad_scale: 16.0 2023-04-30 17:18:29,114 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6658, 2.6289, 1.8373, 2.7296, 2.1082, 2.7750, 2.0896, 2.4085], device='cuda:4'), covar=tensor([0.0288, 0.0368, 0.1317, 0.0180, 0.0779, 0.0471, 0.1259, 0.0686], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0170, 0.0188, 0.0149, 0.0170, 0.0209, 0.0196, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 17:18:30,795 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=177465.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:18:33,653 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=177467.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:18:44,840 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-30 17:19:07,830 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-30 17:19:25,467 INFO [train.py:904] (4/8) Epoch 18, batch 4950, loss[loss=0.2137, simple_loss=0.293, pruned_loss=0.06716, over 12019.00 frames. ], tot_loss[loss=0.1857, simple_loss=0.2748, pruned_loss=0.04826, over 3157832.73 frames. ], batch size: 246, lr: 3.77e-03, grad_scale: 16.0 2023-04-30 17:19:34,762 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=177509.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 17:20:30,519 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.552e+02 2.064e+02 2.400e+02 2.988e+02 6.295e+02, threshold=4.800e+02, percent-clipped=2.0 2023-04-30 17:20:34,084 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4858, 4.4984, 4.4311, 2.8605, 3.8250, 4.3610, 3.8509, 2.2031], device='cuda:4'), covar=tensor([0.0510, 0.0023, 0.0027, 0.0355, 0.0085, 0.0065, 0.0078, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0078, 0.0078, 0.0130, 0.0093, 0.0103, 0.0090, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 17:20:35,814 INFO [train.py:904] (4/8) Epoch 18, batch 5000, loss[loss=0.1973, simple_loss=0.3037, pruned_loss=0.04546, over 16868.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2772, pruned_loss=0.04847, over 3185334.57 frames. ], batch size: 96, lr: 3.77e-03, grad_scale: 16.0 2023-04-30 17:20:37,350 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=177553.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:20:41,850 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=177557.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:21:47,168 INFO [train.py:904] (4/8) Epoch 18, batch 5050, loss[loss=0.1691, simple_loss=0.2633, pruned_loss=0.03746, over 16844.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.278, pruned_loss=0.04841, over 3186694.95 frames. ], batch size: 83, lr: 3.77e-03, grad_scale: 16.0 2023-04-30 17:21:48,861 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8087, 1.8992, 2.4203, 2.8096, 2.7271, 3.2132, 2.0036, 3.1926], device='cuda:4'), covar=tensor([0.0181, 0.0479, 0.0316, 0.0275, 0.0265, 0.0147, 0.0512, 0.0113], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0190, 0.0178, 0.0181, 0.0189, 0.0147, 0.0194, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:22:04,599 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=177614.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:22:18,238 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-30 17:22:54,211 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.495e+02 2.030e+02 2.394e+02 2.756e+02 4.573e+02, threshold=4.789e+02, percent-clipped=0.0 2023-04-30 17:22:58,430 INFO [train.py:904] (4/8) Epoch 18, batch 5100, loss[loss=0.1685, simple_loss=0.2593, pruned_loss=0.03883, over 16540.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2764, pruned_loss=0.0478, over 3190310.41 frames. ], batch size: 68, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:23:16,157 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=177664.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:24:10,477 INFO [train.py:904] (4/8) Epoch 18, batch 5150, loss[loss=0.1837, simple_loss=0.2805, pruned_loss=0.04351, over 16719.00 frames. ], tot_loss[loss=0.1845, simple_loss=0.2756, pruned_loss=0.04671, over 3195190.78 frames. ], batch size: 124, lr: 3.77e-03, grad_scale: 8.0 2023-04-30 17:24:46,116 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=177725.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:25:20,502 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.353e+02 1.888e+02 2.241e+02 2.680e+02 5.710e+02, threshold=4.482e+02, percent-clipped=2.0 2023-04-30 17:25:24,673 INFO [train.py:904] (4/8) Epoch 18, batch 5200, loss[loss=0.2229, simple_loss=0.3051, pruned_loss=0.07033, over 16293.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2744, pruned_loss=0.04634, over 3186205.08 frames. ], batch size: 165, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:25:44,138 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=177765.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:25:47,770 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=177767.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:26:06,676 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=177780.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:26:31,555 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=177797.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:26:39,155 INFO [train.py:904] (4/8) Epoch 18, batch 5250, loss[loss=0.1906, simple_loss=0.2715, pruned_loss=0.05486, over 12508.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2718, pruned_loss=0.04602, over 3186745.41 frames. ], batch size: 248, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:26:55,360 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=177813.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:26:58,849 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=177815.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:27:05,449 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-30 17:27:06,278 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8404, 2.8142, 2.8865, 4.5880, 3.4949, 4.1761, 1.7684, 3.2217], device='cuda:4'), covar=tensor([0.1298, 0.0744, 0.1087, 0.0110, 0.0256, 0.0350, 0.1537, 0.0724], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0169, 0.0191, 0.0180, 0.0204, 0.0214, 0.0196, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 17:27:37,225 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=177841.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:27:47,050 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5927, 5.9793, 5.6402, 5.7564, 5.4163, 5.2886, 5.3052, 6.0513], device='cuda:4'), covar=tensor([0.1182, 0.0769, 0.0918, 0.0656, 0.0796, 0.0629, 0.1017, 0.0861], device='cuda:4'), in_proj_covar=tensor([0.0625, 0.0772, 0.0633, 0.0576, 0.0487, 0.0495, 0.0644, 0.0601], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:27:48,908 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.558e+02 1.963e+02 2.196e+02 2.536e+02 4.359e+02, threshold=4.392e+02, percent-clipped=1.0 2023-04-30 17:27:52,846 INFO [train.py:904] (4/8) Epoch 18, batch 5300, loss[loss=0.1684, simple_loss=0.2491, pruned_loss=0.04392, over 16656.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2676, pruned_loss=0.04463, over 3196581.90 frames. ], batch size: 62, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:28:02,073 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=177858.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:28:04,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9577, 4.2873, 3.1606, 2.5840, 2.8453, 2.7475, 4.6692, 3.7962], device='cuda:4'), covar=tensor([0.2559, 0.0599, 0.1728, 0.2334, 0.2595, 0.1716, 0.0367, 0.1094], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0264, 0.0298, 0.0302, 0.0291, 0.0246, 0.0288, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 17:28:14,847 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8813, 2.8486, 2.7894, 1.9113, 2.6674, 2.7583, 2.6182, 1.8502], device='cuda:4'), covar=tensor([0.0432, 0.0060, 0.0060, 0.0353, 0.0113, 0.0099, 0.0118, 0.0419], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0078, 0.0079, 0.0131, 0.0094, 0.0103, 0.0091, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 17:28:28,954 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8414, 1.4789, 1.6818, 1.7623, 1.8523, 1.9889, 1.6148, 1.8578], device='cuda:4'), covar=tensor([0.0218, 0.0368, 0.0199, 0.0273, 0.0251, 0.0172, 0.0387, 0.0131], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0188, 0.0176, 0.0179, 0.0187, 0.0145, 0.0192, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:28:43,625 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6147, 3.6110, 4.1144, 2.0539, 4.3887, 4.3187, 3.1007, 3.1092], device='cuda:4'), covar=tensor([0.0806, 0.0253, 0.0150, 0.1117, 0.0044, 0.0104, 0.0391, 0.0447], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0106, 0.0094, 0.0138, 0.0076, 0.0121, 0.0125, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 17:29:05,780 INFO [train.py:904] (4/8) Epoch 18, batch 5350, loss[loss=0.1818, simple_loss=0.2752, pruned_loss=0.04419, over 16608.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2656, pruned_loss=0.04382, over 3193622.76 frames. ], batch size: 62, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:29:06,799 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3295, 3.5011, 3.6286, 3.6023, 3.6097, 3.4535, 3.4632, 3.4888], device='cuda:4'), covar=tensor([0.0387, 0.0611, 0.0457, 0.0445, 0.0501, 0.0477, 0.0763, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0382, 0.0417, 0.0410, 0.0384, 0.0456, 0.0430, 0.0527, 0.0343], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 17:29:17,491 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=177909.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:29:49,885 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-30 17:29:51,338 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9340, 2.3011, 1.8861, 2.0903, 2.7243, 2.3555, 2.5869, 2.8198], device='cuda:4'), covar=tensor([0.0137, 0.0397, 0.0521, 0.0454, 0.0239, 0.0368, 0.0192, 0.0282], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0226, 0.0218, 0.0216, 0.0227, 0.0226, 0.0226, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:30:09,396 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-30 17:30:10,634 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.92 vs. limit=5.0 2023-04-30 17:30:15,239 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.427e+02 1.886e+02 2.232e+02 2.628e+02 4.605e+02, threshold=4.463e+02, percent-clipped=1.0 2023-04-30 17:30:19,769 INFO [train.py:904] (4/8) Epoch 18, batch 5400, loss[loss=0.1741, simple_loss=0.2725, pruned_loss=0.03787, over 16708.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2684, pruned_loss=0.04444, over 3204494.32 frames. ], batch size: 83, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:31:28,904 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-30 17:31:39,766 INFO [train.py:904] (4/8) Epoch 18, batch 5450, loss[loss=0.2086, simple_loss=0.2997, pruned_loss=0.05879, over 16839.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2715, pruned_loss=0.04579, over 3206974.80 frames. ], batch size: 116, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:32:08,839 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=178020.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:32:41,182 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3568, 4.4164, 4.7842, 4.7463, 4.7570, 4.4055, 4.3909, 4.2724], device='cuda:4'), covar=tensor([0.0302, 0.0507, 0.0335, 0.0377, 0.0405, 0.0405, 0.0920, 0.0498], device='cuda:4'), in_proj_covar=tensor([0.0384, 0.0419, 0.0411, 0.0386, 0.0458, 0.0433, 0.0531, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 17:32:52,432 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.664e+02 2.290e+02 2.936e+02 3.716e+02 7.248e+02, threshold=5.872e+02, percent-clipped=9.0 2023-04-30 17:32:56,916 INFO [train.py:904] (4/8) Epoch 18, batch 5500, loss[loss=0.2422, simple_loss=0.3284, pruned_loss=0.07802, over 16381.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2793, pruned_loss=0.05132, over 3147833.39 frames. ], batch size: 146, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:34:14,385 INFO [train.py:904] (4/8) Epoch 18, batch 5550, loss[loss=0.2123, simple_loss=0.3081, pruned_loss=0.05831, over 16840.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2868, pruned_loss=0.05641, over 3124291.15 frames. ], batch size: 83, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:35:09,070 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=178136.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:35:31,150 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.108e+02 3.250e+02 3.857e+02 5.113e+02 1.112e+03, threshold=7.715e+02, percent-clipped=10.0 2023-04-30 17:35:34,161 INFO [train.py:904] (4/8) Epoch 18, batch 5600, loss[loss=0.2196, simple_loss=0.3033, pruned_loss=0.06791, over 16799.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2932, pruned_loss=0.062, over 3076298.44 frames. ], batch size: 124, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:35:36,603 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=178153.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:36:06,928 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4581, 1.6909, 2.1360, 2.4003, 2.4868, 2.7316, 1.8195, 2.6958], device='cuda:4'), covar=tensor([0.0184, 0.0463, 0.0280, 0.0275, 0.0269, 0.0172, 0.0477, 0.0126], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0187, 0.0176, 0.0179, 0.0186, 0.0145, 0.0191, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:36:34,079 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2688, 2.4563, 2.0421, 2.2095, 2.8681, 2.4131, 2.9450, 3.0446], device='cuda:4'), covar=tensor([0.0123, 0.0380, 0.0472, 0.0416, 0.0220, 0.0379, 0.0198, 0.0231], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0226, 0.0218, 0.0217, 0.0227, 0.0226, 0.0228, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:36:58,586 INFO [train.py:904] (4/8) Epoch 18, batch 5650, loss[loss=0.2139, simple_loss=0.3128, pruned_loss=0.0575, over 16974.00 frames. ], tot_loss[loss=0.2147, simple_loss=0.298, pruned_loss=0.06567, over 3055061.12 frames. ], batch size: 41, lr: 3.76e-03, grad_scale: 4.0 2023-04-30 17:37:09,690 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=178209.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:38:16,449 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.765e+02 3.197e+02 3.818e+02 4.692e+02 7.543e+02, threshold=7.636e+02, percent-clipped=0.0 2023-04-30 17:38:17,793 INFO [train.py:904] (4/8) Epoch 18, batch 5700, loss[loss=0.2132, simple_loss=0.3092, pruned_loss=0.05856, over 16357.00 frames. ], tot_loss[loss=0.2165, simple_loss=0.2996, pruned_loss=0.06669, over 3058638.03 frames. ], batch size: 146, lr: 3.76e-03, grad_scale: 4.0 2023-04-30 17:38:20,145 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3288, 3.9076, 3.7616, 2.4251, 3.4707, 3.8630, 3.4953, 2.0939], device='cuda:4'), covar=tensor([0.0506, 0.0036, 0.0059, 0.0414, 0.0097, 0.0103, 0.0093, 0.0432], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0078, 0.0079, 0.0130, 0.0093, 0.0103, 0.0090, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 17:38:25,290 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=178257.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:39:39,090 INFO [train.py:904] (4/8) Epoch 18, batch 5750, loss[loss=0.2145, simple_loss=0.2978, pruned_loss=0.06564, over 17096.00 frames. ], tot_loss[loss=0.2184, simple_loss=0.3013, pruned_loss=0.06773, over 3041753.87 frames. ], batch size: 47, lr: 3.76e-03, grad_scale: 4.0 2023-04-30 17:40:08,752 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=178320.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:40:56,642 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-30 17:40:59,319 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.776e+02 3.189e+02 3.915e+02 4.745e+02 1.114e+03, threshold=7.831e+02, percent-clipped=3.0 2023-04-30 17:41:00,704 INFO [train.py:904] (4/8) Epoch 18, batch 5800, loss[loss=0.2423, simple_loss=0.3081, pruned_loss=0.08822, over 12170.00 frames. ], tot_loss[loss=0.2166, simple_loss=0.3004, pruned_loss=0.06642, over 3040847.33 frames. ], batch size: 248, lr: 3.76e-03, grad_scale: 4.0 2023-04-30 17:41:28,156 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=178368.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:42:19,739 INFO [train.py:904] (4/8) Epoch 18, batch 5850, loss[loss=0.2119, simple_loss=0.3025, pruned_loss=0.06064, over 16194.00 frames. ], tot_loss[loss=0.2137, simple_loss=0.2984, pruned_loss=0.06452, over 3064016.00 frames. ], batch size: 165, lr: 3.76e-03, grad_scale: 4.0 2023-04-30 17:43:16,101 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=178436.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:43:39,873 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.729e+02 2.654e+02 3.099e+02 3.804e+02 9.220e+02, threshold=6.197e+02, percent-clipped=1.0 2023-04-30 17:43:41,471 INFO [train.py:904] (4/8) Epoch 18, batch 5900, loss[loss=0.2235, simple_loss=0.311, pruned_loss=0.06805, over 16638.00 frames. ], tot_loss[loss=0.2135, simple_loss=0.298, pruned_loss=0.06451, over 3059673.37 frames. ], batch size: 62, lr: 3.76e-03, grad_scale: 4.0 2023-04-30 17:43:42,228 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=178452.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:43:44,114 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=178453.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:44:32,738 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=178484.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:44:51,611 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9473, 3.0770, 3.1848, 2.0359, 2.9969, 3.1924, 3.0452, 1.8495], device='cuda:4'), covar=tensor([0.0551, 0.0089, 0.0069, 0.0443, 0.0106, 0.0111, 0.0091, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0079, 0.0080, 0.0132, 0.0094, 0.0105, 0.0091, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 17:44:58,646 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=178501.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:44:59,546 INFO [train.py:904] (4/8) Epoch 18, batch 5950, loss[loss=0.2175, simple_loss=0.3058, pruned_loss=0.06461, over 15386.00 frames. ], tot_loss[loss=0.2131, simple_loss=0.299, pruned_loss=0.06361, over 3051573.25 frames. ], batch size: 190, lr: 3.76e-03, grad_scale: 4.0 2023-04-30 17:45:18,613 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=178513.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:45:21,272 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4297, 3.4230, 2.1263, 3.8135, 2.6259, 3.8159, 2.2407, 2.7813], device='cuda:4'), covar=tensor([0.0261, 0.0365, 0.1519, 0.0217, 0.0781, 0.0638, 0.1434, 0.0720], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0171, 0.0190, 0.0151, 0.0173, 0.0211, 0.0198, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 17:45:29,116 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0476, 5.5756, 5.7686, 5.5159, 5.5849, 6.1253, 5.5967, 5.4492], device='cuda:4'), covar=tensor([0.0871, 0.1737, 0.2053, 0.1772, 0.2201, 0.0812, 0.1546, 0.2260], device='cuda:4'), in_proj_covar=tensor([0.0390, 0.0561, 0.0616, 0.0471, 0.0631, 0.0642, 0.0488, 0.0636], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 17:46:17,845 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.944e+02 2.727e+02 3.457e+02 4.156e+02 1.353e+03, threshold=6.914e+02, percent-clipped=3.0 2023-04-30 17:46:19,086 INFO [train.py:904] (4/8) Epoch 18, batch 6000, loss[loss=0.2607, simple_loss=0.3261, pruned_loss=0.09771, over 11591.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.298, pruned_loss=0.06328, over 3056241.92 frames. ], batch size: 248, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:46:19,086 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 17:46:29,951 INFO [train.py:938] (4/8) Epoch 18, validation: loss=0.1531, simple_loss=0.2661, pruned_loss=0.02007, over 944034.00 frames. 2023-04-30 17:46:29,951 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 17:47:48,044 INFO [train.py:904] (4/8) Epoch 18, batch 6050, loss[loss=0.2045, simple_loss=0.3042, pruned_loss=0.05233, over 17005.00 frames. ], tot_loss[loss=0.2109, simple_loss=0.2965, pruned_loss=0.06259, over 3053909.54 frames. ], batch size: 55, lr: 3.76e-03, grad_scale: 8.0 2023-04-30 17:48:14,275 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=178619.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:48:55,311 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7045, 1.8149, 1.6176, 1.5346, 1.9474, 1.5938, 1.6278, 1.8923], device='cuda:4'), covar=tensor([0.0177, 0.0247, 0.0351, 0.0292, 0.0190, 0.0240, 0.0153, 0.0176], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0224, 0.0217, 0.0216, 0.0226, 0.0224, 0.0226, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:49:06,484 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.761e+02 2.787e+02 3.191e+02 3.956e+02 7.476e+02, threshold=6.381e+02, percent-clipped=1.0 2023-04-30 17:49:06,499 INFO [train.py:904] (4/8) Epoch 18, batch 6100, loss[loss=0.2173, simple_loss=0.2988, pruned_loss=0.06788, over 15282.00 frames. ], tot_loss[loss=0.2102, simple_loss=0.2963, pruned_loss=0.06205, over 3051928.01 frames. ], batch size: 191, lr: 3.76e-03, grad_scale: 4.0 2023-04-30 17:49:24,107 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0115, 2.3145, 2.3634, 2.7369, 2.0097, 3.2304, 1.7533, 2.6823], device='cuda:4'), covar=tensor([0.1130, 0.0671, 0.1047, 0.0178, 0.0140, 0.0395, 0.1481, 0.0719], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0170, 0.0191, 0.0180, 0.0206, 0.0215, 0.0196, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 17:49:51,389 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=178680.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:50:12,319 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=178694.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:50:16,782 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=178697.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 17:50:18,829 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0727, 2.8639, 3.0879, 1.7322, 3.3022, 3.3387, 2.6963, 2.5265], device='cuda:4'), covar=tensor([0.0906, 0.0286, 0.0203, 0.1204, 0.0080, 0.0195, 0.0441, 0.0489], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0106, 0.0095, 0.0139, 0.0076, 0.0122, 0.0125, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 17:50:23,986 INFO [train.py:904] (4/8) Epoch 18, batch 6150, loss[loss=0.2475, simple_loss=0.3132, pruned_loss=0.09088, over 11633.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2941, pruned_loss=0.0611, over 3064202.89 frames. ], batch size: 246, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 17:51:34,493 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.79 vs. limit=5.0 2023-04-30 17:51:39,656 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.052e+02 2.816e+02 3.285e+02 4.160e+02 8.412e+02, threshold=6.570e+02, percent-clipped=4.0 2023-04-30 17:51:39,671 INFO [train.py:904] (4/8) Epoch 18, batch 6200, loss[loss=0.212, simple_loss=0.2967, pruned_loss=0.06364, over 16353.00 frames. ], tot_loss[loss=0.2069, simple_loss=0.2922, pruned_loss=0.06074, over 3057493.01 frames. ], batch size: 35, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 17:51:45,787 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=178755.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 17:51:46,931 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0468, 5.0591, 4.8832, 4.1909, 4.9883, 2.0623, 4.7275, 4.7069], device='cuda:4'), covar=tensor([0.0083, 0.0073, 0.0168, 0.0398, 0.0093, 0.2440, 0.0119, 0.0194], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0142, 0.0188, 0.0173, 0.0163, 0.0198, 0.0178, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:51:49,676 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=178758.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 17:52:08,979 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=178770.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:52:12,813 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8179, 3.7032, 3.8663, 3.9704, 4.0533, 3.7033, 4.0157, 4.0643], device='cuda:4'), covar=tensor([0.1605, 0.1159, 0.1333, 0.0715, 0.0647, 0.1713, 0.0817, 0.0672], device='cuda:4'), in_proj_covar=tensor([0.0602, 0.0741, 0.0871, 0.0759, 0.0564, 0.0597, 0.0610, 0.0705], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 17:52:56,568 INFO [train.py:904] (4/8) Epoch 18, batch 6250, loss[loss=0.172, simple_loss=0.276, pruned_loss=0.03399, over 16900.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.2913, pruned_loss=0.06022, over 3079054.58 frames. ], batch size: 96, lr: 3.75e-03, grad_scale: 2.0 2023-04-30 17:53:07,992 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=178808.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:53:41,414 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=178831.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:54:15,682 INFO [train.py:904] (4/8) Epoch 18, batch 6300, loss[loss=0.2048, simple_loss=0.2863, pruned_loss=0.06161, over 16599.00 frames. ], tot_loss[loss=0.2049, simple_loss=0.2909, pruned_loss=0.05943, over 3083727.88 frames. ], batch size: 62, lr: 3.75e-03, grad_scale: 2.0 2023-04-30 17:54:17,523 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.018e+02 2.791e+02 3.273e+02 3.863e+02 9.337e+02, threshold=6.545e+02, percent-clipped=2.0 2023-04-30 17:55:22,821 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=178894.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 17:55:34,230 INFO [train.py:904] (4/8) Epoch 18, batch 6350, loss[loss=0.1759, simple_loss=0.2729, pruned_loss=0.03941, over 16841.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2917, pruned_loss=0.06017, over 3087097.44 frames. ], batch size: 102, lr: 3.75e-03, grad_scale: 2.0 2023-04-30 17:56:52,036 INFO [train.py:904] (4/8) Epoch 18, batch 6400, loss[loss=0.2215, simple_loss=0.3028, pruned_loss=0.07007, over 16339.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2914, pruned_loss=0.06064, over 3109096.77 frames. ], batch size: 146, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 17:56:53,836 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.717e+02 3.036e+02 3.548e+02 4.325e+02 9.483e+02, threshold=7.097e+02, percent-clipped=4.0 2023-04-30 17:56:57,523 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=178955.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 17:57:28,085 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=178975.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 17:58:07,507 INFO [train.py:904] (4/8) Epoch 18, batch 6450, loss[loss=0.1952, simple_loss=0.2819, pruned_loss=0.05422, over 16443.00 frames. ], tot_loss[loss=0.2057, simple_loss=0.2913, pruned_loss=0.06004, over 3110458.24 frames. ], batch size: 68, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 17:59:24,076 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=179050.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 17:59:26,157 INFO [train.py:904] (4/8) Epoch 18, batch 6500, loss[loss=0.2262, simple_loss=0.3031, pruned_loss=0.07462, over 16728.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2898, pruned_loss=0.05981, over 3114554.77 frames. ], batch size: 134, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 17:59:27,317 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.795e+02 2.918e+02 3.772e+02 4.451e+02 8.133e+02, threshold=7.544e+02, percent-clipped=2.0 2023-04-30 17:59:28,316 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=179053.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 18:00:44,327 INFO [train.py:904] (4/8) Epoch 18, batch 6550, loss[loss=0.2016, simple_loss=0.3023, pruned_loss=0.0504, over 16675.00 frames. ], tot_loss[loss=0.2074, simple_loss=0.2928, pruned_loss=0.06099, over 3116625.35 frames. ], batch size: 134, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:00:54,328 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=179108.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:01:20,178 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=179126.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:01:58,890 INFO [train.py:904] (4/8) Epoch 18, batch 6600, loss[loss=0.2157, simple_loss=0.3148, pruned_loss=0.0583, over 16777.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2946, pruned_loss=0.06149, over 3089199.77 frames. ], batch size: 124, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:02:00,668 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.896e+02 2.804e+02 3.237e+02 3.890e+02 6.888e+02, threshold=6.473e+02, percent-clipped=0.0 2023-04-30 18:02:02,443 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8933, 2.0712, 2.3768, 3.1710, 2.1514, 2.2972, 2.2884, 2.2220], device='cuda:4'), covar=tensor([0.1248, 0.3171, 0.2273, 0.0638, 0.3929, 0.2442, 0.2954, 0.3112], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0427, 0.0352, 0.0319, 0.0427, 0.0493, 0.0397, 0.0500], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:02:05,100 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=179156.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:03:07,469 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7716, 3.8553, 4.1302, 4.0761, 4.0875, 3.8483, 3.8609, 3.8118], device='cuda:4'), covar=tensor([0.0342, 0.0615, 0.0422, 0.0451, 0.0490, 0.0453, 0.0898, 0.0582], device='cuda:4'), in_proj_covar=tensor([0.0384, 0.0419, 0.0409, 0.0386, 0.0458, 0.0430, 0.0527, 0.0345], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 18:03:17,504 INFO [train.py:904] (4/8) Epoch 18, batch 6650, loss[loss=0.1842, simple_loss=0.2714, pruned_loss=0.04846, over 16955.00 frames. ], tot_loss[loss=0.2101, simple_loss=0.2948, pruned_loss=0.06268, over 3076908.43 frames. ], batch size: 109, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:04:07,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4435, 4.4822, 4.8319, 4.7858, 4.7935, 4.4723, 4.4758, 4.2446], device='cuda:4'), covar=tensor([0.0322, 0.0487, 0.0386, 0.0379, 0.0473, 0.0400, 0.0984, 0.0585], device='cuda:4'), in_proj_covar=tensor([0.0387, 0.0422, 0.0411, 0.0389, 0.0461, 0.0433, 0.0531, 0.0348], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 18:04:30,597 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=179250.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 18:04:32,473 INFO [train.py:904] (4/8) Epoch 18, batch 6700, loss[loss=0.1915, simple_loss=0.2782, pruned_loss=0.05245, over 16656.00 frames. ], tot_loss[loss=0.2092, simple_loss=0.2934, pruned_loss=0.06249, over 3087047.64 frames. ], batch size: 62, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:04:34,183 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.091e+02 2.690e+02 3.432e+02 4.163e+02 9.246e+02, threshold=6.864e+02, percent-clipped=3.0 2023-04-30 18:05:09,290 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=179275.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:05:24,213 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9748, 2.7837, 2.5491, 4.6248, 3.3201, 4.0235, 1.8083, 2.7911], device='cuda:4'), covar=tensor([0.1318, 0.0867, 0.1356, 0.0162, 0.0341, 0.0478, 0.1653, 0.1047], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0167, 0.0189, 0.0176, 0.0204, 0.0212, 0.0194, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 18:05:48,860 INFO [train.py:904] (4/8) Epoch 18, batch 6750, loss[loss=0.1736, simple_loss=0.2672, pruned_loss=0.04002, over 16835.00 frames. ], tot_loss[loss=0.2079, simple_loss=0.292, pruned_loss=0.06192, over 3097169.88 frames. ], batch size: 102, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:06:19,478 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=179323.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:06:21,433 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3836, 3.3229, 3.3915, 3.4769, 3.5173, 3.2859, 3.4727, 3.5537], device='cuda:4'), covar=tensor([0.1305, 0.0940, 0.1143, 0.0624, 0.0644, 0.2111, 0.1010, 0.0730], device='cuda:4'), in_proj_covar=tensor([0.0600, 0.0735, 0.0866, 0.0751, 0.0562, 0.0593, 0.0609, 0.0699], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:06:25,194 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6312, 2.5784, 1.8952, 2.6595, 2.1191, 2.7496, 2.0760, 2.3394], device='cuda:4'), covar=tensor([0.0338, 0.0375, 0.1273, 0.0244, 0.0685, 0.0515, 0.1373, 0.0671], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0172, 0.0192, 0.0153, 0.0174, 0.0213, 0.0200, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 18:06:36,816 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3426, 4.3217, 4.2045, 3.4636, 4.2799, 1.6694, 4.0393, 3.8170], device='cuda:4'), covar=tensor([0.0100, 0.0083, 0.0172, 0.0338, 0.0089, 0.2907, 0.0127, 0.0248], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0142, 0.0188, 0.0172, 0.0163, 0.0199, 0.0178, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:07:00,949 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=179350.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:07:03,096 INFO [train.py:904] (4/8) Epoch 18, batch 6800, loss[loss=0.2291, simple_loss=0.3021, pruned_loss=0.07811, over 11618.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2925, pruned_loss=0.06198, over 3079962.44 frames. ], batch size: 247, lr: 3.75e-03, grad_scale: 8.0 2023-04-30 18:07:04,928 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.988e+02 2.962e+02 3.412e+02 4.074e+02 1.001e+03, threshold=6.824e+02, percent-clipped=4.0 2023-04-30 18:07:05,954 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=179353.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 18:07:18,272 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=179361.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:08:16,242 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=179398.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:08:20,458 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=179401.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 18:08:21,233 INFO [train.py:904] (4/8) Epoch 18, batch 6850, loss[loss=0.2049, simple_loss=0.3027, pruned_loss=0.05355, over 16929.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.294, pruned_loss=0.06179, over 3103696.26 frames. ], batch size: 109, lr: 3.75e-03, grad_scale: 8.0 2023-04-30 18:08:47,282 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-30 18:08:50,182 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=179422.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:08:56,277 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=179426.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:09:24,169 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=179444.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 18:09:34,744 INFO [train.py:904] (4/8) Epoch 18, batch 6900, loss[loss=0.2085, simple_loss=0.2995, pruned_loss=0.05873, over 16434.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.2963, pruned_loss=0.06113, over 3118225.17 frames. ], batch size: 146, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:09:38,467 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.677e+02 2.748e+02 3.355e+02 3.841e+02 5.597e+02, threshold=6.710e+02, percent-clipped=0.0 2023-04-30 18:09:47,267 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9828, 4.0208, 4.3197, 4.2809, 4.2834, 4.0174, 4.0356, 3.9813], device='cuda:4'), covar=tensor([0.0339, 0.0668, 0.0418, 0.0427, 0.0463, 0.0462, 0.0899, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0390, 0.0425, 0.0414, 0.0391, 0.0464, 0.0436, 0.0534, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 18:10:10,141 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=179474.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:10:48,211 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2908, 3.4589, 3.6098, 3.5729, 3.5876, 3.3829, 3.4362, 3.4610], device='cuda:4'), covar=tensor([0.0411, 0.0655, 0.0459, 0.0447, 0.0517, 0.0579, 0.0858, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0393, 0.0429, 0.0417, 0.0394, 0.0468, 0.0440, 0.0539, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 18:10:53,461 INFO [train.py:904] (4/8) Epoch 18, batch 6950, loss[loss=0.2523, simple_loss=0.3144, pruned_loss=0.09512, over 11268.00 frames. ], tot_loss[loss=0.2126, simple_loss=0.2985, pruned_loss=0.06331, over 3100142.70 frames. ], batch size: 248, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:10:58,859 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=179505.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 18:11:34,349 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0112, 2.3266, 2.2839, 2.6515, 1.8317, 3.1742, 1.7455, 2.6549], device='cuda:4'), covar=tensor([0.1181, 0.0637, 0.1115, 0.0187, 0.0137, 0.0393, 0.1553, 0.0709], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0167, 0.0189, 0.0176, 0.0203, 0.0211, 0.0194, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 18:12:07,078 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=179550.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 18:12:09,855 INFO [train.py:904] (4/8) Epoch 18, batch 7000, loss[loss=0.1989, simple_loss=0.2721, pruned_loss=0.06289, over 11569.00 frames. ], tot_loss[loss=0.2118, simple_loss=0.2983, pruned_loss=0.06261, over 3108857.70 frames. ], batch size: 247, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:12:12,174 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.911e+02 2.791e+02 3.383e+02 4.252e+02 7.699e+02, threshold=6.767e+02, percent-clipped=2.0 2023-04-30 18:12:55,904 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-30 18:13:16,940 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=179598.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 18:13:21,688 INFO [train.py:904] (4/8) Epoch 18, batch 7050, loss[loss=0.2487, simple_loss=0.3077, pruned_loss=0.09482, over 11529.00 frames. ], tot_loss[loss=0.2122, simple_loss=0.2994, pruned_loss=0.06252, over 3107051.41 frames. ], batch size: 248, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:13:35,914 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6666, 3.7012, 2.8459, 2.2101, 2.6090, 2.4018, 4.0708, 3.4182], device='cuda:4'), covar=tensor([0.2870, 0.0775, 0.1814, 0.2717, 0.2529, 0.2055, 0.0427, 0.1293], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0265, 0.0300, 0.0306, 0.0292, 0.0249, 0.0287, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:14:12,904 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=179636.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:14:37,717 INFO [train.py:904] (4/8) Epoch 18, batch 7100, loss[loss=0.2028, simple_loss=0.2906, pruned_loss=0.05755, over 16452.00 frames. ], tot_loss[loss=0.2108, simple_loss=0.2974, pruned_loss=0.06207, over 3109238.36 frames. ], batch size: 146, lr: 3.75e-03, grad_scale: 4.0 2023-04-30 18:14:40,300 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.117e+02 2.863e+02 3.448e+02 4.199e+02 1.001e+03, threshold=6.895e+02, percent-clipped=2.0 2023-04-30 18:15:33,588 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6630, 1.8049, 2.3060, 2.6030, 2.6400, 2.9737, 1.9252, 2.9694], device='cuda:4'), covar=tensor([0.0215, 0.0479, 0.0323, 0.0303, 0.0274, 0.0165, 0.0521, 0.0128], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0189, 0.0175, 0.0178, 0.0187, 0.0145, 0.0191, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:15:46,861 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=179697.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:15:53,685 INFO [train.py:904] (4/8) Epoch 18, batch 7150, loss[loss=0.2058, simple_loss=0.2917, pruned_loss=0.05992, over 15376.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2953, pruned_loss=0.06179, over 3110467.31 frames. ], batch size: 190, lr: 3.74e-03, grad_scale: 4.0 2023-04-30 18:15:55,565 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.26 vs. limit=5.0 2023-04-30 18:16:09,088 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0087, 4.1357, 3.1748, 2.5085, 3.0143, 2.6216, 4.6951, 3.7943], device='cuda:4'), covar=tensor([0.2573, 0.0678, 0.1667, 0.2375, 0.2576, 0.1883, 0.0395, 0.1143], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0265, 0.0299, 0.0305, 0.0292, 0.0249, 0.0287, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:16:16,355 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=179717.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:16:27,326 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.51 vs. limit=2.0 2023-04-30 18:17:00,455 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8003, 3.8995, 2.9817, 2.3335, 2.7256, 2.4967, 4.2845, 3.5561], device='cuda:4'), covar=tensor([0.2769, 0.0740, 0.1794, 0.2672, 0.2558, 0.2004, 0.0471, 0.1206], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0265, 0.0299, 0.0305, 0.0291, 0.0249, 0.0286, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:17:08,205 INFO [train.py:904] (4/8) Epoch 18, batch 7200, loss[loss=0.1715, simple_loss=0.2608, pruned_loss=0.04117, over 16534.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2925, pruned_loss=0.06002, over 3103358.32 frames. ], batch size: 68, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:17:10,635 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.743e+02 2.765e+02 3.366e+02 4.196e+02 7.871e+02, threshold=6.733e+02, percent-clipped=4.0 2023-04-30 18:17:29,416 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-30 18:17:30,278 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3834, 2.5214, 2.2679, 2.2816, 2.9019, 2.5405, 3.0114, 3.0813], device='cuda:4'), covar=tensor([0.0113, 0.0423, 0.0476, 0.0457, 0.0249, 0.0385, 0.0250, 0.0246], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0222, 0.0217, 0.0217, 0.0225, 0.0222, 0.0225, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:17:54,684 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-30 18:18:17,290 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=179796.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:18:24,741 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=179800.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 18:18:27,048 INFO [train.py:904] (4/8) Epoch 18, batch 7250, loss[loss=0.1909, simple_loss=0.2725, pruned_loss=0.05463, over 15171.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.29, pruned_loss=0.05865, over 3112529.99 frames. ], batch size: 190, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:18:29,310 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2006, 4.2862, 4.6116, 4.5504, 4.5672, 4.2983, 4.3134, 4.1583], device='cuda:4'), covar=tensor([0.0321, 0.0525, 0.0376, 0.0444, 0.0442, 0.0386, 0.0812, 0.0514], device='cuda:4'), in_proj_covar=tensor([0.0390, 0.0426, 0.0416, 0.0393, 0.0467, 0.0438, 0.0534, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 18:19:41,454 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8131, 1.3080, 1.7016, 1.6297, 1.7896, 1.9404, 1.5706, 1.7707], device='cuda:4'), covar=tensor([0.0225, 0.0347, 0.0187, 0.0238, 0.0225, 0.0147, 0.0355, 0.0110], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0186, 0.0173, 0.0176, 0.0185, 0.0143, 0.0189, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:19:42,142 INFO [train.py:904] (4/8) Epoch 18, batch 7300, loss[loss=0.2325, simple_loss=0.3041, pruned_loss=0.08045, over 11162.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2897, pruned_loss=0.05874, over 3106878.49 frames. ], batch size: 248, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:19:45,258 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.839e+02 2.793e+02 3.453e+02 4.292e+02 8.148e+02, threshold=6.907e+02, percent-clipped=1.0 2023-04-30 18:19:51,008 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=179857.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:20:33,716 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1674, 5.8420, 6.0105, 5.6806, 5.7996, 6.2827, 5.7468, 5.5116], device='cuda:4'), covar=tensor([0.0789, 0.1530, 0.1681, 0.1674, 0.1840, 0.0773, 0.1442, 0.2185], device='cuda:4'), in_proj_covar=tensor([0.0390, 0.0559, 0.0618, 0.0470, 0.0628, 0.0644, 0.0486, 0.0631], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:20:58,393 INFO [train.py:904] (4/8) Epoch 18, batch 7350, loss[loss=0.2318, simple_loss=0.3021, pruned_loss=0.08071, over 11182.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2913, pruned_loss=0.06041, over 3072853.01 frames. ], batch size: 248, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:21:05,519 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=179906.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 18:21:32,184 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=179924.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:22:13,196 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5406, 4.6042, 4.4217, 4.1454, 4.0974, 4.5060, 4.2349, 4.2207], device='cuda:4'), covar=tensor([0.0705, 0.0677, 0.0324, 0.0328, 0.0912, 0.0512, 0.0670, 0.0777], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0387, 0.0317, 0.0309, 0.0332, 0.0361, 0.0218, 0.0382], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:22:16,602 INFO [train.py:904] (4/8) Epoch 18, batch 7400, loss[loss=0.2089, simple_loss=0.297, pruned_loss=0.06039, over 17048.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2926, pruned_loss=0.06156, over 3057910.25 frames. ], batch size: 55, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:22:19,965 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.039e+02 2.864e+02 3.444e+02 4.186e+02 8.589e+02, threshold=6.889e+02, percent-clipped=1.0 2023-04-30 18:22:41,164 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=179967.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 18:22:45,954 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-04-30 18:23:09,638 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=179985.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:23:21,185 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=179992.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:23:39,778 INFO [train.py:904] (4/8) Epoch 18, batch 7450, loss[loss=0.1976, simple_loss=0.2948, pruned_loss=0.05015, over 16852.00 frames. ], tot_loss[loss=0.21, simple_loss=0.294, pruned_loss=0.063, over 3042146.15 frames. ], batch size: 102, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:24:05,903 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180017.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:24:19,901 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=180026.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:24:38,676 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7899, 3.7892, 2.6292, 2.3852, 2.6249, 2.3668, 4.0589, 3.3794], device='cuda:4'), covar=tensor([0.3012, 0.0909, 0.2331, 0.2771, 0.2890, 0.2233, 0.0654, 0.1413], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0264, 0.0299, 0.0304, 0.0291, 0.0247, 0.0286, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:24:50,302 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4144, 3.3750, 2.6186, 2.1461, 2.2585, 2.2335, 3.4566, 3.1249], device='cuda:4'), covar=tensor([0.2918, 0.0718, 0.1810, 0.2622, 0.2430, 0.2139, 0.0489, 0.1239], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0263, 0.0299, 0.0304, 0.0291, 0.0247, 0.0285, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:24:51,824 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 18:25:01,491 INFO [train.py:904] (4/8) Epoch 18, batch 7500, loss[loss=0.197, simple_loss=0.2814, pruned_loss=0.05634, over 16459.00 frames. ], tot_loss[loss=0.2087, simple_loss=0.2941, pruned_loss=0.06166, over 3073138.08 frames. ], batch size: 68, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:25:04,519 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.958e+02 3.002e+02 3.628e+02 4.895e+02 8.341e+02, threshold=7.256e+02, percent-clipped=3.0 2023-04-30 18:25:22,815 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=180065.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:25:33,708 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5017, 3.5405, 3.3284, 3.0224, 3.1701, 3.4427, 3.2578, 3.2802], device='cuda:4'), covar=tensor([0.0561, 0.0564, 0.0270, 0.0261, 0.0529, 0.0412, 0.1528, 0.0505], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0387, 0.0317, 0.0309, 0.0331, 0.0360, 0.0218, 0.0381], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:25:56,698 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=180087.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:26:16,424 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180100.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 18:26:19,022 INFO [train.py:904] (4/8) Epoch 18, batch 7550, loss[loss=0.1993, simple_loss=0.2878, pruned_loss=0.05542, over 16825.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2933, pruned_loss=0.06198, over 3074339.28 frames. ], batch size: 116, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:26:47,583 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-04-30 18:27:18,245 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7645, 3.8214, 3.9272, 3.6799, 3.8426, 4.2290, 3.8536, 3.5679], device='cuda:4'), covar=tensor([0.2001, 0.2092, 0.2416, 0.2448, 0.2574, 0.1660, 0.1836, 0.2804], device='cuda:4'), in_proj_covar=tensor([0.0387, 0.0559, 0.0619, 0.0469, 0.0627, 0.0644, 0.0488, 0.0631], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:27:30,331 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=180148.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 18:27:36,543 INFO [train.py:904] (4/8) Epoch 18, batch 7600, loss[loss=0.2084, simple_loss=0.2922, pruned_loss=0.06232, over 16937.00 frames. ], tot_loss[loss=0.2074, simple_loss=0.2916, pruned_loss=0.06159, over 3072362.52 frames. ], batch size: 109, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:27:36,875 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=180152.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:27:39,406 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.837e+02 2.698e+02 3.315e+02 4.368e+02 9.372e+02, threshold=6.630e+02, percent-clipped=3.0 2023-04-30 18:28:55,649 INFO [train.py:904] (4/8) Epoch 18, batch 7650, loss[loss=0.2306, simple_loss=0.311, pruned_loss=0.07508, over 16377.00 frames. ], tot_loss[loss=0.2091, simple_loss=0.2931, pruned_loss=0.06256, over 3074382.67 frames. ], batch size: 146, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:29:00,823 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=180205.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:29:05,857 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0280, 2.3292, 2.3792, 2.8112, 1.9466, 3.1588, 1.7857, 2.7049], device='cuda:4'), covar=tensor([0.1201, 0.0684, 0.1007, 0.0222, 0.0131, 0.0369, 0.1541, 0.0724], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0170, 0.0191, 0.0178, 0.0206, 0.0214, 0.0196, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 18:30:13,287 INFO [train.py:904] (4/8) Epoch 18, batch 7700, loss[loss=0.1951, simple_loss=0.2806, pruned_loss=0.0548, over 17250.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.2929, pruned_loss=0.06285, over 3073434.41 frames. ], batch size: 52, lr: 3.74e-03, grad_scale: 4.0 2023-04-30 18:30:18,205 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.849e+02 3.108e+02 3.616e+02 4.495e+02 6.527e+02, threshold=7.232e+02, percent-clipped=0.0 2023-04-30 18:30:20,076 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1278, 2.9375, 3.0782, 1.8248, 3.2250, 3.2945, 2.6791, 2.5669], device='cuda:4'), covar=tensor([0.0895, 0.0262, 0.0216, 0.1262, 0.0104, 0.0220, 0.0480, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0105, 0.0094, 0.0139, 0.0075, 0.0121, 0.0125, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 18:30:29,231 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=180262.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 18:30:35,669 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=180266.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:30:57,474 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=180280.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:31:15,917 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180292.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:31:30,823 INFO [train.py:904] (4/8) Epoch 18, batch 7750, loss[loss=0.2068, simple_loss=0.2925, pruned_loss=0.06055, over 16242.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.292, pruned_loss=0.06174, over 3090805.34 frames. ], batch size: 165, lr: 3.74e-03, grad_scale: 4.0 2023-04-30 18:32:23,805 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=180336.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:32:29,343 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=180340.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:32:46,606 INFO [train.py:904] (4/8) Epoch 18, batch 7800, loss[loss=0.2125, simple_loss=0.2973, pruned_loss=0.06388, over 15335.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2924, pruned_loss=0.06198, over 3090964.78 frames. ], batch size: 190, lr: 3.74e-03, grad_scale: 4.0 2023-04-30 18:32:51,025 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.798e+02 2.996e+02 3.455e+02 4.260e+02 9.369e+02, threshold=6.911e+02, percent-clipped=1.0 2023-04-30 18:32:55,336 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9765, 4.0000, 3.9528, 3.2067, 3.9628, 1.6839, 3.7534, 3.6081], device='cuda:4'), covar=tensor([0.0151, 0.0127, 0.0194, 0.0337, 0.0116, 0.2824, 0.0170, 0.0245], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0138, 0.0184, 0.0168, 0.0159, 0.0195, 0.0173, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:33:33,618 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=180382.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:33:55,561 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=180397.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:33:56,872 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5249, 3.3427, 3.7212, 1.7102, 3.8815, 3.9134, 2.9798, 2.8341], device='cuda:4'), covar=tensor([0.0783, 0.0261, 0.0192, 0.1320, 0.0074, 0.0157, 0.0424, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0105, 0.0094, 0.0138, 0.0075, 0.0120, 0.0124, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 18:34:02,033 INFO [train.py:904] (4/8) Epoch 18, batch 7850, loss[loss=0.2448, simple_loss=0.315, pruned_loss=0.08726, over 11291.00 frames. ], tot_loss[loss=0.2089, simple_loss=0.2934, pruned_loss=0.06215, over 3077474.23 frames. ], batch size: 246, lr: 3.74e-03, grad_scale: 4.0 2023-04-30 18:34:09,647 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9307, 4.1770, 3.9805, 4.0431, 3.6962, 3.8042, 3.8338, 4.1475], device='cuda:4'), covar=tensor([0.1125, 0.0926, 0.1024, 0.0826, 0.0823, 0.1483, 0.0935, 0.1108], device='cuda:4'), in_proj_covar=tensor([0.0618, 0.0760, 0.0624, 0.0569, 0.0475, 0.0490, 0.0636, 0.0591], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:34:59,672 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1896, 5.4858, 5.2025, 5.2573, 5.0063, 4.9143, 4.8224, 5.5936], device='cuda:4'), covar=tensor([0.1209, 0.0826, 0.0994, 0.0881, 0.0731, 0.0793, 0.1159, 0.0863], device='cuda:4'), in_proj_covar=tensor([0.0619, 0.0761, 0.0626, 0.0570, 0.0475, 0.0490, 0.0637, 0.0591], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:35:16,217 INFO [train.py:904] (4/8) Epoch 18, batch 7900, loss[loss=0.2471, simple_loss=0.3122, pruned_loss=0.09099, over 11383.00 frames. ], tot_loss[loss=0.2085, simple_loss=0.2932, pruned_loss=0.06185, over 3079056.58 frames. ], batch size: 248, lr: 3.74e-03, grad_scale: 4.0 2023-04-30 18:35:17,215 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180452.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:35:20,361 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.862e+02 2.825e+02 3.445e+02 4.253e+02 7.547e+02, threshold=6.890e+02, percent-clipped=0.0 2023-04-30 18:36:25,849 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-30 18:36:31,922 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=180500.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:36:34,734 INFO [train.py:904] (4/8) Epoch 18, batch 7950, loss[loss=0.1922, simple_loss=0.2835, pruned_loss=0.05042, over 17123.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2933, pruned_loss=0.06215, over 3076657.20 frames. ], batch size: 48, lr: 3.74e-03, grad_scale: 4.0 2023-04-30 18:37:35,947 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-30 18:37:52,807 INFO [train.py:904] (4/8) Epoch 18, batch 8000, loss[loss=0.224, simple_loss=0.3063, pruned_loss=0.07088, over 16707.00 frames. ], tot_loss[loss=0.2108, simple_loss=0.2946, pruned_loss=0.06345, over 3062440.56 frames. ], batch size: 134, lr: 3.74e-03, grad_scale: 8.0 2023-04-30 18:37:57,084 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.803e+02 2.710e+02 3.198e+02 3.720e+02 8.207e+02, threshold=6.397e+02, percent-clipped=2.0 2023-04-30 18:38:07,591 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=180561.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:38:08,927 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180562.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 18:38:21,367 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=180570.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:38:36,572 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180580.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:39:10,314 INFO [train.py:904] (4/8) Epoch 18, batch 8050, loss[loss=0.2431, simple_loss=0.3112, pruned_loss=0.08754, over 12204.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2941, pruned_loss=0.06239, over 3079765.92 frames. ], batch size: 248, lr: 3.74e-03, grad_scale: 4.0 2023-04-30 18:39:11,809 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 18:39:23,082 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=180610.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 18:39:50,349 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=180628.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:39:54,584 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=180631.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:40:09,066 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7951, 2.3425, 1.7619, 2.1646, 2.6908, 2.3978, 2.6949, 2.8897], device='cuda:4'), covar=tensor([0.0193, 0.0392, 0.0569, 0.0421, 0.0240, 0.0334, 0.0207, 0.0214], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0222, 0.0217, 0.0217, 0.0225, 0.0222, 0.0225, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:40:26,588 INFO [train.py:904] (4/8) Epoch 18, batch 8100, loss[loss=0.2068, simple_loss=0.29, pruned_loss=0.06175, over 16891.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2938, pruned_loss=0.06175, over 3087779.99 frames. ], batch size: 109, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:40:32,030 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.097e+02 2.839e+02 3.233e+02 3.954e+02 8.532e+02, threshold=6.466e+02, percent-clipped=3.0 2023-04-30 18:40:41,361 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5592, 4.5628, 4.4198, 3.6860, 4.4588, 1.6946, 4.2226, 4.1342], device='cuda:4'), covar=tensor([0.0090, 0.0082, 0.0175, 0.0336, 0.0093, 0.2678, 0.0132, 0.0225], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0139, 0.0185, 0.0170, 0.0160, 0.0196, 0.0174, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:41:00,841 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4152, 3.1783, 2.6652, 2.1422, 2.1992, 2.2659, 3.3499, 2.9805], device='cuda:4'), covar=tensor([0.3055, 0.0961, 0.1895, 0.2767, 0.2688, 0.2109, 0.0579, 0.1550], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0265, 0.0301, 0.0306, 0.0294, 0.0250, 0.0288, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:41:11,990 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180682.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:41:26,889 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=180692.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:41:41,226 INFO [train.py:904] (4/8) Epoch 18, batch 8150, loss[loss=0.2002, simple_loss=0.2893, pruned_loss=0.05554, over 16866.00 frames. ], tot_loss[loss=0.2072, simple_loss=0.2921, pruned_loss=0.06111, over 3086317.77 frames. ], batch size: 116, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:42:03,626 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.18 vs. limit=5.0 2023-04-30 18:42:24,165 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=180730.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:42:25,730 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.22 vs. limit=5.0 2023-04-30 18:42:26,686 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=180732.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:42:56,046 INFO [train.py:904] (4/8) Epoch 18, batch 8200, loss[loss=0.2276, simple_loss=0.3228, pruned_loss=0.06622, over 15405.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2896, pruned_loss=0.0605, over 3099881.20 frames. ], batch size: 190, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:43:02,072 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.971e+02 2.834e+02 3.334e+02 4.232e+02 1.685e+03, threshold=6.669e+02, percent-clipped=1.0 2023-04-30 18:43:12,599 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7567, 1.8521, 2.3883, 2.7342, 2.6494, 3.0729, 2.1555, 3.0711], device='cuda:4'), covar=tensor([0.0186, 0.0457, 0.0303, 0.0268, 0.0291, 0.0166, 0.0435, 0.0138], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0187, 0.0173, 0.0175, 0.0184, 0.0143, 0.0188, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:43:51,539 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5808, 4.6468, 4.4596, 4.1412, 4.1340, 4.5490, 4.3195, 4.2251], device='cuda:4'), covar=tensor([0.0586, 0.0457, 0.0307, 0.0316, 0.0941, 0.0417, 0.0465, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0275, 0.0392, 0.0320, 0.0310, 0.0332, 0.0362, 0.0220, 0.0382], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:44:01,860 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=180793.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:44:03,784 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8907, 4.9587, 4.7722, 4.3920, 4.3956, 4.8434, 4.7239, 4.5074], device='cuda:4'), covar=tensor([0.0576, 0.0386, 0.0298, 0.0322, 0.1056, 0.0407, 0.0333, 0.0695], device='cuda:4'), in_proj_covar=tensor([0.0275, 0.0392, 0.0320, 0.0310, 0.0332, 0.0362, 0.0220, 0.0382], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:44:09,984 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.32 vs. limit=5.0 2023-04-30 18:44:14,882 INFO [train.py:904] (4/8) Epoch 18, batch 8250, loss[loss=0.1925, simple_loss=0.2948, pruned_loss=0.04505, over 15305.00 frames. ], tot_loss[loss=0.2019, simple_loss=0.288, pruned_loss=0.0579, over 3075960.81 frames. ], batch size: 191, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:44:44,068 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9533, 3.2161, 3.6072, 2.0108, 3.0138, 2.2446, 3.4978, 3.3742], device='cuda:4'), covar=tensor([0.0263, 0.0826, 0.0462, 0.2041, 0.0754, 0.1013, 0.0619, 0.0966], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0158, 0.0163, 0.0149, 0.0141, 0.0127, 0.0140, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 18:44:53,333 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=180825.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:45:23,727 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5438, 3.5295, 3.4922, 2.6330, 3.3410, 1.9636, 3.1485, 2.8251], device='cuda:4'), covar=tensor([0.0144, 0.0126, 0.0152, 0.0205, 0.0109, 0.2384, 0.0133, 0.0209], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0139, 0.0184, 0.0169, 0.0159, 0.0195, 0.0173, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:45:37,338 INFO [train.py:904] (4/8) Epoch 18, batch 8300, loss[loss=0.194, simple_loss=0.2866, pruned_loss=0.05073, over 16951.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2851, pruned_loss=0.0549, over 3063734.85 frames. ], batch size: 116, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:45:43,843 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.692e+02 2.270e+02 2.749e+02 3.299e+02 7.574e+02, threshold=5.499e+02, percent-clipped=1.0 2023-04-30 18:45:52,480 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180861.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:46:32,570 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=180886.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:46:58,357 INFO [train.py:904] (4/8) Epoch 18, batch 8350, loss[loss=0.1942, simple_loss=0.2908, pruned_loss=0.04881, over 16213.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.284, pruned_loss=0.05271, over 3065838.21 frames. ], batch size: 165, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:47:09,894 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=180909.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:47:35,929 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=180926.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:47:39,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5036, 4.7135, 4.4608, 4.2187, 3.9598, 4.6306, 4.3510, 4.2234], device='cuda:4'), covar=tensor([0.0794, 0.0632, 0.0391, 0.0363, 0.1307, 0.0548, 0.0506, 0.0751], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0390, 0.0318, 0.0307, 0.0328, 0.0360, 0.0218, 0.0380], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:48:16,300 INFO [train.py:904] (4/8) Epoch 18, batch 8400, loss[loss=0.1635, simple_loss=0.2598, pruned_loss=0.03364, over 16208.00 frames. ], tot_loss[loss=0.191, simple_loss=0.281, pruned_loss=0.05053, over 3063075.19 frames. ], batch size: 165, lr: 3.73e-03, grad_scale: 8.0 2023-04-30 18:48:22,150 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.789e+02 2.220e+02 2.634e+02 3.245e+02 6.969e+02, threshold=5.268e+02, percent-clipped=3.0 2023-04-30 18:49:16,573 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=180992.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:49:31,280 INFO [train.py:904] (4/8) Epoch 18, batch 8450, loss[loss=0.1724, simple_loss=0.256, pruned_loss=0.04444, over 12296.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.279, pruned_loss=0.04918, over 3045961.22 frames. ], batch size: 248, lr: 3.73e-03, grad_scale: 8.0 2023-04-30 18:50:31,762 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=181040.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:50:50,496 INFO [train.py:904] (4/8) Epoch 18, batch 8500, loss[loss=0.1647, simple_loss=0.2437, pruned_loss=0.04292, over 11734.00 frames. ], tot_loss[loss=0.1845, simple_loss=0.2754, pruned_loss=0.04687, over 3047102.93 frames. ], batch size: 246, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:50:58,701 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.304e+02 3.083e+02 3.822e+02 7.945e+02, threshold=6.166e+02, percent-clipped=7.0 2023-04-30 18:51:48,654 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=181088.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:52:13,297 INFO [train.py:904] (4/8) Epoch 18, batch 8550, loss[loss=0.1709, simple_loss=0.2691, pruned_loss=0.03637, over 16738.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2732, pruned_loss=0.04606, over 3027295.35 frames. ], batch size: 76, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:53:50,720 INFO [train.py:904] (4/8) Epoch 18, batch 8600, loss[loss=0.1781, simple_loss=0.2763, pruned_loss=0.04001, over 15436.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2739, pruned_loss=0.04485, over 3043714.33 frames. ], batch size: 191, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:54:01,221 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.718e+02 2.216e+02 2.624e+02 3.286e+02 8.410e+02, threshold=5.248e+02, percent-clipped=1.0 2023-04-30 18:54:49,991 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=181181.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:55:07,079 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4630, 3.4665, 2.7549, 2.1153, 2.1732, 2.2373, 3.5981, 3.0740], device='cuda:4'), covar=tensor([0.3077, 0.0709, 0.1796, 0.3054, 0.2917, 0.2228, 0.0469, 0.1405], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0258, 0.0293, 0.0298, 0.0285, 0.0244, 0.0281, 0.0319], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:55:24,936 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4032, 3.0389, 2.6831, 2.2629, 2.1339, 2.2558, 3.0018, 2.8429], device='cuda:4'), covar=tensor([0.2534, 0.0735, 0.1598, 0.2766, 0.2599, 0.2117, 0.0469, 0.1393], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0258, 0.0293, 0.0298, 0.0285, 0.0243, 0.0280, 0.0319], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 18:55:29,852 INFO [train.py:904] (4/8) Epoch 18, batch 8650, loss[loss=0.1544, simple_loss=0.2471, pruned_loss=0.03084, over 12222.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.272, pruned_loss=0.04327, over 3055278.20 frames. ], batch size: 248, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:55:59,189 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0498, 4.0463, 3.9822, 3.2021, 3.9629, 1.6585, 3.7768, 3.6815], device='cuda:4'), covar=tensor([0.0108, 0.0108, 0.0149, 0.0257, 0.0105, 0.2844, 0.0137, 0.0244], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0138, 0.0182, 0.0166, 0.0158, 0.0195, 0.0172, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 18:56:25,675 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=181226.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:57:15,664 INFO [train.py:904] (4/8) Epoch 18, batch 8700, loss[loss=0.1757, simple_loss=0.2645, pruned_loss=0.04346, over 12477.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2696, pruned_loss=0.04228, over 3058238.53 frames. ], batch size: 247, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:57:25,076 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.665e+02 2.245e+02 2.629e+02 3.109e+02 5.522e+02, threshold=5.258e+02, percent-clipped=1.0 2023-04-30 18:57:56,765 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=181274.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:58:39,687 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=181297.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 18:58:48,988 INFO [train.py:904] (4/8) Epoch 18, batch 8750, loss[loss=0.1935, simple_loss=0.2922, pruned_loss=0.0474, over 16731.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.269, pruned_loss=0.04167, over 3042936.08 frames. ], batch size: 134, lr: 3.73e-03, grad_scale: 4.0 2023-04-30 18:59:49,994 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-04-30 19:00:11,042 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4648, 4.2607, 4.5366, 4.6493, 4.8005, 4.3367, 4.8110, 4.7990], device='cuda:4'), covar=tensor([0.1851, 0.1326, 0.1461, 0.0700, 0.0491, 0.1004, 0.0575, 0.0573], device='cuda:4'), in_proj_covar=tensor([0.0577, 0.0712, 0.0833, 0.0729, 0.0542, 0.0575, 0.0590, 0.0683], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:00:41,584 INFO [train.py:904] (4/8) Epoch 18, batch 8800, loss[loss=0.1696, simple_loss=0.2675, pruned_loss=0.03588, over 16249.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.268, pruned_loss=0.04073, over 3065381.99 frames. ], batch size: 165, lr: 3.73e-03, grad_scale: 8.0 2023-04-30 19:00:51,215 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.344e+02 2.185e+02 2.552e+02 3.082e+02 8.545e+02, threshold=5.104e+02, percent-clipped=2.0 2023-04-30 19:00:54,106 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=181358.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:01:02,318 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.98 vs. limit=5.0 2023-04-30 19:01:11,734 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.23 vs. limit=5.0 2023-04-30 19:01:57,983 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=181388.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:02:18,022 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8018, 3.8286, 4.0011, 3.8209, 3.9332, 4.3302, 4.0032, 3.7771], device='cuda:4'), covar=tensor([0.1883, 0.2023, 0.1555, 0.2189, 0.2488, 0.1349, 0.1502, 0.2279], device='cuda:4'), in_proj_covar=tensor([0.0370, 0.0536, 0.0594, 0.0450, 0.0599, 0.0621, 0.0471, 0.0603], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 19:02:27,190 INFO [train.py:904] (4/8) Epoch 18, batch 8850, loss[loss=0.189, simple_loss=0.2893, pruned_loss=0.04434, over 16596.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2704, pruned_loss=0.04028, over 3056533.74 frames. ], batch size: 75, lr: 3.73e-03, grad_scale: 8.0 2023-04-30 19:03:44,879 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=181436.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:04:15,168 INFO [train.py:904] (4/8) Epoch 18, batch 8900, loss[loss=0.1742, simple_loss=0.2686, pruned_loss=0.03995, over 16349.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2706, pruned_loss=0.03982, over 3048388.05 frames. ], batch size: 166, lr: 3.73e-03, grad_scale: 8.0 2023-04-30 19:04:25,753 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 2.184e+02 2.557e+02 3.261e+02 8.427e+02, threshold=5.113e+02, percent-clipped=4.0 2023-04-30 19:05:23,234 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=181481.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:06:10,711 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4716, 1.9706, 1.4581, 1.5960, 2.2593, 1.9684, 2.1501, 2.4699], device='cuda:4'), covar=tensor([0.0202, 0.0607, 0.0755, 0.0668, 0.0395, 0.0487, 0.0242, 0.0336], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0221, 0.0215, 0.0215, 0.0222, 0.0220, 0.0222, 0.0214], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:06:18,289 INFO [train.py:904] (4/8) Epoch 18, batch 8950, loss[loss=0.1445, simple_loss=0.2447, pruned_loss=0.02211, over 16906.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.27, pruned_loss=0.04022, over 3064333.70 frames. ], batch size: 90, lr: 3.73e-03, grad_scale: 8.0 2023-04-30 19:07:11,530 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8898, 4.8803, 4.6605, 3.5224, 4.7160, 1.7138, 4.3591, 4.4576], device='cuda:4'), covar=tensor([0.0109, 0.0098, 0.0221, 0.0612, 0.0119, 0.3257, 0.0199, 0.0313], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0136, 0.0178, 0.0163, 0.0156, 0.0192, 0.0170, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:07:17,050 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=181529.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:07:19,768 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=181530.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:08:08,279 INFO [train.py:904] (4/8) Epoch 18, batch 9000, loss[loss=0.1623, simple_loss=0.2593, pruned_loss=0.03262, over 15421.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2668, pruned_loss=0.03854, over 3063308.91 frames. ], batch size: 192, lr: 3.73e-03, grad_scale: 8.0 2023-04-30 19:08:08,280 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 19:08:17,832 INFO [train.py:938] (4/8) Epoch 18, validation: loss=0.1475, simple_loss=0.2516, pruned_loss=0.02169, over 944034.00 frames. 2023-04-30 19:08:17,833 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 19:08:18,821 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1113, 2.0958, 2.1992, 3.5226, 2.0764, 2.3590, 2.2207, 2.2072], device='cuda:4'), covar=tensor([0.1153, 0.3691, 0.2908, 0.0567, 0.4325, 0.2783, 0.3505, 0.3872], device='cuda:4'), in_proj_covar=tensor([0.0381, 0.0424, 0.0352, 0.0314, 0.0425, 0.0485, 0.0394, 0.0493], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:08:27,942 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.232e+02 2.270e+02 2.606e+02 3.112e+02 6.247e+02, threshold=5.212e+02, percent-clipped=1.0 2023-04-30 19:09:39,043 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=181591.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:10:01,194 INFO [train.py:904] (4/8) Epoch 18, batch 9050, loss[loss=0.1582, simple_loss=0.2483, pruned_loss=0.03406, over 15381.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2674, pruned_loss=0.03925, over 3054215.36 frames. ], batch size: 191, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:11:46,764 INFO [train.py:904] (4/8) Epoch 18, batch 9100, loss[loss=0.1738, simple_loss=0.274, pruned_loss=0.03685, over 16362.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2671, pruned_loss=0.03984, over 3066375.23 frames. ], batch size: 146, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:11:50,071 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=181653.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:11:55,850 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.522e+02 2.401e+02 2.871e+02 3.524e+02 6.480e+02, threshold=5.743e+02, percent-clipped=6.0 2023-04-30 19:13:34,500 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1573, 2.0673, 2.0716, 3.7184, 2.0788, 2.3761, 2.1892, 2.2264], device='cuda:4'), covar=tensor([0.1174, 0.3681, 0.3194, 0.0514, 0.4271, 0.2738, 0.3524, 0.3763], device='cuda:4'), in_proj_covar=tensor([0.0378, 0.0420, 0.0348, 0.0311, 0.0421, 0.0481, 0.0390, 0.0488], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:13:43,587 INFO [train.py:904] (4/8) Epoch 18, batch 9150, loss[loss=0.1795, simple_loss=0.2642, pruned_loss=0.04738, over 12012.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2675, pruned_loss=0.03922, over 3062232.30 frames. ], batch size: 250, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:14:10,423 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5260, 3.5244, 3.5196, 2.7687, 3.4121, 1.9518, 3.2231, 2.8959], device='cuda:4'), covar=tensor([0.0181, 0.0142, 0.0187, 0.0203, 0.0141, 0.2408, 0.0155, 0.0246], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0135, 0.0178, 0.0162, 0.0156, 0.0192, 0.0169, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:15:27,777 INFO [train.py:904] (4/8) Epoch 18, batch 9200, loss[loss=0.1501, simple_loss=0.2331, pruned_loss=0.03356, over 12285.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2632, pruned_loss=0.03825, over 3059398.32 frames. ], batch size: 248, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:15:36,945 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.469e+02 2.118e+02 2.449e+02 2.968e+02 5.097e+02, threshold=4.898e+02, percent-clipped=0.0 2023-04-30 19:15:57,040 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.29 vs. limit=5.0 2023-04-30 19:16:44,308 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4851, 3.4033, 3.5244, 3.6101, 3.6270, 3.3105, 3.5899, 3.6634], device='cuda:4'), covar=tensor([0.1160, 0.0859, 0.1022, 0.0575, 0.0616, 0.2718, 0.0912, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0566, 0.0697, 0.0816, 0.0715, 0.0534, 0.0565, 0.0582, 0.0670], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:17:05,940 INFO [train.py:904] (4/8) Epoch 18, batch 9250, loss[loss=0.1438, simple_loss=0.2465, pruned_loss=0.02059, over 16915.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.263, pruned_loss=0.03812, over 3074161.52 frames. ], batch size: 102, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:18:57,426 INFO [train.py:904] (4/8) Epoch 18, batch 9300, loss[loss=0.1596, simple_loss=0.2585, pruned_loss=0.0303, over 15342.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2614, pruned_loss=0.03759, over 3059225.66 frames. ], batch size: 191, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:19:05,578 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8837, 1.8882, 2.3471, 2.8493, 2.6923, 3.1770, 2.1141, 3.1531], device='cuda:4'), covar=tensor([0.0174, 0.0518, 0.0357, 0.0263, 0.0304, 0.0182, 0.0474, 0.0140], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0182, 0.0169, 0.0169, 0.0180, 0.0138, 0.0183, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:19:06,080 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.346e+02 2.677e+02 3.465e+02 6.012e+02, threshold=5.355e+02, percent-clipped=4.0 2023-04-30 19:19:21,015 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-30 19:20:12,908 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=181886.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:20:41,424 INFO [train.py:904] (4/8) Epoch 18, batch 9350, loss[loss=0.1896, simple_loss=0.266, pruned_loss=0.05655, over 12003.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2612, pruned_loss=0.0377, over 3053960.32 frames. ], batch size: 248, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:20:47,185 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1859, 1.5305, 1.8953, 2.1393, 2.2592, 2.3373, 1.7875, 2.2881], device='cuda:4'), covar=tensor([0.0196, 0.0465, 0.0284, 0.0303, 0.0287, 0.0204, 0.0458, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0183, 0.0170, 0.0170, 0.0181, 0.0139, 0.0185, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:22:24,031 INFO [train.py:904] (4/8) Epoch 18, batch 9400, loss[loss=0.1675, simple_loss=0.2734, pruned_loss=0.03073, over 16165.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2613, pruned_loss=0.03756, over 3042911.78 frames. ], batch size: 165, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:22:28,143 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=181953.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:22:33,105 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.435e+02 2.111e+02 2.455e+02 3.021e+02 7.291e+02, threshold=4.911e+02, percent-clipped=2.0 2023-04-30 19:22:36,231 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7922, 5.0813, 4.8723, 4.8267, 4.6411, 4.5666, 4.5341, 5.1558], device='cuda:4'), covar=tensor([0.1116, 0.0918, 0.0969, 0.0806, 0.0761, 0.0969, 0.1100, 0.0899], device='cuda:4'), in_proj_covar=tensor([0.0612, 0.0745, 0.0613, 0.0557, 0.0469, 0.0483, 0.0623, 0.0579], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:23:16,379 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=181978.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:23:59,454 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-30 19:24:03,759 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8076, 3.8425, 4.1524, 4.1117, 4.1230, 3.9085, 3.9114, 3.9260], device='cuda:4'), covar=tensor([0.0469, 0.1030, 0.0649, 0.0898, 0.0709, 0.0889, 0.1055, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0366, 0.0397, 0.0389, 0.0367, 0.0434, 0.0409, 0.0497, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 19:24:05,164 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=182001.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:24:06,042 INFO [train.py:904] (4/8) Epoch 18, batch 9450, loss[loss=0.1612, simple_loss=0.2609, pruned_loss=0.03077, over 16876.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2628, pruned_loss=0.03793, over 3017570.23 frames. ], batch size: 102, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:24:10,457 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.29 vs. limit=5.0 2023-04-30 19:24:28,873 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1026, 4.1028, 4.4755, 4.4263, 4.4183, 4.1839, 4.1517, 4.1364], device='cuda:4'), covar=tensor([0.0353, 0.0851, 0.0427, 0.0514, 0.0529, 0.0517, 0.0876, 0.0452], device='cuda:4'), in_proj_covar=tensor([0.0365, 0.0396, 0.0388, 0.0367, 0.0433, 0.0408, 0.0496, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 19:25:19,186 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=182039.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:25:41,958 INFO [train.py:904] (4/8) Epoch 18, batch 9500, loss[loss=0.1535, simple_loss=0.2441, pruned_loss=0.03143, over 12909.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2616, pruned_loss=0.03722, over 3039138.46 frames. ], batch size: 247, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:25:55,086 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.398e+02 2.182e+02 2.603e+02 3.134e+02 5.554e+02, threshold=5.207e+02, percent-clipped=1.0 2023-04-30 19:26:15,889 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.01 vs. limit=5.0 2023-04-30 19:27:27,077 INFO [train.py:904] (4/8) Epoch 18, batch 9550, loss[loss=0.164, simple_loss=0.2659, pruned_loss=0.03105, over 16859.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2616, pruned_loss=0.03758, over 3041648.51 frames. ], batch size: 96, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:28:08,125 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.15 vs. limit=5.0 2023-04-30 19:29:06,815 INFO [train.py:904] (4/8) Epoch 18, batch 9600, loss[loss=0.1688, simple_loss=0.2528, pruned_loss=0.04238, over 12614.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2633, pruned_loss=0.03868, over 3035538.65 frames. ], batch size: 250, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:29:15,380 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.236e+02 2.461e+02 2.915e+02 3.491e+02 6.374e+02, threshold=5.830e+02, percent-clipped=5.0 2023-04-30 19:30:14,198 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=182186.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:30:22,127 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-30 19:30:52,019 INFO [train.py:904] (4/8) Epoch 18, batch 9650, loss[loss=0.183, simple_loss=0.2791, pruned_loss=0.04345, over 15342.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2654, pruned_loss=0.0387, over 3053391.37 frames. ], batch size: 191, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:32:00,666 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=182234.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:32:37,723 INFO [train.py:904] (4/8) Epoch 18, batch 9700, loss[loss=0.1638, simple_loss=0.2653, pruned_loss=0.03116, over 16693.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2636, pruned_loss=0.03811, over 3050558.05 frames. ], batch size: 76, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:32:45,794 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.506e+02 2.275e+02 2.741e+02 3.246e+02 5.302e+02, threshold=5.483e+02, percent-clipped=0.0 2023-04-30 19:32:58,871 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-30 19:33:37,412 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=182281.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:33:44,440 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7871, 2.2723, 1.8682, 1.9604, 2.5420, 2.2652, 2.3661, 2.6639], device='cuda:4'), covar=tensor([0.0136, 0.0391, 0.0504, 0.0449, 0.0264, 0.0350, 0.0182, 0.0248], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0221, 0.0215, 0.0215, 0.0222, 0.0220, 0.0218, 0.0212], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:33:58,301 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9098, 2.7247, 2.8870, 2.1595, 2.7011, 2.1141, 2.6858, 2.8755], device='cuda:4'), covar=tensor([0.0265, 0.0885, 0.0490, 0.1772, 0.0746, 0.1008, 0.0610, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0151, 0.0159, 0.0146, 0.0138, 0.0124, 0.0137, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 19:34:17,889 INFO [train.py:904] (4/8) Epoch 18, batch 9750, loss[loss=0.1658, simple_loss=0.2611, pruned_loss=0.03523, over 16468.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.263, pruned_loss=0.03833, over 3055556.43 frames. ], batch size: 68, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:35:19,722 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5944, 1.8291, 2.2291, 2.5130, 2.5107, 2.8479, 1.9612, 2.8785], device='cuda:4'), covar=tensor([0.0205, 0.0473, 0.0334, 0.0310, 0.0317, 0.0210, 0.0474, 0.0143], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0180, 0.0167, 0.0168, 0.0178, 0.0137, 0.0183, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:35:24,325 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=182334.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:35:39,015 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=182342.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:35:56,047 INFO [train.py:904] (4/8) Epoch 18, batch 9800, loss[loss=0.1831, simple_loss=0.2799, pruned_loss=0.04319, over 16900.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2634, pruned_loss=0.03758, over 3074883.01 frames. ], batch size: 109, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:36:05,429 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.374e+02 1.961e+02 2.421e+02 2.847e+02 6.201e+02, threshold=4.843e+02, percent-clipped=1.0 2023-04-30 19:37:39,078 INFO [train.py:904] (4/8) Epoch 18, batch 9850, loss[loss=0.1621, simple_loss=0.2578, pruned_loss=0.03319, over 16653.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2643, pruned_loss=0.03723, over 3084565.68 frames. ], batch size: 134, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:38:34,413 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7116, 3.1329, 3.3660, 1.9256, 2.8312, 2.1721, 3.1628, 3.2522], device='cuda:4'), covar=tensor([0.0261, 0.0808, 0.0494, 0.2029, 0.0798, 0.1007, 0.0689, 0.0870], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0151, 0.0159, 0.0146, 0.0137, 0.0123, 0.0137, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 19:38:34,435 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4740, 3.5836, 2.8160, 2.0855, 2.3006, 2.2464, 3.7556, 3.2416], device='cuda:4'), covar=tensor([0.2892, 0.0595, 0.1647, 0.2956, 0.2836, 0.2177, 0.0342, 0.1272], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0252, 0.0287, 0.0291, 0.0274, 0.0239, 0.0275, 0.0310], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:39:05,017 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0408, 3.2246, 2.5974, 4.8665, 3.5816, 4.3244, 1.5696, 3.3932], device='cuda:4'), covar=tensor([0.1307, 0.0642, 0.1255, 0.0143, 0.0167, 0.0345, 0.1692, 0.0633], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0166, 0.0189, 0.0172, 0.0195, 0.0208, 0.0193, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 19:39:29,921 INFO [train.py:904] (4/8) Epoch 18, batch 9900, loss[loss=0.1703, simple_loss=0.273, pruned_loss=0.03384, over 16734.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2651, pruned_loss=0.03722, over 3074444.24 frames. ], batch size: 83, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:39:40,097 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0733, 2.7456, 2.4469, 4.9210, 3.3799, 4.2833, 1.7319, 3.1704], device='cuda:4'), covar=tensor([0.1287, 0.0859, 0.1356, 0.0128, 0.0195, 0.0375, 0.1600, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0167, 0.0189, 0.0173, 0.0195, 0.0209, 0.0193, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 19:39:40,681 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.546e+02 2.082e+02 2.451e+02 2.864e+02 7.377e+02, threshold=4.903e+02, percent-clipped=2.0 2023-04-30 19:39:42,100 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4913, 3.3810, 3.7201, 1.8374, 3.8844, 3.9692, 3.0085, 2.9604], device='cuda:4'), covar=tensor([0.0700, 0.0250, 0.0194, 0.1205, 0.0065, 0.0119, 0.0400, 0.0407], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0101, 0.0089, 0.0134, 0.0072, 0.0114, 0.0120, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 19:41:28,674 INFO [train.py:904] (4/8) Epoch 18, batch 9950, loss[loss=0.1697, simple_loss=0.2722, pruned_loss=0.0336, over 16184.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.267, pruned_loss=0.03776, over 3062189.27 frames. ], batch size: 165, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:41:35,620 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4316, 4.5483, 4.2985, 3.9520, 3.9810, 4.4220, 4.2470, 4.1417], device='cuda:4'), covar=tensor([0.0591, 0.0553, 0.0337, 0.0324, 0.0940, 0.0534, 0.0483, 0.0591], device='cuda:4'), in_proj_covar=tensor([0.0264, 0.0373, 0.0310, 0.0297, 0.0316, 0.0347, 0.0213, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-30 19:41:58,885 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=182514.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:43:31,283 INFO [train.py:904] (4/8) Epoch 18, batch 10000, loss[loss=0.1715, simple_loss=0.2572, pruned_loss=0.04289, over 13106.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2659, pruned_loss=0.03758, over 3065069.11 frames. ], batch size: 250, lr: 3.72e-03, grad_scale: 8.0 2023-04-30 19:43:42,214 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.446e+02 2.117e+02 2.346e+02 2.932e+02 5.510e+02, threshold=4.691e+02, percent-clipped=3.0 2023-04-30 19:44:16,783 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=182575.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:45:02,984 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.39 vs. limit=5.0 2023-04-30 19:45:14,018 INFO [train.py:904] (4/8) Epoch 18, batch 10050, loss[loss=0.1732, simple_loss=0.2693, pruned_loss=0.03856, over 16687.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2665, pruned_loss=0.03758, over 3068304.72 frames. ], batch size: 134, lr: 3.71e-03, grad_scale: 8.0 2023-04-30 19:46:15,017 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=182634.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:46:20,803 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=182637.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:46:24,042 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 19:46:46,716 INFO [train.py:904] (4/8) Epoch 18, batch 10100, loss[loss=0.1605, simple_loss=0.2516, pruned_loss=0.03469, over 16155.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2662, pruned_loss=0.03737, over 3063934.57 frames. ], batch size: 165, lr: 3.71e-03, grad_scale: 8.0 2023-04-30 19:46:55,747 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.437e+02 2.131e+02 2.554e+02 3.108e+02 6.507e+02, threshold=5.108e+02, percent-clipped=7.0 2023-04-30 19:47:44,811 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=182682.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:48:32,141 INFO [train.py:904] (4/8) Epoch 19, batch 0, loss[loss=0.1986, simple_loss=0.2759, pruned_loss=0.06068, over 16709.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2759, pruned_loss=0.06068, over 16709.00 frames. ], batch size: 89, lr: 3.61e-03, grad_scale: 8.0 2023-04-30 19:48:32,141 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 19:48:39,774 INFO [train.py:938] (4/8) Epoch 19, validation: loss=0.1468, simple_loss=0.2504, pruned_loss=0.0216, over 944034.00 frames. 2023-04-30 19:48:39,774 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 19:49:37,130 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8888, 5.2796, 4.9944, 5.0091, 4.7451, 4.7095, 4.7186, 5.3870], device='cuda:4'), covar=tensor([0.1310, 0.0970, 0.1252, 0.0923, 0.0810, 0.1042, 0.1293, 0.0935], device='cuda:4'), in_proj_covar=tensor([0.0612, 0.0750, 0.0615, 0.0557, 0.0471, 0.0483, 0.0626, 0.0580], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 19:49:50,013 INFO [train.py:904] (4/8) Epoch 19, batch 50, loss[loss=0.1805, simple_loss=0.2778, pruned_loss=0.04161, over 17254.00 frames. ], tot_loss[loss=0.1892, simple_loss=0.274, pruned_loss=0.05225, over 753653.85 frames. ], batch size: 52, lr: 3.61e-03, grad_scale: 1.0 2023-04-30 19:49:59,029 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.571e+02 3.042e+02 3.790e+02 6.505e+02, threshold=6.085e+02, percent-clipped=6.0 2023-04-30 19:50:55,406 INFO [train.py:904] (4/8) Epoch 19, batch 100, loss[loss=0.176, simple_loss=0.2718, pruned_loss=0.04008, over 17103.00 frames. ], tot_loss[loss=0.1857, simple_loss=0.2709, pruned_loss=0.05028, over 1319994.46 frames. ], batch size: 49, lr: 3.61e-03, grad_scale: 1.0 2023-04-30 19:52:03,139 INFO [train.py:904] (4/8) Epoch 19, batch 150, loss[loss=0.1635, simple_loss=0.261, pruned_loss=0.03303, over 17034.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2677, pruned_loss=0.04771, over 1772553.32 frames. ], batch size: 50, lr: 3.61e-03, grad_scale: 1.0 2023-04-30 19:52:10,226 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=182856.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:52:15,179 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.475e+02 2.297e+02 2.720e+02 3.223e+02 8.471e+02, threshold=5.441e+02, percent-clipped=1.0 2023-04-30 19:52:18,025 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=182862.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:52:28,157 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=182870.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:53:02,269 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3646, 3.2647, 2.6582, 5.2331, 4.4068, 4.4059, 1.9137, 3.3247], device='cuda:4'), covar=tensor([0.1061, 0.0621, 0.1164, 0.0211, 0.0215, 0.0504, 0.1361, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0168, 0.0189, 0.0175, 0.0196, 0.0210, 0.0194, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 19:53:13,450 INFO [train.py:904] (4/8) Epoch 19, batch 200, loss[loss=0.1632, simple_loss=0.2594, pruned_loss=0.03347, over 17238.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2666, pruned_loss=0.04775, over 2121112.31 frames. ], batch size: 45, lr: 3.61e-03, grad_scale: 1.0 2023-04-30 19:53:34,347 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=182917.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:53:34,784 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-30 19:53:42,636 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=182923.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:54:00,396 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=182937.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:54:21,114 INFO [train.py:904] (4/8) Epoch 19, batch 250, loss[loss=0.1676, simple_loss=0.246, pruned_loss=0.04458, over 16652.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2648, pruned_loss=0.04718, over 2383684.41 frames. ], batch size: 89, lr: 3.61e-03, grad_scale: 1.0 2023-04-30 19:54:32,987 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.641e+02 2.306e+02 2.752e+02 3.234e+02 1.372e+03, threshold=5.503e+02, percent-clipped=2.0 2023-04-30 19:55:08,834 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=182985.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:55:22,690 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-30 19:55:30,553 INFO [train.py:904] (4/8) Epoch 19, batch 300, loss[loss=0.1435, simple_loss=0.2286, pruned_loss=0.02919, over 16747.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2622, pruned_loss=0.04636, over 2583508.86 frames. ], batch size: 39, lr: 3.61e-03, grad_scale: 1.0 2023-04-30 19:55:31,644 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5122, 3.5484, 3.8164, 2.8199, 3.5926, 3.8644, 3.6364, 2.0138], device='cuda:4'), covar=tensor([0.0515, 0.0280, 0.0054, 0.0362, 0.0106, 0.0114, 0.0100, 0.0549], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0078, 0.0077, 0.0129, 0.0092, 0.0101, 0.0089, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 19:56:41,128 INFO [train.py:904] (4/8) Epoch 19, batch 350, loss[loss=0.1637, simple_loss=0.2616, pruned_loss=0.03295, over 17226.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2599, pruned_loss=0.045, over 2752434.85 frames. ], batch size: 45, lr: 3.61e-03, grad_scale: 1.0 2023-04-30 19:56:52,339 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.344e+02 2.193e+02 2.670e+02 3.333e+02 1.052e+03, threshold=5.340e+02, percent-clipped=4.0 2023-04-30 19:57:51,265 INFO [train.py:904] (4/8) Epoch 19, batch 400, loss[loss=0.1986, simple_loss=0.266, pruned_loss=0.06563, over 16792.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2583, pruned_loss=0.04408, over 2866504.21 frames. ], batch size: 124, lr: 3.61e-03, grad_scale: 2.0 2023-04-30 19:57:57,754 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3668, 3.6934, 3.8992, 2.1111, 3.1739, 2.6753, 3.8279, 3.8033], device='cuda:4'), covar=tensor([0.0282, 0.0864, 0.0550, 0.2091, 0.0855, 0.0945, 0.0735, 0.1039], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0154, 0.0161, 0.0148, 0.0139, 0.0125, 0.0139, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 19:58:18,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7498, 4.7870, 5.1820, 5.1783, 5.1635, 4.8701, 4.8037, 4.6483], device='cuda:4'), covar=tensor([0.0377, 0.0641, 0.0434, 0.0426, 0.0568, 0.0439, 0.0951, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0387, 0.0421, 0.0411, 0.0389, 0.0456, 0.0433, 0.0523, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 19:58:43,839 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7718, 3.8594, 2.4640, 4.4934, 3.0668, 4.4378, 2.5578, 3.1334], device='cuda:4'), covar=tensor([0.0304, 0.0430, 0.1613, 0.0259, 0.0860, 0.0547, 0.1558, 0.0817], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0172, 0.0191, 0.0152, 0.0173, 0.0209, 0.0199, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 19:59:03,169 INFO [train.py:904] (4/8) Epoch 19, batch 450, loss[loss=0.1604, simple_loss=0.2399, pruned_loss=0.04048, over 16888.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2575, pruned_loss=0.04362, over 2964371.05 frames. ], batch size: 90, lr: 3.61e-03, grad_scale: 2.0 2023-04-30 19:59:14,099 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.565e+02 2.162e+02 2.492e+02 3.004e+02 8.634e+02, threshold=4.984e+02, percent-clipped=1.0 2023-04-30 19:59:28,166 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=183170.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 19:59:47,906 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8557, 4.9334, 5.3339, 5.3131, 5.3325, 4.9987, 4.9445, 4.7290], device='cuda:4'), covar=tensor([0.0368, 0.0570, 0.0478, 0.0460, 0.0505, 0.0411, 0.0887, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0424, 0.0413, 0.0391, 0.0459, 0.0436, 0.0525, 0.0348], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 19:59:56,155 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-30 20:00:12,283 INFO [train.py:904] (4/8) Epoch 19, batch 500, loss[loss=0.1533, simple_loss=0.2316, pruned_loss=0.03749, over 16674.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2561, pruned_loss=0.0427, over 3047061.72 frames. ], batch size: 83, lr: 3.61e-03, grad_scale: 2.0 2023-04-30 20:00:28,366 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=183212.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:00:35,916 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=183218.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:00:35,929 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=183218.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:00:38,747 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.98 vs. limit=5.0 2023-04-30 20:00:51,562 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-30 20:01:05,151 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8494, 4.3578, 4.4670, 3.2464, 3.6547, 4.2902, 3.9610, 2.4402], device='cuda:4'), covar=tensor([0.0478, 0.0064, 0.0037, 0.0332, 0.0120, 0.0097, 0.0079, 0.0474], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0079, 0.0079, 0.0130, 0.0094, 0.0103, 0.0090, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 20:01:23,246 INFO [train.py:904] (4/8) Epoch 19, batch 550, loss[loss=0.1533, simple_loss=0.235, pruned_loss=0.03579, over 17006.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2548, pruned_loss=0.04243, over 3106978.91 frames. ], batch size: 41, lr: 3.61e-03, grad_scale: 2.0 2023-04-30 20:01:34,915 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.382e+02 2.167e+02 2.560e+02 2.885e+02 5.610e+02, threshold=5.120e+02, percent-clipped=1.0 2023-04-30 20:01:52,738 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9425, 4.2956, 4.4464, 3.2229, 3.6297, 4.3080, 3.9130, 2.4412], device='cuda:4'), covar=tensor([0.0450, 0.0063, 0.0036, 0.0338, 0.0131, 0.0093, 0.0085, 0.0486], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0080, 0.0079, 0.0131, 0.0094, 0.0103, 0.0091, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 20:02:32,735 INFO [train.py:904] (4/8) Epoch 19, batch 600, loss[loss=0.1739, simple_loss=0.2641, pruned_loss=0.04182, over 16700.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2542, pruned_loss=0.04252, over 3146402.25 frames. ], batch size: 57, lr: 3.61e-03, grad_scale: 2.0 2023-04-30 20:02:53,743 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4182, 4.4423, 4.8074, 4.7706, 4.7948, 4.4974, 4.4926, 4.3627], device='cuda:4'), covar=tensor([0.0423, 0.0691, 0.0444, 0.0463, 0.0529, 0.0476, 0.0860, 0.0624], device='cuda:4'), in_proj_covar=tensor([0.0391, 0.0427, 0.0416, 0.0392, 0.0463, 0.0439, 0.0527, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 20:03:42,736 INFO [train.py:904] (4/8) Epoch 19, batch 650, loss[loss=0.1484, simple_loss=0.2426, pruned_loss=0.02708, over 17234.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2534, pruned_loss=0.04231, over 3189665.68 frames. ], batch size: 45, lr: 3.61e-03, grad_scale: 2.0 2023-04-30 20:03:54,610 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.479e+02 2.241e+02 2.597e+02 3.196e+02 6.464e+02, threshold=5.194e+02, percent-clipped=2.0 2023-04-30 20:04:22,916 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9421, 4.0026, 2.3790, 4.6724, 3.2230, 4.5908, 2.5222, 3.4451], device='cuda:4'), covar=tensor([0.0275, 0.0379, 0.1671, 0.0276, 0.0752, 0.0483, 0.1493, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0174, 0.0193, 0.0154, 0.0174, 0.0211, 0.0200, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:04:53,074 INFO [train.py:904] (4/8) Epoch 19, batch 700, loss[loss=0.1873, simple_loss=0.2642, pruned_loss=0.05515, over 16853.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2537, pruned_loss=0.04235, over 3224173.36 frames. ], batch size: 116, lr: 3.61e-03, grad_scale: 2.0 2023-04-30 20:05:01,086 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=183408.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:05:09,311 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0842, 2.4788, 2.6369, 1.8932, 2.8024, 2.7996, 2.4473, 2.4460], device='cuda:4'), covar=tensor([0.0723, 0.0260, 0.0234, 0.0976, 0.0125, 0.0263, 0.0463, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0106, 0.0095, 0.0140, 0.0077, 0.0122, 0.0126, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 20:05:49,616 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.53 vs. limit=2.0 2023-04-30 20:05:59,956 INFO [train.py:904] (4/8) Epoch 19, batch 750, loss[loss=0.1742, simple_loss=0.2492, pruned_loss=0.0496, over 16790.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2544, pruned_loss=0.04267, over 3241503.09 frames. ], batch size: 83, lr: 3.61e-03, grad_scale: 2.0 2023-04-30 20:06:11,898 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.590e+02 2.150e+02 2.519e+02 3.057e+02 7.230e+02, threshold=5.039e+02, percent-clipped=2.0 2023-04-30 20:06:25,301 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=183469.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:07:10,059 INFO [train.py:904] (4/8) Epoch 19, batch 800, loss[loss=0.1684, simple_loss=0.2523, pruned_loss=0.04224, over 16395.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2541, pruned_loss=0.04255, over 3248776.85 frames. ], batch size: 75, lr: 3.61e-03, grad_scale: 4.0 2023-04-30 20:07:24,166 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=183512.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:07:32,537 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=183518.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:07:50,745 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-30 20:08:19,002 INFO [train.py:904] (4/8) Epoch 19, batch 850, loss[loss=0.1647, simple_loss=0.2597, pruned_loss=0.03483, over 16750.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2533, pruned_loss=0.04208, over 3266853.95 frames. ], batch size: 62, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:08:29,786 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.473e+02 2.149e+02 2.547e+02 3.022e+02 7.911e+02, threshold=5.094e+02, percent-clipped=1.0 2023-04-30 20:08:30,864 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=183560.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:08:38,499 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=183566.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:09:28,459 INFO [train.py:904] (4/8) Epoch 19, batch 900, loss[loss=0.1448, simple_loss=0.2259, pruned_loss=0.03184, over 16949.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2518, pruned_loss=0.04121, over 3276638.92 frames. ], batch size: 41, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:09:44,202 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=183614.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:09:55,838 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-30 20:10:35,713 INFO [train.py:904] (4/8) Epoch 19, batch 950, loss[loss=0.1637, simple_loss=0.2545, pruned_loss=0.03648, over 17077.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2517, pruned_loss=0.04099, over 3290555.69 frames. ], batch size: 55, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:10:45,891 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.532e+02 1.990e+02 2.288e+02 2.656e+02 5.552e+02, threshold=4.575e+02, percent-clipped=1.0 2023-04-30 20:11:08,228 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=183675.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:11:42,836 INFO [train.py:904] (4/8) Epoch 19, batch 1000, loss[loss=0.1848, simple_loss=0.2537, pruned_loss=0.05796, over 16899.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2501, pruned_loss=0.04113, over 3276742.47 frames. ], batch size: 116, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:12:26,024 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7349, 2.6253, 2.1318, 2.3516, 2.9973, 2.6489, 3.5000, 3.2704], device='cuda:4'), covar=tensor([0.0155, 0.0489, 0.0646, 0.0554, 0.0336, 0.0492, 0.0270, 0.0295], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0235, 0.0226, 0.0227, 0.0235, 0.0234, 0.0236, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:12:50,597 INFO [train.py:904] (4/8) Epoch 19, batch 1050, loss[loss=0.138, simple_loss=0.2345, pruned_loss=0.02074, over 17122.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2504, pruned_loss=0.04167, over 3286621.82 frames. ], batch size: 47, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:13:01,769 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.633e+02 2.214e+02 2.595e+02 3.147e+02 6.979e+02, threshold=5.191e+02, percent-clipped=4.0 2023-04-30 20:13:06,522 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=183764.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:14:00,539 INFO [train.py:904] (4/8) Epoch 19, batch 1100, loss[loss=0.1679, simple_loss=0.2588, pruned_loss=0.03852, over 17047.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2502, pruned_loss=0.04164, over 3297331.73 frames. ], batch size: 53, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:14:03,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3252, 5.2952, 5.1224, 4.5735, 5.1329, 2.1027, 4.9216, 5.1552], device='cuda:4'), covar=tensor([0.0082, 0.0080, 0.0186, 0.0406, 0.0099, 0.2608, 0.0145, 0.0178], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0146, 0.0190, 0.0174, 0.0168, 0.0204, 0.0182, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:14:47,510 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2288, 3.9460, 4.4180, 2.2173, 4.5787, 4.6975, 3.4541, 3.6022], device='cuda:4'), covar=tensor([0.0670, 0.0255, 0.0245, 0.1203, 0.0071, 0.0156, 0.0392, 0.0380], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0107, 0.0096, 0.0140, 0.0077, 0.0123, 0.0126, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 20:14:51,542 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9932, 2.0377, 2.5319, 2.8772, 2.6528, 3.4409, 2.2009, 3.3918], device='cuda:4'), covar=tensor([0.0251, 0.0501, 0.0342, 0.0323, 0.0354, 0.0173, 0.0518, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0190, 0.0178, 0.0179, 0.0190, 0.0147, 0.0193, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:14:55,839 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=183842.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:15:08,304 INFO [train.py:904] (4/8) Epoch 19, batch 1150, loss[loss=0.1581, simple_loss=0.2492, pruned_loss=0.03355, over 16608.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2502, pruned_loss=0.04106, over 3304856.63 frames. ], batch size: 57, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:15:20,304 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.320e+02 2.057e+02 2.490e+02 3.144e+02 5.373e+02, threshold=4.980e+02, percent-clipped=1.0 2023-04-30 20:15:26,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9948, 2.0967, 2.5353, 2.8730, 2.7581, 3.2053, 2.1213, 3.2039], device='cuda:4'), covar=tensor([0.0194, 0.0458, 0.0315, 0.0299, 0.0305, 0.0201, 0.0517, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0190, 0.0177, 0.0178, 0.0189, 0.0146, 0.0192, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:16:18,545 INFO [train.py:904] (4/8) Epoch 19, batch 1200, loss[loss=0.1422, simple_loss=0.231, pruned_loss=0.02672, over 16729.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2491, pruned_loss=0.04065, over 3299234.20 frames. ], batch size: 39, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:16:20,033 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=183903.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:16:43,908 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1676, 3.2227, 3.3564, 2.2123, 2.9086, 2.4489, 3.6537, 3.6119], device='cuda:4'), covar=tensor([0.0233, 0.0872, 0.0593, 0.1788, 0.0827, 0.0963, 0.0479, 0.0848], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0159, 0.0165, 0.0150, 0.0142, 0.0127, 0.0143, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:17:25,271 INFO [train.py:904] (4/8) Epoch 19, batch 1250, loss[loss=0.1674, simple_loss=0.2468, pruned_loss=0.04401, over 12333.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2489, pruned_loss=0.04079, over 3306033.41 frames. ], batch size: 247, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:17:35,896 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.556e+02 2.248e+02 2.515e+02 3.055e+02 4.782e+02, threshold=5.031e+02, percent-clipped=0.0 2023-04-30 20:17:38,217 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3733, 5.3485, 5.2069, 4.6828, 4.7781, 5.2384, 5.2486, 4.8643], device='cuda:4'), covar=tensor([0.0563, 0.0550, 0.0308, 0.0370, 0.1128, 0.0483, 0.0308, 0.0747], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0415, 0.0343, 0.0331, 0.0351, 0.0386, 0.0234, 0.0405], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:17:47,526 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6637, 6.0119, 5.7697, 5.8220, 5.4176, 5.4039, 5.4118, 6.1142], device='cuda:4'), covar=tensor([0.1293, 0.0817, 0.1064, 0.0887, 0.0979, 0.0679, 0.1168, 0.0929], device='cuda:4'), in_proj_covar=tensor([0.0662, 0.0812, 0.0667, 0.0604, 0.0508, 0.0519, 0.0676, 0.0631], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:17:50,866 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=183970.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:18:37,489 INFO [train.py:904] (4/8) Epoch 19, batch 1300, loss[loss=0.1746, simple_loss=0.2665, pruned_loss=0.04134, over 17055.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2489, pruned_loss=0.04086, over 3303798.44 frames. ], batch size: 55, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:18:39,176 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2697, 3.6078, 3.9960, 1.9771, 3.1010, 2.5597, 3.6764, 3.7284], device='cuda:4'), covar=tensor([0.0305, 0.0894, 0.0459, 0.2138, 0.0842, 0.0971, 0.0717, 0.1136], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0158, 0.0164, 0.0150, 0.0142, 0.0127, 0.0142, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:19:44,263 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=184050.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:19:46,323 INFO [train.py:904] (4/8) Epoch 19, batch 1350, loss[loss=0.1421, simple_loss=0.2327, pruned_loss=0.02574, over 17211.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2487, pruned_loss=0.04018, over 3307845.90 frames. ], batch size: 45, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:19:55,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.519e+02 2.201e+02 2.541e+02 3.098e+02 6.218e+02, threshold=5.081e+02, percent-clipped=6.0 2023-04-30 20:20:03,864 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=184064.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:20:55,153 INFO [train.py:904] (4/8) Epoch 19, batch 1400, loss[loss=0.1464, simple_loss=0.2341, pruned_loss=0.02935, over 16531.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2491, pruned_loss=0.04043, over 3312976.70 frames. ], batch size: 68, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:21:09,615 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=184111.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:21:10,485 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=184112.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:22:05,080 INFO [train.py:904] (4/8) Epoch 19, batch 1450, loss[loss=0.1928, simple_loss=0.2911, pruned_loss=0.04723, over 16702.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2484, pruned_loss=0.04051, over 3314545.42 frames. ], batch size: 57, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:22:15,458 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.550e+02 2.293e+02 2.595e+02 3.219e+02 5.738e+02, threshold=5.191e+02, percent-clipped=1.0 2023-04-30 20:22:19,243 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.07 vs. limit=5.0 2023-04-30 20:22:25,816 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5289, 3.8415, 4.0446, 2.2099, 3.2661, 2.5797, 4.1107, 4.0539], device='cuda:4'), covar=tensor([0.0239, 0.0821, 0.0473, 0.1980, 0.0770, 0.0994, 0.0536, 0.0931], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0159, 0.0164, 0.0150, 0.0142, 0.0127, 0.0142, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:23:08,678 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=184198.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:23:13,289 INFO [train.py:904] (4/8) Epoch 19, batch 1500, loss[loss=0.1714, simple_loss=0.2656, pruned_loss=0.03866, over 17053.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2494, pruned_loss=0.04139, over 3323595.79 frames. ], batch size: 50, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:24:21,782 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2787, 3.5661, 3.5894, 2.1879, 3.0848, 2.6293, 3.7491, 3.8414], device='cuda:4'), covar=tensor([0.0263, 0.0824, 0.0588, 0.1879, 0.0803, 0.0933, 0.0534, 0.0826], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0159, 0.0164, 0.0150, 0.0142, 0.0127, 0.0142, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:24:22,396 INFO [train.py:904] (4/8) Epoch 19, batch 1550, loss[loss=0.1883, simple_loss=0.2596, pruned_loss=0.05851, over 16490.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2512, pruned_loss=0.04272, over 3318334.85 frames. ], batch size: 75, lr: 3.60e-03, grad_scale: 4.0 2023-04-30 20:24:34,823 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.543e+02 2.353e+02 2.741e+02 3.128e+02 4.649e+02, threshold=5.482e+02, percent-clipped=0.0 2023-04-30 20:24:47,514 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=184270.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:25:31,172 INFO [train.py:904] (4/8) Epoch 19, batch 1600, loss[loss=0.183, simple_loss=0.254, pruned_loss=0.05598, over 16807.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2531, pruned_loss=0.04326, over 3317171.95 frames. ], batch size: 124, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:25:53,347 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=184318.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:26:39,948 INFO [train.py:904] (4/8) Epoch 19, batch 1650, loss[loss=0.1926, simple_loss=0.2635, pruned_loss=0.06088, over 16875.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2533, pruned_loss=0.04288, over 3324955.70 frames. ], batch size: 116, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:26:50,251 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=184359.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:26:52,312 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.549e+02 2.220e+02 2.674e+02 3.187e+02 5.456e+02, threshold=5.348e+02, percent-clipped=0.0 2023-04-30 20:27:49,674 INFO [train.py:904] (4/8) Epoch 19, batch 1700, loss[loss=0.2223, simple_loss=0.3033, pruned_loss=0.07064, over 16539.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2555, pruned_loss=0.04366, over 3317030.93 frames. ], batch size: 68, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:27:55,053 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=184406.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:28:14,027 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=184420.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:28:58,451 INFO [train.py:904] (4/8) Epoch 19, batch 1750, loss[loss=0.1679, simple_loss=0.2631, pruned_loss=0.03634, over 16717.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2561, pruned_loss=0.04304, over 3318455.31 frames. ], batch size: 62, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:29:10,989 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.495e+02 2.342e+02 2.772e+02 3.206e+02 7.183e+02, threshold=5.544e+02, percent-clipped=2.0 2023-04-30 20:29:45,756 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=184486.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:30:02,694 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=184498.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:30:07,526 INFO [train.py:904] (4/8) Epoch 19, batch 1800, loss[loss=0.1619, simple_loss=0.2625, pruned_loss=0.03066, over 17136.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2569, pruned_loss=0.0428, over 3316925.17 frames. ], batch size: 49, lr: 3.60e-03, grad_scale: 8.0 2023-04-30 20:31:07,262 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=184546.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:31:09,284 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=184547.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:31:16,691 INFO [train.py:904] (4/8) Epoch 19, batch 1850, loss[loss=0.1731, simple_loss=0.2686, pruned_loss=0.03878, over 16724.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2582, pruned_loss=0.04315, over 3319042.69 frames. ], batch size: 62, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:31:29,852 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.319e+02 2.167e+02 2.563e+02 3.002e+02 5.511e+02, threshold=5.126e+02, percent-clipped=0.0 2023-04-30 20:32:24,059 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-30 20:32:26,186 INFO [train.py:904] (4/8) Epoch 19, batch 1900, loss[loss=0.1374, simple_loss=0.2203, pruned_loss=0.02726, over 16975.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2572, pruned_loss=0.04228, over 3320280.40 frames. ], batch size: 41, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:32:26,585 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9044, 4.6569, 4.9340, 5.1287, 5.3418, 4.6567, 5.3631, 5.3175], device='cuda:4'), covar=tensor([0.1712, 0.1416, 0.1822, 0.0815, 0.0528, 0.0932, 0.0503, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0638, 0.0786, 0.0920, 0.0800, 0.0596, 0.0628, 0.0645, 0.0749], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:32:27,819 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4567, 4.2849, 4.5186, 4.6867, 4.8208, 4.3012, 4.6991, 4.7941], device='cuda:4'), covar=tensor([0.1612, 0.1418, 0.1589, 0.0859, 0.0671, 0.1163, 0.1531, 0.0938], device='cuda:4'), in_proj_covar=tensor([0.0638, 0.0786, 0.0920, 0.0800, 0.0596, 0.0628, 0.0645, 0.0749], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:33:20,939 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0920, 3.9871, 4.3517, 2.1837, 4.5413, 4.6660, 3.3650, 3.5269], device='cuda:4'), covar=tensor([0.0742, 0.0237, 0.0247, 0.1178, 0.0090, 0.0180, 0.0425, 0.0412], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0108, 0.0097, 0.0141, 0.0079, 0.0125, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:33:28,921 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-30 20:33:35,921 INFO [train.py:904] (4/8) Epoch 19, batch 1950, loss[loss=0.1628, simple_loss=0.2593, pruned_loss=0.03318, over 17188.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2572, pruned_loss=0.04161, over 3328110.36 frames. ], batch size: 46, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:33:48,870 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.475e+02 2.129e+02 2.542e+02 2.951e+02 4.781e+02, threshold=5.085e+02, percent-clipped=0.0 2023-04-30 20:34:14,648 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2492, 3.4264, 3.7548, 2.2015, 3.1889, 2.5044, 3.7000, 3.7005], device='cuda:4'), covar=tensor([0.0278, 0.0962, 0.0540, 0.1977, 0.0805, 0.1000, 0.0593, 0.0983], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0161, 0.0165, 0.0151, 0.0143, 0.0128, 0.0143, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:34:47,042 INFO [train.py:904] (4/8) Epoch 19, batch 2000, loss[loss=0.1846, simple_loss=0.2537, pruned_loss=0.05772, over 16906.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2573, pruned_loss=0.04182, over 3327091.88 frames. ], batch size: 109, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:34:50,512 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-30 20:34:53,116 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=184706.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:35:06,418 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=184715.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:35:29,097 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1153, 2.4641, 2.6324, 1.9081, 2.7443, 2.7846, 2.4217, 2.4136], device='cuda:4'), covar=tensor([0.0680, 0.0253, 0.0223, 0.0910, 0.0112, 0.0268, 0.0460, 0.0407], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0108, 0.0096, 0.0141, 0.0078, 0.0125, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:35:56,926 INFO [train.py:904] (4/8) Epoch 19, batch 2050, loss[loss=0.1756, simple_loss=0.2703, pruned_loss=0.04052, over 17057.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2567, pruned_loss=0.04222, over 3314563.45 frames. ], batch size: 53, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:35:59,925 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=184754.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:36:10,010 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 2.173e+02 2.577e+02 2.933e+02 5.913e+02, threshold=5.153e+02, percent-clipped=1.0 2023-04-30 20:36:57,673 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6851, 4.5814, 4.5959, 4.3162, 4.3172, 4.6600, 4.4138, 4.3981], device='cuda:4'), covar=tensor([0.0633, 0.0741, 0.0290, 0.0262, 0.0801, 0.0460, 0.0500, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0423, 0.0347, 0.0337, 0.0359, 0.0394, 0.0237, 0.0414], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:37:02,196 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-30 20:37:10,295 INFO [train.py:904] (4/8) Epoch 19, batch 2100, loss[loss=0.1885, simple_loss=0.2686, pruned_loss=0.05418, over 16687.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2579, pruned_loss=0.04305, over 3302496.42 frames. ], batch size: 134, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:37:27,011 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-30 20:38:05,058 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=184842.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:38:18,977 INFO [train.py:904] (4/8) Epoch 19, batch 2150, loss[loss=0.1888, simple_loss=0.2656, pruned_loss=0.05601, over 16635.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2581, pruned_loss=0.04299, over 3316781.39 frames. ], batch size: 134, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:38:30,739 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.463e+02 2.154e+02 2.741e+02 3.179e+02 7.758e+02, threshold=5.482e+02, percent-clipped=2.0 2023-04-30 20:38:38,181 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=184866.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:39:07,257 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0888, 2.4579, 2.6135, 1.9153, 2.7699, 2.7314, 2.3339, 2.3883], device='cuda:4'), covar=tensor([0.0714, 0.0289, 0.0267, 0.0921, 0.0118, 0.0281, 0.0528, 0.0427], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0096, 0.0141, 0.0078, 0.0125, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:39:11,598 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-30 20:39:27,103 INFO [train.py:904] (4/8) Epoch 19, batch 2200, loss[loss=0.153, simple_loss=0.2356, pruned_loss=0.03517, over 16235.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2587, pruned_loss=0.04371, over 3317442.10 frames. ], batch size: 36, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:39:57,059 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-30 20:40:01,193 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=184927.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:40:17,865 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=184939.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:40:34,450 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-30 20:40:35,941 INFO [train.py:904] (4/8) Epoch 19, batch 2250, loss[loss=0.1537, simple_loss=0.2501, pruned_loss=0.02869, over 17099.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.26, pruned_loss=0.04419, over 3318568.72 frames. ], batch size: 47, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:40:40,818 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.55 vs. limit=5.0 2023-04-30 20:40:48,428 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.653e+02 2.330e+02 2.636e+02 3.046e+02 4.750e+02, threshold=5.271e+02, percent-clipped=0.0 2023-04-30 20:41:44,996 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=185000.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:41:46,800 INFO [train.py:904] (4/8) Epoch 19, batch 2300, loss[loss=0.183, simple_loss=0.2738, pruned_loss=0.0461, over 16760.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2595, pruned_loss=0.04389, over 3329572.02 frames. ], batch size: 62, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:42:05,840 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=185015.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:42:16,422 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0234, 3.4574, 2.8423, 5.2021, 4.3145, 4.5531, 1.9118, 3.3557], device='cuda:4'), covar=tensor([0.1280, 0.0623, 0.1102, 0.0154, 0.0234, 0.0401, 0.1520, 0.0711], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0172, 0.0192, 0.0183, 0.0204, 0.0215, 0.0197, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:42:40,340 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-30 20:42:56,743 INFO [train.py:904] (4/8) Epoch 19, batch 2350, loss[loss=0.1966, simple_loss=0.2658, pruned_loss=0.06377, over 16906.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2595, pruned_loss=0.04353, over 3326825.01 frames. ], batch size: 116, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:43:08,898 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=185060.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:43:09,638 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.650e+02 2.245e+02 2.599e+02 3.315e+02 5.469e+02, threshold=5.198e+02, percent-clipped=1.0 2023-04-30 20:43:12,099 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=185063.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:43:21,905 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5567, 3.6403, 3.3150, 2.9514, 3.1985, 3.5063, 3.3286, 3.3445], device='cuda:4'), covar=tensor([0.0617, 0.0559, 0.0279, 0.0247, 0.0476, 0.0451, 0.1134, 0.0508], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0425, 0.0350, 0.0340, 0.0359, 0.0395, 0.0239, 0.0415], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 20:43:39,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4736, 3.7167, 4.0602, 2.3207, 3.2380, 2.6489, 3.9769, 3.9226], device='cuda:4'), covar=tensor([0.0280, 0.0959, 0.0494, 0.1900, 0.0819, 0.0941, 0.0612, 0.0943], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0160, 0.0165, 0.0151, 0.0142, 0.0127, 0.0142, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:44:06,533 INFO [train.py:904] (4/8) Epoch 19, batch 2400, loss[loss=0.1902, simple_loss=0.2891, pruned_loss=0.04568, over 16730.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2605, pruned_loss=0.04418, over 3321524.45 frames. ], batch size: 62, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:44:33,443 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=185121.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 20:44:43,170 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=185129.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:45:02,156 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=185142.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:45:15,142 INFO [train.py:904] (4/8) Epoch 19, batch 2450, loss[loss=0.1889, simple_loss=0.2704, pruned_loss=0.0537, over 16693.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2607, pruned_loss=0.04365, over 3323337.80 frames. ], batch size: 134, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:45:27,049 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.487e+02 2.259e+02 2.760e+02 3.178e+02 5.977e+02, threshold=5.520e+02, percent-clipped=3.0 2023-04-30 20:46:01,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0612, 3.1187, 3.3518, 2.1511, 2.9391, 2.2371, 3.5357, 3.4428], device='cuda:4'), covar=tensor([0.0231, 0.0993, 0.0577, 0.1847, 0.0823, 0.1023, 0.0548, 0.0924], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0160, 0.0164, 0.0151, 0.0142, 0.0127, 0.0142, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:46:04,414 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7811, 1.9616, 2.3508, 2.7279, 2.7243, 2.8009, 1.9643, 2.9873], device='cuda:4'), covar=tensor([0.0163, 0.0423, 0.0310, 0.0242, 0.0268, 0.0241, 0.0477, 0.0135], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0192, 0.0179, 0.0181, 0.0191, 0.0150, 0.0195, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:46:08,234 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=185190.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:46:08,466 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=185190.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:46:23,967 INFO [train.py:904] (4/8) Epoch 19, batch 2500, loss[loss=0.1638, simple_loss=0.2484, pruned_loss=0.03963, over 16988.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2602, pruned_loss=0.04333, over 3324256.80 frames. ], batch size: 41, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:46:51,193 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=185222.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:47:14,070 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5889, 2.3242, 2.2780, 4.4223, 2.2572, 2.7299, 2.4460, 2.4436], device='cuda:4'), covar=tensor([0.1181, 0.3646, 0.3067, 0.0454, 0.4283, 0.2695, 0.3405, 0.3871], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0439, 0.0362, 0.0327, 0.0434, 0.0506, 0.0408, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:47:21,427 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8762, 3.7152, 4.2517, 1.9888, 4.4365, 4.5003, 3.2154, 3.3627], device='cuda:4'), covar=tensor([0.0719, 0.0266, 0.0196, 0.1241, 0.0078, 0.0167, 0.0435, 0.0415], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0097, 0.0141, 0.0078, 0.0125, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:47:32,254 INFO [train.py:904] (4/8) Epoch 19, batch 2550, loss[loss=0.1443, simple_loss=0.2285, pruned_loss=0.03009, over 16838.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2601, pruned_loss=0.04325, over 3323732.70 frames. ], batch size: 42, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:47:44,040 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.637e+02 2.145e+02 2.534e+02 2.988e+02 8.366e+02, threshold=5.067e+02, percent-clipped=1.0 2023-04-30 20:48:31,912 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=185295.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:48:40,161 INFO [train.py:904] (4/8) Epoch 19, batch 2600, loss[loss=0.1743, simple_loss=0.2745, pruned_loss=0.03698, over 17076.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2598, pruned_loss=0.04285, over 3325207.68 frames. ], batch size: 53, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:49:16,566 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8219, 1.9973, 2.5058, 2.8319, 2.7058, 3.3096, 2.1832, 3.2928], device='cuda:4'), covar=tensor([0.0247, 0.0457, 0.0311, 0.0323, 0.0299, 0.0177, 0.0490, 0.0140], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0193, 0.0179, 0.0182, 0.0192, 0.0151, 0.0195, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:49:21,431 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7404, 3.7663, 2.9080, 2.2612, 2.4945, 2.3924, 3.8979, 3.3909], device='cuda:4'), covar=tensor([0.2462, 0.0582, 0.1562, 0.2960, 0.2826, 0.1913, 0.0485, 0.1237], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0266, 0.0300, 0.0303, 0.0294, 0.0251, 0.0288, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 20:49:50,076 INFO [train.py:904] (4/8) Epoch 19, batch 2650, loss[loss=0.1966, simple_loss=0.2728, pruned_loss=0.06017, over 15554.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2607, pruned_loss=0.04247, over 3334571.84 frames. ], batch size: 191, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:50:03,486 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 2.113e+02 2.462e+02 3.036e+02 8.000e+02, threshold=4.924e+02, percent-clipped=5.0 2023-04-30 20:50:39,034 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5047, 3.3537, 3.7653, 1.8369, 3.9017, 3.9097, 3.0361, 2.9605], device='cuda:4'), covar=tensor([0.0793, 0.0250, 0.0172, 0.1268, 0.0084, 0.0175, 0.0433, 0.0410], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0097, 0.0141, 0.0079, 0.0126, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:50:59,019 INFO [train.py:904] (4/8) Epoch 19, batch 2700, loss[loss=0.1816, simple_loss=0.2686, pruned_loss=0.04727, over 16242.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2613, pruned_loss=0.04257, over 3334373.72 frames. ], batch size: 35, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:51:19,124 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=185416.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 20:52:08,646 INFO [train.py:904] (4/8) Epoch 19, batch 2750, loss[loss=0.1697, simple_loss=0.2655, pruned_loss=0.03692, over 17061.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2612, pruned_loss=0.04198, over 3334953.35 frames. ], batch size: 53, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:52:09,837 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8536, 3.9028, 2.9521, 2.3538, 2.6126, 2.5259, 4.1765, 3.4996], device='cuda:4'), covar=tensor([0.2531, 0.0644, 0.1753, 0.2850, 0.2729, 0.1957, 0.0457, 0.1344], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0267, 0.0301, 0.0305, 0.0294, 0.0251, 0.0289, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 20:52:20,528 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.496e+02 2.075e+02 2.462e+02 3.061e+02 7.702e+02, threshold=4.923e+02, percent-clipped=3.0 2023-04-30 20:52:30,838 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=185468.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:52:54,434 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=185485.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:53:18,343 INFO [train.py:904] (4/8) Epoch 19, batch 2800, loss[loss=0.1909, simple_loss=0.2782, pruned_loss=0.05184, over 12224.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2611, pruned_loss=0.04209, over 3325545.89 frames. ], batch size: 246, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:53:34,592 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.06 vs. limit=5.0 2023-04-30 20:53:46,625 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=185522.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:53:56,371 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=185529.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:54:19,828 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0896, 5.4255, 5.1766, 5.1897, 4.9406, 4.8681, 4.8379, 5.5110], device='cuda:4'), covar=tensor([0.1206, 0.0879, 0.1043, 0.0815, 0.0833, 0.0941, 0.1291, 0.0985], device='cuda:4'), in_proj_covar=tensor([0.0674, 0.0829, 0.0686, 0.0615, 0.0521, 0.0528, 0.0690, 0.0642], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:54:28,292 INFO [train.py:904] (4/8) Epoch 19, batch 2850, loss[loss=0.1871, simple_loss=0.2615, pruned_loss=0.05632, over 16908.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2601, pruned_loss=0.04221, over 3327416.90 frames. ], batch size: 109, lr: 3.59e-03, grad_scale: 8.0 2023-04-30 20:54:41,477 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.535e+02 2.166e+02 2.620e+02 3.152e+02 5.145e+02, threshold=5.240e+02, percent-clipped=1.0 2023-04-30 20:54:53,415 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=185570.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:54:58,018 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6538, 2.5557, 2.1309, 2.2476, 2.8905, 2.6562, 3.3285, 3.1507], device='cuda:4'), covar=tensor([0.0171, 0.0490, 0.0634, 0.0548, 0.0318, 0.0426, 0.0262, 0.0320], device='cuda:4'), in_proj_covar=tensor([0.0203, 0.0234, 0.0224, 0.0225, 0.0235, 0.0233, 0.0238, 0.0229], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:55:10,241 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5240, 3.7366, 4.0510, 2.2839, 3.3047, 2.6292, 4.0267, 3.9771], device='cuda:4'), covar=tensor([0.0283, 0.0938, 0.0501, 0.1941, 0.0793, 0.0939, 0.0597, 0.1116], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0160, 0.0164, 0.0150, 0.0142, 0.0127, 0.0142, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 20:55:10,507 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-30 20:55:28,879 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=185595.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:55:39,063 INFO [train.py:904] (4/8) Epoch 19, batch 2900, loss[loss=0.1733, simple_loss=0.2599, pruned_loss=0.04334, over 17084.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.259, pruned_loss=0.04224, over 3322922.36 frames. ], batch size: 53, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 20:56:32,514 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-30 20:56:36,416 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=185643.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:56:47,750 INFO [train.py:904] (4/8) Epoch 19, batch 2950, loss[loss=0.1781, simple_loss=0.2695, pruned_loss=0.04336, over 16511.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2597, pruned_loss=0.04303, over 3318790.79 frames. ], batch size: 68, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 20:57:00,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.894e+02 2.308e+02 2.753e+02 3.431e+02 9.141e+02, threshold=5.506e+02, percent-clipped=4.0 2023-04-30 20:57:36,168 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.73 vs. limit=5.0 2023-04-30 20:57:38,243 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9222, 1.9374, 2.4825, 2.7610, 2.7260, 2.8441, 1.9542, 2.9967], device='cuda:4'), covar=tensor([0.0162, 0.0472, 0.0297, 0.0273, 0.0280, 0.0287, 0.0553, 0.0151], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0193, 0.0180, 0.0182, 0.0193, 0.0152, 0.0196, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 20:57:58,853 INFO [train.py:904] (4/8) Epoch 19, batch 3000, loss[loss=0.1887, simple_loss=0.2606, pruned_loss=0.05842, over 16907.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2599, pruned_loss=0.04382, over 3317210.66 frames. ], batch size: 116, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 20:57:58,854 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 20:58:07,640 INFO [train.py:938] (4/8) Epoch 19, validation: loss=0.1362, simple_loss=0.2416, pruned_loss=0.01538, over 944034.00 frames. 2023-04-30 20:58:07,641 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 20:58:25,297 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-30 20:58:28,298 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=185716.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 20:58:35,540 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3572, 4.3249, 4.2702, 4.0370, 4.0660, 4.3669, 4.1379, 4.1440], device='cuda:4'), covar=tensor([0.0635, 0.0580, 0.0266, 0.0248, 0.0695, 0.0460, 0.0619, 0.0535], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0424, 0.0350, 0.0341, 0.0360, 0.0395, 0.0239, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 20:59:16,277 INFO [train.py:904] (4/8) Epoch 19, batch 3050, loss[loss=0.1712, simple_loss=0.2546, pruned_loss=0.04386, over 16382.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2603, pruned_loss=0.04411, over 3317687.55 frames. ], batch size: 146, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 20:59:28,110 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=185759.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 20:59:29,997 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.398e+02 2.137e+02 2.455e+02 2.756e+02 4.328e+02, threshold=4.910e+02, percent-clipped=0.0 2023-04-30 20:59:35,051 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=185764.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:00:03,965 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=185785.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:00:23,319 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=185799.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 21:00:26,857 INFO [train.py:904] (4/8) Epoch 19, batch 3100, loss[loss=0.1787, simple_loss=0.2619, pruned_loss=0.04777, over 16569.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2592, pruned_loss=0.04376, over 3306473.04 frames. ], batch size: 62, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 21:00:51,273 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=185820.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 21:00:57,013 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=185824.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:01:09,059 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=185833.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:01:34,487 INFO [train.py:904] (4/8) Epoch 19, batch 3150, loss[loss=0.1817, simple_loss=0.2821, pruned_loss=0.04069, over 17101.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2584, pruned_loss=0.04338, over 3312536.78 frames. ], batch size: 55, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 21:01:46,025 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=185860.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 21:01:47,104 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.651e+02 2.254e+02 2.659e+02 3.271e+02 6.332e+02, threshold=5.317e+02, percent-clipped=3.0 2023-04-30 21:02:04,310 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2015, 4.1697, 4.1329, 3.5740, 4.1434, 1.8044, 3.9681, 3.7151], device='cuda:4'), covar=tensor([0.0133, 0.0111, 0.0180, 0.0280, 0.0097, 0.2738, 0.0132, 0.0236], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0151, 0.0196, 0.0179, 0.0173, 0.0206, 0.0188, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:02:43,075 INFO [train.py:904] (4/8) Epoch 19, batch 3200, loss[loss=0.1466, simple_loss=0.2364, pruned_loss=0.02835, over 16270.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2568, pruned_loss=0.04251, over 3314013.28 frames. ], batch size: 36, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 21:03:26,847 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-30 21:03:37,440 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8060, 4.3032, 2.9785, 2.3444, 2.7450, 2.5689, 4.5467, 3.5730], device='cuda:4'), covar=tensor([0.2651, 0.0486, 0.1760, 0.2675, 0.2576, 0.1915, 0.0331, 0.1291], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0266, 0.0301, 0.0305, 0.0296, 0.0252, 0.0290, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:03:53,534 INFO [train.py:904] (4/8) Epoch 19, batch 3250, loss[loss=0.1427, simple_loss=0.2416, pruned_loss=0.02195, over 17069.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2574, pruned_loss=0.04267, over 3312178.49 frames. ], batch size: 50, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 21:04:06,307 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.442e+02 2.264e+02 2.586e+02 3.072e+02 8.632e+02, threshold=5.173e+02, percent-clipped=3.0 2023-04-30 21:04:35,943 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2633, 2.1870, 2.3622, 4.0801, 2.2064, 2.5928, 2.2823, 2.3917], device='cuda:4'), covar=tensor([0.1314, 0.3551, 0.2759, 0.0539, 0.3761, 0.2400, 0.3716, 0.3056], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0440, 0.0363, 0.0328, 0.0435, 0.0506, 0.0409, 0.0514], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:05:07,473 INFO [train.py:904] (4/8) Epoch 19, batch 3300, loss[loss=0.1944, simple_loss=0.2697, pruned_loss=0.05954, over 16730.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2586, pruned_loss=0.04312, over 3314202.43 frames. ], batch size: 124, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 21:06:15,806 INFO [train.py:904] (4/8) Epoch 19, batch 3350, loss[loss=0.1925, simple_loss=0.2703, pruned_loss=0.05737, over 16677.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2591, pruned_loss=0.04307, over 3320508.27 frames. ], batch size: 134, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 21:06:27,998 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.470e+02 2.170e+02 2.521e+02 2.996e+02 5.638e+02, threshold=5.041e+02, percent-clipped=4.0 2023-04-30 21:06:41,345 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0642, 5.1054, 5.5257, 5.5467, 5.5489, 5.1880, 5.1404, 4.9636], device='cuda:4'), covar=tensor([0.0339, 0.0574, 0.0417, 0.0389, 0.0453, 0.0406, 0.0868, 0.0411], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0450, 0.0436, 0.0407, 0.0482, 0.0460, 0.0555, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 21:07:23,749 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186100.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:07:25,699 INFO [train.py:904] (4/8) Epoch 19, batch 3400, loss[loss=0.1716, simple_loss=0.244, pruned_loss=0.04961, over 16860.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2577, pruned_loss=0.04269, over 3328794.98 frames. ], batch size: 109, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 21:07:44,697 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186115.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 21:07:57,637 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=186124.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:08:34,723 INFO [train.py:904] (4/8) Epoch 19, batch 3450, loss[loss=0.1751, simple_loss=0.2617, pruned_loss=0.04423, over 15534.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2565, pruned_loss=0.04246, over 3328365.83 frames. ], batch size: 191, lr: 3.58e-03, grad_scale: 8.0 2023-04-30 21:08:38,532 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186155.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 21:08:44,162 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186158.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:08:47,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.616e+02 2.301e+02 2.793e+02 3.380e+02 7.495e+02, threshold=5.586e+02, percent-clipped=5.0 2023-04-30 21:08:48,294 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186161.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:09:02,957 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=186172.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:09:27,194 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-30 21:09:45,135 INFO [train.py:904] (4/8) Epoch 19, batch 3500, loss[loss=0.1547, simple_loss=0.2537, pruned_loss=0.0279, over 17107.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2555, pruned_loss=0.04183, over 3324089.55 frames. ], batch size: 47, lr: 3.58e-03, grad_scale: 16.0 2023-04-30 21:10:08,910 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186219.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:10:55,186 INFO [train.py:904] (4/8) Epoch 19, batch 3550, loss[loss=0.1516, simple_loss=0.2448, pruned_loss=0.02918, over 17115.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2553, pruned_loss=0.04152, over 3331437.89 frames. ], batch size: 48, lr: 3.58e-03, grad_scale: 16.0 2023-04-30 21:11:05,136 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186259.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:11:06,980 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.235e+02 2.073e+02 2.370e+02 2.856e+02 5.323e+02, threshold=4.740e+02, percent-clipped=0.0 2023-04-30 21:11:08,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4981, 3.4906, 3.4692, 2.7963, 3.3129, 2.0907, 3.1113, 2.7389], device='cuda:4'), covar=tensor([0.0154, 0.0124, 0.0187, 0.0229, 0.0105, 0.2410, 0.0140, 0.0277], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0151, 0.0197, 0.0180, 0.0174, 0.0206, 0.0189, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:12:03,396 INFO [train.py:904] (4/8) Epoch 19, batch 3600, loss[loss=0.1529, simple_loss=0.2437, pruned_loss=0.03102, over 17202.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2545, pruned_loss=0.04149, over 3323056.76 frames. ], batch size: 46, lr: 3.58e-03, grad_scale: 16.0 2023-04-30 21:12:03,771 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7243, 5.0414, 4.7772, 4.8040, 4.6102, 4.5349, 4.4696, 5.0800], device='cuda:4'), covar=tensor([0.1160, 0.0814, 0.1140, 0.0850, 0.0813, 0.1164, 0.1246, 0.0950], device='cuda:4'), in_proj_covar=tensor([0.0677, 0.0832, 0.0687, 0.0619, 0.0524, 0.0530, 0.0696, 0.0643], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:12:28,382 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186320.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:13:13,925 INFO [train.py:904] (4/8) Epoch 19, batch 3650, loss[loss=0.181, simple_loss=0.2597, pruned_loss=0.05117, over 15509.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2535, pruned_loss=0.04248, over 3313950.28 frames. ], batch size: 191, lr: 3.58e-03, grad_scale: 16.0 2023-04-30 21:13:27,661 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.643e+02 2.244e+02 2.696e+02 3.128e+02 5.239e+02, threshold=5.393e+02, percent-clipped=3.0 2023-04-30 21:14:28,525 INFO [train.py:904] (4/8) Epoch 19, batch 3700, loss[loss=0.1753, simple_loss=0.2453, pruned_loss=0.05267, over 16864.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2527, pruned_loss=0.04388, over 3293846.60 frames. ], batch size: 116, lr: 3.58e-03, grad_scale: 16.0 2023-04-30 21:14:48,253 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-30 21:14:48,830 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=186415.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:14:57,113 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186421.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:15:41,466 INFO [train.py:904] (4/8) Epoch 19, batch 3750, loss[loss=0.2059, simple_loss=0.2756, pruned_loss=0.06813, over 16258.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2537, pruned_loss=0.04551, over 3277001.99 frames. ], batch size: 165, lr: 3.58e-03, grad_scale: 16.0 2023-04-30 21:15:46,591 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=186455.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 21:15:47,762 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186456.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:15:54,456 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5302, 3.6917, 3.8085, 2.1627, 3.1515, 2.5434, 3.9826, 4.0732], device='cuda:4'), covar=tensor([0.0192, 0.0688, 0.0580, 0.2037, 0.0818, 0.0884, 0.0499, 0.0731], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0162, 0.0166, 0.0151, 0.0144, 0.0127, 0.0143, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:15:55,023 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.254e+02 2.571e+02 3.173e+02 4.680e+02, threshold=5.142e+02, percent-clipped=0.0 2023-04-30 21:15:57,662 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=186463.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:16:15,871 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186475.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:16:26,531 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186482.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:16:31,281 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0136, 4.2345, 2.6778, 4.8625, 3.3063, 4.8837, 2.7323, 3.2596], device='cuda:4'), covar=tensor([0.0215, 0.0256, 0.1432, 0.0122, 0.0714, 0.0279, 0.1371, 0.0711], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0178, 0.0195, 0.0164, 0.0177, 0.0220, 0.0202, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:16:35,076 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.51 vs. limit=5.0 2023-04-30 21:16:38,343 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-30 21:16:49,591 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3667, 3.6933, 3.8851, 2.6407, 3.5767, 3.9548, 3.6556, 2.1742], device='cuda:4'), covar=tensor([0.0525, 0.0119, 0.0051, 0.0389, 0.0106, 0.0091, 0.0082, 0.0488], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0080, 0.0080, 0.0132, 0.0095, 0.0105, 0.0092, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:16:53,853 INFO [train.py:904] (4/8) Epoch 19, batch 3800, loss[loss=0.1588, simple_loss=0.2387, pruned_loss=0.03946, over 16890.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2545, pruned_loss=0.04655, over 3285915.55 frames. ], batch size: 90, lr: 3.58e-03, grad_scale: 16.0 2023-04-30 21:16:55,938 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=186503.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 21:17:11,595 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186514.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:17:15,491 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186517.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:17:43,642 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186536.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:17:47,957 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186539.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:18:05,818 INFO [train.py:904] (4/8) Epoch 19, batch 3850, loss[loss=0.1521, simple_loss=0.238, pruned_loss=0.03308, over 16497.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2547, pruned_loss=0.04723, over 3284095.75 frames. ], batch size: 68, lr: 3.58e-03, grad_scale: 4.0 2023-04-30 21:18:22,160 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.100e+02 2.475e+02 3.097e+02 5.903e+02, threshold=4.949e+02, percent-clipped=1.0 2023-04-30 21:18:25,352 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3562, 3.2267, 3.4943, 2.0067, 3.6104, 3.5625, 2.9173, 2.7857], device='cuda:4'), covar=tensor([0.0780, 0.0240, 0.0171, 0.1048, 0.0094, 0.0202, 0.0378, 0.0420], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0097, 0.0138, 0.0078, 0.0125, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 21:18:28,802 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186567.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:18:44,973 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186578.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:18:56,134 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186586.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:19:16,902 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186600.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:19:18,738 INFO [train.py:904] (4/8) Epoch 19, batch 3900, loss[loss=0.1742, simple_loss=0.2593, pruned_loss=0.04457, over 16805.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2548, pruned_loss=0.0477, over 3280039.89 frames. ], batch size: 39, lr: 3.58e-03, grad_scale: 4.0 2023-04-30 21:19:37,487 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186615.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:19:55,104 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186628.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:20:20,793 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186647.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:20:28,329 INFO [train.py:904] (4/8) Epoch 19, batch 3950, loss[loss=0.171, simple_loss=0.2496, pruned_loss=0.04623, over 16389.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2541, pruned_loss=0.0479, over 3284022.96 frames. ], batch size: 146, lr: 3.57e-03, grad_scale: 4.0 2023-04-30 21:20:41,094 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3036, 3.7594, 3.8750, 2.7501, 3.6392, 3.9995, 3.6297, 2.1483], device='cuda:4'), covar=tensor([0.0500, 0.0110, 0.0052, 0.0351, 0.0091, 0.0083, 0.0083, 0.0468], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0079, 0.0080, 0.0131, 0.0095, 0.0105, 0.0092, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:20:42,972 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.601e+02 2.399e+02 2.822e+02 3.355e+02 8.994e+02, threshold=5.644e+02, percent-clipped=5.0 2023-04-30 21:20:50,387 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-30 21:21:27,854 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5520, 3.4207, 3.8726, 1.9549, 4.0307, 3.9886, 3.1271, 3.0176], device='cuda:4'), covar=tensor([0.0764, 0.0251, 0.0174, 0.1149, 0.0067, 0.0168, 0.0355, 0.0434], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0109, 0.0098, 0.0140, 0.0079, 0.0126, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:21:39,185 INFO [train.py:904] (4/8) Epoch 19, batch 4000, loss[loss=0.1631, simple_loss=0.2439, pruned_loss=0.04113, over 17030.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.254, pruned_loss=0.04829, over 3285647.70 frames. ], batch size: 55, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:21:49,001 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186709.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:22:42,375 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8956, 3.1618, 3.4391, 1.9799, 2.8654, 2.1993, 3.4699, 3.4180], device='cuda:4'), covar=tensor([0.0197, 0.0807, 0.0506, 0.2011, 0.0878, 0.0971, 0.0525, 0.0803], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0163, 0.0167, 0.0153, 0.0144, 0.0129, 0.0144, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:22:49,923 INFO [train.py:904] (4/8) Epoch 19, batch 4050, loss[loss=0.1725, simple_loss=0.2619, pruned_loss=0.04161, over 16700.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2544, pruned_loss=0.04741, over 3283444.43 frames. ], batch size: 134, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:22:57,348 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=186756.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:23:06,009 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.223e+02 1.873e+02 2.216e+02 2.589e+02 4.495e+02, threshold=4.432e+02, percent-clipped=0.0 2023-04-30 21:23:16,230 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186770.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:23:27,095 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186777.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:23:48,440 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2209, 3.9102, 3.7979, 2.6190, 3.4934, 3.8850, 3.4867, 2.0677], device='cuda:4'), covar=tensor([0.0489, 0.0044, 0.0047, 0.0340, 0.0097, 0.0090, 0.0096, 0.0459], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0080, 0.0081, 0.0132, 0.0095, 0.0105, 0.0092, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:23:58,560 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7711, 3.7337, 3.8961, 3.6304, 3.8241, 4.2291, 3.8778, 3.5613], device='cuda:4'), covar=tensor([0.2280, 0.2223, 0.2107, 0.2421, 0.2544, 0.1828, 0.1563, 0.2482], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0587, 0.0646, 0.0490, 0.0655, 0.0684, 0.0508, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:24:03,623 INFO [train.py:904] (4/8) Epoch 19, batch 4100, loss[loss=0.1984, simple_loss=0.2781, pruned_loss=0.05933, over 16884.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2556, pruned_loss=0.04674, over 3276776.78 frames. ], batch size: 109, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:24:07,123 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=186804.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:24:13,805 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3645, 4.2165, 4.0916, 2.7800, 3.7146, 4.1938, 3.6755, 2.2448], device='cuda:4'), covar=tensor([0.0495, 0.0027, 0.0043, 0.0354, 0.0086, 0.0074, 0.0079, 0.0437], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0080, 0.0081, 0.0132, 0.0095, 0.0105, 0.0092, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:24:21,800 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=186814.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:24:33,210 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5369, 4.6131, 4.9105, 4.8653, 4.8962, 4.5849, 4.5536, 4.4247], device='cuda:4'), covar=tensor([0.0303, 0.0470, 0.0299, 0.0366, 0.0459, 0.0348, 0.0915, 0.0454], device='cuda:4'), in_proj_covar=tensor([0.0404, 0.0445, 0.0432, 0.0404, 0.0475, 0.0456, 0.0549, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 21:24:46,725 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186831.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:24:59,810 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186840.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:25:08,833 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9251, 2.0335, 2.1538, 3.4779, 1.9638, 2.3113, 2.1821, 2.1797], device='cuda:4'), covar=tensor([0.1419, 0.3784, 0.2763, 0.0603, 0.4373, 0.2456, 0.3558, 0.3555], device='cuda:4'), in_proj_covar=tensor([0.0396, 0.0440, 0.0360, 0.0326, 0.0432, 0.0507, 0.0408, 0.0514], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:25:14,220 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0594, 4.2394, 4.4398, 4.4016, 4.4442, 4.1734, 4.0410, 4.1497], device='cuda:4'), covar=tensor([0.0468, 0.0557, 0.0484, 0.0552, 0.0566, 0.0518, 0.1256, 0.0533], device='cuda:4'), in_proj_covar=tensor([0.0400, 0.0440, 0.0428, 0.0399, 0.0470, 0.0452, 0.0544, 0.0357], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 21:25:17,893 INFO [train.py:904] (4/8) Epoch 19, batch 4150, loss[loss=0.2147, simple_loss=0.3083, pruned_loss=0.06058, over 15471.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2623, pruned_loss=0.0491, over 3233300.89 frames. ], batch size: 190, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:25:34,161 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=186862.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:25:34,923 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.347e+02 2.016e+02 2.433e+02 2.890e+02 5.187e+02, threshold=4.866e+02, percent-clipped=4.0 2023-04-30 21:25:49,714 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186873.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:25:53,414 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-30 21:26:10,487 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7337, 2.5105, 2.2480, 3.3338, 2.2937, 3.5036, 1.5845, 2.6342], device='cuda:4'), covar=tensor([0.1386, 0.0756, 0.1279, 0.0168, 0.0171, 0.0378, 0.1672, 0.0830], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0172, 0.0191, 0.0184, 0.0205, 0.0215, 0.0196, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:26:23,061 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186895.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:26:31,963 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186901.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 21:26:32,669 INFO [train.py:904] (4/8) Epoch 19, batch 4200, loss[loss=0.2104, simple_loss=0.3096, pruned_loss=0.05564, over 16462.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2693, pruned_loss=0.05066, over 3222029.49 frames. ], batch size: 146, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:26:47,060 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=186911.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:26:48,321 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9012, 4.9324, 4.7782, 4.4801, 4.4186, 4.8507, 4.7129, 4.4839], device='cuda:4'), covar=tensor([0.0554, 0.0528, 0.0279, 0.0275, 0.0942, 0.0420, 0.0369, 0.0670], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0414, 0.0342, 0.0331, 0.0352, 0.0385, 0.0233, 0.0406], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:26:50,134 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0005, 5.4955, 5.6994, 5.3920, 5.4165, 6.0038, 5.5645, 5.2683], device='cuda:4'), covar=tensor([0.0894, 0.1665, 0.1503, 0.1745, 0.2243, 0.0823, 0.1299, 0.2176], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0580, 0.0638, 0.0486, 0.0645, 0.0676, 0.0501, 0.0647], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:26:53,367 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=186915.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:27:06,366 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186923.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:27:34,736 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=186942.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:27:34,933 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3794, 2.1770, 1.7988, 1.9138, 2.4012, 2.0843, 2.1766, 2.5241], device='cuda:4'), covar=tensor([0.0186, 0.0411, 0.0542, 0.0471, 0.0253, 0.0387, 0.0166, 0.0263], device='cuda:4'), in_proj_covar=tensor([0.0203, 0.0232, 0.0222, 0.0225, 0.0234, 0.0232, 0.0237, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:27:44,529 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0082, 2.0300, 1.7324, 1.6870, 2.1737, 1.8480, 1.8259, 2.2705], device='cuda:4'), covar=tensor([0.0192, 0.0359, 0.0493, 0.0468, 0.0247, 0.0381, 0.0170, 0.0259], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0231, 0.0222, 0.0224, 0.0233, 0.0231, 0.0236, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:27:49,658 INFO [train.py:904] (4/8) Epoch 19, batch 4250, loss[loss=0.1825, simple_loss=0.2772, pruned_loss=0.0439, over 15233.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2733, pruned_loss=0.05104, over 3181044.70 frames. ], batch size: 191, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:28:05,618 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.577e+02 2.275e+02 2.568e+02 3.131e+02 7.157e+02, threshold=5.135e+02, percent-clipped=7.0 2023-04-30 21:28:06,528 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=186963.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:28:19,188 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=186972.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:28:27,752 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-30 21:28:56,053 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 21:29:02,895 INFO [train.py:904] (4/8) Epoch 19, batch 4300, loss[loss=0.1908, simple_loss=0.2845, pruned_loss=0.04852, over 16727.00 frames. ], tot_loss[loss=0.1872, simple_loss=0.2745, pruned_loss=0.04993, over 3193652.58 frames. ], batch size: 83, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:29:37,788 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=187025.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:29:43,835 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4001, 2.3900, 2.4012, 4.1199, 2.2496, 2.7586, 2.4816, 2.5150], device='cuda:4'), covar=tensor([0.1173, 0.3076, 0.2542, 0.0451, 0.3802, 0.2230, 0.2786, 0.3219], device='cuda:4'), in_proj_covar=tensor([0.0395, 0.0439, 0.0359, 0.0324, 0.0432, 0.0506, 0.0407, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:29:56,636 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=187038.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:30:16,621 INFO [train.py:904] (4/8) Epoch 19, batch 4350, loss[loss=0.1903, simple_loss=0.2864, pruned_loss=0.04713, over 16914.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2782, pruned_loss=0.05107, over 3199265.72 frames. ], batch size: 96, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:30:32,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.681e+02 2.200e+02 2.562e+02 2.922e+02 5.186e+02, threshold=5.123e+02, percent-clipped=1.0 2023-04-30 21:30:35,926 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187065.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:30:54,555 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187077.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:31:07,466 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=187086.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 21:31:26,314 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=187099.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:31:31,214 INFO [train.py:904] (4/8) Epoch 19, batch 4400, loss[loss=0.1942, simple_loss=0.2804, pruned_loss=0.05404, over 16546.00 frames. ], tot_loss[loss=0.1927, simple_loss=0.2805, pruned_loss=0.05246, over 3189010.11 frames. ], batch size: 62, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:31:35,940 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4042, 1.6586, 2.1461, 2.4248, 2.4119, 2.7600, 1.7529, 2.6689], device='cuda:4'), covar=tensor([0.0227, 0.0465, 0.0286, 0.0299, 0.0288, 0.0171, 0.0499, 0.0121], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0188, 0.0177, 0.0179, 0.0190, 0.0148, 0.0192, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:31:39,075 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4103, 5.3975, 5.2708, 4.9163, 4.8737, 5.2889, 5.1743, 4.9327], device='cuda:4'), covar=tensor([0.0483, 0.0268, 0.0235, 0.0235, 0.0906, 0.0293, 0.0288, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0408, 0.0338, 0.0327, 0.0347, 0.0379, 0.0230, 0.0400], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:31:46,404 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2289, 4.0428, 3.8269, 2.4084, 3.5123, 3.9445, 3.5557, 2.1061], device='cuda:4'), covar=tensor([0.0538, 0.0027, 0.0052, 0.0428, 0.0088, 0.0075, 0.0082, 0.0466], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0080, 0.0081, 0.0133, 0.0096, 0.0106, 0.0093, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:31:52,412 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3108, 4.1842, 4.3776, 4.4867, 4.5994, 4.1958, 4.5683, 4.6337], device='cuda:4'), covar=tensor([0.1495, 0.1014, 0.1286, 0.0630, 0.0482, 0.1061, 0.0699, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0622, 0.0769, 0.0903, 0.0787, 0.0587, 0.0615, 0.0632, 0.0737], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:32:05,653 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187125.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:32:13,624 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187131.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:32:43,362 INFO [train.py:904] (4/8) Epoch 19, batch 4450, loss[loss=0.2062, simple_loss=0.3009, pruned_loss=0.05574, over 16682.00 frames. ], tot_loss[loss=0.196, simple_loss=0.2843, pruned_loss=0.05389, over 3211903.20 frames. ], batch size: 134, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:32:47,739 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1994, 4.2269, 4.4966, 4.4604, 4.5221, 4.1988, 4.2167, 4.1203], device='cuda:4'), covar=tensor([0.0302, 0.0510, 0.0352, 0.0366, 0.0370, 0.0379, 0.0871, 0.0468], device='cuda:4'), in_proj_covar=tensor([0.0393, 0.0432, 0.0419, 0.0392, 0.0461, 0.0443, 0.0534, 0.0349], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 21:33:00,621 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.478e+02 2.070e+02 2.320e+02 2.761e+02 4.804e+02, threshold=4.641e+02, percent-clipped=0.0 2023-04-30 21:33:15,031 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187173.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:33:24,880 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187179.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:33:47,940 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187195.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:33:49,990 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187196.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 21:33:57,509 INFO [train.py:904] (4/8) Epoch 19, batch 4500, loss[loss=0.208, simple_loss=0.3067, pruned_loss=0.05466, over 16588.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2851, pruned_loss=0.05459, over 3222406.61 frames. ], batch size: 75, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:34:24,639 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187221.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:34:27,166 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187223.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:34:39,792 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6694, 4.8763, 4.5982, 3.0683, 4.0452, 4.6789, 4.0547, 2.7325], device='cuda:4'), covar=tensor([0.0520, 0.0017, 0.0034, 0.0379, 0.0078, 0.0066, 0.0073, 0.0400], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0080, 0.0081, 0.0133, 0.0095, 0.0106, 0.0093, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:34:54,311 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187242.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:34:55,416 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187243.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:35:08,491 INFO [train.py:904] (4/8) Epoch 19, batch 4550, loss[loss=0.2257, simple_loss=0.3088, pruned_loss=0.0713, over 16312.00 frames. ], tot_loss[loss=0.198, simple_loss=0.2857, pruned_loss=0.05516, over 3230330.10 frames. ], batch size: 35, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:35:24,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.475e+02 1.917e+02 2.180e+02 2.712e+02 4.446e+02, threshold=4.359e+02, percent-clipped=0.0 2023-04-30 21:35:30,227 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187267.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:35:36,697 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187271.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:35:43,915 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8803, 2.0813, 2.2361, 3.4052, 1.9796, 2.3251, 2.1936, 2.1696], device='cuda:4'), covar=tensor([0.1394, 0.3490, 0.2668, 0.0620, 0.4368, 0.2572, 0.3296, 0.3468], device='cuda:4'), in_proj_covar=tensor([0.0395, 0.0438, 0.0358, 0.0323, 0.0432, 0.0505, 0.0407, 0.0511], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:36:05,135 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187290.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:36:20,300 INFO [train.py:904] (4/8) Epoch 19, batch 4600, loss[loss=0.1816, simple_loss=0.2689, pruned_loss=0.04713, over 16778.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2864, pruned_loss=0.05525, over 3234500.14 frames. ], batch size: 57, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:37:33,043 INFO [train.py:904] (4/8) Epoch 19, batch 4650, loss[loss=0.1864, simple_loss=0.2788, pruned_loss=0.04695, over 15496.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2857, pruned_loss=0.05571, over 3236396.63 frames. ], batch size: 191, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:37:49,309 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.228e+02 1.930e+02 2.188e+02 2.546e+02 5.637e+02, threshold=4.377e+02, percent-clipped=1.0 2023-04-30 21:37:52,208 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187365.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:38:01,572 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=187372.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:38:15,576 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187381.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 21:38:15,681 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9225, 3.5769, 3.4286, 2.1993, 3.2986, 3.5050, 3.2234, 1.8360], device='cuda:4'), covar=tensor([0.0600, 0.0045, 0.0058, 0.0445, 0.0084, 0.0106, 0.0100, 0.0526], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0079, 0.0081, 0.0132, 0.0095, 0.0106, 0.0092, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:38:33,235 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187394.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:38:44,848 INFO [train.py:904] (4/8) Epoch 19, batch 4700, loss[loss=0.1838, simple_loss=0.2736, pruned_loss=0.047, over 17043.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2836, pruned_loss=0.05474, over 3218680.90 frames. ], batch size: 53, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:39:01,715 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187413.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:39:28,151 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 21:39:29,993 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=187433.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:39:55,445 INFO [train.py:904] (4/8) Epoch 19, batch 4750, loss[loss=0.1678, simple_loss=0.2566, pruned_loss=0.03943, over 16722.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2798, pruned_loss=0.05262, over 3213724.82 frames. ], batch size: 89, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:40:11,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.275e+02 1.903e+02 2.161e+02 2.490e+02 3.592e+02, threshold=4.322e+02, percent-clipped=0.0 2023-04-30 21:40:57,626 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187496.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:40:59,595 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5311, 4.3692, 4.3179, 3.0044, 3.7378, 4.3315, 3.7438, 2.3958], device='cuda:4'), covar=tensor([0.0462, 0.0032, 0.0033, 0.0312, 0.0085, 0.0081, 0.0085, 0.0401], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0079, 0.0080, 0.0132, 0.0095, 0.0105, 0.0092, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 21:41:05,658 INFO [train.py:904] (4/8) Epoch 19, batch 4800, loss[loss=0.2132, simple_loss=0.298, pruned_loss=0.06416, over 16710.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2757, pruned_loss=0.05038, over 3224317.47 frames. ], batch size: 134, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:41:21,478 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=187512.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:41:30,306 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=187519.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:41:37,752 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=187523.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:41:42,029 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9264, 3.2187, 3.3187, 1.8582, 2.7748, 2.1955, 3.4136, 3.3337], device='cuda:4'), covar=tensor([0.0268, 0.0797, 0.0658, 0.2168, 0.0918, 0.1036, 0.0642, 0.0960], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0161, 0.0165, 0.0150, 0.0142, 0.0127, 0.0142, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:42:08,456 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187544.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:42:19,532 INFO [train.py:904] (4/8) Epoch 19, batch 4850, loss[loss=0.2162, simple_loss=0.3089, pruned_loss=0.06169, over 16919.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.276, pruned_loss=0.04993, over 3191218.80 frames. ], batch size: 109, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:42:24,830 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9824, 5.0033, 4.8389, 4.4659, 4.4687, 4.8857, 4.7494, 4.6050], device='cuda:4'), covar=tensor([0.0559, 0.0350, 0.0295, 0.0301, 0.1029, 0.0467, 0.0378, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0401, 0.0331, 0.0322, 0.0342, 0.0373, 0.0228, 0.0393], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:42:36,205 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.385e+02 1.857e+02 2.110e+02 2.587e+02 7.641e+02, threshold=4.221e+02, percent-clipped=2.0 2023-04-30 21:42:41,923 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187567.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:42:50,829 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=187573.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:43:01,971 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=187580.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:43:07,563 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=187584.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:43:07,656 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5326, 2.9004, 3.0591, 1.8145, 2.7044, 2.0890, 3.1622, 3.1279], device='cuda:4'), covar=tensor([0.0273, 0.0746, 0.0619, 0.2072, 0.0874, 0.0988, 0.0609, 0.0841], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0160, 0.0164, 0.0150, 0.0142, 0.0127, 0.0142, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:43:32,170 INFO [train.py:904] (4/8) Epoch 19, batch 4900, loss[loss=0.1599, simple_loss=0.259, pruned_loss=0.03038, over 16702.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.275, pruned_loss=0.04863, over 3171110.08 frames. ], batch size: 89, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:43:43,276 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6914, 4.7400, 4.5326, 4.1743, 4.1798, 4.5968, 4.4740, 4.2907], device='cuda:4'), covar=tensor([0.0608, 0.0360, 0.0315, 0.0315, 0.1012, 0.0482, 0.0445, 0.0684], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0399, 0.0330, 0.0320, 0.0340, 0.0371, 0.0226, 0.0391], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:43:51,074 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187615.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:44:43,104 INFO [train.py:904] (4/8) Epoch 19, batch 4950, loss[loss=0.1867, simple_loss=0.2791, pruned_loss=0.04719, over 16425.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.2745, pruned_loss=0.04759, over 3187784.35 frames. ], batch size: 68, lr: 3.57e-03, grad_scale: 8.0 2023-04-30 21:44:58,341 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.400e+02 2.032e+02 2.342e+02 2.691e+02 5.269e+02, threshold=4.685e+02, percent-clipped=1.0 2023-04-30 21:45:24,750 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187681.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 21:45:38,498 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4337, 5.7246, 5.3941, 5.5433, 5.2314, 4.9960, 5.1214, 5.8261], device='cuda:4'), covar=tensor([0.1140, 0.0791, 0.1015, 0.0681, 0.0748, 0.0789, 0.1048, 0.0816], device='cuda:4'), in_proj_covar=tensor([0.0642, 0.0793, 0.0651, 0.0587, 0.0499, 0.0505, 0.0656, 0.0611], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:45:40,998 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9706, 2.8490, 2.3637, 2.6357, 3.2972, 2.8866, 3.5395, 3.5107], device='cuda:4'), covar=tensor([0.0067, 0.0375, 0.0510, 0.0378, 0.0221, 0.0377, 0.0189, 0.0219], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0226, 0.0219, 0.0219, 0.0228, 0.0227, 0.0230, 0.0222], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:45:43,961 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=187694.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:45:55,107 INFO [train.py:904] (4/8) Epoch 19, batch 5000, loss[loss=0.1918, simple_loss=0.2816, pruned_loss=0.051, over 15318.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2757, pruned_loss=0.0476, over 3202667.81 frames. ], batch size: 190, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:46:16,831 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1639, 4.0569, 4.2270, 4.3384, 4.4739, 4.0759, 4.4194, 4.5056], device='cuda:4'), covar=tensor([0.1507, 0.1054, 0.1312, 0.0690, 0.0488, 0.1221, 0.0685, 0.0546], device='cuda:4'), in_proj_covar=tensor([0.0606, 0.0747, 0.0879, 0.0767, 0.0571, 0.0599, 0.0614, 0.0716], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:46:32,201 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187728.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:46:33,445 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187729.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:46:52,594 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=187742.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:47:07,083 INFO [train.py:904] (4/8) Epoch 19, batch 5050, loss[loss=0.1792, simple_loss=0.2722, pruned_loss=0.04306, over 16406.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.276, pruned_loss=0.04748, over 3205444.51 frames. ], batch size: 68, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:47:19,577 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4165, 5.7222, 5.4158, 5.5598, 5.2286, 5.0790, 5.1956, 5.8068], device='cuda:4'), covar=tensor([0.1184, 0.0810, 0.1028, 0.0737, 0.0847, 0.0735, 0.1030, 0.0856], device='cuda:4'), in_proj_covar=tensor([0.0645, 0.0796, 0.0655, 0.0590, 0.0501, 0.0508, 0.0660, 0.0614], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:47:21,709 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.446e+02 2.150e+02 2.460e+02 2.787e+02 5.516e+02, threshold=4.919e+02, percent-clipped=2.0 2023-04-30 21:48:17,850 INFO [train.py:904] (4/8) Epoch 19, batch 5100, loss[loss=0.2197, simple_loss=0.3012, pruned_loss=0.06905, over 12036.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2743, pruned_loss=0.04677, over 3213480.44 frames. ], batch size: 248, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:49:05,147 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6594, 4.9494, 4.7266, 4.7851, 4.5001, 4.4535, 4.3725, 5.0323], device='cuda:4'), covar=tensor([0.1168, 0.0787, 0.0932, 0.0713, 0.0694, 0.1052, 0.1052, 0.0789], device='cuda:4'), in_proj_covar=tensor([0.0642, 0.0792, 0.0652, 0.0588, 0.0499, 0.0504, 0.0657, 0.0610], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:49:30,691 INFO [train.py:904] (4/8) Epoch 19, batch 5150, loss[loss=0.1871, simple_loss=0.2818, pruned_loss=0.04619, over 16737.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2748, pruned_loss=0.04631, over 3213878.28 frames. ], batch size: 124, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:49:47,460 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.389e+02 2.013e+02 2.291e+02 2.768e+02 7.535e+02, threshold=4.582e+02, percent-clipped=1.0 2023-04-30 21:49:54,399 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187868.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:50:04,810 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187875.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:50:11,410 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=187879.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:50:24,430 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-30 21:50:41,937 INFO [train.py:904] (4/8) Epoch 19, batch 5200, loss[loss=0.1737, simple_loss=0.2661, pruned_loss=0.04069, over 16883.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2728, pruned_loss=0.04554, over 3215788.98 frames. ], batch size: 96, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:50:55,129 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1820, 4.2456, 2.8470, 4.9969, 3.4674, 4.8850, 3.0556, 3.4533], device='cuda:4'), covar=tensor([0.0194, 0.0267, 0.1296, 0.0154, 0.0698, 0.0292, 0.1172, 0.0635], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0173, 0.0190, 0.0153, 0.0172, 0.0211, 0.0197, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:51:53,321 INFO [train.py:904] (4/8) Epoch 19, batch 5250, loss[loss=0.1696, simple_loss=0.2555, pruned_loss=0.04188, over 16496.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2705, pruned_loss=0.0453, over 3214780.76 frames. ], batch size: 68, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:52:08,308 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.295e+02 1.924e+02 2.321e+02 2.643e+02 4.137e+02, threshold=4.643e+02, percent-clipped=0.0 2023-04-30 21:53:08,598 INFO [train.py:904] (4/8) Epoch 19, batch 5300, loss[loss=0.176, simple_loss=0.2605, pruned_loss=0.04572, over 16714.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2671, pruned_loss=0.04421, over 3220077.07 frames. ], batch size: 134, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:53:43,828 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-30 21:53:46,806 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=188028.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:53:49,772 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4580, 4.4702, 4.8323, 4.7804, 4.8008, 4.4893, 4.4441, 4.3675], device='cuda:4'), covar=tensor([0.0286, 0.0571, 0.0348, 0.0381, 0.0448, 0.0382, 0.0904, 0.0469], device='cuda:4'), in_proj_covar=tensor([0.0391, 0.0431, 0.0417, 0.0391, 0.0464, 0.0440, 0.0528, 0.0347], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 21:53:53,780 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7412, 2.6644, 2.5108, 4.1301, 2.8075, 3.9860, 1.4908, 2.8859], device='cuda:4'), covar=tensor([0.1216, 0.0714, 0.1192, 0.0134, 0.0194, 0.0374, 0.1593, 0.0759], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0173, 0.0193, 0.0184, 0.0205, 0.0215, 0.0198, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:54:21,262 INFO [train.py:904] (4/8) Epoch 19, batch 5350, loss[loss=0.1818, simple_loss=0.2667, pruned_loss=0.04844, over 17054.00 frames. ], tot_loss[loss=0.176, simple_loss=0.265, pruned_loss=0.04346, over 3231889.17 frames. ], batch size: 53, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:54:38,011 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.400e+02 1.961e+02 2.159e+02 2.589e+02 4.705e+02, threshold=4.317e+02, percent-clipped=1.0 2023-04-30 21:54:57,073 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=188076.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:55:30,507 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4039, 4.6530, 4.4952, 4.4944, 4.2367, 4.1809, 4.1832, 4.7157], device='cuda:4'), covar=tensor([0.1123, 0.0780, 0.0829, 0.0746, 0.0725, 0.1414, 0.0976, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0642, 0.0794, 0.0653, 0.0589, 0.0500, 0.0504, 0.0658, 0.0612], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:55:31,774 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1666, 3.1779, 1.8880, 3.4744, 2.4027, 3.4944, 2.1113, 2.6181], device='cuda:4'), covar=tensor([0.0310, 0.0375, 0.1676, 0.0165, 0.0918, 0.0557, 0.1514, 0.0791], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0194, 0.0155, 0.0175, 0.0213, 0.0200, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:55:35,754 INFO [train.py:904] (4/8) Epoch 19, batch 5400, loss[loss=0.2005, simple_loss=0.2841, pruned_loss=0.05844, over 11578.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.268, pruned_loss=0.04436, over 3215775.00 frames. ], batch size: 246, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:55:48,118 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0164, 4.1662, 2.7003, 4.9912, 3.3896, 4.8231, 2.9706, 3.4184], device='cuda:4'), covar=tensor([0.0237, 0.0291, 0.1493, 0.0131, 0.0742, 0.0425, 0.1231, 0.0622], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0194, 0.0155, 0.0175, 0.0214, 0.0200, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 21:55:48,134 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1572, 2.2834, 2.8703, 3.1768, 3.0202, 3.7998, 2.4121, 3.7179], device='cuda:4'), covar=tensor([0.0210, 0.0438, 0.0305, 0.0262, 0.0273, 0.0114, 0.0468, 0.0112], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0189, 0.0178, 0.0181, 0.0190, 0.0149, 0.0192, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:56:54,053 INFO [train.py:904] (4/8) Epoch 19, batch 5450, loss[loss=0.1964, simple_loss=0.2827, pruned_loss=0.0551, over 16753.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2714, pruned_loss=0.04597, over 3206066.58 frames. ], batch size: 76, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:56:59,146 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7297, 1.8747, 2.4178, 2.6704, 2.5642, 3.1016, 2.0010, 3.1294], device='cuda:4'), covar=tensor([0.0216, 0.0461, 0.0293, 0.0316, 0.0291, 0.0163, 0.0451, 0.0129], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0189, 0.0178, 0.0181, 0.0190, 0.0149, 0.0192, 0.0144], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:57:11,919 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.414e+02 1.998e+02 2.677e+02 3.507e+02 6.066e+02, threshold=5.353e+02, percent-clipped=8.0 2023-04-30 21:57:19,747 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=188168.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:57:32,208 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=188175.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:57:38,479 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=188179.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:58:14,499 INFO [train.py:904] (4/8) Epoch 19, batch 5500, loss[loss=0.2009, simple_loss=0.2939, pruned_loss=0.05388, over 17103.00 frames. ], tot_loss[loss=0.1892, simple_loss=0.2782, pruned_loss=0.0501, over 3165464.83 frames. ], batch size: 49, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:58:32,301 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5666, 2.6170, 2.0681, 2.4247, 3.0302, 2.5896, 3.2156, 3.1970], device='cuda:4'), covar=tensor([0.0096, 0.0354, 0.0501, 0.0408, 0.0230, 0.0367, 0.0185, 0.0230], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0228, 0.0219, 0.0220, 0.0229, 0.0227, 0.0230, 0.0223], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:58:38,577 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=188216.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:58:39,777 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5266, 3.5230, 3.4854, 2.8698, 3.4503, 2.0577, 3.2352, 2.8202], device='cuda:4'), covar=tensor([0.0163, 0.0131, 0.0173, 0.0253, 0.0108, 0.2296, 0.0140, 0.0241], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0144, 0.0190, 0.0174, 0.0166, 0.0199, 0.0180, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 21:58:48,994 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=188223.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:58:55,240 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=188227.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 21:59:35,425 INFO [train.py:904] (4/8) Epoch 19, batch 5550, loss[loss=0.2034, simple_loss=0.2905, pruned_loss=0.05817, over 17259.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2861, pruned_loss=0.05608, over 3125283.85 frames. ], batch size: 52, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 21:59:53,551 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.484e+02 3.263e+02 3.931e+02 4.573e+02 1.048e+03, threshold=7.861e+02, percent-clipped=6.0 2023-04-30 22:00:26,046 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=188283.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:00:55,096 INFO [train.py:904] (4/8) Epoch 19, batch 5600, loss[loss=0.2434, simple_loss=0.3197, pruned_loss=0.08353, over 15457.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2916, pruned_loss=0.06072, over 3079531.62 frames. ], batch size: 190, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:02:07,270 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=188344.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:02:19,465 INFO [train.py:904] (4/8) Epoch 19, batch 5650, loss[loss=0.2408, simple_loss=0.3211, pruned_loss=0.0803, over 15333.00 frames. ], tot_loss[loss=0.214, simple_loss=0.2968, pruned_loss=0.06561, over 3026376.74 frames. ], batch size: 190, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:02:36,991 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.386e+02 3.211e+02 3.758e+02 5.159e+02 1.051e+03, threshold=7.516e+02, percent-clipped=3.0 2023-04-30 22:03:36,670 INFO [train.py:904] (4/8) Epoch 19, batch 5700, loss[loss=0.189, simple_loss=0.2876, pruned_loss=0.04518, over 16372.00 frames. ], tot_loss[loss=0.2154, simple_loss=0.2976, pruned_loss=0.06659, over 3028421.95 frames. ], batch size: 165, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:04:38,936 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-30 22:04:44,712 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0880, 5.7199, 5.8882, 5.5283, 5.7043, 6.2174, 5.6939, 5.4221], device='cuda:4'), covar=tensor([0.0866, 0.1670, 0.2105, 0.1939, 0.2100, 0.0834, 0.1461, 0.2353], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0568, 0.0626, 0.0478, 0.0636, 0.0660, 0.0490, 0.0642], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:04:55,839 INFO [train.py:904] (4/8) Epoch 19, batch 5750, loss[loss=0.1814, simple_loss=0.2776, pruned_loss=0.04263, over 16884.00 frames. ], tot_loss[loss=0.2185, simple_loss=0.3003, pruned_loss=0.06836, over 2999334.86 frames. ], batch size: 96, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:05:12,679 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.963e+02 2.939e+02 3.429e+02 4.149e+02 7.338e+02, threshold=6.857e+02, percent-clipped=0.0 2023-04-30 22:05:45,246 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=188481.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:05:46,470 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0080, 3.0633, 3.1282, 2.2073, 2.9376, 3.1567, 3.0265, 1.8473], device='cuda:4'), covar=tensor([0.0506, 0.0072, 0.0067, 0.0415, 0.0118, 0.0106, 0.0095, 0.0502], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0079, 0.0079, 0.0131, 0.0095, 0.0105, 0.0091, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:06:05,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9459, 3.1045, 3.1470, 2.1300, 2.9696, 3.1602, 3.0167, 1.8544], device='cuda:4'), covar=tensor([0.0557, 0.0072, 0.0068, 0.0435, 0.0112, 0.0106, 0.0098, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0078, 0.0079, 0.0131, 0.0094, 0.0105, 0.0091, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:06:19,318 INFO [train.py:904] (4/8) Epoch 19, batch 5800, loss[loss=0.1993, simple_loss=0.2753, pruned_loss=0.06171, over 12322.00 frames. ], tot_loss[loss=0.2162, simple_loss=0.2993, pruned_loss=0.06654, over 3009447.58 frames. ], batch size: 250, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:07:24,008 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=188542.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:07:38,949 INFO [train.py:904] (4/8) Epoch 19, batch 5850, loss[loss=0.2035, simple_loss=0.2939, pruned_loss=0.05659, over 16211.00 frames. ], tot_loss[loss=0.2131, simple_loss=0.2971, pruned_loss=0.06454, over 3029485.57 frames. ], batch size: 165, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:07:57,311 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.796e+02 2.681e+02 3.202e+02 3.753e+02 8.228e+02, threshold=6.404e+02, percent-clipped=1.0 2023-04-30 22:08:00,951 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0307, 2.3708, 2.3532, 2.8899, 2.0314, 3.2175, 1.8939, 2.7186], device='cuda:4'), covar=tensor([0.1042, 0.0554, 0.0995, 0.0181, 0.0124, 0.0362, 0.1302, 0.0639], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0171, 0.0192, 0.0182, 0.0203, 0.0212, 0.0196, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 22:08:05,687 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-30 22:08:08,357 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8706, 5.1879, 5.3114, 5.1174, 5.1346, 5.6953, 5.0913, 4.8225], device='cuda:4'), covar=tensor([0.1066, 0.1708, 0.2113, 0.1825, 0.2286, 0.0910, 0.1582, 0.2479], device='cuda:4'), in_proj_covar=tensor([0.0400, 0.0575, 0.0634, 0.0482, 0.0641, 0.0667, 0.0494, 0.0647], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:08:14,916 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6805, 3.7693, 2.9636, 2.3082, 2.5991, 2.4583, 4.0889, 3.3907], device='cuda:4'), covar=tensor([0.2764, 0.0692, 0.1718, 0.2626, 0.2633, 0.1933, 0.0482, 0.1223], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0264, 0.0298, 0.0304, 0.0292, 0.0248, 0.0288, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:08:59,599 INFO [train.py:904] (4/8) Epoch 19, batch 5900, loss[loss=0.2076, simple_loss=0.3045, pruned_loss=0.05529, over 16164.00 frames. ], tot_loss[loss=0.2115, simple_loss=0.2962, pruned_loss=0.06345, over 3056668.55 frames. ], batch size: 35, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:09:02,422 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8462, 4.9054, 4.7083, 4.3370, 4.3458, 4.7846, 4.6211, 4.4771], device='cuda:4'), covar=tensor([0.0998, 0.1283, 0.0494, 0.0478, 0.1011, 0.0669, 0.0857, 0.1165], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0406, 0.0332, 0.0322, 0.0342, 0.0376, 0.0227, 0.0396], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:10:01,851 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=188639.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:10:19,632 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-30 22:10:21,847 INFO [train.py:904] (4/8) Epoch 19, batch 5950, loss[loss=0.215, simple_loss=0.2935, pruned_loss=0.06824, over 17042.00 frames. ], tot_loss[loss=0.2109, simple_loss=0.2967, pruned_loss=0.06261, over 3043733.41 frames. ], batch size: 55, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:10:22,482 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4839, 3.0519, 3.1906, 1.9619, 2.7599, 2.0860, 3.1922, 3.2750], device='cuda:4'), covar=tensor([0.0279, 0.0738, 0.0588, 0.2069, 0.0877, 0.1044, 0.0663, 0.0838], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0161, 0.0167, 0.0152, 0.0144, 0.0129, 0.0144, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 22:10:40,535 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.478e+02 2.950e+02 3.832e+02 6.444e+02, threshold=5.900e+02, percent-clipped=1.0 2023-04-30 22:10:40,872 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4069, 4.4831, 4.5924, 4.4431, 4.5154, 5.0163, 4.5490, 4.2930], device='cuda:4'), covar=tensor([0.1511, 0.2012, 0.2577, 0.2026, 0.2297, 0.1050, 0.1646, 0.2421], device='cuda:4'), in_proj_covar=tensor([0.0399, 0.0574, 0.0634, 0.0482, 0.0639, 0.0667, 0.0494, 0.0644], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:11:33,327 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5834, 3.6980, 2.8312, 2.2688, 2.5294, 2.3975, 4.0298, 3.4655], device='cuda:4'), covar=tensor([0.2986, 0.0720, 0.1795, 0.2549, 0.2522, 0.2016, 0.0440, 0.1224], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0263, 0.0297, 0.0303, 0.0290, 0.0247, 0.0287, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:11:41,785 INFO [train.py:904] (4/8) Epoch 19, batch 6000, loss[loss=0.2157, simple_loss=0.2891, pruned_loss=0.07113, over 11404.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2954, pruned_loss=0.06175, over 3052185.48 frames. ], batch size: 248, lr: 3.56e-03, grad_scale: 8.0 2023-04-30 22:11:41,786 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 22:11:52,558 INFO [train.py:938] (4/8) Epoch 19, validation: loss=0.1523, simple_loss=0.2653, pruned_loss=0.01968, over 944034.00 frames. 2023-04-30 22:11:52,558 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 22:12:06,393 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-30 22:13:13,860 INFO [train.py:904] (4/8) Epoch 19, batch 6050, loss[loss=0.1999, simple_loss=0.2959, pruned_loss=0.05197, over 16770.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.2936, pruned_loss=0.06085, over 3065761.91 frames. ], batch size: 124, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:13:33,103 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.872e+02 2.719e+02 3.422e+02 3.983e+02 7.831e+02, threshold=6.844e+02, percent-clipped=4.0 2023-04-30 22:14:32,938 INFO [train.py:904] (4/8) Epoch 19, batch 6100, loss[loss=0.1928, simple_loss=0.278, pruned_loss=0.05384, over 17165.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2929, pruned_loss=0.05956, over 3096174.15 frames. ], batch size: 46, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:15:33,786 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=188837.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:15:57,521 INFO [train.py:904] (4/8) Epoch 19, batch 6150, loss[loss=0.21, simple_loss=0.2953, pruned_loss=0.06242, over 16448.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2907, pruned_loss=0.05824, over 3116764.10 frames. ], batch size: 146, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:16:00,999 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6756, 4.8509, 5.0037, 4.8178, 4.8935, 5.4119, 4.8833, 4.6081], device='cuda:4'), covar=tensor([0.1146, 0.1859, 0.2133, 0.2200, 0.2656, 0.1007, 0.1651, 0.2671], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0578, 0.0638, 0.0487, 0.0644, 0.0670, 0.0501, 0.0649], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:16:16,882 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.816e+02 2.888e+02 3.402e+02 4.182e+02 1.052e+03, threshold=6.804e+02, percent-clipped=3.0 2023-04-30 22:17:16,339 INFO [train.py:904] (4/8) Epoch 19, batch 6200, loss[loss=0.2101, simple_loss=0.2918, pruned_loss=0.06416, over 15251.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2891, pruned_loss=0.05788, over 3128938.08 frames. ], batch size: 190, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:17:58,787 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=188929.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:18:01,719 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=188931.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:18:14,958 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=188939.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:18:34,122 INFO [train.py:904] (4/8) Epoch 19, batch 6250, loss[loss=0.2387, simple_loss=0.3084, pruned_loss=0.08447, over 11276.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.2886, pruned_loss=0.05779, over 3136730.70 frames. ], batch size: 248, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:18:52,889 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.881e+02 2.827e+02 3.410e+02 4.287e+02 1.283e+03, threshold=6.821e+02, percent-clipped=4.0 2023-04-30 22:19:28,070 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=188987.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:19:31,014 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9969, 3.0711, 3.0364, 5.1718, 3.9982, 4.4634, 1.9505, 3.2366], device='cuda:4'), covar=tensor([0.1365, 0.0812, 0.1126, 0.0204, 0.0447, 0.0453, 0.1620, 0.0858], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0172, 0.0194, 0.0184, 0.0205, 0.0214, 0.0197, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 22:19:32,248 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=188990.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:19:36,052 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=188992.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:19:51,028 INFO [train.py:904] (4/8) Epoch 19, batch 6300, loss[loss=0.176, simple_loss=0.2632, pruned_loss=0.04437, over 17094.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2881, pruned_loss=0.05681, over 3155830.04 frames. ], batch size: 49, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:20:35,803 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189030.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:21:05,468 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-30 22:21:09,036 INFO [train.py:904] (4/8) Epoch 19, batch 6350, loss[loss=0.1864, simple_loss=0.2774, pruned_loss=0.0477, over 17010.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2884, pruned_loss=0.05753, over 3154606.49 frames. ], batch size: 50, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:21:27,080 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.790e+02 2.691e+02 3.246e+02 4.132e+02 1.028e+03, threshold=6.492e+02, percent-clipped=3.0 2023-04-30 22:21:28,331 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189064.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:21:42,607 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.76 vs. limit=5.0 2023-04-30 22:21:51,870 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189079.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:22:10,455 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189091.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:22:26,882 INFO [train.py:904] (4/8) Epoch 19, batch 6400, loss[loss=0.2599, simple_loss=0.3279, pruned_loss=0.09599, over 11072.00 frames. ], tot_loss[loss=0.2032, simple_loss=0.2889, pruned_loss=0.05881, over 3144569.02 frames. ], batch size: 248, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:23:02,735 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189125.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:23:21,428 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=189137.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:23:25,616 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189140.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:23:25,891 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-30 22:23:43,274 INFO [train.py:904] (4/8) Epoch 19, batch 6450, loss[loss=0.1982, simple_loss=0.2952, pruned_loss=0.05059, over 16833.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2889, pruned_loss=0.05816, over 3155807.59 frames. ], batch size: 83, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:24:00,999 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.834e+02 2.902e+02 3.401e+02 4.159e+02 9.749e+02, threshold=6.803e+02, percent-clipped=3.0 2023-04-30 22:24:34,446 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=189185.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:25:00,511 INFO [train.py:904] (4/8) Epoch 19, batch 6500, loss[loss=0.2365, simple_loss=0.2968, pruned_loss=0.08805, over 11431.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2871, pruned_loss=0.05828, over 3132900.64 frames. ], batch size: 246, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:26:14,609 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189248.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:26:19,680 INFO [train.py:904] (4/8) Epoch 19, batch 6550, loss[loss=0.2145, simple_loss=0.3133, pruned_loss=0.05785, over 16538.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2899, pruned_loss=0.05959, over 3115780.27 frames. ], batch size: 62, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:26:37,017 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.957e+02 2.752e+02 3.137e+02 4.130e+02 8.401e+02, threshold=6.273e+02, percent-clipped=2.0 2023-04-30 22:27:09,601 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189285.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:27:12,558 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189287.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:27:26,493 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-30 22:27:34,958 INFO [train.py:904] (4/8) Epoch 19, batch 6600, loss[loss=0.2253, simple_loss=0.3121, pruned_loss=0.06929, over 16729.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2924, pruned_loss=0.05994, over 3126550.33 frames. ], batch size: 124, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:27:45,615 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189309.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:28:19,794 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6261, 3.7304, 2.8733, 2.2365, 2.4956, 2.3672, 4.0198, 3.4109], device='cuda:4'), covar=tensor([0.2883, 0.0714, 0.1713, 0.2590, 0.2611, 0.2089, 0.0419, 0.1121], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0264, 0.0299, 0.0304, 0.0292, 0.0250, 0.0288, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:28:23,915 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8503, 5.2404, 5.4008, 5.1870, 5.2909, 5.8060, 5.2985, 5.0444], device='cuda:4'), covar=tensor([0.0976, 0.1778, 0.2072, 0.1850, 0.2149, 0.0833, 0.1476, 0.2213], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0576, 0.0637, 0.0484, 0.0642, 0.0668, 0.0498, 0.0647], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:28:44,282 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189347.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:28:50,942 INFO [train.py:904] (4/8) Epoch 19, batch 6650, loss[loss=0.2069, simple_loss=0.2911, pruned_loss=0.06134, over 16694.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2938, pruned_loss=0.06113, over 3129891.49 frames. ], batch size: 124, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:29:08,220 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.792e+02 2.870e+02 3.524e+02 4.277e+02 9.308e+02, threshold=7.047e+02, percent-clipped=5.0 2023-04-30 22:29:42,431 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189386.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:30:06,163 INFO [train.py:904] (4/8) Epoch 19, batch 6700, loss[loss=0.1972, simple_loss=0.2837, pruned_loss=0.05539, over 15307.00 frames. ], tot_loss[loss=0.2069, simple_loss=0.2921, pruned_loss=0.06088, over 3116146.32 frames. ], batch size: 191, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:30:12,899 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3155, 5.6147, 5.3634, 5.3936, 5.0972, 4.9959, 5.0582, 5.7611], device='cuda:4'), covar=tensor([0.1181, 0.0863, 0.1003, 0.0854, 0.0769, 0.0757, 0.1071, 0.0771], device='cuda:4'), in_proj_covar=tensor([0.0642, 0.0782, 0.0645, 0.0584, 0.0490, 0.0502, 0.0651, 0.0604], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:30:15,404 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189408.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:30:33,775 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189420.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:30:44,552 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.64 vs. limit=5.0 2023-04-30 22:30:56,600 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189435.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:31:21,912 INFO [train.py:904] (4/8) Epoch 19, batch 6750, loss[loss=0.2611, simple_loss=0.3222, pruned_loss=0.09995, over 11738.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2914, pruned_loss=0.06085, over 3118524.48 frames. ], batch size: 248, lr: 3.55e-03, grad_scale: 4.0 2023-04-30 22:31:42,235 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.872e+02 2.911e+02 3.292e+02 3.890e+02 7.012e+02, threshold=6.585e+02, percent-clipped=0.0 2023-04-30 22:32:29,192 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4870, 2.2030, 1.8225, 1.9655, 2.5219, 2.1455, 2.3580, 2.6538], device='cuda:4'), covar=tensor([0.0183, 0.0401, 0.0487, 0.0447, 0.0233, 0.0379, 0.0195, 0.0228], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0224, 0.0217, 0.0217, 0.0225, 0.0224, 0.0225, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:32:38,330 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189501.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:32:39,013 INFO [train.py:904] (4/8) Epoch 19, batch 6800, loss[loss=0.2362, simple_loss=0.3234, pruned_loss=0.07454, over 16832.00 frames. ], tot_loss[loss=0.2072, simple_loss=0.2916, pruned_loss=0.06133, over 3108250.98 frames. ], batch size: 39, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:33:28,341 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5098, 3.7116, 2.5816, 2.1657, 2.5988, 2.3122, 3.7824, 3.3724], device='cuda:4'), covar=tensor([0.3035, 0.0664, 0.1978, 0.2502, 0.2412, 0.2064, 0.0610, 0.1181], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0264, 0.0298, 0.0303, 0.0292, 0.0250, 0.0288, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:33:58,200 INFO [train.py:904] (4/8) Epoch 19, batch 6850, loss[loss=0.2194, simple_loss=0.3209, pruned_loss=0.05895, over 16711.00 frames. ], tot_loss[loss=0.2083, simple_loss=0.2934, pruned_loss=0.06157, over 3108906.02 frames. ], batch size: 62, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:34:13,553 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189562.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:34:17,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.892e+02 2.732e+02 3.346e+02 4.467e+02 6.913e+02, threshold=6.693e+02, percent-clipped=4.0 2023-04-30 22:34:23,590 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4438, 2.9112, 3.0208, 1.9190, 2.6982, 2.1442, 3.0356, 3.1740], device='cuda:4'), covar=tensor([0.0274, 0.0773, 0.0626, 0.2052, 0.0874, 0.1044, 0.0692, 0.0885], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0162, 0.0168, 0.0152, 0.0145, 0.0130, 0.0144, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 22:34:35,691 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-30 22:34:47,118 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=189585.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:34:50,443 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=189587.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:35:04,590 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-30 22:35:12,995 INFO [train.py:904] (4/8) Epoch 19, batch 6900, loss[loss=0.251, simple_loss=0.3177, pruned_loss=0.09216, over 11709.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2958, pruned_loss=0.06162, over 3104210.72 frames. ], batch size: 246, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:35:17,328 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189604.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:35:40,464 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.49 vs. limit=5.0 2023-04-30 22:36:00,957 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=189633.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:36:04,231 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=189635.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:36:12,151 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189640.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:36:29,234 INFO [train.py:904] (4/8) Epoch 19, batch 6950, loss[loss=0.2705, simple_loss=0.3296, pruned_loss=0.1057, over 11374.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.2976, pruned_loss=0.06346, over 3085614.91 frames. ], batch size: 248, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:36:48,877 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.017e+02 3.156e+02 3.874e+02 4.949e+02 9.528e+02, threshold=7.748e+02, percent-clipped=6.0 2023-04-30 22:37:01,271 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9671, 3.4158, 3.3828, 1.9977, 2.8479, 2.4266, 3.4389, 3.5953], device='cuda:4'), covar=tensor([0.0320, 0.0766, 0.0699, 0.2161, 0.0972, 0.1005, 0.0756, 0.0973], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0161, 0.0167, 0.0152, 0.0144, 0.0129, 0.0143, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 22:37:21,357 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=189686.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:37:43,891 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189701.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:37:44,543 INFO [train.py:904] (4/8) Epoch 19, batch 7000, loss[loss=0.1938, simple_loss=0.2992, pruned_loss=0.04423, over 17125.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2971, pruned_loss=0.06196, over 3110571.71 frames. ], batch size: 47, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:37:47,741 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189703.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:37:49,124 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189704.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:38:13,451 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=189720.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:38:25,592 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=189728.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:38:34,807 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=189734.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:38:36,107 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=189735.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:38:59,649 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-30 22:39:01,643 INFO [train.py:904] (4/8) Epoch 19, batch 7050, loss[loss=0.209, simple_loss=0.2944, pruned_loss=0.06185, over 16263.00 frames. ], tot_loss[loss=0.2119, simple_loss=0.2987, pruned_loss=0.06258, over 3092688.41 frames. ], batch size: 165, lr: 3.55e-03, grad_scale: 8.0 2023-04-30 22:39:09,277 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-30 22:39:22,079 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.118e+02 2.966e+02 3.403e+02 4.255e+02 9.294e+02, threshold=6.806e+02, percent-clipped=3.0 2023-04-30 22:39:23,216 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189765.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:39:27,397 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=189768.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:39:50,863 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=189783.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:40:00,679 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=189789.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:40:19,983 INFO [train.py:904] (4/8) Epoch 19, batch 7100, loss[loss=0.2191, simple_loss=0.3113, pruned_loss=0.0634, over 16411.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2967, pruned_loss=0.06167, over 3099878.13 frames. ], batch size: 146, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:40:25,933 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0303, 5.0004, 4.8204, 4.1676, 4.9180, 1.9482, 4.6943, 4.6140], device='cuda:4'), covar=tensor([0.0073, 0.0067, 0.0164, 0.0334, 0.0075, 0.2447, 0.0108, 0.0192], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0144, 0.0190, 0.0173, 0.0165, 0.0199, 0.0179, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:40:37,466 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7878, 3.1599, 3.2745, 1.9552, 2.7761, 1.9910, 3.4323, 3.3921], device='cuda:4'), covar=tensor([0.0256, 0.0776, 0.0609, 0.2065, 0.0895, 0.1095, 0.0606, 0.0887], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0161, 0.0166, 0.0152, 0.0144, 0.0129, 0.0143, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 22:41:38,322 INFO [train.py:904] (4/8) Epoch 19, batch 7150, loss[loss=0.2118, simple_loss=0.289, pruned_loss=0.06731, over 16581.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2942, pruned_loss=0.061, over 3096560.17 frames. ], batch size: 62, lr: 3.54e-03, grad_scale: 4.0 2023-04-30 22:41:45,504 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189857.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:41:58,949 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.902e+02 2.755e+02 3.301e+02 3.879e+02 9.632e+02, threshold=6.601e+02, percent-clipped=3.0 2023-04-30 22:42:07,497 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-30 22:42:12,688 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4565, 4.4448, 4.3035, 3.6097, 4.3266, 1.7688, 4.1224, 3.9830], device='cuda:4'), covar=tensor([0.0097, 0.0089, 0.0176, 0.0377, 0.0111, 0.2675, 0.0132, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0144, 0.0190, 0.0173, 0.0165, 0.0199, 0.0179, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:42:21,026 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8674, 2.7348, 2.6285, 1.9314, 2.5511, 2.6592, 2.5760, 1.8689], device='cuda:4'), covar=tensor([0.0402, 0.0073, 0.0071, 0.0363, 0.0127, 0.0108, 0.0112, 0.0384], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0078, 0.0078, 0.0132, 0.0094, 0.0105, 0.0090, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:42:51,645 INFO [train.py:904] (4/8) Epoch 19, batch 7200, loss[loss=0.1932, simple_loss=0.2749, pruned_loss=0.05574, over 11469.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2924, pruned_loss=0.05988, over 3062750.52 frames. ], batch size: 247, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:42:55,742 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=189904.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:44:12,493 INFO [train.py:904] (4/8) Epoch 19, batch 7250, loss[loss=0.1875, simple_loss=0.2696, pruned_loss=0.05273, over 16616.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2898, pruned_loss=0.05836, over 3083083.36 frames. ], batch size: 62, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:44:12,829 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=189952.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:44:34,219 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.477e+02 2.873e+02 3.623e+02 8.553e+02, threshold=5.746e+02, percent-clipped=4.0 2023-04-30 22:44:43,103 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5912, 1.7359, 2.2517, 2.5396, 2.4547, 2.8931, 1.9748, 2.7630], device='cuda:4'), covar=tensor([0.0207, 0.0500, 0.0292, 0.0310, 0.0326, 0.0163, 0.0463, 0.0132], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0189, 0.0174, 0.0178, 0.0190, 0.0148, 0.0191, 0.0141], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:45:19,910 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=189996.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:45:32,080 INFO [train.py:904] (4/8) Epoch 19, batch 7300, loss[loss=0.2177, simple_loss=0.3051, pruned_loss=0.06511, over 16329.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.289, pruned_loss=0.05773, over 3096098.97 frames. ], batch size: 146, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:45:33,920 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=190003.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:45:34,315 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.06 vs. limit=5.0 2023-04-30 22:46:45,701 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6991, 3.6732, 2.4948, 4.3698, 2.9112, 4.3477, 2.5132, 2.9933], device='cuda:4'), covar=tensor([0.0246, 0.0388, 0.1430, 0.0126, 0.0782, 0.0357, 0.1375, 0.0751], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0174, 0.0192, 0.0154, 0.0173, 0.0212, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-30 22:46:48,531 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=190051.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:46:49,321 INFO [train.py:904] (4/8) Epoch 19, batch 7350, loss[loss=0.2171, simple_loss=0.2873, pruned_loss=0.0734, over 11022.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2889, pruned_loss=0.05786, over 3114357.11 frames. ], batch size: 250, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:47:01,677 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=190060.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:47:10,691 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.253e+02 3.016e+02 3.386e+02 4.070e+02 1.285e+03, threshold=6.773e+02, percent-clipped=7.0 2023-04-30 22:47:29,498 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5869, 4.8448, 4.6224, 4.6126, 4.3230, 4.3268, 4.3264, 4.9027], device='cuda:4'), covar=tensor([0.1023, 0.0788, 0.0947, 0.0859, 0.0815, 0.1209, 0.1109, 0.0810], device='cuda:4'), in_proj_covar=tensor([0.0631, 0.0769, 0.0636, 0.0575, 0.0483, 0.0496, 0.0643, 0.0593], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:47:41,034 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=190084.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:48:08,858 INFO [train.py:904] (4/8) Epoch 19, batch 7400, loss[loss=0.1849, simple_loss=0.2805, pruned_loss=0.04466, over 16675.00 frames. ], tot_loss[loss=0.2035, simple_loss=0.2901, pruned_loss=0.05846, over 3115882.19 frames. ], batch size: 76, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:48:45,678 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.29 vs. limit=5.0 2023-04-30 22:49:09,856 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=190139.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 22:49:30,153 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=190151.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:49:30,830 INFO [train.py:904] (4/8) Epoch 19, batch 7450, loss[loss=0.2175, simple_loss=0.31, pruned_loss=0.06247, over 15292.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2918, pruned_loss=0.05976, over 3116896.39 frames. ], batch size: 190, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:49:37,121 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.85 vs. limit=5.0 2023-04-30 22:49:41,025 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=190157.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:49:55,755 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.975e+02 2.938e+02 3.493e+02 4.428e+02 1.014e+03, threshold=6.986e+02, percent-clipped=6.0 2023-04-30 22:50:04,924 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.80 vs. limit=5.0 2023-04-30 22:50:42,335 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-30 22:50:51,765 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=190200.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 22:50:54,217 INFO [train.py:904] (4/8) Epoch 19, batch 7500, loss[loss=0.2002, simple_loss=0.2903, pruned_loss=0.05504, over 16730.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2921, pruned_loss=0.05945, over 3094180.52 frames. ], batch size: 89, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:50:59,307 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=190205.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:51:00,696 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4938, 3.5065, 3.4512, 2.7320, 3.4013, 2.0640, 3.1554, 2.7376], device='cuda:4'), covar=tensor([0.0156, 0.0125, 0.0186, 0.0217, 0.0105, 0.2208, 0.0121, 0.0224], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0143, 0.0189, 0.0172, 0.0164, 0.0199, 0.0178, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:51:10,844 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=190212.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:52:00,588 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-30 22:52:13,612 INFO [train.py:904] (4/8) Epoch 19, batch 7550, loss[loss=0.2297, simple_loss=0.3057, pruned_loss=0.07688, over 16688.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.291, pruned_loss=0.06, over 3068413.49 frames. ], batch size: 57, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:52:34,940 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.804e+02 2.643e+02 3.256e+02 4.054e+02 9.532e+02, threshold=6.511e+02, percent-clipped=2.0 2023-04-30 22:53:22,450 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=190296.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:53:31,290 INFO [train.py:904] (4/8) Epoch 19, batch 7600, loss[loss=0.2344, simple_loss=0.2986, pruned_loss=0.08511, over 11386.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2897, pruned_loss=0.05986, over 3082110.09 frames. ], batch size: 246, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:53:53,505 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=190316.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:54:36,981 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=190344.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:54:48,372 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1701, 4.1885, 4.0855, 3.3080, 4.1378, 1.5900, 3.9089, 3.7026], device='cuda:4'), covar=tensor([0.0132, 0.0099, 0.0197, 0.0351, 0.0097, 0.2869, 0.0148, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0143, 0.0189, 0.0172, 0.0165, 0.0199, 0.0178, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 22:54:48,951 INFO [train.py:904] (4/8) Epoch 19, batch 7650, loss[loss=0.2251, simple_loss=0.3156, pruned_loss=0.06733, over 16222.00 frames. ], tot_loss[loss=0.2057, simple_loss=0.2909, pruned_loss=0.06022, over 3093899.09 frames. ], batch size: 165, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:55:02,303 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=190360.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:55:10,291 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.023e+02 2.803e+02 3.556e+02 4.194e+02 6.713e+02, threshold=7.113e+02, percent-clipped=1.0 2023-04-30 22:55:27,579 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=190377.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:55:38,582 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=190384.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:56:02,469 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=190399.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:56:06,244 INFO [train.py:904] (4/8) Epoch 19, batch 7700, loss[loss=0.2146, simple_loss=0.3012, pruned_loss=0.06404, over 16555.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2917, pruned_loss=0.06149, over 3069661.94 frames. ], batch size: 75, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:56:15,129 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=190408.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:56:45,744 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2689, 3.4420, 3.6044, 3.5758, 3.5885, 3.3920, 3.4445, 3.5012], device='cuda:4'), covar=tensor([0.0439, 0.0768, 0.0479, 0.0490, 0.0572, 0.0610, 0.0898, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0389, 0.0428, 0.0415, 0.0389, 0.0463, 0.0436, 0.0530, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 22:56:50,414 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=190432.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:57:21,407 INFO [train.py:904] (4/8) Epoch 19, batch 7750, loss[loss=0.2187, simple_loss=0.3029, pruned_loss=0.06724, over 16725.00 frames. ], tot_loss[loss=0.207, simple_loss=0.2917, pruned_loss=0.06118, over 3072470.39 frames. ], batch size: 83, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:57:35,048 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=190460.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:57:43,808 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.766e+02 2.725e+02 3.462e+02 4.075e+02 8.569e+02, threshold=6.924e+02, percent-clipped=2.0 2023-04-30 22:58:30,671 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=190495.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 22:58:39,265 INFO [train.py:904] (4/8) Epoch 19, batch 7800, loss[loss=0.1825, simple_loss=0.2784, pruned_loss=0.04326, over 16427.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2926, pruned_loss=0.06153, over 3068961.72 frames. ], batch size: 68, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 22:58:47,775 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=190507.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 22:58:56,693 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-30 22:59:55,261 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3054, 3.8192, 3.8156, 2.5077, 3.5461, 3.8742, 3.5119, 2.1651], device='cuda:4'), covar=tensor([0.0509, 0.0055, 0.0049, 0.0384, 0.0091, 0.0099, 0.0090, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0080, 0.0081, 0.0135, 0.0095, 0.0108, 0.0093, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 22:59:57,134 INFO [train.py:904] (4/8) Epoch 19, batch 7850, loss[loss=0.2254, simple_loss=0.3073, pruned_loss=0.07172, over 15301.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2935, pruned_loss=0.06122, over 3080848.83 frames. ], batch size: 191, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 23:00:17,838 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.643e+02 2.838e+02 3.506e+02 4.307e+02 8.316e+02, threshold=7.012e+02, percent-clipped=1.0 2023-04-30 23:00:56,950 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=190591.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 23:01:12,372 INFO [train.py:904] (4/8) Epoch 19, batch 7900, loss[loss=0.2257, simple_loss=0.3183, pruned_loss=0.06653, over 16685.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2921, pruned_loss=0.06005, over 3107430.85 frames. ], batch size: 134, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 23:01:23,409 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=190609.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:01:56,068 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-30 23:02:32,481 INFO [train.py:904] (4/8) Epoch 19, batch 7950, loss[loss=0.2114, simple_loss=0.2966, pruned_loss=0.06314, over 16238.00 frames. ], tot_loss[loss=0.2069, simple_loss=0.2924, pruned_loss=0.0607, over 3093008.53 frames. ], batch size: 165, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 23:02:32,973 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=190652.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 23:02:54,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.686e+02 2.695e+02 3.262e+02 3.966e+02 7.542e+02, threshold=6.523e+02, percent-clipped=2.0 2023-04-30 23:03:01,381 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=190670.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 23:03:04,351 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=190672.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:03:08,670 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=190675.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:03:49,816 INFO [train.py:904] (4/8) Epoch 19, batch 8000, loss[loss=0.2009, simple_loss=0.2821, pruned_loss=0.05983, over 16626.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.2929, pruned_loss=0.06122, over 3092489.41 frames. ], batch size: 76, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 23:04:44,419 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=190736.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:05:07,643 INFO [train.py:904] (4/8) Epoch 19, batch 8050, loss[loss=0.2292, simple_loss=0.2957, pruned_loss=0.08135, over 11551.00 frames. ], tot_loss[loss=0.2066, simple_loss=0.2925, pruned_loss=0.06033, over 3102389.91 frames. ], batch size: 246, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 23:05:12,104 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=190755.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:05:28,126 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.906e+02 2.693e+02 3.232e+02 4.056e+02 6.686e+02, threshold=6.463e+02, percent-clipped=1.0 2023-04-30 23:06:12,490 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=190795.0, num_to_drop=1, layers_to_drop={1} 2023-04-30 23:06:22,312 INFO [train.py:904] (4/8) Epoch 19, batch 8100, loss[loss=0.2025, simple_loss=0.2707, pruned_loss=0.06715, over 11498.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.2917, pruned_loss=0.05992, over 3108727.26 frames. ], batch size: 246, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 23:06:31,312 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=190807.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:07:25,608 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=190843.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 23:07:37,463 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3700, 2.4531, 2.3851, 4.3080, 2.4140, 2.7395, 2.4891, 2.5448], device='cuda:4'), covar=tensor([0.1237, 0.3149, 0.2660, 0.0421, 0.3576, 0.2263, 0.3185, 0.2923], device='cuda:4'), in_proj_covar=tensor([0.0389, 0.0433, 0.0355, 0.0319, 0.0431, 0.0497, 0.0404, 0.0505], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:07:39,349 INFO [train.py:904] (4/8) Epoch 19, batch 8150, loss[loss=0.1931, simple_loss=0.2772, pruned_loss=0.05447, over 15294.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2892, pruned_loss=0.0588, over 3122935.94 frames. ], batch size: 190, lr: 3.54e-03, grad_scale: 8.0 2023-04-30 23:07:44,596 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=190855.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:08:01,318 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.839e+02 2.760e+02 3.332e+02 4.142e+02 6.995e+02, threshold=6.664e+02, percent-clipped=2.0 2023-04-30 23:08:22,418 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4996, 5.8675, 5.5804, 5.6451, 5.2741, 5.1886, 5.3086, 6.0013], device='cuda:4'), covar=tensor([0.1152, 0.0786, 0.0910, 0.0798, 0.0727, 0.0690, 0.1043, 0.0783], device='cuda:4'), in_proj_covar=tensor([0.0638, 0.0772, 0.0642, 0.0581, 0.0485, 0.0498, 0.0648, 0.0598], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:08:57,387 INFO [train.py:904] (4/8) Epoch 19, batch 8200, loss[loss=0.1705, simple_loss=0.2664, pruned_loss=0.03729, over 16864.00 frames. ], tot_loss[loss=0.2013, simple_loss=0.2866, pruned_loss=0.058, over 3111669.98 frames. ], batch size: 96, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:10:11,509 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=190947.0, num_to_drop=1, layers_to_drop={3} 2023-04-30 23:10:19,456 INFO [train.py:904] (4/8) Epoch 19, batch 8250, loss[loss=0.1999, simple_loss=0.2926, pruned_loss=0.05364, over 16849.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2854, pruned_loss=0.05578, over 3085066.71 frames. ], batch size: 116, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:10:39,903 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=190965.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 23:10:41,153 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.632e+02 2.606e+02 3.069e+02 3.589e+02 7.796e+02, threshold=6.137e+02, percent-clipped=2.0 2023-04-30 23:10:51,836 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=190972.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:10:55,293 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-30 23:11:02,534 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5831, 3.6099, 3.4241, 3.1597, 3.1877, 3.5383, 3.3091, 3.3252], device='cuda:4'), covar=tensor([0.0561, 0.0637, 0.0279, 0.0276, 0.0509, 0.0478, 0.1414, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0277, 0.0403, 0.0326, 0.0317, 0.0335, 0.0371, 0.0225, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:11:39,003 INFO [train.py:904] (4/8) Epoch 19, batch 8300, loss[loss=0.1776, simple_loss=0.2609, pruned_loss=0.04716, over 11775.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2829, pruned_loss=0.05315, over 3081096.00 frames. ], batch size: 247, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:12:10,264 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=191020.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:12:27,700 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=191031.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:13:00,825 INFO [train.py:904] (4/8) Epoch 19, batch 8350, loss[loss=0.1656, simple_loss=0.2598, pruned_loss=0.03573, over 17192.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2818, pruned_loss=0.05092, over 3074020.17 frames. ], batch size: 44, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:13:06,536 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-30 23:13:08,041 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=191055.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:13:24,344 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.632e+02 2.343e+02 2.792e+02 3.340e+02 8.341e+02, threshold=5.585e+02, percent-clipped=1.0 2023-04-30 23:13:28,640 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=191068.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:14:22,693 INFO [train.py:904] (4/8) Epoch 19, batch 8400, loss[loss=0.1641, simple_loss=0.2527, pruned_loss=0.03776, over 12221.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.279, pruned_loss=0.04891, over 3062297.55 frames. ], batch size: 247, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:14:24,464 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=191103.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:15:06,919 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=191129.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:15:11,013 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=191132.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:15:42,699 INFO [train.py:904] (4/8) Epoch 19, batch 8450, loss[loss=0.173, simple_loss=0.2689, pruned_loss=0.03859, over 16243.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2775, pruned_loss=0.04758, over 3068999.10 frames. ], batch size: 165, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:15:57,737 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1235, 3.1381, 1.9302, 3.3282, 2.3716, 3.3466, 2.1164, 2.6469], device='cuda:4'), covar=tensor([0.0259, 0.0353, 0.1528, 0.0284, 0.0782, 0.0580, 0.1497, 0.0660], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0170, 0.0189, 0.0151, 0.0170, 0.0208, 0.0196, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 23:16:06,334 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.247e+02 2.599e+02 3.228e+02 6.628e+02, threshold=5.197e+02, percent-clipped=1.0 2023-04-30 23:16:13,909 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9075, 2.7321, 2.9544, 2.0884, 2.7713, 2.1847, 2.7218, 2.9303], device='cuda:4'), covar=tensor([0.0271, 0.0829, 0.0435, 0.1802, 0.0699, 0.0952, 0.0613, 0.0831], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0156, 0.0162, 0.0147, 0.0141, 0.0125, 0.0140, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 23:16:21,238 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4014, 3.4563, 2.1101, 3.7494, 2.6048, 3.7591, 2.2518, 2.8639], device='cuda:4'), covar=tensor([0.0272, 0.0312, 0.1435, 0.0230, 0.0716, 0.0469, 0.1422, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0170, 0.0188, 0.0150, 0.0170, 0.0207, 0.0196, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 23:16:37,007 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=191185.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:16:48,374 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8603, 2.2781, 1.9702, 1.9455, 2.5845, 2.2337, 2.4212, 2.7034], device='cuda:4'), covar=tensor([0.0165, 0.0391, 0.0479, 0.0485, 0.0253, 0.0368, 0.0236, 0.0271], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0223, 0.0215, 0.0215, 0.0224, 0.0221, 0.0223, 0.0217], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:16:50,121 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=191193.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:17:04,704 INFO [train.py:904] (4/8) Epoch 19, batch 8500, loss[loss=0.158, simple_loss=0.2512, pruned_loss=0.03241, over 16770.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2742, pruned_loss=0.04567, over 3061574.96 frames. ], batch size: 124, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:18:22,853 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=191246.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:18:24,091 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=191247.0, num_to_drop=1, layers_to_drop={2} 2023-04-30 23:18:32,471 INFO [train.py:904] (4/8) Epoch 19, batch 8550, loss[loss=0.1618, simple_loss=0.2545, pruned_loss=0.03453, over 17203.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2718, pruned_loss=0.04505, over 3029125.01 frames. ], batch size: 44, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:18:55,733 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9127, 2.7883, 2.6044, 1.9093, 2.5880, 2.7473, 2.6641, 1.9230], device='cuda:4'), covar=tensor([0.0427, 0.0071, 0.0064, 0.0347, 0.0120, 0.0095, 0.0096, 0.0406], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0078, 0.0079, 0.0132, 0.0093, 0.0105, 0.0090, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-30 23:18:57,563 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=191265.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:19:00,615 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.506e+02 2.304e+02 2.724e+02 3.153e+02 7.683e+02, threshold=5.448e+02, percent-clipped=1.0 2023-04-30 23:19:44,562 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-30 23:20:00,162 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=191295.0, num_to_drop=1, layers_to_drop={0} 2023-04-30 23:20:13,437 INFO [train.py:904] (4/8) Epoch 19, batch 8600, loss[loss=0.1581, simple_loss=0.2547, pruned_loss=0.03082, over 16762.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2721, pruned_loss=0.0439, over 3027008.60 frames. ], batch size: 76, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:20:37,121 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=191313.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:21:07,067 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-04-30 23:21:12,190 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=191331.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:21:47,574 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.90 vs. limit=5.0 2023-04-30 23:21:52,931 INFO [train.py:904] (4/8) Epoch 19, batch 8650, loss[loss=0.1621, simple_loss=0.2662, pruned_loss=0.02903, over 16460.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2701, pruned_loss=0.04248, over 3029238.18 frames. ], batch size: 68, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:22:25,737 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=191365.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:22:29,946 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.572e+02 2.144e+02 2.571e+02 3.162e+02 5.065e+02, threshold=5.143e+02, percent-clipped=0.0 2023-04-30 23:22:55,703 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=191379.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:23:04,874 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8289, 1.8880, 2.3495, 2.7673, 2.6387, 3.1226, 2.1315, 3.1168], device='cuda:4'), covar=tensor([0.0181, 0.0528, 0.0304, 0.0272, 0.0296, 0.0159, 0.0452, 0.0122], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0186, 0.0171, 0.0175, 0.0187, 0.0145, 0.0188, 0.0139], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:23:39,020 INFO [train.py:904] (4/8) Epoch 19, batch 8700, loss[loss=0.1829, simple_loss=0.2802, pruned_loss=0.04277, over 16847.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2678, pruned_loss=0.0412, over 3037124.44 frames. ], batch size: 124, lr: 3.53e-03, grad_scale: 2.0 2023-04-30 23:24:20,671 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=191424.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:24:23,613 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=191426.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:25:14,445 INFO [train.py:904] (4/8) Epoch 19, batch 8750, loss[loss=0.1755, simple_loss=0.2777, pruned_loss=0.03661, over 16529.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2676, pruned_loss=0.04081, over 3044076.60 frames. ], batch size: 68, lr: 3.53e-03, grad_scale: 2.0 2023-04-30 23:25:47,579 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-30 23:25:58,340 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.289e+02 2.109e+02 2.725e+02 3.287e+02 5.129e+02, threshold=5.449e+02, percent-clipped=0.0 2023-04-30 23:26:40,173 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=191488.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:26:53,475 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7433, 3.7748, 2.3546, 4.4015, 2.8307, 4.2807, 2.4102, 3.0280], device='cuda:4'), covar=tensor([0.0239, 0.0322, 0.1447, 0.0155, 0.0784, 0.0431, 0.1456, 0.0709], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0168, 0.0188, 0.0149, 0.0169, 0.0206, 0.0195, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 23:27:07,632 INFO [train.py:904] (4/8) Epoch 19, batch 8800, loss[loss=0.17, simple_loss=0.268, pruned_loss=0.036, over 16219.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2655, pruned_loss=0.03938, over 3047937.82 frames. ], batch size: 165, lr: 3.53e-03, grad_scale: 4.0 2023-04-30 23:28:28,983 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=191541.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:28:51,736 INFO [train.py:904] (4/8) Epoch 19, batch 8850, loss[loss=0.1857, simple_loss=0.2845, pruned_loss=0.04341, over 16358.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2689, pruned_loss=0.03943, over 3048529.20 frames. ], batch size: 146, lr: 3.53e-03, grad_scale: 4.0 2023-04-30 23:29:28,628 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.662e+02 2.105e+02 2.539e+02 3.345e+02 5.950e+02, threshold=5.078e+02, percent-clipped=2.0 2023-04-30 23:30:39,776 INFO [train.py:904] (4/8) Epoch 19, batch 8900, loss[loss=0.1666, simple_loss=0.2599, pruned_loss=0.03664, over 12581.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2698, pruned_loss=0.03927, over 3050735.91 frames. ], batch size: 246, lr: 3.53e-03, grad_scale: 4.0 2023-04-30 23:32:45,267 INFO [train.py:904] (4/8) Epoch 19, batch 8950, loss[loss=0.1721, simple_loss=0.2728, pruned_loss=0.03571, over 16989.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2688, pruned_loss=0.03953, over 3041918.61 frames. ], batch size: 109, lr: 3.53e-03, grad_scale: 4.0 2023-04-30 23:32:50,366 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-30 23:33:21,158 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.530e+02 2.162e+02 2.456e+02 2.877e+02 5.499e+02, threshold=4.911e+02, percent-clipped=2.0 2023-04-30 23:34:34,392 INFO [train.py:904] (4/8) Epoch 19, batch 9000, loss[loss=0.1582, simple_loss=0.2544, pruned_loss=0.03103, over 16282.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.266, pruned_loss=0.03835, over 3059198.92 frames. ], batch size: 146, lr: 3.53e-03, grad_scale: 4.0 2023-04-30 23:34:34,392 INFO [train.py:929] (4/8) Computing validation loss 2023-04-30 23:34:41,199 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([6.3246, 6.5846, 6.3833, 6.5487, 6.2082, 6.2023, 6.0749, 6.6250], device='cuda:4'), covar=tensor([0.0782, 0.0553, 0.0583, 0.0467, 0.0481, 0.0184, 0.0788, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0626, 0.0763, 0.0628, 0.0569, 0.0479, 0.0491, 0.0638, 0.0587], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:34:44,207 INFO [train.py:938] (4/8) Epoch 19, validation: loss=0.1462, simple_loss=0.2506, pruned_loss=0.02087, over 944034.00 frames. 2023-04-30 23:34:44,208 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-04-30 23:35:24,065 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-30 23:35:25,456 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=191721.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:35:31,565 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=191724.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:35:53,832 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3154, 2.1778, 2.1570, 4.1087, 2.1940, 2.5378, 2.2650, 2.3561], device='cuda:4'), covar=tensor([0.1140, 0.3743, 0.3074, 0.0429, 0.4163, 0.2495, 0.3682, 0.3451], device='cuda:4'), in_proj_covar=tensor([0.0381, 0.0425, 0.0351, 0.0312, 0.0423, 0.0487, 0.0397, 0.0495], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:36:27,796 INFO [train.py:904] (4/8) Epoch 19, batch 9050, loss[loss=0.167, simple_loss=0.2498, pruned_loss=0.04207, over 16504.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2667, pruned_loss=0.03899, over 3052275.31 frames. ], batch size: 62, lr: 3.53e-03, grad_scale: 4.0 2023-04-30 23:37:04,254 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 2.085e+02 2.454e+02 3.075e+02 7.905e+02, threshold=4.907e+02, percent-clipped=4.0 2023-04-30 23:37:09,313 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=191772.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:37:20,905 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6578, 3.6993, 3.4753, 3.1774, 3.2287, 3.6213, 3.3579, 3.4095], device='cuda:4'), covar=tensor([0.0518, 0.0621, 0.0323, 0.0261, 0.0560, 0.0492, 0.1400, 0.0530], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0390, 0.0316, 0.0308, 0.0323, 0.0359, 0.0219, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:37:39,996 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=191788.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:38:10,475 INFO [train.py:904] (4/8) Epoch 19, batch 9100, loss[loss=0.1747, simple_loss=0.2748, pruned_loss=0.03726, over 15357.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2662, pruned_loss=0.03907, over 3063212.67 frames. ], batch size: 191, lr: 3.53e-03, grad_scale: 4.0 2023-04-30 23:38:30,700 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.99 vs. limit=5.0 2023-04-30 23:39:32,420 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=191836.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:39:44,001 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=191841.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:40:08,191 INFO [train.py:904] (4/8) Epoch 19, batch 9150, loss[loss=0.1661, simple_loss=0.2596, pruned_loss=0.03632, over 15349.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2667, pruned_loss=0.03864, over 3052244.31 frames. ], batch size: 190, lr: 3.53e-03, grad_scale: 4.0 2023-04-30 23:40:46,342 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.543e+02 2.243e+02 2.682e+02 3.704e+02 5.734e+02, threshold=5.364e+02, percent-clipped=7.0 2023-04-30 23:41:30,832 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=191889.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:41:52,732 INFO [train.py:904] (4/8) Epoch 19, batch 9200, loss[loss=0.1671, simple_loss=0.2562, pruned_loss=0.03901, over 16972.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2629, pruned_loss=0.03768, over 3075117.68 frames. ], batch size: 109, lr: 3.53e-03, grad_scale: 8.0 2023-04-30 23:42:22,620 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1049, 2.5518, 2.5789, 1.9171, 2.7582, 2.8554, 2.4683, 2.4990], device='cuda:4'), covar=tensor([0.0578, 0.0248, 0.0254, 0.0934, 0.0098, 0.0191, 0.0415, 0.0401], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0102, 0.0091, 0.0134, 0.0074, 0.0115, 0.0121, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-30 23:42:39,163 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-30 23:43:29,321 INFO [train.py:904] (4/8) Epoch 19, batch 9250, loss[loss=0.1534, simple_loss=0.2513, pruned_loss=0.02778, over 16202.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2629, pruned_loss=0.03772, over 3067662.90 frames. ], batch size: 165, lr: 3.52e-03, grad_scale: 8.0 2023-04-30 23:44:05,909 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.629e+02 2.331e+02 2.686e+02 3.317e+02 7.389e+02, threshold=5.371e+02, percent-clipped=4.0 2023-04-30 23:44:42,358 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8135, 4.8605, 4.6887, 4.2788, 4.3430, 4.7545, 4.5908, 4.4497], device='cuda:4'), covar=tensor([0.0559, 0.0517, 0.0273, 0.0277, 0.0887, 0.0517, 0.0380, 0.0605], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0388, 0.0315, 0.0306, 0.0322, 0.0359, 0.0219, 0.0375], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:45:23,367 INFO [train.py:904] (4/8) Epoch 19, batch 9300, loss[loss=0.158, simple_loss=0.244, pruned_loss=0.03598, over 12347.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2611, pruned_loss=0.03743, over 3045818.85 frames. ], batch size: 250, lr: 3.52e-03, grad_scale: 4.0 2023-04-30 23:46:09,691 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=192021.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:47:09,785 INFO [train.py:904] (4/8) Epoch 19, batch 9350, loss[loss=0.1865, simple_loss=0.2753, pruned_loss=0.04889, over 16795.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.261, pruned_loss=0.03733, over 3060274.85 frames. ], batch size: 124, lr: 3.52e-03, grad_scale: 4.0 2023-04-30 23:47:28,507 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7889, 3.8410, 4.1303, 4.1244, 4.1161, 3.9090, 3.8968, 3.9540], device='cuda:4'), covar=tensor([0.0568, 0.1425, 0.0745, 0.0775, 0.0741, 0.0903, 0.1118, 0.0636], device='cuda:4'), in_proj_covar=tensor([0.0378, 0.0410, 0.0404, 0.0379, 0.0446, 0.0422, 0.0511, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 23:47:46,334 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=192069.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:47:47,130 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.346e+02 1.981e+02 2.500e+02 3.039e+02 5.486e+02, threshold=4.999e+02, percent-clipped=1.0 2023-04-30 23:48:22,360 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3601, 2.2298, 2.1975, 4.0431, 2.2017, 2.5915, 2.3238, 2.3657], device='cuda:4'), covar=tensor([0.1128, 0.3711, 0.3021, 0.0467, 0.3891, 0.2564, 0.3509, 0.3279], device='cuda:4'), in_proj_covar=tensor([0.0380, 0.0423, 0.0351, 0.0311, 0.0422, 0.0485, 0.0395, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:48:37,353 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4066, 3.4625, 2.2651, 3.9576, 2.5831, 3.8494, 2.1546, 2.7506], device='cuda:4'), covar=tensor([0.0313, 0.0428, 0.1483, 0.0200, 0.0910, 0.0597, 0.1605, 0.0817], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0168, 0.0187, 0.0149, 0.0169, 0.0203, 0.0194, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 23:48:49,308 INFO [train.py:904] (4/8) Epoch 19, batch 9400, loss[loss=0.1558, simple_loss=0.2419, pruned_loss=0.03482, over 12319.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2601, pruned_loss=0.03688, over 3035209.99 frames. ], batch size: 247, lr: 3.52e-03, grad_scale: 4.0 2023-04-30 23:48:50,434 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192102.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:48:50,487 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4480, 2.8577, 3.1354, 1.9242, 2.7519, 2.1094, 3.0482, 3.1008], device='cuda:4'), covar=tensor([0.0343, 0.0859, 0.0565, 0.2070, 0.0834, 0.1086, 0.0639, 0.0942], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0152, 0.0159, 0.0146, 0.0139, 0.0124, 0.0138, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-04-30 23:48:58,447 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-30 23:50:29,958 INFO [train.py:904] (4/8) Epoch 19, batch 9450, loss[loss=0.1719, simple_loss=0.268, pruned_loss=0.03796, over 16392.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2624, pruned_loss=0.03721, over 3046715.27 frames. ], batch size: 146, lr: 3.52e-03, grad_scale: 4.0 2023-04-30 23:50:35,892 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192155.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:50:43,846 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3871, 4.4194, 4.7983, 4.7902, 4.7891, 4.5198, 4.4875, 4.4209], device='cuda:4'), covar=tensor([0.0545, 0.1311, 0.0766, 0.0903, 0.0887, 0.0893, 0.1163, 0.0535], device='cuda:4'), in_proj_covar=tensor([0.0378, 0.0410, 0.0403, 0.0378, 0.0445, 0.0421, 0.0509, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-04-30 23:50:51,019 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192163.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:51:05,238 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.652e+02 2.168e+02 2.516e+02 3.148e+02 5.649e+02, threshold=5.031e+02, percent-clipped=1.0 2023-04-30 23:51:36,370 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-30 23:52:10,463 INFO [train.py:904] (4/8) Epoch 19, batch 9500, loss[loss=0.1583, simple_loss=0.2447, pruned_loss=0.03592, over 12561.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2617, pruned_loss=0.03671, over 3075720.97 frames. ], batch size: 248, lr: 3.52e-03, grad_scale: 4.0 2023-04-30 23:52:39,610 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192216.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:52:45,059 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192219.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:52:47,798 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.70 vs. limit=2.0 2023-04-30 23:52:49,328 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192221.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:53:27,521 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0441, 1.8425, 1.7059, 1.4956, 1.9866, 1.6565, 1.6109, 1.9894], device='cuda:4'), covar=tensor([0.0153, 0.0284, 0.0361, 0.0357, 0.0187, 0.0249, 0.0154, 0.0195], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0223, 0.0215, 0.0216, 0.0225, 0.0222, 0.0221, 0.0214], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:53:55,149 INFO [train.py:904] (4/8) Epoch 19, batch 9550, loss[loss=0.2041, simple_loss=0.2962, pruned_loss=0.05602, over 16098.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2614, pruned_loss=0.03682, over 3069165.12 frames. ], batch size: 165, lr: 3.52e-03, grad_scale: 4.0 2023-04-30 23:54:34,513 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.443e+02 2.006e+02 2.351e+02 2.778e+02 5.702e+02, threshold=4.701e+02, percent-clipped=1.0 2023-04-30 23:54:55,546 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192280.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:55:01,202 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192282.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:55:11,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5518, 2.5751, 2.2530, 2.2482, 2.8703, 2.5727, 3.0272, 3.1412], device='cuda:4'), covar=tensor([0.0120, 0.0415, 0.0470, 0.0467, 0.0292, 0.0389, 0.0241, 0.0281], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0224, 0.0215, 0.0217, 0.0225, 0.0222, 0.0221, 0.0215], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-30 23:55:38,438 INFO [train.py:904] (4/8) Epoch 19, batch 9600, loss[loss=0.1888, simple_loss=0.2805, pruned_loss=0.04854, over 12582.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.264, pruned_loss=0.03788, over 3073571.07 frames. ], batch size: 250, lr: 3.52e-03, grad_scale: 8.0 2023-04-30 23:56:04,282 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8879, 2.2608, 2.2279, 3.0393, 1.9027, 3.2182, 1.6823, 2.7163], device='cuda:4'), covar=tensor([0.1425, 0.0718, 0.1198, 0.0178, 0.0081, 0.0385, 0.1739, 0.0757], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0167, 0.0189, 0.0175, 0.0196, 0.0207, 0.0194, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-30 23:56:53,193 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9199, 4.8957, 4.7580, 4.3239, 4.4195, 4.8322, 4.7528, 4.4864], device='cuda:4'), covar=tensor([0.0590, 0.0665, 0.0328, 0.0327, 0.1027, 0.0582, 0.0323, 0.0742], device='cuda:4'), in_proj_covar=tensor([0.0267, 0.0382, 0.0311, 0.0302, 0.0318, 0.0355, 0.0216, 0.0371], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-30 23:57:27,639 INFO [train.py:904] (4/8) Epoch 19, batch 9650, loss[loss=0.1676, simple_loss=0.2668, pruned_loss=0.03419, over 15401.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2651, pruned_loss=0.03831, over 3042301.71 frames. ], batch size: 191, lr: 3.52e-03, grad_scale: 8.0 2023-04-30 23:58:09,762 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.637e+02 2.165e+02 2.448e+02 2.941e+02 5.888e+02, threshold=4.896e+02, percent-clipped=1.0 2023-04-30 23:58:54,551 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192392.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:59:09,399 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192399.0, num_to_drop=0, layers_to_drop=set() 2023-04-30 23:59:15,172 INFO [train.py:904] (4/8) Epoch 19, batch 9700, loss[loss=0.1621, simple_loss=0.2463, pruned_loss=0.03891, over 12154.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2634, pruned_loss=0.03777, over 3036340.64 frames. ], batch size: 248, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:00:14,533 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9816, 2.0567, 2.2957, 3.1925, 2.1903, 2.2211, 2.2690, 2.1262], device='cuda:4'), covar=tensor([0.1227, 0.4040, 0.2702, 0.0710, 0.4597, 0.3055, 0.3768, 0.4053], device='cuda:4'), in_proj_covar=tensor([0.0380, 0.0423, 0.0350, 0.0309, 0.0422, 0.0484, 0.0395, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:00:19,294 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192433.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:00:57,137 INFO [train.py:904] (4/8) Epoch 19, batch 9750, loss[loss=0.1589, simple_loss=0.2517, pruned_loss=0.03307, over 16481.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2626, pruned_loss=0.03806, over 3028696.87 frames. ], batch size: 68, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:00:58,982 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192453.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:01:09,807 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=192458.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:01:12,217 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0009, 3.2764, 3.1682, 2.0815, 2.9804, 3.2508, 3.1648, 1.8455], device='cuda:4'), covar=tensor([0.0560, 0.0047, 0.0060, 0.0448, 0.0103, 0.0085, 0.0075, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0077, 0.0077, 0.0131, 0.0092, 0.0103, 0.0089, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 00:01:14,072 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192460.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:01:31,653 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 2.106e+02 2.517e+02 3.092e+02 5.794e+02, threshold=5.033e+02, percent-clipped=3.0 2023-05-01 00:02:23,619 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192494.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:02:37,067 INFO [train.py:904] (4/8) Epoch 19, batch 9800, loss[loss=0.1815, simple_loss=0.2848, pruned_loss=0.03914, over 16319.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2625, pruned_loss=0.03729, over 3026446.08 frames. ], batch size: 146, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:02:48,130 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0977, 2.3296, 1.9898, 2.0357, 2.7086, 2.3428, 2.5306, 2.8332], device='cuda:4'), covar=tensor([0.0142, 0.0439, 0.0531, 0.0546, 0.0276, 0.0427, 0.0229, 0.0274], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0225, 0.0217, 0.0218, 0.0227, 0.0225, 0.0222, 0.0217], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:02:55,410 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=192511.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:04:23,519 INFO [train.py:904] (4/8) Epoch 19, batch 9850, loss[loss=0.1833, simple_loss=0.2717, pruned_loss=0.04745, over 16703.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2634, pruned_loss=0.03679, over 3040712.10 frames. ], batch size: 134, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:05:00,568 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.479e+02 1.961e+02 2.506e+02 2.957e+02 4.630e+02, threshold=5.011e+02, percent-clipped=0.0 2023-05-01 00:05:11,432 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=192575.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:05:13,872 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.14 vs. limit=5.0 2023-05-01 00:05:15,528 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=192577.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:06:08,047 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 00:06:09,279 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1952, 3.0210, 3.1433, 1.8076, 3.2793, 3.3747, 2.7432, 2.6887], device='cuda:4'), covar=tensor([0.0756, 0.0253, 0.0183, 0.1190, 0.0094, 0.0173, 0.0424, 0.0449], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0101, 0.0088, 0.0131, 0.0073, 0.0113, 0.0119, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 00:06:14,628 INFO [train.py:904] (4/8) Epoch 19, batch 9900, loss[loss=0.1878, simple_loss=0.2855, pruned_loss=0.04503, over 15296.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2643, pruned_loss=0.03678, over 3060573.72 frames. ], batch size: 191, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:06:46,507 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192616.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:08:02,212 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-05-01 00:08:13,280 INFO [train.py:904] (4/8) Epoch 19, batch 9950, loss[loss=0.1779, simple_loss=0.2777, pruned_loss=0.0391, over 16423.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2668, pruned_loss=0.03726, over 3062592.52 frames. ], batch size: 146, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:08:48,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7296, 3.8156, 2.1916, 4.2232, 2.8213, 4.1364, 2.2911, 3.0436], device='cuda:4'), covar=tensor([0.0241, 0.0313, 0.1615, 0.0229, 0.0829, 0.0455, 0.1744, 0.0688], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0167, 0.0185, 0.0147, 0.0168, 0.0202, 0.0194, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 00:08:54,784 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.596e+02 2.121e+02 2.565e+02 3.086e+02 6.326e+02, threshold=5.129e+02, percent-clipped=1.0 2023-05-01 00:09:04,353 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192673.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:09:14,950 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192677.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:10:14,416 INFO [train.py:904] (4/8) Epoch 19, batch 10000, loss[loss=0.1632, simple_loss=0.2731, pruned_loss=0.02668, over 15254.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2654, pruned_loss=0.03683, over 3076286.27 frames. ], batch size: 190, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:11:22,414 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192734.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:11:25,778 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192735.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:11:50,827 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=192748.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:11:56,957 INFO [train.py:904] (4/8) Epoch 19, batch 10050, loss[loss=0.203, simple_loss=0.2966, pruned_loss=0.05465, over 16225.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2653, pruned_loss=0.03682, over 3066557.19 frames. ], batch size: 165, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:12:02,952 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=192755.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:12:09,601 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=192758.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:12:32,889 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.427e+02 2.015e+02 2.519e+02 2.909e+02 8.183e+02, threshold=5.037e+02, percent-clipped=1.0 2023-05-01 00:13:08,086 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=192789.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:13:08,231 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2239, 3.3808, 3.3593, 2.3363, 3.0830, 3.4063, 3.2229, 1.9521], device='cuda:4'), covar=tensor([0.0478, 0.0060, 0.0055, 0.0380, 0.0111, 0.0102, 0.0090, 0.0504], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0077, 0.0077, 0.0129, 0.0092, 0.0102, 0.0088, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 00:13:21,285 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192796.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:13:30,579 INFO [train.py:904] (4/8) Epoch 19, batch 10100, loss[loss=0.1618, simple_loss=0.2513, pruned_loss=0.03615, over 16932.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2652, pruned_loss=0.03677, over 3070112.33 frames. ], batch size: 116, lr: 3.52e-03, grad_scale: 8.0 2023-05-01 00:13:38,855 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=192806.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:13:48,411 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=192811.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:14:28,783 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=192831.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 00:15:13,688 INFO [train.py:904] (4/8) Epoch 20, batch 0, loss[loss=0.2148, simple_loss=0.3063, pruned_loss=0.06168, over 16980.00 frames. ], tot_loss[loss=0.2148, simple_loss=0.3063, pruned_loss=0.06168, over 16980.00 frames. ], batch size: 55, lr: 3.43e-03, grad_scale: 8.0 2023-05-01 00:15:13,688 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 00:15:21,160 INFO [train.py:938] (4/8) Epoch 20, validation: loss=0.146, simple_loss=0.2496, pruned_loss=0.02121, over 944034.00 frames. 2023-05-01 00:15:21,161 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 00:15:32,535 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=192859.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:15:36,711 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7553, 3.6693, 4.0287, 2.1055, 4.1665, 4.1399, 3.1328, 3.1967], device='cuda:4'), covar=tensor([0.0683, 0.0212, 0.0154, 0.1149, 0.0064, 0.0154, 0.0392, 0.0398], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0101, 0.0089, 0.0133, 0.0074, 0.0114, 0.0120, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 00:15:49,198 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.713e+02 2.412e+02 2.946e+02 3.638e+02 7.163e+02, threshold=5.893e+02, percent-clipped=6.0 2023-05-01 00:15:53,700 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=192875.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:15:56,057 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=192877.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:16:17,605 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=192892.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 00:16:31,046 INFO [train.py:904] (4/8) Epoch 20, batch 50, loss[loss=0.1891, simple_loss=0.2694, pruned_loss=0.05439, over 16436.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2736, pruned_loss=0.05111, over 749919.71 frames. ], batch size: 146, lr: 3.43e-03, grad_scale: 2.0 2023-05-01 00:17:00,246 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=192923.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:17:03,081 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=192925.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:17:21,276 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0771, 5.0139, 4.8396, 4.4726, 4.8810, 1.8312, 4.6814, 4.7407], device='cuda:4'), covar=tensor([0.0103, 0.0110, 0.0236, 0.0361, 0.0129, 0.2757, 0.0162, 0.0233], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0142, 0.0184, 0.0163, 0.0161, 0.0198, 0.0174, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:17:38,938 INFO [train.py:904] (4/8) Epoch 20, batch 100, loss[loss=0.1812, simple_loss=0.2597, pruned_loss=0.05138, over 16791.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2691, pruned_loss=0.04923, over 1313053.19 frames. ], batch size: 83, lr: 3.43e-03, grad_scale: 2.0 2023-05-01 00:17:48,062 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3885, 5.7870, 5.5077, 5.5740, 5.1824, 5.2137, 5.1613, 5.9123], device='cuda:4'), covar=tensor([0.1284, 0.1025, 0.1259, 0.0907, 0.0976, 0.0687, 0.1206, 0.1013], device='cuda:4'), in_proj_covar=tensor([0.0627, 0.0774, 0.0627, 0.0573, 0.0484, 0.0494, 0.0645, 0.0590], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:18:07,333 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.526e+02 2.214e+02 2.624e+02 3.260e+02 6.598e+02, threshold=5.249e+02, percent-clipped=3.0 2023-05-01 00:18:07,673 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=192972.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:18:13,029 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7623, 4.9973, 5.1647, 4.9597, 5.0029, 5.5989, 5.0711, 4.7682], device='cuda:4'), covar=tensor([0.1399, 0.2066, 0.2338, 0.2421, 0.2928, 0.1175, 0.1787, 0.2766], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0558, 0.0618, 0.0467, 0.0622, 0.0649, 0.0485, 0.0619], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 00:18:37,596 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-01 00:18:41,871 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9315, 3.9853, 2.2739, 4.6786, 3.0365, 4.6010, 2.6337, 3.4000], device='cuda:4'), covar=tensor([0.0329, 0.0426, 0.1874, 0.0252, 0.0952, 0.0528, 0.1550, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0171, 0.0189, 0.0152, 0.0172, 0.0207, 0.0198, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:18:48,440 INFO [train.py:904] (4/8) Epoch 20, batch 150, loss[loss=0.1837, simple_loss=0.2654, pruned_loss=0.05098, over 16522.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2665, pruned_loss=0.04738, over 1758138.71 frames. ], batch size: 75, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:19:01,766 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-05-01 00:19:27,145 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193029.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:19:43,992 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=193041.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:19:52,830 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193048.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:19:58,179 INFO [train.py:904] (4/8) Epoch 20, batch 200, loss[loss=0.1856, simple_loss=0.2612, pruned_loss=0.05498, over 16431.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2659, pruned_loss=0.04687, over 2106303.53 frames. ], batch size: 146, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:20:03,473 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193055.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:20:05,084 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-05-01 00:20:27,491 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.769e+02 2.323e+02 2.688e+02 3.539e+02 1.444e+03, threshold=5.377e+02, percent-clipped=5.0 2023-05-01 00:20:50,799 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193089.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:20:53,767 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193091.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:21:00,959 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193096.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:21:08,806 INFO [train.py:904] (4/8) Epoch 20, batch 250, loss[loss=0.1461, simple_loss=0.2335, pruned_loss=0.02935, over 17217.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2636, pruned_loss=0.0463, over 2378365.26 frames. ], batch size: 44, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:21:09,318 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=193102.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:21:10,234 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193103.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:21:10,605 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.67 vs. limit=5.0 2023-05-01 00:21:12,324 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2668, 2.9763, 2.6329, 2.2006, 2.1970, 2.2696, 2.9590, 2.7373], device='cuda:4'), covar=tensor([0.2614, 0.0726, 0.1743, 0.2387, 0.2422, 0.2070, 0.0490, 0.1292], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0260, 0.0295, 0.0300, 0.0283, 0.0248, 0.0282, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 00:21:58,203 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193137.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:22:16,428 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7915, 3.8402, 3.0355, 2.2550, 2.4856, 2.3795, 3.9084, 3.3374], device='cuda:4'), covar=tensor([0.2508, 0.0556, 0.1486, 0.2797, 0.2673, 0.2113, 0.0478, 0.1377], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0261, 0.0297, 0.0301, 0.0285, 0.0249, 0.0283, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 00:22:17,118 INFO [train.py:904] (4/8) Epoch 20, batch 300, loss[loss=0.1814, simple_loss=0.2547, pruned_loss=0.05408, over 16849.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2613, pruned_loss=0.04484, over 2592161.65 frames. ], batch size: 116, lr: 3.42e-03, grad_scale: 1.0 2023-05-01 00:22:46,393 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.636e+02 2.161e+02 2.496e+02 2.884e+02 5.011e+02, threshold=4.993e+02, percent-clipped=0.0 2023-05-01 00:22:52,936 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7134, 3.5798, 4.0207, 2.0077, 4.1719, 4.1919, 3.1771, 3.1637], device='cuda:4'), covar=tensor([0.0738, 0.0278, 0.0205, 0.1232, 0.0086, 0.0191, 0.0412, 0.0429], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0105, 0.0093, 0.0137, 0.0076, 0.0119, 0.0124, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 00:23:05,532 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193187.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 00:23:28,161 INFO [train.py:904] (4/8) Epoch 20, batch 350, loss[loss=0.1572, simple_loss=0.251, pruned_loss=0.03168, over 16750.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2595, pruned_loss=0.0439, over 2751403.53 frames. ], batch size: 57, lr: 3.42e-03, grad_scale: 1.0 2023-05-01 00:23:39,539 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=193211.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:24:37,894 INFO [train.py:904] (4/8) Epoch 20, batch 400, loss[loss=0.1736, simple_loss=0.2539, pruned_loss=0.04661, over 16287.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2581, pruned_loss=0.0435, over 2883124.34 frames. ], batch size: 165, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:25:05,932 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193272.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:25:06,039 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=193272.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 00:25:06,669 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.612e+02 2.085e+02 2.558e+02 2.961e+02 5.266e+02, threshold=5.116e+02, percent-clipped=1.0 2023-05-01 00:25:23,367 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=193284.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:25:46,656 INFO [train.py:904] (4/8) Epoch 20, batch 450, loss[loss=0.164, simple_loss=0.2578, pruned_loss=0.03511, over 17273.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2561, pruned_loss=0.04248, over 2983553.26 frames. ], batch size: 52, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:26:05,326 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9098, 2.8423, 2.8723, 5.0264, 3.9689, 4.3314, 1.9470, 3.0427], device='cuda:4'), covar=tensor([0.1406, 0.0890, 0.1196, 0.0226, 0.0300, 0.0484, 0.1607, 0.0917], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0171, 0.0191, 0.0181, 0.0199, 0.0212, 0.0197, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:26:13,646 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193320.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:26:26,090 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193329.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:26:46,072 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7557, 3.8474, 2.6015, 4.5666, 3.0247, 4.4494, 2.5695, 3.1376], device='cuda:4'), covar=tensor([0.0310, 0.0457, 0.1460, 0.0229, 0.0884, 0.0497, 0.1444, 0.0774], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0176, 0.0193, 0.0157, 0.0175, 0.0212, 0.0201, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:26:48,824 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=193345.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:26:57,283 INFO [train.py:904] (4/8) Epoch 20, batch 500, loss[loss=0.1509, simple_loss=0.2344, pruned_loss=0.03367, over 16952.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2549, pruned_loss=0.04201, over 3061868.44 frames. ], batch size: 41, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:27:26,048 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.420e+02 2.231e+02 2.613e+02 3.252e+02 5.197e+02, threshold=5.226e+02, percent-clipped=2.0 2023-05-01 00:27:31,763 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193377.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:27:51,494 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193391.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:27:59,426 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193397.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:28:07,092 INFO [train.py:904] (4/8) Epoch 20, batch 550, loss[loss=0.2006, simple_loss=0.2742, pruned_loss=0.06353, over 16709.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2537, pruned_loss=0.04134, over 3126015.71 frames. ], batch size: 124, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:28:21,966 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 00:28:32,323 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=193421.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:28:56,787 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193439.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:29:14,811 INFO [train.py:904] (4/8) Epoch 20, batch 600, loss[loss=0.1426, simple_loss=0.2305, pruned_loss=0.02732, over 17223.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2536, pruned_loss=0.04188, over 3178014.29 frames. ], batch size: 43, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:29:18,163 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=193454.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 00:29:43,211 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.108e+02 2.456e+02 2.826e+02 6.166e+02, threshold=4.913e+02, percent-clipped=1.0 2023-05-01 00:29:55,238 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=193482.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 00:30:03,357 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193487.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 00:30:21,023 INFO [train.py:904] (4/8) Epoch 20, batch 650, loss[loss=0.162, simple_loss=0.2603, pruned_loss=0.03186, over 17137.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.252, pruned_loss=0.04158, over 3218942.79 frames. ], batch size: 48, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:30:39,976 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=193515.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 00:31:07,549 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193535.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 00:31:29,637 INFO [train.py:904] (4/8) Epoch 20, batch 700, loss[loss=0.1828, simple_loss=0.2751, pruned_loss=0.04528, over 16735.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2517, pruned_loss=0.04118, over 3244223.47 frames. ], batch size: 57, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:31:48,972 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193567.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 00:31:57,448 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.555e+02 2.109e+02 2.476e+02 3.086e+02 5.855e+02, threshold=4.951e+02, percent-clipped=1.0 2023-05-01 00:32:35,677 INFO [train.py:904] (4/8) Epoch 20, batch 750, loss[loss=0.1372, simple_loss=0.2307, pruned_loss=0.02188, over 16872.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2527, pruned_loss=0.04178, over 3261241.59 frames. ], batch size: 42, lr: 3.42e-03, grad_scale: 2.0 2023-05-01 00:32:41,067 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-01 00:32:54,597 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=193615.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:33:09,542 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-05-01 00:33:26,215 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0352, 4.8006, 5.0628, 5.2282, 5.4291, 4.7341, 5.4308, 5.4133], device='cuda:4'), covar=tensor([0.1780, 0.1304, 0.1698, 0.0838, 0.0510, 0.0922, 0.0527, 0.0587], device='cuda:4'), in_proj_covar=tensor([0.0630, 0.0782, 0.0913, 0.0800, 0.0595, 0.0622, 0.0646, 0.0747], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:33:28,438 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193640.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:33:42,945 INFO [train.py:904] (4/8) Epoch 20, batch 800, loss[loss=0.1739, simple_loss=0.2534, pruned_loss=0.04721, over 16310.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2522, pruned_loss=0.0415, over 3276457.49 frames. ], batch size: 165, lr: 3.42e-03, grad_scale: 4.0 2023-05-01 00:34:00,630 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3980, 4.4099, 4.7418, 4.7529, 4.7769, 4.5056, 4.4919, 4.3828], device='cuda:4'), covar=tensor([0.0361, 0.0717, 0.0399, 0.0389, 0.0505, 0.0420, 0.0803, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0399, 0.0437, 0.0424, 0.0397, 0.0473, 0.0446, 0.0538, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 00:34:10,661 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.331e+02 2.187e+02 2.610e+02 3.095e+02 1.121e+03, threshold=5.220e+02, percent-clipped=2.0 2023-05-01 00:34:16,595 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=193676.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:34:44,992 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193697.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:34:52,173 INFO [train.py:904] (4/8) Epoch 20, batch 850, loss[loss=0.1661, simple_loss=0.2408, pruned_loss=0.04574, over 16689.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2518, pruned_loss=0.04111, over 3288708.16 frames. ], batch size: 124, lr: 3.42e-03, grad_scale: 4.0 2023-05-01 00:35:17,107 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0249, 4.9509, 4.9064, 4.5153, 4.5611, 4.9573, 4.7803, 4.6319], device='cuda:4'), covar=tensor([0.0625, 0.0738, 0.0276, 0.0306, 0.0844, 0.0408, 0.0433, 0.0671], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0417, 0.0338, 0.0331, 0.0348, 0.0387, 0.0234, 0.0405], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:35:28,196 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-05-01 00:35:51,454 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193745.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:35:59,348 INFO [train.py:904] (4/8) Epoch 20, batch 900, loss[loss=0.1706, simple_loss=0.259, pruned_loss=0.04113, over 17035.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2514, pruned_loss=0.0407, over 3299117.08 frames. ], batch size: 53, lr: 3.42e-03, grad_scale: 4.0 2023-05-01 00:36:05,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2759, 4.5962, 4.5943, 3.3723, 3.7982, 4.5029, 4.0805, 2.7579], device='cuda:4'), covar=tensor([0.0392, 0.0047, 0.0039, 0.0318, 0.0121, 0.0088, 0.0076, 0.0423], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0082, 0.0081, 0.0135, 0.0096, 0.0107, 0.0092, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 00:36:12,523 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6588, 2.5920, 1.7969, 2.7508, 2.0618, 2.8325, 2.1231, 2.3653], device='cuda:4'), covar=tensor([0.0298, 0.0374, 0.1419, 0.0294, 0.0694, 0.0464, 0.1260, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0180, 0.0197, 0.0163, 0.0179, 0.0218, 0.0206, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:36:28,230 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.015e+02 2.385e+02 2.719e+02 5.424e+02, threshold=4.769e+02, percent-clipped=3.0 2023-05-01 00:36:33,794 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193777.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 00:37:09,295 INFO [train.py:904] (4/8) Epoch 20, batch 950, loss[loss=0.1499, simple_loss=0.2359, pruned_loss=0.03196, over 16795.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2507, pruned_loss=0.04033, over 3300671.06 frames. ], batch size: 39, lr: 3.42e-03, grad_scale: 4.0 2023-05-01 00:37:20,552 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193810.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 00:37:21,728 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3239, 5.2581, 5.1430, 4.5959, 4.7391, 5.1695, 5.2050, 4.7487], device='cuda:4'), covar=tensor([0.0599, 0.0495, 0.0304, 0.0388, 0.1186, 0.0483, 0.0284, 0.0812], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0420, 0.0340, 0.0334, 0.0351, 0.0391, 0.0236, 0.0408], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:38:17,860 INFO [train.py:904] (4/8) Epoch 20, batch 1000, loss[loss=0.1865, simple_loss=0.2516, pruned_loss=0.06068, over 16887.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.249, pruned_loss=0.04003, over 3300921.18 frames. ], batch size: 109, lr: 3.42e-03, grad_scale: 4.0 2023-05-01 00:38:22,258 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-05-01 00:38:39,197 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0102, 5.5215, 5.5961, 5.3780, 5.3618, 6.0221, 5.4786, 5.2274], device='cuda:4'), covar=tensor([0.1078, 0.2007, 0.2279, 0.2157, 0.3107, 0.1131, 0.1590, 0.2375], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0584, 0.0645, 0.0486, 0.0650, 0.0674, 0.0503, 0.0647], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 00:38:39,309 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193867.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 00:38:45,938 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.639e+02 2.059e+02 2.451e+02 3.046e+02 5.383e+02, threshold=4.901e+02, percent-clipped=2.0 2023-05-01 00:39:00,647 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=193884.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:39:24,779 INFO [train.py:904] (4/8) Epoch 20, batch 1050, loss[loss=0.1696, simple_loss=0.247, pruned_loss=0.04613, over 16804.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2489, pruned_loss=0.04028, over 3306083.94 frames. ], batch size: 96, lr: 3.42e-03, grad_scale: 4.0 2023-05-01 00:39:43,672 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193915.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:40:03,378 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7292, 2.6537, 2.3858, 2.8152, 3.1369, 2.9607, 3.4414, 3.3658], device='cuda:4'), covar=tensor([0.0152, 0.0446, 0.0509, 0.0387, 0.0268, 0.0369, 0.0227, 0.0248], device='cuda:4'), in_proj_covar=tensor([0.0205, 0.0237, 0.0225, 0.0229, 0.0239, 0.0237, 0.0238, 0.0230], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:40:18,174 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=193940.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:40:25,143 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=193945.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:40:35,455 INFO [train.py:904] (4/8) Epoch 20, batch 1100, loss[loss=0.1775, simple_loss=0.2731, pruned_loss=0.0409, over 16690.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.249, pruned_loss=0.03979, over 3303931.85 frames. ], batch size: 57, lr: 3.42e-03, grad_scale: 4.0 2023-05-01 00:40:42,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2673, 5.2284, 5.0008, 4.4800, 5.0417, 2.0713, 4.8170, 4.9379], device='cuda:4'), covar=tensor([0.0080, 0.0082, 0.0224, 0.0394, 0.0118, 0.2583, 0.0159, 0.0220], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0149, 0.0194, 0.0173, 0.0170, 0.0205, 0.0183, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:41:01,675 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=193971.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:41:04,217 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.456e+02 2.207e+02 2.619e+02 3.357e+02 2.000e+03, threshold=5.237e+02, percent-clipped=3.0 2023-05-01 00:41:25,068 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=193988.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:41:46,949 INFO [train.py:904] (4/8) Epoch 20, batch 1150, loss[loss=0.1669, simple_loss=0.246, pruned_loss=0.04391, over 16488.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2486, pruned_loss=0.03983, over 3307876.22 frames. ], batch size: 146, lr: 3.42e-03, grad_scale: 4.0 2023-05-01 00:42:56,045 INFO [train.py:904] (4/8) Epoch 20, batch 1200, loss[loss=0.1525, simple_loss=0.2304, pruned_loss=0.03727, over 15677.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2484, pruned_loss=0.0394, over 3314474.62 frames. ], batch size: 191, lr: 3.42e-03, grad_scale: 8.0 2023-05-01 00:43:25,035 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=194072.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:43:25,793 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.319e+02 2.146e+02 2.588e+02 3.131e+02 5.620e+02, threshold=5.177e+02, percent-clipped=2.0 2023-05-01 00:43:33,083 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=194077.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 00:43:45,920 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8628, 2.9950, 2.6803, 4.9858, 4.0351, 4.2968, 1.6771, 3.1827], device='cuda:4'), covar=tensor([0.1303, 0.0749, 0.1241, 0.0193, 0.0202, 0.0417, 0.1626, 0.0765], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0173, 0.0193, 0.0184, 0.0201, 0.0214, 0.0198, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:44:06,932 INFO [train.py:904] (4/8) Epoch 20, batch 1250, loss[loss=0.1709, simple_loss=0.2658, pruned_loss=0.03801, over 16714.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2481, pruned_loss=0.03973, over 3320990.80 frames. ], batch size: 57, lr: 3.42e-03, grad_scale: 8.0 2023-05-01 00:44:17,568 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=194110.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 00:44:38,137 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=194125.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:44:38,380 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7107, 2.5126, 2.4707, 3.8834, 3.1774, 4.0047, 1.5400, 2.9208], device='cuda:4'), covar=tensor([0.1397, 0.0757, 0.1211, 0.0200, 0.0178, 0.0372, 0.1600, 0.0845], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0173, 0.0192, 0.0184, 0.0200, 0.0213, 0.0198, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:44:49,863 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=194133.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:45:02,554 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-05-01 00:45:15,926 INFO [train.py:904] (4/8) Epoch 20, batch 1300, loss[loss=0.146, simple_loss=0.2359, pruned_loss=0.02801, over 16852.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2481, pruned_loss=0.03937, over 3325403.03 frames. ], batch size: 42, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 00:45:26,571 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=194158.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 00:45:46,645 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.297e+02 2.207e+02 2.541e+02 3.000e+02 4.873e+02, threshold=5.083e+02, percent-clipped=0.0 2023-05-01 00:46:27,314 INFO [train.py:904] (4/8) Epoch 20, batch 1350, loss[loss=0.1797, simple_loss=0.2618, pruned_loss=0.04881, over 16486.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2476, pruned_loss=0.03886, over 3324827.13 frames. ], batch size: 68, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:46:46,819 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6730, 2.3084, 2.3137, 4.4437, 2.3580, 2.7501, 2.4708, 2.5344], device='cuda:4'), covar=tensor([0.1183, 0.3904, 0.3170, 0.0453, 0.4253, 0.2821, 0.3427, 0.3755], device='cuda:4'), in_proj_covar=tensor([0.0396, 0.0439, 0.0363, 0.0325, 0.0433, 0.0505, 0.0408, 0.0515], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:47:05,232 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6068, 3.6207, 3.9215, 2.1028, 4.1864, 4.1625, 3.0294, 3.1653], device='cuda:4'), covar=tensor([0.0845, 0.0270, 0.0242, 0.1234, 0.0091, 0.0194, 0.0475, 0.0473], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0097, 0.0140, 0.0079, 0.0124, 0.0128, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:47:20,600 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=194240.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:47:36,939 INFO [train.py:904] (4/8) Epoch 20, batch 1400, loss[loss=0.1829, simple_loss=0.254, pruned_loss=0.05589, over 16844.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.247, pruned_loss=0.039, over 3326755.75 frames. ], batch size: 90, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:47:57,380 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.94 vs. limit=5.0 2023-05-01 00:48:03,505 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=194271.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:48:06,430 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.050e+02 2.370e+02 3.098e+02 5.438e+02, threshold=4.739e+02, percent-clipped=2.0 2023-05-01 00:48:06,883 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2955, 4.3986, 4.7256, 4.7181, 4.7363, 4.4111, 4.4414, 4.3300], device='cuda:4'), covar=tensor([0.0447, 0.0686, 0.0473, 0.0477, 0.0539, 0.0454, 0.0899, 0.0676], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0445, 0.0429, 0.0403, 0.0481, 0.0453, 0.0545, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 00:48:44,297 INFO [train.py:904] (4/8) Epoch 20, batch 1450, loss[loss=0.1928, simple_loss=0.2772, pruned_loss=0.05426, over 17083.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2468, pruned_loss=0.03923, over 3326896.40 frames. ], batch size: 49, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:48:48,921 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6639, 3.5052, 3.8194, 2.0399, 3.9531, 3.9585, 3.0641, 2.9795], device='cuda:4'), covar=tensor([0.0719, 0.0234, 0.0171, 0.1157, 0.0091, 0.0181, 0.0398, 0.0444], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0108, 0.0097, 0.0140, 0.0079, 0.0124, 0.0127, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:49:06,294 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=194318.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:49:08,663 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=194319.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:49:41,303 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-05-01 00:49:54,001 INFO [train.py:904] (4/8) Epoch 20, batch 1500, loss[loss=0.1576, simple_loss=0.2394, pruned_loss=0.03794, over 12406.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2472, pruned_loss=0.03959, over 3317799.92 frames. ], batch size: 246, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:49:55,471 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=194353.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:50:01,280 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.95 vs. limit=5.0 2023-05-01 00:50:08,000 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=194362.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 00:50:24,391 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.600e+02 2.182e+02 2.566e+02 3.379e+02 8.536e+02, threshold=5.133e+02, percent-clipped=4.0 2023-05-01 00:50:31,301 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=194379.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:51:03,654 INFO [train.py:904] (4/8) Epoch 20, batch 1550, loss[loss=0.1825, simple_loss=0.2579, pruned_loss=0.05352, over 16429.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.25, pruned_loss=0.04082, over 3320568.00 frames. ], batch size: 146, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:51:20,831 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=194414.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:51:34,182 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=194423.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 00:51:40,607 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=194428.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:52:13,192 INFO [train.py:904] (4/8) Epoch 20, batch 1600, loss[loss=0.1594, simple_loss=0.2378, pruned_loss=0.04046, over 16810.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.251, pruned_loss=0.04125, over 3325562.95 frames. ], batch size: 102, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 00:52:29,858 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.36 vs. limit=5.0 2023-05-01 00:52:34,842 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-05-01 00:52:43,859 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.299e+02 2.983e+02 3.595e+02 1.383e+03, threshold=5.966e+02, percent-clipped=6.0 2023-05-01 00:53:22,707 INFO [train.py:904] (4/8) Epoch 20, batch 1650, loss[loss=0.1635, simple_loss=0.2429, pruned_loss=0.04208, over 16683.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2527, pruned_loss=0.04243, over 3317410.35 frames. ], batch size: 89, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 00:54:09,763 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2257, 2.6799, 2.1708, 2.4067, 3.0033, 2.6961, 3.0735, 3.1667], device='cuda:4'), covar=tensor([0.0227, 0.0367, 0.0507, 0.0438, 0.0233, 0.0362, 0.0266, 0.0225], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0239, 0.0228, 0.0231, 0.0241, 0.0240, 0.0240, 0.0232], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:54:16,672 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=194540.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:54:33,460 INFO [train.py:904] (4/8) Epoch 20, batch 1700, loss[loss=0.1689, simple_loss=0.2684, pruned_loss=0.03468, over 17107.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2558, pruned_loss=0.04286, over 3312855.58 frames. ], batch size: 49, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:54:45,443 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8135, 2.8506, 2.5165, 2.7417, 3.1858, 2.8649, 3.3685, 3.3590], device='cuda:4'), covar=tensor([0.0144, 0.0390, 0.0481, 0.0444, 0.0298, 0.0410, 0.0297, 0.0271], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0239, 0.0228, 0.0231, 0.0241, 0.0239, 0.0239, 0.0232], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:54:48,063 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.99 vs. limit=5.0 2023-05-01 00:55:05,696 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.524e+02 2.259e+02 2.640e+02 3.354e+02 1.280e+03, threshold=5.281e+02, percent-clipped=2.0 2023-05-01 00:55:25,665 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=194588.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:55:43,639 INFO [train.py:904] (4/8) Epoch 20, batch 1750, loss[loss=0.1642, simple_loss=0.2471, pruned_loss=0.04061, over 16546.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2569, pruned_loss=0.043, over 3311135.94 frames. ], batch size: 75, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:55:44,018 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6754, 4.4188, 4.5847, 4.9239, 5.0470, 4.4679, 5.1175, 5.0863], device='cuda:4'), covar=tensor([0.2185, 0.1730, 0.2289, 0.1001, 0.0918, 0.1201, 0.0951, 0.0885], device='cuda:4'), in_proj_covar=tensor([0.0648, 0.0803, 0.0940, 0.0823, 0.0612, 0.0642, 0.0663, 0.0766], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:56:27,153 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=194634.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:56:51,747 INFO [train.py:904] (4/8) Epoch 20, batch 1800, loss[loss=0.1732, simple_loss=0.2568, pruned_loss=0.04475, over 16675.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2583, pruned_loss=0.04309, over 3313334.98 frames. ], batch size: 134, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:57:06,064 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0960, 3.8228, 4.3110, 2.1786, 4.6065, 4.6506, 3.2805, 3.5146], device='cuda:4'), covar=tensor([0.0677, 0.0292, 0.0245, 0.1157, 0.0071, 0.0155, 0.0413, 0.0412], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0098, 0.0141, 0.0080, 0.0125, 0.0128, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 00:57:10,173 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0916, 5.1098, 5.5629, 5.5498, 5.5957, 5.2044, 5.1387, 4.9711], device='cuda:4'), covar=tensor([0.0385, 0.0576, 0.0439, 0.0486, 0.0539, 0.0405, 0.1082, 0.0453], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0452, 0.0435, 0.0410, 0.0488, 0.0462, 0.0553, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 00:57:22,480 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=194674.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:57:23,424 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.509e+02 2.398e+02 2.901e+02 3.350e+02 9.637e+02, threshold=5.801e+02, percent-clipped=10.0 2023-05-01 00:57:37,299 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3517, 4.7211, 4.7128, 3.4916, 3.9871, 4.6914, 4.0196, 3.2305], device='cuda:4'), covar=tensor([0.0397, 0.0056, 0.0041, 0.0301, 0.0125, 0.0073, 0.0087, 0.0325], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0082, 0.0082, 0.0135, 0.0097, 0.0108, 0.0094, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0004], device='cuda:4') 2023-05-01 00:57:50,526 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=194695.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:57:58,019 INFO [train.py:904] (4/8) Epoch 20, batch 1850, loss[loss=0.1841, simple_loss=0.2646, pruned_loss=0.05185, over 16808.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2579, pruned_loss=0.04259, over 3314688.43 frames. ], batch size: 83, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:58:08,682 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=194709.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:58:21,439 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=194718.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 00:58:34,469 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=194728.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 00:58:56,334 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3261, 5.6798, 5.4165, 5.4903, 5.0747, 5.1320, 5.0567, 5.8189], device='cuda:4'), covar=tensor([0.1271, 0.0884, 0.1124, 0.0786, 0.0901, 0.0799, 0.1153, 0.0859], device='cuda:4'), in_proj_covar=tensor([0.0672, 0.0828, 0.0675, 0.0616, 0.0518, 0.0525, 0.0693, 0.0635], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 00:59:06,818 INFO [train.py:904] (4/8) Epoch 20, batch 1900, loss[loss=0.1718, simple_loss=0.2565, pruned_loss=0.04354, over 16165.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2576, pruned_loss=0.04242, over 3294821.12 frames. ], batch size: 165, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 00:59:38,381 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 2.030e+02 2.433e+02 3.010e+02 6.888e+02, threshold=4.866e+02, percent-clipped=1.0 2023-05-01 00:59:41,138 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=194776.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:00:16,323 INFO [train.py:904] (4/8) Epoch 20, batch 1950, loss[loss=0.1973, simple_loss=0.289, pruned_loss=0.05285, over 12236.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2577, pruned_loss=0.04203, over 3303818.33 frames. ], batch size: 246, lr: 3.41e-03, grad_scale: 4.0 2023-05-01 01:00:19,395 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 01:00:44,329 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-01 01:00:49,246 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2062, 5.1565, 5.0621, 4.5426, 4.6373, 5.1322, 5.1123, 4.7225], device='cuda:4'), covar=tensor([0.0615, 0.0470, 0.0322, 0.0403, 0.1177, 0.0461, 0.0311, 0.0771], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0428, 0.0346, 0.0340, 0.0357, 0.0399, 0.0239, 0.0414], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:01:17,149 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=194847.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:01:23,588 INFO [train.py:904] (4/8) Epoch 20, batch 2000, loss[loss=0.1527, simple_loss=0.2448, pruned_loss=0.03028, over 17236.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2569, pruned_loss=0.04179, over 3313861.75 frames. ], batch size: 44, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:01:54,983 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.555e+02 2.120e+02 2.565e+02 3.052e+02 4.628e+02, threshold=5.130e+02, percent-clipped=0.0 2023-05-01 01:02:32,415 INFO [train.py:904] (4/8) Epoch 20, batch 2050, loss[loss=0.1893, simple_loss=0.2667, pruned_loss=0.05601, over 16456.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2572, pruned_loss=0.0422, over 3314641.30 frames. ], batch size: 75, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:02:41,420 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=194908.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:02:55,027 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7503, 4.0519, 2.9735, 2.3202, 2.7023, 2.5179, 4.3864, 3.4921], device='cuda:4'), covar=tensor([0.2890, 0.0675, 0.1859, 0.2787, 0.2762, 0.2141, 0.0432, 0.1353], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0267, 0.0300, 0.0304, 0.0293, 0.0253, 0.0289, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 01:03:11,917 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 01:03:18,572 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8084, 3.8525, 2.1313, 4.4561, 3.0112, 4.3695, 2.0028, 3.0931], device='cuda:4'), covar=tensor([0.0273, 0.0400, 0.1889, 0.0276, 0.0775, 0.0437, 0.2068, 0.0759], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0180, 0.0196, 0.0164, 0.0178, 0.0220, 0.0203, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:03:27,310 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0483, 2.2020, 2.6788, 3.0199, 2.8703, 3.5306, 2.4753, 3.4684], device='cuda:4'), covar=tensor([0.0227, 0.0459, 0.0310, 0.0311, 0.0312, 0.0178, 0.0404, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0193, 0.0178, 0.0181, 0.0195, 0.0152, 0.0195, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:03:41,679 INFO [train.py:904] (4/8) Epoch 20, batch 2100, loss[loss=0.1627, simple_loss=0.2441, pruned_loss=0.04067, over 15932.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2582, pruned_loss=0.04303, over 3315987.59 frames. ], batch size: 35, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:04:12,825 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=194974.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:04:13,685 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.402e+02 2.067e+02 2.444e+02 3.012e+02 5.275e+02, threshold=4.887e+02, percent-clipped=1.0 2023-05-01 01:04:34,697 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=194990.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:04:45,188 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-05-01 01:04:50,814 INFO [train.py:904] (4/8) Epoch 20, batch 2150, loss[loss=0.1826, simple_loss=0.2611, pruned_loss=0.05201, over 16489.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2591, pruned_loss=0.04387, over 3309791.53 frames. ], batch size: 146, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:05:01,409 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=195009.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:05:04,390 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=195011.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:05:13,360 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=195018.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 01:05:18,550 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=195022.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:05:49,746 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=195044.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 01:05:55,265 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6614, 2.3776, 2.2908, 4.4546, 2.3610, 2.7713, 2.3913, 2.6122], device='cuda:4'), covar=tensor([0.1164, 0.3627, 0.3147, 0.0447, 0.4051, 0.2620, 0.3495, 0.3576], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0440, 0.0363, 0.0326, 0.0433, 0.0506, 0.0409, 0.0515], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:06:01,906 INFO [train.py:904] (4/8) Epoch 20, batch 2200, loss[loss=0.1894, simple_loss=0.272, pruned_loss=0.05341, over 16508.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2595, pruned_loss=0.0438, over 3319802.93 frames. ], batch size: 68, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:06:09,187 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=195057.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:06:21,179 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=195066.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 01:06:28,214 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=195071.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:06:30,115 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=195072.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 01:06:33,765 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.892e+02 2.410e+02 2.761e+02 3.300e+02 5.006e+02, threshold=5.521e+02, percent-clipped=1.0 2023-05-01 01:07:10,884 INFO [train.py:904] (4/8) Epoch 20, batch 2250, loss[loss=0.1724, simple_loss=0.2519, pruned_loss=0.04647, over 16496.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2607, pruned_loss=0.04423, over 3313678.85 frames. ], batch size: 68, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:07:11,370 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4736, 3.6763, 3.7440, 1.9649, 2.9656, 2.0900, 3.9886, 3.9823], device='cuda:4'), covar=tensor([0.0262, 0.0943, 0.0627, 0.2527, 0.1051, 0.1456, 0.0603, 0.1025], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0167, 0.0152, 0.0145, 0.0129, 0.0144, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:07:15,402 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=195105.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 01:07:37,886 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2200, 2.6469, 2.1389, 2.4144, 2.9524, 2.7348, 3.1491, 3.1217], device='cuda:4'), covar=tensor([0.0227, 0.0399, 0.0545, 0.0459, 0.0279, 0.0362, 0.0234, 0.0256], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0240, 0.0227, 0.0230, 0.0240, 0.0239, 0.0240, 0.0232], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:07:54,113 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=195132.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:07:58,455 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-01 01:08:21,142 INFO [train.py:904] (4/8) Epoch 20, batch 2300, loss[loss=0.1751, simple_loss=0.2552, pruned_loss=0.04752, over 16706.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.26, pruned_loss=0.04392, over 3311954.92 frames. ], batch size: 124, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:08:51,382 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.574e+02 2.280e+02 2.663e+02 3.175e+02 5.300e+02, threshold=5.327e+02, percent-clipped=0.0 2023-05-01 01:08:54,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1466, 1.8982, 2.7099, 3.0525, 2.9093, 3.6123, 2.1815, 3.5876], device='cuda:4'), covar=tensor([0.0212, 0.0642, 0.0315, 0.0304, 0.0326, 0.0183, 0.0602, 0.0156], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0195, 0.0180, 0.0182, 0.0196, 0.0154, 0.0196, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:09:29,599 INFO [train.py:904] (4/8) Epoch 20, batch 2350, loss[loss=0.1864, simple_loss=0.2597, pruned_loss=0.05653, over 16431.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2594, pruned_loss=0.04389, over 3319562.91 frames. ], batch size: 146, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:09:31,699 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=195203.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:10:36,693 INFO [train.py:904] (4/8) Epoch 20, batch 2400, loss[loss=0.1803, simple_loss=0.2742, pruned_loss=0.04324, over 17082.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2602, pruned_loss=0.0436, over 3324702.89 frames. ], batch size: 53, lr: 3.41e-03, grad_scale: 8.0 2023-05-01 01:11:07,015 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.647e+02 2.253e+02 2.639e+02 3.077e+02 5.540e+02, threshold=5.278e+02, percent-clipped=1.0 2023-05-01 01:11:29,091 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=195290.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:11:45,804 INFO [train.py:904] (4/8) Epoch 20, batch 2450, loss[loss=0.1583, simple_loss=0.257, pruned_loss=0.02977, over 17113.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2612, pruned_loss=0.04332, over 3324922.79 frames. ], batch size: 49, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:12:35,686 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=195338.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:12:54,075 INFO [train.py:904] (4/8) Epoch 20, batch 2500, loss[loss=0.1455, simple_loss=0.2317, pruned_loss=0.02967, over 17007.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2608, pruned_loss=0.0432, over 3327241.31 frames. ], batch size: 41, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:12:55,405 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-05-01 01:13:07,197 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-05-01 01:13:15,893 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=195367.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 01:13:26,925 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.442e+02 2.185e+02 2.562e+02 3.023e+02 6.708e+02, threshold=5.124e+02, percent-clipped=4.0 2023-05-01 01:13:50,882 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=195392.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:14:01,552 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=195400.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 01:14:04,077 INFO [train.py:904] (4/8) Epoch 20, batch 2550, loss[loss=0.1779, simple_loss=0.2578, pruned_loss=0.04899, over 16747.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2616, pruned_loss=0.04386, over 3325076.23 frames. ], batch size: 134, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:14:38,117 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=195427.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:15:11,782 INFO [train.py:904] (4/8) Epoch 20, batch 2600, loss[loss=0.1931, simple_loss=0.272, pruned_loss=0.0571, over 12232.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2618, pruned_loss=0.04392, over 3319793.52 frames. ], batch size: 246, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:15:13,360 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=195453.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:15:40,208 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4918, 4.7959, 4.6133, 4.6172, 4.3847, 4.3214, 4.3210, 4.8769], device='cuda:4'), covar=tensor([0.1244, 0.0970, 0.1079, 0.0867, 0.0802, 0.1268, 0.1136, 0.0847], device='cuda:4'), in_proj_covar=tensor([0.0661, 0.0817, 0.0665, 0.0610, 0.0511, 0.0515, 0.0681, 0.0624], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:15:42,987 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.574e+02 2.224e+02 2.625e+02 3.381e+02 7.141e+02, threshold=5.251e+02, percent-clipped=4.0 2023-05-01 01:16:01,494 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7354, 4.6365, 4.5884, 4.2998, 4.3302, 4.6696, 4.5027, 4.3997], device='cuda:4'), covar=tensor([0.0677, 0.0789, 0.0300, 0.0295, 0.0850, 0.0502, 0.0449, 0.0665], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0432, 0.0349, 0.0344, 0.0361, 0.0400, 0.0241, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:16:20,745 INFO [train.py:904] (4/8) Epoch 20, batch 2650, loss[loss=0.1621, simple_loss=0.2595, pruned_loss=0.03236, over 17060.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2617, pruned_loss=0.04296, over 3326776.78 frames. ], batch size: 50, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:16:22,194 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=195503.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:16:24,520 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6873, 1.7985, 2.2823, 2.4811, 2.6019, 2.6515, 1.9189, 2.7630], device='cuda:4'), covar=tensor([0.0165, 0.0434, 0.0309, 0.0278, 0.0285, 0.0244, 0.0494, 0.0148], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0194, 0.0179, 0.0183, 0.0195, 0.0153, 0.0196, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:16:48,636 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2308, 4.1897, 4.5817, 4.5486, 4.5867, 4.2860, 4.3018, 4.2029], device='cuda:4'), covar=tensor([0.0367, 0.0631, 0.0392, 0.0443, 0.0511, 0.0429, 0.0821, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0448, 0.0432, 0.0406, 0.0480, 0.0457, 0.0548, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 01:17:03,652 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 01:17:20,708 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-05-01 01:17:28,727 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=195551.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:17:29,628 INFO [train.py:904] (4/8) Epoch 20, batch 2700, loss[loss=0.164, simple_loss=0.2596, pruned_loss=0.0342, over 17111.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2618, pruned_loss=0.04262, over 3331824.01 frames. ], batch size: 49, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:18:00,661 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.486e+02 2.029e+02 2.571e+02 3.039e+02 8.808e+02, threshold=5.142e+02, percent-clipped=4.0 2023-05-01 01:18:31,030 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.59 vs. limit=2.0 2023-05-01 01:18:39,436 INFO [train.py:904] (4/8) Epoch 20, batch 2750, loss[loss=0.166, simple_loss=0.2643, pruned_loss=0.03382, over 17247.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2616, pruned_loss=0.04208, over 3330350.00 frames. ], batch size: 52, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:19:40,413 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3322, 4.6551, 4.4712, 4.4749, 4.2164, 4.1709, 4.2072, 4.7075], device='cuda:4'), covar=tensor([0.1221, 0.0942, 0.1048, 0.0862, 0.0761, 0.1595, 0.1141, 0.0886], device='cuda:4'), in_proj_covar=tensor([0.0667, 0.0822, 0.0670, 0.0615, 0.0515, 0.0519, 0.0686, 0.0628], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:19:47,489 INFO [train.py:904] (4/8) Epoch 20, batch 2800, loss[loss=0.1502, simple_loss=0.2393, pruned_loss=0.03052, over 17217.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2609, pruned_loss=0.04186, over 3330111.86 frames. ], batch size: 44, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:20:07,203 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=195667.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 01:20:18,614 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.320e+02 2.150e+02 2.557e+02 3.074e+02 6.713e+02, threshold=5.114e+02, percent-clipped=2.0 2023-05-01 01:20:39,435 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9288, 4.1673, 2.6586, 4.7240, 3.1579, 4.6673, 2.7554, 3.3402], device='cuda:4'), covar=tensor([0.0270, 0.0308, 0.1488, 0.0240, 0.0759, 0.0380, 0.1349, 0.0697], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0179, 0.0195, 0.0165, 0.0178, 0.0219, 0.0203, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:20:52,433 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=195700.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 01:20:54,302 INFO [train.py:904] (4/8) Epoch 20, batch 2850, loss[loss=0.1663, simple_loss=0.2565, pruned_loss=0.0381, over 16779.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2611, pruned_loss=0.04207, over 3332100.06 frames. ], batch size: 83, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:21:13,265 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=195715.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:21:28,575 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-05-01 01:21:29,679 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=195727.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:21:56,478 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=195748.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 01:21:56,485 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=195748.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:22:02,475 INFO [train.py:904] (4/8) Epoch 20, batch 2900, loss[loss=0.1858, simple_loss=0.2762, pruned_loss=0.04773, over 15994.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2601, pruned_loss=0.0428, over 3321114.97 frames. ], batch size: 35, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:22:33,087 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.525e+02 2.294e+02 2.743e+02 3.539e+02 8.268e+02, threshold=5.486e+02, percent-clipped=5.0 2023-05-01 01:22:33,331 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=195775.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:22:44,746 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5719, 3.6070, 3.8654, 2.6957, 3.5081, 3.9357, 3.6220, 2.2400], device='cuda:4'), covar=tensor([0.0512, 0.0278, 0.0058, 0.0387, 0.0137, 0.0111, 0.0111, 0.0508], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0083, 0.0082, 0.0134, 0.0097, 0.0109, 0.0094, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0004], device='cuda:4') 2023-05-01 01:23:10,989 INFO [train.py:904] (4/8) Epoch 20, batch 2950, loss[loss=0.1864, simple_loss=0.2609, pruned_loss=0.05599, over 16877.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2599, pruned_loss=0.04409, over 3324255.90 frames. ], batch size: 116, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:23:38,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4886, 2.9098, 3.0763, 2.1133, 2.7101, 2.1544, 3.0611, 3.2064], device='cuda:4'), covar=tensor([0.0299, 0.0851, 0.0552, 0.1779, 0.0846, 0.0973, 0.0615, 0.0791], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0164, 0.0166, 0.0151, 0.0144, 0.0129, 0.0144, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:24:12,448 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0054, 4.7875, 5.0447, 5.2236, 5.4465, 4.7070, 5.4067, 5.4277], device='cuda:4'), covar=tensor([0.1856, 0.1357, 0.1681, 0.0803, 0.0549, 0.0945, 0.0612, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0663, 0.0821, 0.0957, 0.0840, 0.0627, 0.0656, 0.0674, 0.0778], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:24:18,942 INFO [train.py:904] (4/8) Epoch 20, batch 3000, loss[loss=0.1943, simple_loss=0.271, pruned_loss=0.05884, over 16400.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2602, pruned_loss=0.04431, over 3307864.63 frames. ], batch size: 146, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:24:18,942 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 01:24:27,132 INFO [train.py:938] (4/8) Epoch 20, validation: loss=0.1354, simple_loss=0.2409, pruned_loss=0.01492, over 944034.00 frames. 2023-05-01 01:24:27,133 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 01:24:31,445 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6918, 3.7203, 2.2436, 4.3414, 2.8766, 4.3401, 2.4811, 3.0415], device='cuda:4'), covar=tensor([0.0299, 0.0410, 0.1712, 0.0488, 0.0791, 0.0488, 0.1473, 0.0775], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0179, 0.0195, 0.0165, 0.0178, 0.0220, 0.0203, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:24:49,752 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6883, 6.0330, 5.7923, 5.8137, 5.4717, 5.4804, 5.4577, 6.1278], device='cuda:4'), covar=tensor([0.1320, 0.0942, 0.0994, 0.0854, 0.0879, 0.0630, 0.1103, 0.0887], device='cuda:4'), in_proj_covar=tensor([0.0663, 0.0822, 0.0667, 0.0612, 0.0514, 0.0515, 0.0684, 0.0628], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:24:58,716 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.513e+02 2.268e+02 2.640e+02 3.074e+02 4.713e+02, threshold=5.280e+02, percent-clipped=0.0 2023-05-01 01:25:01,992 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7558, 4.2559, 3.1309, 2.3260, 2.7357, 2.5774, 4.6461, 3.6221], device='cuda:4'), covar=tensor([0.2858, 0.0584, 0.1619, 0.2554, 0.2660, 0.1978, 0.0314, 0.1319], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0270, 0.0302, 0.0306, 0.0296, 0.0255, 0.0292, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 01:25:38,246 INFO [train.py:904] (4/8) Epoch 20, batch 3050, loss[loss=0.1852, simple_loss=0.2588, pruned_loss=0.0558, over 16915.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2591, pruned_loss=0.0438, over 3313072.20 frames. ], batch size: 116, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:25:49,049 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.11 vs. limit=2.0 2023-05-01 01:26:06,368 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2662, 4.1153, 4.4462, 2.2738, 4.7282, 4.7646, 3.3655, 3.7023], device='cuda:4'), covar=tensor([0.0643, 0.0243, 0.0240, 0.1129, 0.0067, 0.0145, 0.0406, 0.0378], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0110, 0.0099, 0.0141, 0.0081, 0.0126, 0.0129, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:26:39,391 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-05-01 01:26:46,789 INFO [train.py:904] (4/8) Epoch 20, batch 3100, loss[loss=0.1821, simple_loss=0.2728, pruned_loss=0.04566, over 17071.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2584, pruned_loss=0.04318, over 3320766.93 frames. ], batch size: 55, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:26:48,738 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-05-01 01:27:08,214 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3798, 3.4684, 3.6875, 2.4752, 3.3763, 3.7558, 3.4735, 2.0457], device='cuda:4'), covar=tensor([0.0526, 0.0156, 0.0066, 0.0415, 0.0116, 0.0103, 0.0103, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0083, 0.0083, 0.0134, 0.0097, 0.0109, 0.0094, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0004], device='cuda:4') 2023-05-01 01:27:16,965 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 2.170e+02 2.517e+02 3.010e+02 4.589e+02, threshold=5.034e+02, percent-clipped=0.0 2023-05-01 01:27:55,510 INFO [train.py:904] (4/8) Epoch 20, batch 3150, loss[loss=0.1686, simple_loss=0.2532, pruned_loss=0.04197, over 16451.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2574, pruned_loss=0.04231, over 3329119.28 frames. ], batch size: 68, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:28:44,338 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0126, 4.7586, 5.0672, 5.2196, 5.4342, 4.7261, 5.4092, 5.4012], device='cuda:4'), covar=tensor([0.1839, 0.1370, 0.1591, 0.0760, 0.0480, 0.0944, 0.0514, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0665, 0.0824, 0.0960, 0.0842, 0.0629, 0.0657, 0.0673, 0.0782], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:28:57,097 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=196048.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:29:03,217 INFO [train.py:904] (4/8) Epoch 20, batch 3200, loss[loss=0.171, simple_loss=0.2542, pruned_loss=0.04388, over 16835.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2572, pruned_loss=0.04281, over 3319500.31 frames. ], batch size: 102, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:29:35,517 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 2.205e+02 2.543e+02 3.042e+02 4.560e+02, threshold=5.087e+02, percent-clipped=0.0 2023-05-01 01:29:42,517 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2460, 5.0332, 5.2050, 5.4435, 5.5919, 4.8450, 5.5610, 5.5994], device='cuda:4'), covar=tensor([0.1885, 0.1420, 0.1865, 0.0879, 0.0723, 0.0978, 0.0674, 0.0739], device='cuda:4'), in_proj_covar=tensor([0.0666, 0.0825, 0.0961, 0.0843, 0.0630, 0.0658, 0.0673, 0.0783], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:30:04,479 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=196096.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:30:09,166 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2734, 4.1269, 4.3311, 4.4516, 4.5586, 4.1019, 4.3408, 4.5453], device='cuda:4'), covar=tensor([0.1574, 0.1139, 0.1282, 0.0676, 0.0598, 0.1228, 0.2005, 0.0672], device='cuda:4'), in_proj_covar=tensor([0.0666, 0.0825, 0.0961, 0.0843, 0.0630, 0.0657, 0.0673, 0.0783], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:30:11,565 INFO [train.py:904] (4/8) Epoch 20, batch 3250, loss[loss=0.2266, simple_loss=0.2926, pruned_loss=0.08031, over 16927.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.257, pruned_loss=0.04271, over 3318408.03 frames. ], batch size: 109, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:30:30,781 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2133, 3.3196, 3.4085, 2.2794, 2.9918, 2.3098, 3.7414, 3.7095], device='cuda:4'), covar=tensor([0.0208, 0.0892, 0.0601, 0.1705, 0.0792, 0.1008, 0.0457, 0.0746], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0163, 0.0165, 0.0150, 0.0143, 0.0128, 0.0143, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:30:37,762 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.34 vs. limit=5.0 2023-05-01 01:31:19,979 INFO [train.py:904] (4/8) Epoch 20, batch 3300, loss[loss=0.1892, simple_loss=0.2712, pruned_loss=0.05361, over 16296.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2585, pruned_loss=0.04327, over 3311541.11 frames. ], batch size: 165, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:31:52,354 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.603e+02 2.123e+02 2.550e+02 3.106e+02 4.546e+02, threshold=5.100e+02, percent-clipped=0.0 2023-05-01 01:31:53,967 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2599, 3.2741, 2.0871, 3.4343, 2.5871, 3.4516, 2.2347, 2.6789], device='cuda:4'), covar=tensor([0.0274, 0.0419, 0.1531, 0.0318, 0.0771, 0.0790, 0.1345, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0180, 0.0197, 0.0167, 0.0179, 0.0222, 0.0204, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:31:56,406 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7784, 4.2433, 4.2458, 3.0049, 3.5609, 4.2386, 3.8661, 2.5035], device='cuda:4'), covar=tensor([0.0476, 0.0070, 0.0049, 0.0362, 0.0133, 0.0094, 0.0081, 0.0444], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0083, 0.0083, 0.0134, 0.0097, 0.0109, 0.0094, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0004], device='cuda:4') 2023-05-01 01:32:28,252 INFO [train.py:904] (4/8) Epoch 20, batch 3350, loss[loss=0.1469, simple_loss=0.2346, pruned_loss=0.02957, over 15815.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2596, pruned_loss=0.04373, over 3308580.25 frames. ], batch size: 35, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:33:35,777 INFO [train.py:904] (4/8) Epoch 20, batch 3400, loss[loss=0.1848, simple_loss=0.2623, pruned_loss=0.05371, over 16263.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.259, pruned_loss=0.04314, over 3317601.62 frames. ], batch size: 165, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:34:06,840 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.504e+02 2.175e+02 2.675e+02 3.072e+02 5.322e+02, threshold=5.351e+02, percent-clipped=1.0 2023-05-01 01:34:40,193 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9235, 2.0777, 2.5363, 2.9635, 2.7319, 3.4331, 2.3988, 3.3624], device='cuda:4'), covar=tensor([0.0229, 0.0495, 0.0324, 0.0301, 0.0332, 0.0176, 0.0437, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0194, 0.0178, 0.0183, 0.0196, 0.0154, 0.0195, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:34:44,339 INFO [train.py:904] (4/8) Epoch 20, batch 3450, loss[loss=0.1563, simple_loss=0.2518, pruned_loss=0.0304, over 17134.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.258, pruned_loss=0.04239, over 3312507.74 frames. ], batch size: 49, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:35:10,244 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=196321.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:35:31,757 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9753, 3.1026, 3.1790, 2.0375, 2.7098, 2.2004, 3.4973, 3.4569], device='cuda:4'), covar=tensor([0.0227, 0.0899, 0.0633, 0.1911, 0.0857, 0.1017, 0.0495, 0.0801], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0164, 0.0166, 0.0151, 0.0144, 0.0129, 0.0144, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:35:50,466 INFO [train.py:904] (4/8) Epoch 20, batch 3500, loss[loss=0.1759, simple_loss=0.2686, pruned_loss=0.0416, over 16715.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2571, pruned_loss=0.04196, over 3314145.36 frames. ], batch size: 62, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:36:03,315 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=196360.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:36:23,665 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.372e+02 2.052e+02 2.465e+02 2.837e+02 5.038e+02, threshold=4.930e+02, percent-clipped=0.0 2023-05-01 01:36:35,840 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=196382.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:36:51,470 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7659, 4.5548, 4.7928, 4.9347, 5.1534, 4.4794, 5.0978, 5.1248], device='cuda:4'), covar=tensor([0.1858, 0.1422, 0.1742, 0.0917, 0.0657, 0.1192, 0.0762, 0.0816], device='cuda:4'), in_proj_covar=tensor([0.0664, 0.0824, 0.0963, 0.0844, 0.0628, 0.0657, 0.0673, 0.0781], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:37:01,667 INFO [train.py:904] (4/8) Epoch 20, batch 3550, loss[loss=0.1963, simple_loss=0.2655, pruned_loss=0.0635, over 16864.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.256, pruned_loss=0.04192, over 3312147.37 frames. ], batch size: 116, lr: 3.40e-03, grad_scale: 8.0 2023-05-01 01:37:28,345 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=196421.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:37:45,335 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8463, 3.9794, 2.2850, 4.5901, 2.9998, 4.4800, 2.3930, 3.2021], device='cuda:4'), covar=tensor([0.0286, 0.0379, 0.1677, 0.0257, 0.0764, 0.0490, 0.1637, 0.0707], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0180, 0.0195, 0.0166, 0.0178, 0.0221, 0.0204, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:38:10,216 INFO [train.py:904] (4/8) Epoch 20, batch 3600, loss[loss=0.1483, simple_loss=0.2373, pruned_loss=0.02965, over 15855.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2539, pruned_loss=0.04126, over 3302752.71 frames. ], batch size: 35, lr: 3.39e-03, grad_scale: 8.0 2023-05-01 01:38:32,262 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6565, 1.9424, 2.3104, 2.5091, 2.6374, 2.6623, 1.8890, 2.8196], device='cuda:4'), covar=tensor([0.0181, 0.0442, 0.0325, 0.0291, 0.0270, 0.0287, 0.0505, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0193, 0.0178, 0.0182, 0.0196, 0.0154, 0.0194, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:38:33,501 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1250, 4.8188, 5.1575, 5.2803, 5.5333, 4.7681, 5.5126, 5.5098], device='cuda:4'), covar=tensor([0.1867, 0.1448, 0.1723, 0.0926, 0.0537, 0.0940, 0.0495, 0.0587], device='cuda:4'), in_proj_covar=tensor([0.0663, 0.0822, 0.0959, 0.0842, 0.0627, 0.0655, 0.0669, 0.0781], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:38:41,980 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.511e+02 2.224e+02 2.585e+02 3.008e+02 4.978e+02, threshold=5.169e+02, percent-clipped=1.0 2023-05-01 01:39:20,690 INFO [train.py:904] (4/8) Epoch 20, batch 3650, loss[loss=0.176, simple_loss=0.2516, pruned_loss=0.0502, over 15447.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2531, pruned_loss=0.04198, over 3284165.03 frames. ], batch size: 190, lr: 3.39e-03, grad_scale: 8.0 2023-05-01 01:40:14,935 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-05-01 01:40:32,641 INFO [train.py:904] (4/8) Epoch 20, batch 3700, loss[loss=0.1682, simple_loss=0.2403, pruned_loss=0.04802, over 16820.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.252, pruned_loss=0.04345, over 3276700.55 frames. ], batch size: 102, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:41:07,134 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.139e+02 2.467e+02 2.956e+02 4.680e+02, threshold=4.935e+02, percent-clipped=0.0 2023-05-01 01:41:11,824 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=196578.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:41:39,142 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4621, 2.9083, 2.9751, 2.0766, 2.6193, 2.1343, 3.0793, 3.1996], device='cuda:4'), covar=tensor([0.0299, 0.0867, 0.0629, 0.1907, 0.0947, 0.1070, 0.0640, 0.0759], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0162, 0.0164, 0.0150, 0.0142, 0.0127, 0.0143, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:41:47,078 INFO [train.py:904] (4/8) Epoch 20, batch 3750, loss[loss=0.1536, simple_loss=0.2293, pruned_loss=0.03891, over 16803.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2522, pruned_loss=0.04463, over 3256620.19 frames. ], batch size: 102, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:42:25,696 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3916, 5.4087, 5.1783, 4.5169, 5.3541, 2.0577, 5.0791, 4.7785], device='cuda:4'), covar=tensor([0.0055, 0.0044, 0.0156, 0.0308, 0.0061, 0.2583, 0.0097, 0.0229], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0152, 0.0198, 0.0178, 0.0175, 0.0206, 0.0188, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:42:38,983 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=196639.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:42:45,922 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-05-01 01:42:58,327 INFO [train.py:904] (4/8) Epoch 20, batch 3800, loss[loss=0.1908, simple_loss=0.2702, pruned_loss=0.05571, over 16405.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2538, pruned_loss=0.04605, over 3258734.22 frames. ], batch size: 146, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:43:31,146 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.679e+02 2.212e+02 2.504e+02 2.913e+02 5.553e+02, threshold=5.009e+02, percent-clipped=1.0 2023-05-01 01:43:35,322 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=196677.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:43:56,252 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6061, 4.4565, 4.6380, 4.8026, 4.9427, 4.4681, 4.7906, 4.9117], device='cuda:4'), covar=tensor([0.1611, 0.1215, 0.1456, 0.0709, 0.0579, 0.1025, 0.1603, 0.0689], device='cuda:4'), in_proj_covar=tensor([0.0655, 0.0814, 0.0947, 0.0829, 0.0621, 0.0646, 0.0667, 0.0773], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:44:10,772 INFO [train.py:904] (4/8) Epoch 20, batch 3850, loss[loss=0.1716, simple_loss=0.2462, pruned_loss=0.04848, over 16747.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2541, pruned_loss=0.04687, over 3267207.08 frames. ], batch size: 83, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:44:17,184 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-05-01 01:44:32,369 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=196716.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:45:24,170 INFO [train.py:904] (4/8) Epoch 20, batch 3900, loss[loss=0.1639, simple_loss=0.2474, pruned_loss=0.04019, over 17112.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2542, pruned_loss=0.04728, over 3267565.75 frames. ], batch size: 48, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:45:57,453 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.527e+02 2.027e+02 2.552e+02 2.919e+02 4.484e+02, threshold=5.103e+02, percent-clipped=0.0 2023-05-01 01:46:04,915 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=196779.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:46:36,464 INFO [train.py:904] (4/8) Epoch 20, batch 3950, loss[loss=0.1833, simple_loss=0.2519, pruned_loss=0.05733, over 16728.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2534, pruned_loss=0.04744, over 3268481.59 frames. ], batch size: 124, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:46:58,099 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3829, 3.9435, 4.0032, 2.7674, 3.6446, 4.0706, 3.6874, 2.1718], device='cuda:4'), covar=tensor([0.0509, 0.0144, 0.0052, 0.0356, 0.0096, 0.0098, 0.0095, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0082, 0.0081, 0.0132, 0.0096, 0.0107, 0.0093, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 01:47:11,104 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8421, 2.7310, 2.5797, 4.0288, 3.3776, 4.0508, 1.6227, 2.9300], device='cuda:4'), covar=tensor([0.1259, 0.0650, 0.1082, 0.0198, 0.0134, 0.0341, 0.1489, 0.0780], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0173, 0.0192, 0.0186, 0.0205, 0.0215, 0.0198, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:47:32,094 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=196840.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:47:48,960 INFO [train.py:904] (4/8) Epoch 20, batch 4000, loss[loss=0.1871, simple_loss=0.2565, pruned_loss=0.05882, over 16842.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2531, pruned_loss=0.0474, over 3276866.25 frames. ], batch size: 116, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:47:52,321 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=196854.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:48:21,943 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.289e+02 2.579e+02 3.040e+02 5.791e+02, threshold=5.158e+02, percent-clipped=1.0 2023-05-01 01:48:32,425 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=196881.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:48:38,650 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7655, 2.3567, 2.3235, 3.0711, 2.2212, 3.5220, 1.6201, 2.6288], device='cuda:4'), covar=tensor([0.1280, 0.0804, 0.1158, 0.0152, 0.0163, 0.0315, 0.1531, 0.0897], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0173, 0.0192, 0.0187, 0.0205, 0.0215, 0.0198, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:49:01,021 INFO [train.py:904] (4/8) Epoch 20, batch 4050, loss[loss=0.1604, simple_loss=0.2452, pruned_loss=0.03776, over 17116.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2541, pruned_loss=0.04696, over 3270402.49 frames. ], batch size: 48, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:49:17,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7418, 4.8601, 5.1512, 5.0961, 5.1215, 4.8107, 4.7949, 4.6412], device='cuda:4'), covar=tensor([0.0268, 0.0519, 0.0311, 0.0401, 0.0417, 0.0335, 0.0788, 0.0422], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0453, 0.0436, 0.0408, 0.0484, 0.0462, 0.0550, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 01:49:19,949 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=196915.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:49:47,740 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=196934.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:49:58,034 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=196942.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:50:12,374 INFO [train.py:904] (4/8) Epoch 20, batch 4100, loss[loss=0.1911, simple_loss=0.2749, pruned_loss=0.05369, over 16557.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2559, pruned_loss=0.04657, over 3267718.96 frames. ], batch size: 57, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:50:23,463 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-05-01 01:50:34,498 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4970, 4.5396, 4.7265, 4.5096, 4.5733, 5.1439, 4.6687, 4.3630], device='cuda:4'), covar=tensor([0.1353, 0.1872, 0.1920, 0.1868, 0.2413, 0.0935, 0.1430, 0.2278], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0600, 0.0658, 0.0502, 0.0664, 0.0693, 0.0513, 0.0669], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 01:50:35,013 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.76 vs. limit=5.0 2023-05-01 01:50:44,593 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.448e+02 1.746e+02 2.102e+02 2.519e+02 4.672e+02, threshold=4.204e+02, percent-clipped=0.0 2023-05-01 01:50:48,591 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=196977.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:51:07,835 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2437, 3.9340, 3.8076, 2.6028, 3.4194, 3.8448, 3.4734, 1.9838], device='cuda:4'), covar=tensor([0.0550, 0.0041, 0.0059, 0.0392, 0.0113, 0.0121, 0.0108, 0.0509], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0082, 0.0081, 0.0133, 0.0096, 0.0107, 0.0093, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 01:51:23,807 INFO [train.py:904] (4/8) Epoch 20, batch 4150, loss[loss=0.2108, simple_loss=0.2955, pruned_loss=0.06301, over 16536.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2631, pruned_loss=0.04915, over 3249454.52 frames. ], batch size: 75, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:51:45,576 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=197016.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:52:00,049 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=197025.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:52:37,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7576, 3.8825, 2.9481, 2.2929, 2.6712, 2.6329, 4.3336, 3.4227], device='cuda:4'), covar=tensor([0.2681, 0.0637, 0.1680, 0.2550, 0.2625, 0.1800, 0.0395, 0.1205], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0270, 0.0303, 0.0308, 0.0298, 0.0254, 0.0293, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 01:52:40,715 INFO [train.py:904] (4/8) Epoch 20, batch 4200, loss[loss=0.2179, simple_loss=0.3024, pruned_loss=0.06675, over 11651.00 frames. ], tot_loss[loss=0.1857, simple_loss=0.27, pruned_loss=0.05073, over 3216273.87 frames. ], batch size: 247, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:52:58,762 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=197064.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:53:12,969 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.581e+02 2.329e+02 2.798e+02 3.249e+02 9.837e+02, threshold=5.595e+02, percent-clipped=2.0 2023-05-01 01:53:51,290 INFO [train.py:904] (4/8) Epoch 20, batch 4250, loss[loss=0.1838, simple_loss=0.28, pruned_loss=0.04375, over 16319.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2731, pruned_loss=0.05031, over 3194938.59 frames. ], batch size: 165, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:54:15,103 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1172, 2.0358, 1.7398, 1.7553, 2.2808, 1.9518, 1.9217, 2.3589], device='cuda:4'), covar=tensor([0.0197, 0.0345, 0.0473, 0.0412, 0.0213, 0.0300, 0.0163, 0.0229], device='cuda:4'), in_proj_covar=tensor([0.0205, 0.0233, 0.0223, 0.0225, 0.0234, 0.0233, 0.0235, 0.0229], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 01:54:40,508 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=197135.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:55:04,298 INFO [train.py:904] (4/8) Epoch 20, batch 4300, loss[loss=0.2068, simple_loss=0.2903, pruned_loss=0.06167, over 11674.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2745, pruned_loss=0.04958, over 3194593.26 frames. ], batch size: 248, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:55:25,950 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.12 vs. limit=5.0 2023-05-01 01:55:37,987 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.176e+02 2.515e+02 3.039e+02 5.119e+02, threshold=5.030e+02, percent-clipped=0.0 2023-05-01 01:55:45,560 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-05-01 01:56:12,629 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1625, 2.8887, 3.0574, 1.7804, 3.2718, 3.3163, 2.6179, 2.5954], device='cuda:4'), covar=tensor([0.0837, 0.0281, 0.0254, 0.1156, 0.0090, 0.0158, 0.0482, 0.0466], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0097, 0.0139, 0.0079, 0.0124, 0.0127, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:56:18,918 INFO [train.py:904] (4/8) Epoch 20, batch 4350, loss[loss=0.1896, simple_loss=0.2819, pruned_loss=0.04868, over 16782.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2775, pruned_loss=0.05057, over 3193657.00 frames. ], batch size: 83, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:56:21,039 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-05-01 01:56:29,917 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=197210.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:57:05,402 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=197234.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:57:09,112 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=197237.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:57:31,880 INFO [train.py:904] (4/8) Epoch 20, batch 4400, loss[loss=0.1745, simple_loss=0.2674, pruned_loss=0.04084, over 16555.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2793, pruned_loss=0.05165, over 3179039.45 frames. ], batch size: 68, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:57:48,691 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6286, 2.5758, 1.8567, 2.7597, 2.0876, 2.7889, 2.1221, 2.3944], device='cuda:4'), covar=tensor([0.0271, 0.0331, 0.1260, 0.0197, 0.0639, 0.0363, 0.1033, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0176, 0.0193, 0.0160, 0.0175, 0.0215, 0.0200, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 01:58:04,301 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.492e+02 2.097e+02 2.441e+02 2.762e+02 5.334e+02, threshold=4.883e+02, percent-clipped=1.0 2023-05-01 01:58:05,005 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.22 vs. limit=5.0 2023-05-01 01:58:14,499 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=197282.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 01:58:42,794 INFO [train.py:904] (4/8) Epoch 20, batch 4450, loss[loss=0.2343, simple_loss=0.3042, pruned_loss=0.0822, over 11645.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.283, pruned_loss=0.05295, over 3182798.54 frames. ], batch size: 246, lr: 3.39e-03, grad_scale: 16.0 2023-05-01 01:59:55,983 INFO [train.py:904] (4/8) Epoch 20, batch 4500, loss[loss=0.2233, simple_loss=0.2932, pruned_loss=0.07667, over 11502.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2835, pruned_loss=0.05374, over 3176610.14 frames. ], batch size: 246, lr: 3.39e-03, grad_scale: 8.0 2023-05-01 02:00:15,399 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6135, 2.8337, 2.5032, 2.5740, 3.1589, 2.7717, 3.1805, 3.2789], device='cuda:4'), covar=tensor([0.0076, 0.0308, 0.0386, 0.0376, 0.0185, 0.0293, 0.0193, 0.0205], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0231, 0.0222, 0.0223, 0.0233, 0.0231, 0.0233, 0.0228], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:00:30,608 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.345e+02 1.873e+02 2.163e+02 2.505e+02 3.448e+02, threshold=4.325e+02, percent-clipped=0.0 2023-05-01 02:01:08,131 INFO [train.py:904] (4/8) Epoch 20, batch 4550, loss[loss=0.2217, simple_loss=0.3019, pruned_loss=0.07077, over 12515.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2844, pruned_loss=0.05475, over 3192619.86 frames. ], batch size: 246, lr: 3.39e-03, grad_scale: 8.0 2023-05-01 02:01:27,201 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=197415.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 02:01:56,097 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=197435.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:01:57,315 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1202, 3.9323, 3.7523, 2.4096, 3.5360, 3.8949, 3.4964, 2.0650], device='cuda:4'), covar=tensor([0.0592, 0.0035, 0.0048, 0.0433, 0.0081, 0.0078, 0.0082, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0082, 0.0081, 0.0133, 0.0096, 0.0107, 0.0094, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 02:02:18,308 INFO [train.py:904] (4/8) Epoch 20, batch 4600, loss[loss=0.1915, simple_loss=0.2819, pruned_loss=0.05057, over 16787.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2858, pruned_loss=0.05555, over 3194983.27 frames. ], batch size: 83, lr: 3.39e-03, grad_scale: 8.0 2023-05-01 02:02:51,733 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-01 02:02:52,206 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 1.798e+02 2.178e+02 2.547e+02 5.479e+02, threshold=4.357e+02, percent-clipped=1.0 2023-05-01 02:02:52,807 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=197476.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 02:03:03,000 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=197483.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:03:30,134 INFO [train.py:904] (4/8) Epoch 20, batch 4650, loss[loss=0.1861, simple_loss=0.2726, pruned_loss=0.04982, over 16745.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2839, pruned_loss=0.05503, over 3220098.15 frames. ], batch size: 89, lr: 3.39e-03, grad_scale: 8.0 2023-05-01 02:03:41,261 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=197510.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:03:49,515 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=197516.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:04:20,718 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=197537.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:04:32,444 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=197545.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:04:42,155 INFO [train.py:904] (4/8) Epoch 20, batch 4700, loss[loss=0.1803, simple_loss=0.2705, pruned_loss=0.04505, over 16973.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2815, pruned_loss=0.05368, over 3213739.25 frames. ], batch size: 109, lr: 3.39e-03, grad_scale: 8.0 2023-05-01 02:04:52,429 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=197558.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:05:18,232 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.404e+02 1.893e+02 2.227e+02 2.649e+02 4.562e+02, threshold=4.454e+02, percent-clipped=1.0 2023-05-01 02:05:19,861 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=197577.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 02:05:31,365 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=197585.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:05:55,396 INFO [train.py:904] (4/8) Epoch 20, batch 4750, loss[loss=0.1735, simple_loss=0.2595, pruned_loss=0.04377, over 16954.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2773, pruned_loss=0.05172, over 3216818.54 frames. ], batch size: 109, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:06:01,823 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=197606.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:06:11,387 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.16 vs. limit=5.0 2023-05-01 02:07:01,773 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6742, 3.5116, 4.0605, 1.8113, 4.3080, 4.2709, 3.0005, 2.9972], device='cuda:4'), covar=tensor([0.0788, 0.0297, 0.0166, 0.1314, 0.0056, 0.0122, 0.0406, 0.0514], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0110, 0.0099, 0.0140, 0.0080, 0.0125, 0.0129, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 02:07:08,960 INFO [train.py:904] (4/8) Epoch 20, batch 4800, loss[loss=0.1828, simple_loss=0.277, pruned_loss=0.04426, over 16733.00 frames. ], tot_loss[loss=0.1857, simple_loss=0.2731, pruned_loss=0.04915, over 3230493.56 frames. ], batch size: 124, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:07:45,093 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.779e+02 2.011e+02 2.336e+02 4.336e+02, threshold=4.022e+02, percent-clipped=0.0 2023-05-01 02:08:24,127 INFO [train.py:904] (4/8) Epoch 20, batch 4850, loss[loss=0.1728, simple_loss=0.2675, pruned_loss=0.03903, over 16906.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2747, pruned_loss=0.04895, over 3189827.25 frames. ], batch size: 109, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:08:46,853 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.93 vs. limit=5.0 2023-05-01 02:08:57,123 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5206, 4.3584, 4.5885, 4.7406, 4.9169, 4.3369, 4.8669, 4.9108], device='cuda:4'), covar=tensor([0.1575, 0.1210, 0.1462, 0.0694, 0.0455, 0.1031, 0.0557, 0.0607], device='cuda:4'), in_proj_covar=tensor([0.0620, 0.0767, 0.0891, 0.0782, 0.0585, 0.0613, 0.0628, 0.0727], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:09:40,078 INFO [train.py:904] (4/8) Epoch 20, batch 4900, loss[loss=0.1707, simple_loss=0.2612, pruned_loss=0.04007, over 16622.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2736, pruned_loss=0.0475, over 3186933.32 frames. ], batch size: 89, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:09:55,940 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5543, 2.6388, 2.0893, 2.4192, 3.0256, 2.6947, 3.0386, 3.2414], device='cuda:4'), covar=tensor([0.0088, 0.0423, 0.0596, 0.0470, 0.0242, 0.0395, 0.0271, 0.0272], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0229, 0.0222, 0.0222, 0.0232, 0.0231, 0.0232, 0.0228], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:10:08,660 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=197771.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 02:10:15,936 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.320e+02 1.929e+02 2.192e+02 2.700e+02 5.551e+02, threshold=4.384e+02, percent-clipped=1.0 2023-05-01 02:10:48,290 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5042, 3.6276, 3.3469, 3.0891, 3.0886, 3.5185, 3.2991, 3.1907], device='cuda:4'), covar=tensor([0.0656, 0.0609, 0.0355, 0.0313, 0.0737, 0.0539, 0.1482, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0409, 0.0331, 0.0327, 0.0342, 0.0378, 0.0228, 0.0395], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:10:52,823 INFO [train.py:904] (4/8) Epoch 20, batch 4950, loss[loss=0.1908, simple_loss=0.287, pruned_loss=0.04728, over 16740.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.273, pruned_loss=0.04672, over 3188192.78 frames. ], batch size: 83, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:12:04,523 INFO [train.py:904] (4/8) Epoch 20, batch 5000, loss[loss=0.1737, simple_loss=0.2637, pruned_loss=0.04187, over 17027.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2748, pruned_loss=0.04657, over 3197681.32 frames. ], batch size: 50, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:12:33,417 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=197872.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 02:12:38,857 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.376e+02 1.953e+02 2.351e+02 2.789e+02 5.042e+02, threshold=4.702e+02, percent-clipped=2.0 2023-05-01 02:13:01,260 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=197892.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:13:13,653 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=197901.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:13:14,487 INFO [train.py:904] (4/8) Epoch 20, batch 5050, loss[loss=0.1969, simple_loss=0.2863, pruned_loss=0.05379, over 16686.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2757, pruned_loss=0.04684, over 3196321.03 frames. ], batch size: 57, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:13:35,064 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5237, 4.3862, 4.5633, 4.7287, 4.9299, 4.4222, 4.8948, 4.9266], device='cuda:4'), covar=tensor([0.1851, 0.1169, 0.1526, 0.0753, 0.0451, 0.0961, 0.0502, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0625, 0.0773, 0.0900, 0.0790, 0.0588, 0.0618, 0.0633, 0.0732], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:13:39,541 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-05-01 02:14:25,041 INFO [train.py:904] (4/8) Epoch 20, batch 5100, loss[loss=0.1782, simple_loss=0.2695, pruned_loss=0.0435, over 16723.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2733, pruned_loss=0.04589, over 3207592.24 frames. ], batch size: 124, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:14:27,940 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=197953.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:15:00,608 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.528e+02 1.957e+02 2.274e+02 2.606e+02 4.454e+02, threshold=4.549e+02, percent-clipped=0.0 2023-05-01 02:15:41,431 INFO [train.py:904] (4/8) Epoch 20, batch 5150, loss[loss=0.1832, simple_loss=0.2786, pruned_loss=0.0439, over 11955.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2737, pruned_loss=0.04539, over 3190793.83 frames. ], batch size: 246, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:16:52,361 INFO [train.py:904] (4/8) Epoch 20, batch 5200, loss[loss=0.1871, simple_loss=0.2709, pruned_loss=0.05163, over 16931.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2715, pruned_loss=0.04489, over 3205743.21 frames. ], batch size: 109, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:17:20,354 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=198071.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 02:17:27,747 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.470e+02 1.904e+02 2.236e+02 2.771e+02 4.329e+02, threshold=4.472e+02, percent-clipped=0.0 2023-05-01 02:17:34,813 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=198081.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:17:48,842 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1080, 2.1797, 2.2637, 3.8135, 2.0616, 2.5744, 2.2743, 2.3681], device='cuda:4'), covar=tensor([0.1434, 0.3678, 0.2844, 0.0536, 0.4162, 0.2470, 0.3621, 0.3089], device='cuda:4'), in_proj_covar=tensor([0.0396, 0.0439, 0.0360, 0.0322, 0.0431, 0.0506, 0.0408, 0.0514], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:18:03,836 INFO [train.py:904] (4/8) Epoch 20, batch 5250, loss[loss=0.1653, simple_loss=0.2578, pruned_loss=0.03639, over 16719.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2691, pruned_loss=0.04444, over 3207450.02 frames. ], batch size: 134, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:18:28,862 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=198119.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 02:19:01,630 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=198142.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:19:15,056 INFO [train.py:904] (4/8) Epoch 20, batch 5300, loss[loss=0.1622, simple_loss=0.2492, pruned_loss=0.03756, over 16390.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2654, pruned_loss=0.04316, over 3220341.60 frames. ], batch size: 146, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:19:45,010 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=198172.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 02:19:49,803 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.400e+02 1.825e+02 2.153e+02 2.534e+02 4.591e+02, threshold=4.307e+02, percent-clipped=1.0 2023-05-01 02:20:27,525 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=198201.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:20:28,366 INFO [train.py:904] (4/8) Epoch 20, batch 5350, loss[loss=0.2008, simple_loss=0.2963, pruned_loss=0.05271, over 16445.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2638, pruned_loss=0.04294, over 3213568.07 frames. ], batch size: 146, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:20:52,641 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3303, 2.3866, 2.4365, 4.1427, 2.2468, 2.7799, 2.4280, 2.5914], device='cuda:4'), covar=tensor([0.1277, 0.3348, 0.2722, 0.0484, 0.3747, 0.2375, 0.3522, 0.2906], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0440, 0.0362, 0.0324, 0.0432, 0.0507, 0.0410, 0.0515], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:20:54,788 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=198220.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:21:08,040 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 02:21:21,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0556, 3.3581, 3.5448, 2.1321, 2.9708, 2.3630, 3.4709, 3.6053], device='cuda:4'), covar=tensor([0.0244, 0.0700, 0.0557, 0.1847, 0.0792, 0.0911, 0.0626, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0162, 0.0166, 0.0150, 0.0143, 0.0128, 0.0144, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 02:21:35,323 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=198248.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:21:37,096 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=198249.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:21:40,983 INFO [train.py:904] (4/8) Epoch 20, batch 5400, loss[loss=0.1876, simple_loss=0.28, pruned_loss=0.04759, over 16825.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2662, pruned_loss=0.04351, over 3213131.29 frames. ], batch size: 42, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:22:15,985 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.461e+02 1.945e+02 2.236e+02 2.644e+02 4.931e+02, threshold=4.472e+02, percent-clipped=2.0 2023-05-01 02:22:57,916 INFO [train.py:904] (4/8) Epoch 20, batch 5450, loss[loss=0.2331, simple_loss=0.3195, pruned_loss=0.0733, over 15126.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2706, pruned_loss=0.04556, over 3217973.68 frames. ], batch size: 190, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:24:14,758 INFO [train.py:904] (4/8) Epoch 20, batch 5500, loss[loss=0.2335, simple_loss=0.3115, pruned_loss=0.07781, over 16434.00 frames. ], tot_loss[loss=0.188, simple_loss=0.2774, pruned_loss=0.0493, over 3208605.54 frames. ], batch size: 146, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:24:51,691 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.837e+02 2.873e+02 3.524e+02 4.449e+02 7.452e+02, threshold=7.049e+02, percent-clipped=24.0 2023-05-01 02:25:34,175 INFO [train.py:904] (4/8) Epoch 20, batch 5550, loss[loss=0.2694, simple_loss=0.3331, pruned_loss=0.1029, over 11107.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.285, pruned_loss=0.05512, over 3145368.37 frames. ], batch size: 247, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:25:54,859 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6358, 2.2403, 1.8927, 2.0185, 2.5850, 2.2188, 2.3808, 2.7248], device='cuda:4'), covar=tensor([0.0181, 0.0412, 0.0502, 0.0427, 0.0237, 0.0358, 0.0168, 0.0237], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0231, 0.0223, 0.0223, 0.0234, 0.0232, 0.0233, 0.0229], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:26:02,752 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2762, 5.2471, 5.1413, 4.7542, 4.7841, 5.1693, 5.0779, 4.8526], device='cuda:4'), covar=tensor([0.0631, 0.0507, 0.0265, 0.0294, 0.0970, 0.0464, 0.0393, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0419, 0.0338, 0.0333, 0.0349, 0.0388, 0.0232, 0.0404], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:26:30,782 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=198437.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:26:53,768 INFO [train.py:904] (4/8) Epoch 20, batch 5600, loss[loss=0.2251, simple_loss=0.306, pruned_loss=0.0721, over 16753.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.29, pruned_loss=0.0591, over 3130119.41 frames. ], batch size: 134, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:27:34,780 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.278e+02 3.283e+02 3.655e+02 4.380e+02 7.107e+02, threshold=7.309e+02, percent-clipped=1.0 2023-05-01 02:28:14,132 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=198499.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:28:18,605 INFO [train.py:904] (4/8) Epoch 20, batch 5650, loss[loss=0.2253, simple_loss=0.3038, pruned_loss=0.0734, over 16693.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2948, pruned_loss=0.06287, over 3105409.50 frames. ], batch size: 134, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:28:44,357 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5512, 4.5182, 4.3562, 2.7890, 3.8753, 4.3937, 3.9019, 2.5384], device='cuda:4'), covar=tensor([0.0519, 0.0033, 0.0041, 0.0404, 0.0098, 0.0130, 0.0091, 0.0400], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0081, 0.0081, 0.0133, 0.0095, 0.0107, 0.0092, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 02:29:24,495 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-05-01 02:29:25,767 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 02:29:32,483 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=198548.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:29:36,994 INFO [train.py:904] (4/8) Epoch 20, batch 5700, loss[loss=0.2597, simple_loss=0.3226, pruned_loss=0.09837, over 11337.00 frames. ], tot_loss[loss=0.2127, simple_loss=0.2964, pruned_loss=0.06453, over 3085163.62 frames. ], batch size: 247, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:29:41,371 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5588, 3.3859, 3.8173, 1.9030, 4.0355, 4.0522, 3.0465, 2.8986], device='cuda:4'), covar=tensor([0.0793, 0.0287, 0.0193, 0.1251, 0.0072, 0.0136, 0.0418, 0.0493], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0108, 0.0097, 0.0139, 0.0080, 0.0124, 0.0128, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 02:29:50,289 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=198560.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:29:54,107 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1770, 2.0570, 1.7804, 1.7692, 2.2708, 1.9542, 1.9926, 2.3411], device='cuda:4'), covar=tensor([0.0176, 0.0325, 0.0450, 0.0424, 0.0240, 0.0344, 0.0175, 0.0238], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0229, 0.0222, 0.0222, 0.0232, 0.0230, 0.0231, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:30:11,489 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2201, 4.2196, 4.0692, 3.3273, 4.1591, 1.6139, 3.9367, 3.7599], device='cuda:4'), covar=tensor([0.0108, 0.0099, 0.0202, 0.0373, 0.0108, 0.3034, 0.0141, 0.0267], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0148, 0.0192, 0.0174, 0.0168, 0.0202, 0.0182, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:30:14,597 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.943e+02 3.352e+02 4.045e+02 5.075e+02 1.137e+03, threshold=8.090e+02, percent-clipped=5.0 2023-05-01 02:30:47,281 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=198596.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:30:55,554 INFO [train.py:904] (4/8) Epoch 20, batch 5750, loss[loss=0.2064, simple_loss=0.2885, pruned_loss=0.06221, over 15179.00 frames. ], tot_loss[loss=0.2153, simple_loss=0.299, pruned_loss=0.0658, over 3080717.55 frames. ], batch size: 190, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:31:17,289 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=198615.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:31:19,563 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7152, 3.6544, 4.0318, 2.1558, 3.1995, 2.6293, 4.0722, 4.0064], device='cuda:4'), covar=tensor([0.0198, 0.0770, 0.0518, 0.1981, 0.0771, 0.0907, 0.0530, 0.0923], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0162, 0.0165, 0.0151, 0.0143, 0.0128, 0.0144, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 02:31:32,347 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4928, 3.4879, 3.4772, 2.8162, 3.3867, 2.0978, 3.1734, 2.8293], device='cuda:4'), covar=tensor([0.0137, 0.0118, 0.0173, 0.0231, 0.0097, 0.2200, 0.0122, 0.0223], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0147, 0.0191, 0.0173, 0.0168, 0.0201, 0.0181, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:31:39,109 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-05-01 02:32:16,951 INFO [train.py:904] (4/8) Epoch 20, batch 5800, loss[loss=0.2204, simple_loss=0.2918, pruned_loss=0.0745, over 12273.00 frames. ], tot_loss[loss=0.2142, simple_loss=0.2982, pruned_loss=0.06504, over 3061221.31 frames. ], batch size: 248, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:32:53,794 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.915e+02 2.730e+02 3.384e+02 4.113e+02 5.841e+02, threshold=6.768e+02, percent-clipped=0.0 2023-05-01 02:32:54,350 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=198676.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:32:56,252 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-05-01 02:32:57,337 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9459, 2.1374, 2.4083, 3.1795, 2.2173, 2.3726, 2.3212, 2.2861], device='cuda:4'), covar=tensor([0.1255, 0.3291, 0.2363, 0.0685, 0.3925, 0.2332, 0.3166, 0.3146], device='cuda:4'), in_proj_covar=tensor([0.0394, 0.0436, 0.0359, 0.0321, 0.0431, 0.0503, 0.0406, 0.0511], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:33:15,735 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-05-01 02:33:34,962 INFO [train.py:904] (4/8) Epoch 20, batch 5850, loss[loss=0.1889, simple_loss=0.272, pruned_loss=0.05295, over 17137.00 frames. ], tot_loss[loss=0.2117, simple_loss=0.2965, pruned_loss=0.06344, over 3073033.86 frames. ], batch size: 47, lr: 3.38e-03, grad_scale: 8.0 2023-05-01 02:34:30,869 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=198737.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:34:53,519 INFO [train.py:904] (4/8) Epoch 20, batch 5900, loss[loss=0.2059, simple_loss=0.2848, pruned_loss=0.06349, over 11761.00 frames. ], tot_loss[loss=0.2118, simple_loss=0.2965, pruned_loss=0.06356, over 3068839.26 frames. ], batch size: 246, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:35:29,367 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 02:35:34,248 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.541e+02 2.837e+02 3.258e+02 4.009e+02 8.301e+02, threshold=6.515e+02, percent-clipped=1.0 2023-05-01 02:35:48,534 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=198785.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:36:08,856 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2326, 2.2345, 2.2642, 3.9682, 2.1445, 2.5780, 2.2854, 2.3909], device='cuda:4'), covar=tensor([0.1366, 0.3376, 0.2825, 0.0501, 0.4171, 0.2435, 0.3423, 0.3169], device='cuda:4'), in_proj_covar=tensor([0.0393, 0.0435, 0.0358, 0.0320, 0.0429, 0.0502, 0.0405, 0.0510], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:36:14,567 INFO [train.py:904] (4/8) Epoch 20, batch 5950, loss[loss=0.211, simple_loss=0.2991, pruned_loss=0.06148, over 15263.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2969, pruned_loss=0.06147, over 3087780.09 frames. ], batch size: 190, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:37:01,084 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6363, 2.5816, 1.9007, 2.7413, 2.1807, 2.7883, 2.1083, 2.3524], device='cuda:4'), covar=tensor([0.0310, 0.0402, 0.1252, 0.0236, 0.0626, 0.0495, 0.1159, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0175, 0.0192, 0.0157, 0.0175, 0.0213, 0.0200, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 02:37:31,021 INFO [train.py:904] (4/8) Epoch 20, batch 6000, loss[loss=0.2428, simple_loss=0.3199, pruned_loss=0.08282, over 11613.00 frames. ], tot_loss[loss=0.2092, simple_loss=0.2958, pruned_loss=0.06133, over 3082636.60 frames. ], batch size: 250, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:37:31,021 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 02:37:41,832 INFO [train.py:938] (4/8) Epoch 20, validation: loss=0.1516, simple_loss=0.2644, pruned_loss=0.01942, over 944034.00 frames. 2023-05-01 02:37:41,832 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 02:37:42,323 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=198852.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:37:46,499 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=198855.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:38:09,198 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9880, 5.2856, 5.0306, 5.0257, 4.7713, 4.7538, 4.6499, 5.3733], device='cuda:4'), covar=tensor([0.1228, 0.0804, 0.1029, 0.0920, 0.0802, 0.0918, 0.1247, 0.0811], device='cuda:4'), in_proj_covar=tensor([0.0642, 0.0783, 0.0646, 0.0589, 0.0494, 0.0505, 0.0658, 0.0607], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:38:17,161 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.795e+02 2.764e+02 3.257e+02 3.996e+02 6.001e+02, threshold=6.515e+02, percent-clipped=0.0 2023-05-01 02:38:28,071 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=198883.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:38:56,337 INFO [train.py:904] (4/8) Epoch 20, batch 6050, loss[loss=0.1926, simple_loss=0.2871, pruned_loss=0.04902, over 16279.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2943, pruned_loss=0.06046, over 3092122.72 frames. ], batch size: 165, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:39:14,616 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=198913.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:39:18,558 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9752, 3.9527, 4.6312, 2.2183, 4.8610, 4.8735, 3.6259, 3.5102], device='cuda:4'), covar=tensor([0.0745, 0.0222, 0.0127, 0.1132, 0.0047, 0.0098, 0.0291, 0.0430], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0108, 0.0097, 0.0139, 0.0080, 0.0124, 0.0128, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 02:39:32,914 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.66 vs. limit=5.0 2023-05-01 02:39:36,218 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=198927.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:40:05,155 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=198944.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:40:14,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1456, 1.6370, 1.9102, 2.0993, 2.2262, 2.3786, 1.7818, 2.2879], device='cuda:4'), covar=tensor([0.0220, 0.0470, 0.0258, 0.0350, 0.0294, 0.0197, 0.0474, 0.0138], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0191, 0.0176, 0.0180, 0.0192, 0.0149, 0.0192, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:40:15,346 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-01 02:40:17,596 INFO [train.py:904] (4/8) Epoch 20, batch 6100, loss[loss=0.2049, simple_loss=0.2941, pruned_loss=0.05779, over 15329.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2933, pruned_loss=0.05944, over 3103152.89 frames. ], batch size: 191, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:40:40,844 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.46 vs. limit=5.0 2023-05-01 02:40:46,238 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=198971.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:40:53,272 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.654e+02 2.636e+02 3.021e+02 3.731e+02 7.049e+02, threshold=6.042e+02, percent-clipped=1.0 2023-05-01 02:41:12,459 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=198988.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:41:32,342 INFO [train.py:904] (4/8) Epoch 20, batch 6150, loss[loss=0.1639, simple_loss=0.2507, pruned_loss=0.03852, over 17236.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2908, pruned_loss=0.0586, over 3113802.11 frames. ], batch size: 52, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:42:49,579 INFO [train.py:904] (4/8) Epoch 20, batch 6200, loss[loss=0.2282, simple_loss=0.2985, pruned_loss=0.07896, over 11741.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2879, pruned_loss=0.0577, over 3119461.39 frames. ], batch size: 248, lr: 3.37e-03, grad_scale: 4.0 2023-05-01 02:42:58,425 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-05-01 02:43:28,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.886e+02 2.934e+02 3.552e+02 4.446e+02 1.093e+03, threshold=7.104e+02, percent-clipped=2.0 2023-05-01 02:43:34,416 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6041, 2.7669, 2.2270, 2.5066, 3.1479, 2.8426, 3.2367, 3.3397], device='cuda:4'), covar=tensor([0.0114, 0.0373, 0.0519, 0.0449, 0.0244, 0.0337, 0.0237, 0.0242], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0229, 0.0222, 0.0222, 0.0232, 0.0229, 0.0230, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:43:42,077 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0683, 3.0583, 1.9112, 3.2675, 2.3881, 3.3209, 2.1666, 2.5654], device='cuda:4'), covar=tensor([0.0268, 0.0365, 0.1488, 0.0225, 0.0778, 0.0586, 0.1339, 0.0684], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0175, 0.0192, 0.0157, 0.0174, 0.0212, 0.0200, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 02:44:06,521 INFO [train.py:904] (4/8) Epoch 20, batch 6250, loss[loss=0.1756, simple_loss=0.2723, pruned_loss=0.03946, over 16317.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2874, pruned_loss=0.05725, over 3123664.60 frames. ], batch size: 35, lr: 3.37e-03, grad_scale: 4.0 2023-05-01 02:45:21,381 INFO [train.py:904] (4/8) Epoch 20, batch 6300, loss[loss=0.2553, simple_loss=0.3194, pruned_loss=0.09565, over 11801.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2879, pruned_loss=0.05712, over 3128051.60 frames. ], batch size: 247, lr: 3.37e-03, grad_scale: 4.0 2023-05-01 02:45:26,510 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=199155.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:45:37,247 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-05-01 02:45:59,350 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=199176.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:46:00,084 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.793e+02 2.632e+02 3.175e+02 3.943e+02 7.347e+02, threshold=6.350e+02, percent-clipped=2.0 2023-05-01 02:46:01,764 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3408, 2.1496, 1.7612, 1.9509, 2.4607, 2.1284, 2.1105, 2.5439], device='cuda:4'), covar=tensor([0.0192, 0.0385, 0.0550, 0.0449, 0.0239, 0.0361, 0.0212, 0.0270], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0229, 0.0222, 0.0222, 0.0232, 0.0230, 0.0231, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:46:38,965 INFO [train.py:904] (4/8) Epoch 20, batch 6350, loss[loss=0.2787, simple_loss=0.3333, pruned_loss=0.1121, over 11241.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2887, pruned_loss=0.05815, over 3128423.74 frames. ], batch size: 250, lr: 3.37e-03, grad_scale: 4.0 2023-05-01 02:46:40,592 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=199203.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:46:48,060 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=199208.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:47:30,826 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=199237.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 02:47:34,238 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=199239.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:47:38,969 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-01 02:47:52,850 INFO [train.py:904] (4/8) Epoch 20, batch 6400, loss[loss=0.2039, simple_loss=0.2867, pruned_loss=0.06056, over 16453.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2901, pruned_loss=0.06011, over 3105344.66 frames. ], batch size: 146, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:47:53,821 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.53 vs. limit=5.0 2023-05-01 02:47:55,145 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6433, 2.4658, 2.2307, 3.6638, 2.4012, 3.8567, 1.4927, 2.5798], device='cuda:4'), covar=tensor([0.1534, 0.0984, 0.1446, 0.0236, 0.0263, 0.0405, 0.1941, 0.1056], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0175, 0.0194, 0.0186, 0.0207, 0.0214, 0.0201, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 02:48:21,595 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=199271.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:48:23,502 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-01 02:48:25,690 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-05-01 02:48:29,164 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.234e+02 2.937e+02 3.403e+02 4.348e+02 9.192e+02, threshold=6.807e+02, percent-clipped=3.0 2023-05-01 02:48:39,447 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=199283.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:49:07,106 INFO [train.py:904] (4/8) Epoch 20, batch 6450, loss[loss=0.1988, simple_loss=0.2862, pruned_loss=0.05568, over 16827.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2898, pruned_loss=0.05948, over 3088452.75 frames. ], batch size: 96, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:49:17,256 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.88 vs. limit=5.0 2023-05-01 02:49:33,334 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=199319.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:49:46,549 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.86 vs. limit=5.0 2023-05-01 02:49:59,709 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8110, 1.8070, 2.3800, 2.6564, 2.6891, 3.0818, 1.9876, 3.0052], device='cuda:4'), covar=tensor([0.0203, 0.0538, 0.0294, 0.0332, 0.0276, 0.0184, 0.0517, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0192, 0.0177, 0.0182, 0.0193, 0.0150, 0.0193, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:50:24,668 INFO [train.py:904] (4/8) Epoch 20, batch 6500, loss[loss=0.2279, simple_loss=0.293, pruned_loss=0.08135, over 11450.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2878, pruned_loss=0.05906, over 3083795.17 frames. ], batch size: 247, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:50:43,891 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6088, 4.4600, 4.6785, 4.8249, 4.9784, 4.5025, 4.9945, 4.9815], device='cuda:4'), covar=tensor([0.1862, 0.1260, 0.1447, 0.0645, 0.0544, 0.0990, 0.0519, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0618, 0.0761, 0.0890, 0.0773, 0.0585, 0.0608, 0.0624, 0.0724], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 02:51:02,007 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.751e+02 2.867e+02 3.442e+02 4.204e+02 8.359e+02, threshold=6.884e+02, percent-clipped=2.0 2023-05-01 02:51:39,120 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=199400.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:51:41,756 INFO [train.py:904] (4/8) Epoch 20, batch 6550, loss[loss=0.2072, simple_loss=0.3037, pruned_loss=0.05537, over 16635.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2907, pruned_loss=0.06013, over 3084534.21 frames. ], batch size: 134, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:51:58,376 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.02 vs. limit=5.0 2023-05-01 02:52:36,185 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-01 02:52:56,160 INFO [train.py:904] (4/8) Epoch 20, batch 6600, loss[loss=0.1955, simple_loss=0.283, pruned_loss=0.05397, over 16711.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.2931, pruned_loss=0.0611, over 3043064.36 frames. ], batch size: 124, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:53:05,845 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=199458.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:53:10,044 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=199461.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 02:53:33,376 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.746e+02 2.722e+02 3.309e+02 3.917e+02 9.551e+02, threshold=6.618e+02, percent-clipped=2.0 2023-05-01 02:53:59,296 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=199494.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 02:54:11,565 INFO [train.py:904] (4/8) Epoch 20, batch 6650, loss[loss=0.2427, simple_loss=0.3097, pruned_loss=0.08783, over 11549.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2936, pruned_loss=0.06147, over 3062125.64 frames. ], batch size: 248, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:54:20,059 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=199508.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:54:37,227 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=199519.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 02:54:56,458 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=199532.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 02:55:00,543 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=199535.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 02:55:06,199 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=199539.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:55:25,177 INFO [train.py:904] (4/8) Epoch 20, batch 6700, loss[loss=0.1971, simple_loss=0.286, pruned_loss=0.05407, over 15165.00 frames. ], tot_loss[loss=0.2068, simple_loss=0.2919, pruned_loss=0.06084, over 3079551.51 frames. ], batch size: 190, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:55:30,736 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=199555.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 02:55:32,562 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=199556.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:56:02,759 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.649e+02 2.838e+02 3.650e+02 4.570e+02 8.139e+02, threshold=7.299e+02, percent-clipped=7.0 2023-05-01 02:56:13,430 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=199583.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:56:18,210 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=199587.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:56:29,414 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-05-01 02:56:31,511 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=199596.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 02:56:38,604 INFO [train.py:904] (4/8) Epoch 20, batch 6750, loss[loss=0.2546, simple_loss=0.3208, pruned_loss=0.09423, over 12008.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2911, pruned_loss=0.06074, over 3080879.87 frames. ], batch size: 248, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:56:57,557 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.84 vs. limit=5.0 2023-05-01 02:56:59,550 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5601, 3.6942, 2.5633, 2.1726, 2.5398, 2.2855, 3.9150, 3.3090], device='cuda:4'), covar=tensor([0.3179, 0.0749, 0.2287, 0.2761, 0.2762, 0.2234, 0.0611, 0.1286], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0268, 0.0302, 0.0307, 0.0295, 0.0253, 0.0292, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 02:57:20,690 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=199631.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 02:57:53,091 INFO [train.py:904] (4/8) Epoch 20, batch 6800, loss[loss=0.2159, simple_loss=0.3023, pruned_loss=0.06472, over 16805.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2912, pruned_loss=0.06069, over 3087433.83 frames. ], batch size: 124, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 02:58:29,264 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.005e+02 2.711e+02 3.238e+02 3.835e+02 6.988e+02, threshold=6.476e+02, percent-clipped=0.0 2023-05-01 02:59:05,436 INFO [train.py:904] (4/8) Epoch 20, batch 6850, loss[loss=0.1787, simple_loss=0.2934, pruned_loss=0.03195, over 16734.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2919, pruned_loss=0.06039, over 3099428.37 frames. ], batch size: 83, lr: 3.37e-03, grad_scale: 8.0 2023-05-01 03:00:20,624 INFO [train.py:904] (4/8) Epoch 20, batch 6900, loss[loss=0.1966, simple_loss=0.2957, pruned_loss=0.04873, over 16447.00 frames. ], tot_loss[loss=0.2068, simple_loss=0.2938, pruned_loss=0.05991, over 3110424.25 frames. ], batch size: 35, lr: 3.37e-03, grad_scale: 4.0 2023-05-01 03:00:26,874 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=199756.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 03:00:40,598 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5044, 4.1352, 4.0818, 2.7024, 3.6486, 4.0852, 3.6948, 2.3878], device='cuda:4'), covar=tensor([0.0497, 0.0041, 0.0047, 0.0373, 0.0098, 0.0108, 0.0086, 0.0420], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0081, 0.0080, 0.0133, 0.0095, 0.0107, 0.0092, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 03:00:59,795 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.774e+02 2.558e+02 3.121e+02 3.929e+02 6.424e+02, threshold=6.242e+02, percent-clipped=0.0 2023-05-01 03:01:36,153 INFO [train.py:904] (4/8) Epoch 20, batch 6950, loss[loss=0.2294, simple_loss=0.3197, pruned_loss=0.06957, over 16904.00 frames. ], tot_loss[loss=0.2091, simple_loss=0.2955, pruned_loss=0.06137, over 3107229.59 frames. ], batch size: 109, lr: 3.37e-03, grad_scale: 4.0 2023-05-01 03:01:37,907 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7541, 3.9467, 2.8910, 2.3128, 2.9240, 2.4774, 4.2752, 3.6789], device='cuda:4'), covar=tensor([0.2950, 0.0736, 0.1897, 0.2679, 0.2447, 0.2050, 0.0458, 0.1090], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0267, 0.0301, 0.0307, 0.0294, 0.0253, 0.0292, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 03:01:55,026 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=199814.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 03:02:22,392 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=199832.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:02:23,939 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.79 vs. limit=5.0 2023-05-01 03:02:47,650 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=199850.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 03:02:49,629 INFO [train.py:904] (4/8) Epoch 20, batch 7000, loss[loss=0.1916, simple_loss=0.2906, pruned_loss=0.04629, over 16723.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2958, pruned_loss=0.06089, over 3109165.39 frames. ], batch size: 57, lr: 3.37e-03, grad_scale: 4.0 2023-05-01 03:03:02,118 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-01 03:03:29,997 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.723e+02 2.829e+02 3.376e+02 4.341e+02 9.860e+02, threshold=6.751e+02, percent-clipped=7.0 2023-05-01 03:03:33,438 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=199880.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:03:49,864 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=199891.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 03:04:07,200 INFO [train.py:904] (4/8) Epoch 20, batch 7050, loss[loss=0.2598, simple_loss=0.3258, pruned_loss=0.09692, over 11524.00 frames. ], tot_loss[loss=0.2085, simple_loss=0.2966, pruned_loss=0.06022, over 3133638.76 frames. ], batch size: 249, lr: 3.37e-03, grad_scale: 4.0 2023-05-01 03:04:18,543 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=199909.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:04:18,794 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-01 03:05:25,771 INFO [train.py:904] (4/8) Epoch 20, batch 7100, loss[loss=0.2421, simple_loss=0.3023, pruned_loss=0.09093, over 11423.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2949, pruned_loss=0.05988, over 3120730.13 frames. ], batch size: 248, lr: 3.36e-03, grad_scale: 4.0 2023-05-01 03:05:54,850 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=199970.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:06:05,959 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.882e+02 2.782e+02 3.561e+02 4.163e+02 7.041e+02, threshold=7.121e+02, percent-clipped=1.0 2023-05-01 03:06:45,331 INFO [train.py:904] (4/8) Epoch 20, batch 7150, loss[loss=0.1841, simple_loss=0.2805, pruned_loss=0.04389, over 16888.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2929, pruned_loss=0.05993, over 3118866.81 frames. ], batch size: 96, lr: 3.36e-03, grad_scale: 4.0 2023-05-01 03:07:00,426 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6633, 1.7912, 1.6349, 1.4840, 1.8962, 1.5606, 1.5883, 1.8525], device='cuda:4'), covar=tensor([0.0176, 0.0256, 0.0367, 0.0307, 0.0185, 0.0235, 0.0154, 0.0199], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0229, 0.0221, 0.0222, 0.0231, 0.0229, 0.0228, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:07:59,653 INFO [train.py:904] (4/8) Epoch 20, batch 7200, loss[loss=0.2013, simple_loss=0.2868, pruned_loss=0.05791, over 17222.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2905, pruned_loss=0.05854, over 3107061.14 frames. ], batch size: 45, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:08:06,660 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=200056.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 03:08:26,175 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-05-01 03:08:37,891 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7663, 2.7356, 2.8542, 2.1783, 2.6943, 2.1048, 2.7557, 2.8887], device='cuda:4'), covar=tensor([0.0248, 0.0672, 0.0487, 0.1674, 0.0736, 0.0909, 0.0489, 0.0673], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0161, 0.0165, 0.0151, 0.0143, 0.0129, 0.0143, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 03:08:40,490 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.812e+02 2.598e+02 3.072e+02 3.513e+02 6.350e+02, threshold=6.144e+02, percent-clipped=0.0 2023-05-01 03:08:43,234 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-05-01 03:09:19,965 INFO [train.py:904] (4/8) Epoch 20, batch 7250, loss[loss=0.1842, simple_loss=0.2719, pruned_loss=0.04824, over 16913.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2883, pruned_loss=0.05738, over 3102157.68 frames. ], batch size: 109, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:09:23,329 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=200104.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:09:37,485 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=200114.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:10:31,955 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=200150.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 03:10:35,121 INFO [train.py:904] (4/8) Epoch 20, batch 7300, loss[loss=0.2051, simple_loss=0.2932, pruned_loss=0.05851, over 16671.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2883, pruned_loss=0.05743, over 3107989.00 frames. ], batch size: 62, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:10:50,981 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=200162.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:11:14,650 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=200177.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 03:11:15,883 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.776e+02 2.904e+02 3.413e+02 4.614e+02 7.836e+02, threshold=6.825e+02, percent-clipped=9.0 2023-05-01 03:11:36,960 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=200191.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 03:11:47,791 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=200198.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 03:11:52,585 INFO [train.py:904] (4/8) Epoch 20, batch 7350, loss[loss=0.2089, simple_loss=0.3005, pruned_loss=0.05865, over 16728.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2894, pruned_loss=0.05812, over 3096059.63 frames. ], batch size: 124, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:12:51,364 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=200238.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 03:12:52,206 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=200239.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 03:13:12,413 INFO [train.py:904] (4/8) Epoch 20, batch 7400, loss[loss=0.2229, simple_loss=0.3054, pruned_loss=0.07018, over 15155.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2907, pruned_loss=0.05897, over 3092053.54 frames. ], batch size: 190, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:13:33,057 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=200265.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:13:52,751 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.844e+02 2.893e+02 3.356e+02 4.091e+02 7.315e+02, threshold=6.713e+02, percent-clipped=2.0 2023-05-01 03:13:55,455 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=200279.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:14:32,167 INFO [train.py:904] (4/8) Epoch 20, batch 7450, loss[loss=0.1952, simple_loss=0.2871, pruned_loss=0.05163, over 16531.00 frames. ], tot_loss[loss=0.205, simple_loss=0.2912, pruned_loss=0.05938, over 3113789.20 frames. ], batch size: 68, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:15:36,235 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=200340.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 03:15:54,178 INFO [train.py:904] (4/8) Epoch 20, batch 7500, loss[loss=0.2029, simple_loss=0.2912, pruned_loss=0.05732, over 16312.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.291, pruned_loss=0.05826, over 3119803.13 frames. ], batch size: 146, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:16:34,099 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.040e+02 2.766e+02 3.444e+02 4.330e+02 6.961e+02, threshold=6.888e+02, percent-clipped=1.0 2023-05-01 03:17:11,460 INFO [train.py:904] (4/8) Epoch 20, batch 7550, loss[loss=0.1994, simple_loss=0.2833, pruned_loss=0.05771, over 16408.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2902, pruned_loss=0.0587, over 3099907.51 frames. ], batch size: 146, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:18:26,296 INFO [train.py:904] (4/8) Epoch 20, batch 7600, loss[loss=0.2154, simple_loss=0.2934, pruned_loss=0.06873, over 17027.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2892, pruned_loss=0.05876, over 3112668.12 frames. ], batch size: 41, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:19:06,493 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.884e+02 2.777e+02 3.348e+02 4.097e+02 9.491e+02, threshold=6.697e+02, percent-clipped=3.0 2023-05-01 03:19:40,332 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0822, 1.9728, 2.6073, 3.0249, 2.9063, 3.5148, 2.2098, 3.4658], device='cuda:4'), covar=tensor([0.0215, 0.0541, 0.0334, 0.0314, 0.0312, 0.0173, 0.0536, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0188, 0.0172, 0.0177, 0.0190, 0.0148, 0.0190, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:19:44,054 INFO [train.py:904] (4/8) Epoch 20, batch 7650, loss[loss=0.179, simple_loss=0.2651, pruned_loss=0.0465, over 17259.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2892, pruned_loss=0.05933, over 3105154.83 frames. ], batch size: 52, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:20:33,300 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=200533.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 03:20:53,465 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.8878, 6.1932, 5.8772, 6.0540, 5.5652, 5.4509, 5.6361, 6.3005], device='cuda:4'), covar=tensor([0.1065, 0.0783, 0.1035, 0.0750, 0.0795, 0.0655, 0.1079, 0.0799], device='cuda:4'), in_proj_covar=tensor([0.0645, 0.0792, 0.0653, 0.0595, 0.0497, 0.0512, 0.0661, 0.0611], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:21:02,177 INFO [train.py:904] (4/8) Epoch 20, batch 7700, loss[loss=0.2235, simple_loss=0.3009, pruned_loss=0.07307, over 16414.00 frames. ], tot_loss[loss=0.205, simple_loss=0.2897, pruned_loss=0.06014, over 3093276.09 frames. ], batch size: 35, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:21:08,115 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0003, 4.9582, 4.8056, 3.4169, 4.9060, 1.6969, 4.5029, 4.3098], device='cuda:4'), covar=tensor([0.0210, 0.0182, 0.0287, 0.0857, 0.0159, 0.3584, 0.0259, 0.0469], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0148, 0.0190, 0.0173, 0.0168, 0.0202, 0.0180, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:21:22,850 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=200565.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:21:25,701 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5363, 3.5892, 2.7755, 2.1974, 2.4560, 2.3256, 3.8446, 3.2544], device='cuda:4'), covar=tensor([0.3030, 0.0768, 0.1895, 0.2700, 0.2564, 0.2149, 0.0499, 0.1311], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0269, 0.0304, 0.0309, 0.0297, 0.0256, 0.0293, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 03:21:44,265 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.047e+02 2.918e+02 3.681e+02 4.507e+02 8.362e+02, threshold=7.363e+02, percent-clipped=4.0 2023-05-01 03:21:56,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2988, 3.1708, 3.6665, 1.8514, 3.7909, 3.8832, 2.8960, 2.7509], device='cuda:4'), covar=tensor([0.0850, 0.0300, 0.0207, 0.1214, 0.0092, 0.0170, 0.0414, 0.0503], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0097, 0.0139, 0.0079, 0.0123, 0.0127, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 03:22:04,981 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6331, 3.8325, 2.9206, 2.2912, 2.7936, 2.4766, 4.2534, 3.4548], device='cuda:4'), covar=tensor([0.3108, 0.0754, 0.1988, 0.2802, 0.2558, 0.2099, 0.0474, 0.1278], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0268, 0.0304, 0.0309, 0.0297, 0.0256, 0.0293, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 03:22:21,772 INFO [train.py:904] (4/8) Epoch 20, batch 7750, loss[loss=0.2185, simple_loss=0.3069, pruned_loss=0.06504, over 16991.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2898, pruned_loss=0.05986, over 3109144.02 frames. ], batch size: 109, lr: 3.36e-03, grad_scale: 4.0 2023-05-01 03:22:38,649 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=200613.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:23:11,917 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=200635.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 03:23:28,262 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=200645.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:23:36,962 INFO [train.py:904] (4/8) Epoch 20, batch 7800, loss[loss=0.1679, simple_loss=0.2584, pruned_loss=0.03873, over 16486.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2909, pruned_loss=0.06077, over 3089297.86 frames. ], batch size: 68, lr: 3.36e-03, grad_scale: 4.0 2023-05-01 03:24:18,680 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.931e+02 2.831e+02 3.422e+02 4.035e+02 8.624e+02, threshold=6.845e+02, percent-clipped=1.0 2023-05-01 03:24:27,883 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4675, 1.6211, 2.1293, 2.3758, 2.4075, 2.7987, 1.8415, 2.6897], device='cuda:4'), covar=tensor([0.0203, 0.0539, 0.0320, 0.0331, 0.0315, 0.0173, 0.0502, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0189, 0.0173, 0.0178, 0.0190, 0.0148, 0.0191, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:24:28,271 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-05-01 03:24:49,604 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 03:24:53,416 INFO [train.py:904] (4/8) Epoch 20, batch 7850, loss[loss=0.176, simple_loss=0.2692, pruned_loss=0.04136, over 16721.00 frames. ], tot_loss[loss=0.2067, simple_loss=0.2917, pruned_loss=0.06082, over 3070315.86 frames. ], batch size: 89, lr: 3.36e-03, grad_scale: 4.0 2023-05-01 03:25:00,297 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=200706.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:25:32,349 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8023, 4.6544, 4.9008, 5.0376, 5.2193, 4.7565, 5.2099, 5.2251], device='cuda:4'), covar=tensor([0.1941, 0.1359, 0.1545, 0.0728, 0.0569, 0.0951, 0.0603, 0.0674], device='cuda:4'), in_proj_covar=tensor([0.0619, 0.0761, 0.0888, 0.0776, 0.0588, 0.0608, 0.0628, 0.0726], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:25:32,516 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7373, 3.9653, 3.0041, 2.3683, 2.9172, 2.6126, 4.3917, 3.5699], device='cuda:4'), covar=tensor([0.2810, 0.0712, 0.1757, 0.2662, 0.2389, 0.1898, 0.0418, 0.1188], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0267, 0.0302, 0.0308, 0.0296, 0.0255, 0.0291, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 03:26:08,719 INFO [train.py:904] (4/8) Epoch 20, batch 7900, loss[loss=0.1981, simple_loss=0.2897, pruned_loss=0.05324, over 16514.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.291, pruned_loss=0.06058, over 3063470.82 frames. ], batch size: 75, lr: 3.36e-03, grad_scale: 4.0 2023-05-01 03:26:49,221 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.098e+02 2.835e+02 3.346e+02 4.054e+02 6.354e+02, threshold=6.693e+02, percent-clipped=0.0 2023-05-01 03:27:27,127 INFO [train.py:904] (4/8) Epoch 20, batch 7950, loss[loss=0.2487, simple_loss=0.3151, pruned_loss=0.09119, over 11620.00 frames. ], tot_loss[loss=0.2075, simple_loss=0.2918, pruned_loss=0.06165, over 3043066.19 frames. ], batch size: 248, lr: 3.36e-03, grad_scale: 4.0 2023-05-01 03:28:09,379 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-05-01 03:28:16,251 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=200833.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 03:28:44,612 INFO [train.py:904] (4/8) Epoch 20, batch 8000, loss[loss=0.2005, simple_loss=0.2847, pruned_loss=0.05813, over 16823.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2927, pruned_loss=0.06225, over 3037424.49 frames. ], batch size: 116, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:29:24,823 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.028e+02 3.046e+02 3.376e+02 4.028e+02 7.694e+02, threshold=6.753e+02, percent-clipped=2.0 2023-05-01 03:29:28,342 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=200881.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 03:29:59,852 INFO [train.py:904] (4/8) Epoch 20, batch 8050, loss[loss=0.202, simple_loss=0.2905, pruned_loss=0.05673, over 16464.00 frames. ], tot_loss[loss=0.2074, simple_loss=0.292, pruned_loss=0.06143, over 3057184.81 frames. ], batch size: 146, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:30:08,102 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5385, 3.6004, 3.3359, 3.0730, 3.1766, 3.5017, 3.3134, 3.3261], device='cuda:4'), covar=tensor([0.0634, 0.0718, 0.0290, 0.0275, 0.0517, 0.0495, 0.1369, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0408, 0.0329, 0.0324, 0.0338, 0.0376, 0.0227, 0.0394], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:30:12,822 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 03:30:50,056 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=200935.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:31:15,303 INFO [train.py:904] (4/8) Epoch 20, batch 8100, loss[loss=0.2098, simple_loss=0.2948, pruned_loss=0.06241, over 16669.00 frames. ], tot_loss[loss=0.2057, simple_loss=0.2912, pruned_loss=0.06009, over 3086431.88 frames. ], batch size: 57, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:31:57,146 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.671e+02 2.634e+02 3.129e+02 3.819e+02 6.896e+02, threshold=6.259e+02, percent-clipped=1.0 2023-05-01 03:32:04,948 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=200983.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:32:31,845 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=201001.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:32:33,316 INFO [train.py:904] (4/8) Epoch 20, batch 8150, loss[loss=0.2246, simple_loss=0.2994, pruned_loss=0.07496, over 11704.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2885, pruned_loss=0.05907, over 3090838.68 frames. ], batch size: 247, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:33:49,087 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3090, 2.5185, 2.0907, 2.1790, 2.9151, 2.5395, 2.9065, 3.0802], device='cuda:4'), covar=tensor([0.0151, 0.0406, 0.0567, 0.0502, 0.0250, 0.0394, 0.0256, 0.0272], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0227, 0.0219, 0.0220, 0.0228, 0.0228, 0.0227, 0.0224], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:33:52,111 INFO [train.py:904] (4/8) Epoch 20, batch 8200, loss[loss=0.1794, simple_loss=0.2756, pruned_loss=0.04158, over 16762.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2867, pruned_loss=0.05883, over 3094094.28 frames. ], batch size: 134, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:34:24,247 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=201071.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:34:36,688 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.058e+02 2.732e+02 3.390e+02 4.593e+02 1.143e+03, threshold=6.781e+02, percent-clipped=6.0 2023-05-01 03:35:14,751 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-05-01 03:35:15,217 INFO [train.py:904] (4/8) Epoch 20, batch 8250, loss[loss=0.176, simple_loss=0.279, pruned_loss=0.03646, over 16375.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.285, pruned_loss=0.05582, over 3068146.03 frames. ], batch size: 146, lr: 3.36e-03, grad_scale: 8.0 2023-05-01 03:36:06,026 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=201132.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:36:14,577 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8914, 4.8411, 4.6963, 4.0802, 4.7242, 1.7496, 4.4821, 4.5498], device='cuda:4'), covar=tensor([0.0102, 0.0102, 0.0190, 0.0423, 0.0118, 0.2745, 0.0150, 0.0211], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0147, 0.0189, 0.0172, 0.0166, 0.0200, 0.0179, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:36:38,968 INFO [train.py:904] (4/8) Epoch 20, batch 8300, loss[loss=0.1879, simple_loss=0.2831, pruned_loss=0.04635, over 16885.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.283, pruned_loss=0.0532, over 3080168.31 frames. ], batch size: 116, lr: 3.35e-03, grad_scale: 8.0 2023-05-01 03:37:06,048 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4087, 4.5555, 4.8644, 4.8208, 4.8480, 4.5972, 4.4035, 4.4106], device='cuda:4'), covar=tensor([0.0520, 0.0908, 0.0586, 0.0635, 0.0758, 0.0610, 0.1519, 0.0622], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0449, 0.0432, 0.0402, 0.0482, 0.0455, 0.0549, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 03:37:22,447 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.604e+02 2.376e+02 2.857e+02 3.605e+02 7.537e+02, threshold=5.714e+02, percent-clipped=2.0 2023-05-01 03:38:00,754 INFO [train.py:904] (4/8) Epoch 20, batch 8350, loss[loss=0.1908, simple_loss=0.2887, pruned_loss=0.04644, over 16327.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2825, pruned_loss=0.05172, over 3072454.39 frames. ], batch size: 146, lr: 3.35e-03, grad_scale: 8.0 2023-05-01 03:39:23,232 INFO [train.py:904] (4/8) Epoch 20, batch 8400, loss[loss=0.1724, simple_loss=0.2673, pruned_loss=0.03881, over 16853.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2797, pruned_loss=0.04973, over 3052441.64 frames. ], batch size: 42, lr: 3.35e-03, grad_scale: 8.0 2023-05-01 03:39:26,558 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=201253.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:39:32,707 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2249, 3.7373, 3.7243, 2.4870, 3.4108, 3.7091, 3.4405, 2.1550], device='cuda:4'), covar=tensor([0.0564, 0.0056, 0.0058, 0.0410, 0.0112, 0.0118, 0.0100, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0080, 0.0080, 0.0131, 0.0095, 0.0107, 0.0091, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 03:39:36,337 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-05-01 03:40:00,753 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=201274.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:40:08,501 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.430e+02 2.319e+02 2.651e+02 3.336e+02 5.393e+02, threshold=5.301e+02, percent-clipped=0.0 2023-05-01 03:40:44,263 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=201301.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:40:45,023 INFO [train.py:904] (4/8) Epoch 20, batch 8450, loss[loss=0.1799, simple_loss=0.2678, pruned_loss=0.04601, over 12360.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2779, pruned_loss=0.048, over 3052969.95 frames. ], batch size: 247, lr: 3.35e-03, grad_scale: 8.0 2023-05-01 03:40:45,607 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9688, 3.8284, 4.0449, 4.1310, 4.2668, 3.8498, 4.1933, 4.2813], device='cuda:4'), covar=tensor([0.1739, 0.1308, 0.1352, 0.0738, 0.0572, 0.1413, 0.0750, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0609, 0.0753, 0.0880, 0.0766, 0.0581, 0.0602, 0.0620, 0.0721], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:41:05,817 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=201314.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:41:39,383 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=201335.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:42:02,467 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=201349.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:42:07,348 INFO [train.py:904] (4/8) Epoch 20, batch 8500, loss[loss=0.156, simple_loss=0.2391, pruned_loss=0.03647, over 11839.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2743, pruned_loss=0.04589, over 3029449.69 frames. ], batch size: 246, lr: 3.35e-03, grad_scale: 8.0 2023-05-01 03:42:11,258 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3891, 4.6899, 4.5012, 4.5242, 4.2384, 4.1927, 4.2220, 4.7271], device='cuda:4'), covar=tensor([0.1155, 0.0874, 0.0974, 0.0800, 0.0786, 0.1438, 0.1201, 0.0832], device='cuda:4'), in_proj_covar=tensor([0.0633, 0.0773, 0.0639, 0.0583, 0.0488, 0.0502, 0.0647, 0.0598], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:42:17,338 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8118, 3.9037, 3.9693, 3.7565, 3.8457, 4.2839, 3.9266, 3.5877], device='cuda:4'), covar=tensor([0.1868, 0.2143, 0.2270, 0.2296, 0.2855, 0.1604, 0.1537, 0.2629], device='cuda:4'), in_proj_covar=tensor([0.0389, 0.0566, 0.0625, 0.0469, 0.0623, 0.0656, 0.0490, 0.0635], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 03:42:26,005 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-05-01 03:42:50,125 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.494e+02 2.179e+02 2.765e+02 3.393e+02 6.239e+02, threshold=5.531e+02, percent-clipped=4.0 2023-05-01 03:43:31,052 INFO [train.py:904] (4/8) Epoch 20, batch 8550, loss[loss=0.168, simple_loss=0.2554, pruned_loss=0.04027, over 11666.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2716, pruned_loss=0.04444, over 3012809.67 frames. ], batch size: 247, lr: 3.35e-03, grad_scale: 8.0 2023-05-01 03:44:19,133 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=201427.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:44:29,358 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4277, 3.3572, 3.4819, 3.5299, 3.5788, 3.2941, 3.5263, 3.6177], device='cuda:4'), covar=tensor([0.1265, 0.1069, 0.1145, 0.0685, 0.0685, 0.2852, 0.1204, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0608, 0.0753, 0.0881, 0.0765, 0.0581, 0.0602, 0.0620, 0.0716], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:44:37,608 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0515, 2.0632, 2.0929, 3.6039, 2.0579, 2.4149, 2.1854, 2.2024], device='cuda:4'), covar=tensor([0.1275, 0.3778, 0.3126, 0.0560, 0.4312, 0.2571, 0.3685, 0.3728], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0432, 0.0356, 0.0315, 0.0425, 0.0496, 0.0402, 0.0505], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:45:08,357 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4197, 2.0237, 1.8001, 1.7405, 2.2779, 1.9304, 1.9186, 2.3676], device='cuda:4'), covar=tensor([0.0177, 0.0352, 0.0448, 0.0424, 0.0227, 0.0348, 0.0174, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0222, 0.0215, 0.0215, 0.0223, 0.0223, 0.0221, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:45:09,639 INFO [train.py:904] (4/8) Epoch 20, batch 8600, loss[loss=0.1707, simple_loss=0.2695, pruned_loss=0.03601, over 15301.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2724, pruned_loss=0.0435, over 3016705.76 frames. ], batch size: 191, lr: 3.35e-03, grad_scale: 8.0 2023-05-01 03:45:44,492 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7691, 2.5605, 2.4205, 3.6816, 2.1906, 3.7555, 1.4660, 2.9255], device='cuda:4'), covar=tensor([0.1315, 0.0784, 0.1134, 0.0149, 0.0101, 0.0374, 0.1738, 0.0696], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0171, 0.0192, 0.0183, 0.0205, 0.0211, 0.0199, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 03:46:02,940 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.295e+02 2.180e+02 2.549e+02 3.038e+02 6.489e+02, threshold=5.098e+02, percent-clipped=1.0 2023-05-01 03:46:48,552 INFO [train.py:904] (4/8) Epoch 20, batch 8650, loss[loss=0.1717, simple_loss=0.2594, pruned_loss=0.04205, over 12282.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2698, pruned_loss=0.04198, over 3013034.72 frames. ], batch size: 246, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 03:46:52,602 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-01 03:48:36,196 INFO [train.py:904] (4/8) Epoch 20, batch 8700, loss[loss=0.1685, simple_loss=0.2695, pruned_loss=0.03373, over 15348.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2668, pruned_loss=0.04074, over 2989517.86 frames. ], batch size: 190, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 03:49:28,968 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.047e+02 2.487e+02 3.171e+02 4.995e+02, threshold=4.975e+02, percent-clipped=0.0 2023-05-01 03:50:13,607 INFO [train.py:904] (4/8) Epoch 20, batch 8750, loss[loss=0.1821, simple_loss=0.2795, pruned_loss=0.0423, over 16520.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2664, pruned_loss=0.04008, over 3016640.31 frames. ], batch size: 68, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 03:50:31,548 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=201609.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:51:22,015 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=201630.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:51:51,022 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=201644.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:52:07,671 INFO [train.py:904] (4/8) Epoch 20, batch 8800, loss[loss=0.1684, simple_loss=0.268, pruned_loss=0.03439, over 16190.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2651, pruned_loss=0.0388, over 3038512.46 frames. ], batch size: 165, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 03:52:09,037 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-01 03:52:24,864 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=201661.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:53:06,514 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.591e+02 2.243e+02 2.623e+02 3.212e+02 5.763e+02, threshold=5.246e+02, percent-clipped=4.0 2023-05-01 03:53:51,630 INFO [train.py:904] (4/8) Epoch 20, batch 8850, loss[loss=0.1566, simple_loss=0.2473, pruned_loss=0.03295, over 12228.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2672, pruned_loss=0.03865, over 3009758.77 frames. ], batch size: 248, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 03:53:57,890 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=201705.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:53:59,771 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=201706.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:54:10,702 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0012, 1.8751, 1.9890, 3.6164, 1.8234, 2.0958, 2.0265, 2.0110], device='cuda:4'), covar=tensor([0.1585, 0.4938, 0.3555, 0.0675, 0.5824, 0.3434, 0.4419, 0.4575], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0434, 0.0358, 0.0316, 0.0428, 0.0498, 0.0404, 0.0507], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:54:21,298 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9753, 1.9284, 2.4383, 2.8835, 2.6704, 3.3020, 2.2181, 3.3604], device='cuda:4'), covar=tensor([0.0208, 0.0528, 0.0367, 0.0273, 0.0331, 0.0166, 0.0489, 0.0130], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0187, 0.0172, 0.0176, 0.0189, 0.0146, 0.0190, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:54:34,773 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=201722.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:54:36,626 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8689, 2.1449, 2.1840, 2.9220, 1.5720, 3.2085, 1.6856, 2.6885], device='cuda:4'), covar=tensor([0.1325, 0.0777, 0.1217, 0.0161, 0.0079, 0.0380, 0.1695, 0.0767], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0168, 0.0189, 0.0180, 0.0201, 0.0207, 0.0196, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 03:54:45,904 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=201727.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:55:01,983 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4042, 4.3502, 4.1287, 3.3195, 4.2767, 1.5550, 3.9914, 3.8109], device='cuda:4'), covar=tensor([0.0104, 0.0091, 0.0256, 0.0444, 0.0119, 0.3178, 0.0172, 0.0376], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0146, 0.0187, 0.0169, 0.0165, 0.0201, 0.0178, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 03:55:38,107 INFO [train.py:904] (4/8) Epoch 20, batch 8900, loss[loss=0.1678, simple_loss=0.2555, pruned_loss=0.04003, over 12617.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2681, pruned_loss=0.03791, over 3033709.73 frames. ], batch size: 249, lr: 3.35e-03, grad_scale: 2.0 2023-05-01 03:56:08,863 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=201767.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:56:27,850 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=201775.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 03:56:43,167 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-05-01 03:56:46,606 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.534e+02 2.216e+02 2.705e+02 3.264e+02 5.822e+02, threshold=5.410e+02, percent-clipped=1.0 2023-05-01 03:57:13,690 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.4985, 2.4887, 2.2909, 3.9885, 2.2672, 3.9060, 1.4083, 2.6869], device='cuda:4'), covar=tensor([0.1629, 0.0845, 0.1333, 0.0167, 0.0150, 0.0338, 0.1989, 0.0881], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0169, 0.0190, 0.0180, 0.0201, 0.0207, 0.0197, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 03:57:41,706 INFO [train.py:904] (4/8) Epoch 20, batch 8950, loss[loss=0.1516, simple_loss=0.2502, pruned_loss=0.02653, over 15476.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2672, pruned_loss=0.03785, over 3052500.10 frames. ], batch size: 192, lr: 3.35e-03, grad_scale: 2.0 2023-05-01 03:57:43,296 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.04 vs. limit=5.0 2023-05-01 03:58:51,568 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 03:59:30,704 INFO [train.py:904] (4/8) Epoch 20, batch 9000, loss[loss=0.1517, simple_loss=0.2454, pruned_loss=0.02899, over 16620.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2643, pruned_loss=0.03683, over 3044019.97 frames. ], batch size: 89, lr: 3.35e-03, grad_scale: 2.0 2023-05-01 03:59:30,705 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 03:59:41,115 INFO [train.py:938] (4/8) Epoch 20, validation: loss=0.1464, simple_loss=0.2502, pruned_loss=0.02125, over 944034.00 frames. 2023-05-01 03:59:41,116 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 04:00:35,864 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6588, 2.4604, 2.2671, 3.5701, 1.9882, 3.6367, 1.3569, 2.7646], device='cuda:4'), covar=tensor([0.1482, 0.0763, 0.1289, 0.0150, 0.0106, 0.0355, 0.1909, 0.0779], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0169, 0.0190, 0.0180, 0.0201, 0.0208, 0.0197, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:00:41,793 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.381e+02 2.083e+02 2.552e+02 3.280e+02 1.556e+03, threshold=5.104e+02, percent-clipped=3.0 2023-05-01 04:00:44,210 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8337, 3.9063, 4.0474, 3.8129, 3.9770, 4.3752, 4.0298, 3.6892], device='cuda:4'), covar=tensor([0.2069, 0.2156, 0.1842, 0.2511, 0.2392, 0.1390, 0.1473, 0.2570], device='cuda:4'), in_proj_covar=tensor([0.0379, 0.0552, 0.0607, 0.0458, 0.0605, 0.0636, 0.0476, 0.0614], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 04:01:24,589 INFO [train.py:904] (4/8) Epoch 20, batch 9050, loss[loss=0.1778, simple_loss=0.2668, pruned_loss=0.0444, over 12733.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2647, pruned_loss=0.03708, over 3054243.78 frames. ], batch size: 246, lr: 3.35e-03, grad_scale: 2.0 2023-05-01 04:01:40,868 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=201909.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:02:03,216 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-01 04:02:09,361 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9313, 5.2533, 5.0537, 5.0708, 4.7981, 4.8032, 4.6241, 5.3349], device='cuda:4'), covar=tensor([0.1213, 0.0931, 0.0965, 0.0791, 0.0742, 0.0929, 0.1183, 0.0874], device='cuda:4'), in_proj_covar=tensor([0.0631, 0.0771, 0.0635, 0.0580, 0.0491, 0.0501, 0.0645, 0.0599], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:02:21,211 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=201930.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:02:58,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6783, 2.6077, 1.8371, 2.8088, 2.1241, 2.8094, 2.1228, 2.4126], device='cuda:4'), covar=tensor([0.0267, 0.0331, 0.1225, 0.0254, 0.0650, 0.0448, 0.1212, 0.0591], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0168, 0.0188, 0.0150, 0.0170, 0.0205, 0.0196, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 04:03:09,163 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3100, 3.1451, 3.4622, 1.7357, 3.6342, 3.6642, 2.9039, 2.8195], device='cuda:4'), covar=tensor([0.0769, 0.0282, 0.0214, 0.1224, 0.0081, 0.0163, 0.0423, 0.0461], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0104, 0.0092, 0.0134, 0.0075, 0.0117, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 04:03:12,544 INFO [train.py:904] (4/8) Epoch 20, batch 9100, loss[loss=0.1819, simple_loss=0.2855, pruned_loss=0.03914, over 16393.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2645, pruned_loss=0.03772, over 3042538.19 frames. ], batch size: 146, lr: 3.35e-03, grad_scale: 2.0 2023-05-01 04:03:22,656 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=201957.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:03:30,437 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6819, 2.6354, 1.8124, 2.8302, 2.1469, 2.8219, 2.0910, 2.4005], device='cuda:4'), covar=tensor([0.0286, 0.0346, 0.1276, 0.0245, 0.0662, 0.0432, 0.1194, 0.0609], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0168, 0.0187, 0.0150, 0.0169, 0.0204, 0.0196, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 04:04:14,332 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=201978.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:04:23,055 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.517e+02 2.025e+02 2.515e+02 2.947e+02 5.747e+02, threshold=5.029e+02, percent-clipped=1.0 2023-05-01 04:04:42,952 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=201990.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:04:49,166 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4284, 3.0158, 2.8534, 1.8159, 2.5695, 1.9788, 2.9951, 3.2536], device='cuda:4'), covar=tensor([0.0320, 0.0847, 0.0848, 0.2654, 0.1213, 0.1395, 0.0849, 0.0827], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0154, 0.0159, 0.0147, 0.0139, 0.0125, 0.0139, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:05:11,349 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=202000.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:05:14,319 INFO [train.py:904] (4/8) Epoch 20, batch 9150, loss[loss=0.1679, simple_loss=0.2533, pruned_loss=0.0413, over 12096.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2651, pruned_loss=0.03764, over 3032327.91 frames. ], batch size: 250, lr: 3.35e-03, grad_scale: 2.0 2023-05-01 04:05:17,390 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3896, 5.7117, 5.4821, 5.5283, 5.2224, 5.2062, 5.0682, 5.7780], device='cuda:4'), covar=tensor([0.1132, 0.0831, 0.0849, 0.0685, 0.0686, 0.0667, 0.1199, 0.0796], device='cuda:4'), in_proj_covar=tensor([0.0627, 0.0766, 0.0630, 0.0577, 0.0488, 0.0498, 0.0642, 0.0594], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:05:45,938 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-05-01 04:05:48,713 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=202017.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:06:03,278 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4711, 3.4025, 3.5120, 3.5638, 3.5990, 3.3002, 3.5898, 3.6464], device='cuda:4'), covar=tensor([0.1147, 0.0935, 0.1026, 0.0627, 0.0666, 0.2201, 0.0818, 0.0794], device='cuda:4'), in_proj_covar=tensor([0.0594, 0.0731, 0.0854, 0.0751, 0.0566, 0.0587, 0.0604, 0.0698], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:06:59,060 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=202051.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:06:59,884 INFO [train.py:904] (4/8) Epoch 20, batch 9200, loss[loss=0.1583, simple_loss=0.2554, pruned_loss=0.03061, over 16230.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2612, pruned_loss=0.03689, over 3041065.55 frames. ], batch size: 165, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 04:07:20,314 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=202062.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:07:55,479 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.735e+02 2.302e+02 2.670e+02 3.574e+02 1.158e+03, threshold=5.341e+02, percent-clipped=6.0 2023-05-01 04:08:35,590 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6666, 2.9661, 3.3759, 2.0223, 2.9295, 2.1791, 3.2730, 3.1761], device='cuda:4'), covar=tensor([0.0275, 0.0958, 0.0456, 0.1984, 0.0762, 0.0985, 0.0614, 0.0965], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0153, 0.0158, 0.0146, 0.0138, 0.0125, 0.0138, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:08:37,604 INFO [train.py:904] (4/8) Epoch 20, batch 9250, loss[loss=0.1399, simple_loss=0.2271, pruned_loss=0.02636, over 12224.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.261, pruned_loss=0.03684, over 3034764.08 frames. ], batch size: 248, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 04:09:06,203 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-05-01 04:10:12,029 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4743, 3.3780, 2.7067, 2.1653, 2.1338, 2.2678, 3.4601, 3.0085], device='cuda:4'), covar=tensor([0.2935, 0.0702, 0.1707, 0.2964, 0.3023, 0.2273, 0.0425, 0.1340], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0260, 0.0295, 0.0299, 0.0282, 0.0248, 0.0282, 0.0321], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 04:10:29,963 INFO [train.py:904] (4/8) Epoch 20, batch 9300, loss[loss=0.1491, simple_loss=0.2395, pruned_loss=0.02938, over 17070.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2593, pruned_loss=0.03666, over 3021266.51 frames. ], batch size: 53, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 04:10:30,816 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1082, 2.4890, 2.6452, 1.8710, 2.8393, 2.8680, 2.4914, 2.4765], device='cuda:4'), covar=tensor([0.0671, 0.0283, 0.0232, 0.1014, 0.0108, 0.0233, 0.0461, 0.0433], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0104, 0.0091, 0.0134, 0.0075, 0.0116, 0.0122, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 04:11:37,344 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.443e+02 1.963e+02 2.278e+02 2.874e+02 5.860e+02, threshold=4.556e+02, percent-clipped=2.0 2023-05-01 04:12:16,144 INFO [train.py:904] (4/8) Epoch 20, batch 9350, loss[loss=0.1768, simple_loss=0.2718, pruned_loss=0.04095, over 16576.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2594, pruned_loss=0.03674, over 3030303.40 frames. ], batch size: 57, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 04:12:17,212 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5543, 3.1897, 3.1872, 1.9786, 2.7455, 2.2211, 3.0938, 3.3617], device='cuda:4'), covar=tensor([0.0377, 0.0714, 0.0587, 0.2072, 0.0908, 0.1022, 0.0861, 0.0861], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0154, 0.0159, 0.0147, 0.0139, 0.0125, 0.0139, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:12:33,384 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-05-01 04:12:53,930 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6438, 4.0566, 3.6337, 3.9565, 3.6221, 3.6290, 3.6006, 4.0779], device='cuda:4'), covar=tensor([0.3098, 0.1976, 0.2789, 0.1650, 0.1886, 0.3434, 0.2737, 0.2026], device='cuda:4'), in_proj_covar=tensor([0.0630, 0.0766, 0.0629, 0.0577, 0.0489, 0.0498, 0.0641, 0.0595], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:13:59,184 INFO [train.py:904] (4/8) Epoch 20, batch 9400, loss[loss=0.1575, simple_loss=0.2593, pruned_loss=0.02782, over 15202.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2597, pruned_loss=0.03666, over 3037807.62 frames. ], batch size: 190, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 04:15:00,972 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 2.064e+02 2.476e+02 2.960e+02 4.580e+02, threshold=4.951e+02, percent-clipped=1.0 2023-05-01 04:15:39,445 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=202300.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:15:41,541 INFO [train.py:904] (4/8) Epoch 20, batch 9450, loss[loss=0.1711, simple_loss=0.2643, pruned_loss=0.03897, over 12749.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2623, pruned_loss=0.03673, over 3050624.95 frames. ], batch size: 248, lr: 3.35e-03, grad_scale: 4.0 2023-05-01 04:16:13,606 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=202317.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:17:14,532 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=202346.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:17:17,848 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=202348.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:17:19,795 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9076, 4.0147, 2.4357, 4.5813, 3.0006, 4.4426, 2.5341, 3.1113], device='cuda:4'), covar=tensor([0.0240, 0.0294, 0.1535, 0.0188, 0.0788, 0.0450, 0.1570, 0.0738], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0167, 0.0187, 0.0150, 0.0169, 0.0203, 0.0195, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 04:17:25,743 INFO [train.py:904] (4/8) Epoch 20, batch 9500, loss[loss=0.1749, simple_loss=0.265, pruned_loss=0.04236, over 12784.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2619, pruned_loss=0.03675, over 3064100.05 frames. ], batch size: 246, lr: 3.34e-03, grad_scale: 4.0 2023-05-01 04:17:49,626 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=202362.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:17:55,126 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=202365.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:18:26,630 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.335e+02 2.130e+02 2.550e+02 3.364e+02 1.443e+03, threshold=5.099e+02, percent-clipped=7.0 2023-05-01 04:19:08,447 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9774, 3.6385, 4.0365, 2.1173, 4.2121, 4.2603, 3.1978, 3.1895], device='cuda:4'), covar=tensor([0.0582, 0.0240, 0.0163, 0.1056, 0.0060, 0.0106, 0.0350, 0.0398], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0104, 0.0091, 0.0135, 0.0075, 0.0116, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 04:19:13,220 INFO [train.py:904] (4/8) Epoch 20, batch 9550, loss[loss=0.1662, simple_loss=0.2567, pruned_loss=0.03782, over 12048.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2616, pruned_loss=0.03676, over 3071118.41 frames. ], batch size: 247, lr: 3.34e-03, grad_scale: 4.0 2023-05-01 04:19:18,124 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6411, 4.6277, 4.4821, 3.9736, 4.5366, 1.7137, 4.3262, 4.2206], device='cuda:4'), covar=tensor([0.0079, 0.0079, 0.0165, 0.0262, 0.0088, 0.2570, 0.0121, 0.0214], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0143, 0.0181, 0.0164, 0.0162, 0.0196, 0.0173, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:19:32,642 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=202410.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:20:55,235 INFO [train.py:904] (4/8) Epoch 20, batch 9600, loss[loss=0.1719, simple_loss=0.2679, pruned_loss=0.03793, over 12173.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2637, pruned_loss=0.0376, over 3074749.67 frames. ], batch size: 247, lr: 3.34e-03, grad_scale: 8.0 2023-05-01 04:21:19,888 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-05-01 04:21:53,894 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.693e+02 2.243e+02 2.599e+02 3.368e+02 6.147e+02, threshold=5.198e+02, percent-clipped=3.0 2023-05-01 04:22:41,347 INFO [train.py:904] (4/8) Epoch 20, batch 9650, loss[loss=0.1605, simple_loss=0.2612, pruned_loss=0.02995, over 16448.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2652, pruned_loss=0.038, over 3057731.55 frames. ], batch size: 147, lr: 3.34e-03, grad_scale: 8.0 2023-05-01 04:23:42,260 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7351, 2.4908, 2.3328, 3.5465, 1.9725, 3.6830, 1.4907, 2.7615], device='cuda:4'), covar=tensor([0.1467, 0.0806, 0.1270, 0.0181, 0.0109, 0.0376, 0.1905, 0.0819], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0169, 0.0190, 0.0180, 0.0197, 0.0207, 0.0197, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:24:24,968 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-01 04:24:31,593 INFO [train.py:904] (4/8) Epoch 20, batch 9700, loss[loss=0.1791, simple_loss=0.2683, pruned_loss=0.04495, over 12480.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2639, pruned_loss=0.03763, over 3077880.36 frames. ], batch size: 248, lr: 3.34e-03, grad_scale: 4.0 2023-05-01 04:25:37,512 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 2.240e+02 2.594e+02 3.217e+02 6.091e+02, threshold=5.188e+02, percent-clipped=1.0 2023-05-01 04:25:44,950 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7263, 2.1943, 1.9277, 1.9215, 2.5410, 2.1571, 2.2058, 2.6269], device='cuda:4'), covar=tensor([0.0152, 0.0416, 0.0533, 0.0509, 0.0259, 0.0427, 0.0198, 0.0256], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0224, 0.0215, 0.0215, 0.0223, 0.0223, 0.0218, 0.0216], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:26:10,332 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-05-01 04:26:15,968 INFO [train.py:904] (4/8) Epoch 20, batch 9750, loss[loss=0.1544, simple_loss=0.2427, pruned_loss=0.03304, over 12649.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2626, pruned_loss=0.03779, over 3071253.47 frames. ], batch size: 248, lr: 3.34e-03, grad_scale: 4.0 2023-05-01 04:27:18,551 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-05-01 04:27:25,623 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3832, 4.3762, 4.1885, 3.6605, 4.2855, 1.7430, 4.0853, 3.9081], device='cuda:4'), covar=tensor([0.0068, 0.0069, 0.0157, 0.0202, 0.0079, 0.2521, 0.0099, 0.0219], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0144, 0.0182, 0.0164, 0.0163, 0.0197, 0.0174, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:27:47,501 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=202646.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:27:57,087 INFO [train.py:904] (4/8) Epoch 20, batch 9800, loss[loss=0.1687, simple_loss=0.2704, pruned_loss=0.0335, over 16903.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2621, pruned_loss=0.03681, over 3060826.20 frames. ], batch size: 116, lr: 3.34e-03, grad_scale: 2.0 2023-05-01 04:28:57,878 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.410e+02 2.120e+02 2.499e+02 2.833e+02 7.051e+02, threshold=4.998e+02, percent-clipped=2.0 2023-05-01 04:29:26,526 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=202694.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:29:42,372 INFO [train.py:904] (4/8) Epoch 20, batch 9850, loss[loss=0.1539, simple_loss=0.2457, pruned_loss=0.03103, over 12516.00 frames. ], tot_loss[loss=0.168, simple_loss=0.263, pruned_loss=0.03656, over 3051237.40 frames. ], batch size: 246, lr: 3.34e-03, grad_scale: 2.0 2023-05-01 04:29:49,782 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-05-01 04:30:11,211 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5867, 2.9084, 3.2355, 1.8420, 2.7677, 2.1423, 3.1771, 3.1403], device='cuda:4'), covar=tensor([0.0252, 0.0861, 0.0508, 0.2090, 0.0836, 0.0958, 0.0692, 0.0847], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0153, 0.0159, 0.0147, 0.0138, 0.0124, 0.0139, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:31:34,886 INFO [train.py:904] (4/8) Epoch 20, batch 9900, loss[loss=0.171, simple_loss=0.2573, pruned_loss=0.04233, over 12773.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2635, pruned_loss=0.03659, over 3036624.69 frames. ], batch size: 247, lr: 3.34e-03, grad_scale: 2.0 2023-05-01 04:32:49,188 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.306e+02 2.190e+02 2.483e+02 3.109e+02 6.841e+02, threshold=4.966e+02, percent-clipped=3.0 2023-05-01 04:33:05,856 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-05-01 04:33:30,682 INFO [train.py:904] (4/8) Epoch 20, batch 9950, loss[loss=0.1613, simple_loss=0.2612, pruned_loss=0.03066, over 16187.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2659, pruned_loss=0.0372, over 3038741.75 frames. ], batch size: 165, lr: 3.34e-03, grad_scale: 2.0 2023-05-01 04:34:12,618 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3878, 3.2761, 3.3074, 3.5157, 3.5561, 3.2389, 3.5361, 3.5807], device='cuda:4'), covar=tensor([0.1560, 0.1244, 0.1635, 0.0920, 0.0890, 0.3340, 0.1106, 0.0968], device='cuda:4'), in_proj_covar=tensor([0.0594, 0.0727, 0.0848, 0.0749, 0.0565, 0.0584, 0.0603, 0.0698], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:34:41,291 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=202830.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 04:34:56,042 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-05-01 04:35:16,592 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3207, 2.3063, 2.3621, 4.1920, 2.1595, 2.5783, 2.3516, 2.4421], device='cuda:4'), covar=tensor([0.1241, 0.3692, 0.2993, 0.0406, 0.4104, 0.2734, 0.3873, 0.3096], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0431, 0.0357, 0.0315, 0.0425, 0.0492, 0.0402, 0.0502], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:35:32,300 INFO [train.py:904] (4/8) Epoch 20, batch 10000, loss[loss=0.1942, simple_loss=0.3002, pruned_loss=0.04415, over 15501.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2646, pruned_loss=0.03667, over 3042422.00 frames. ], batch size: 192, lr: 3.34e-03, grad_scale: 4.0 2023-05-01 04:36:36,180 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.432e+02 2.233e+02 2.660e+02 3.205e+02 6.054e+02, threshold=5.319e+02, percent-clipped=4.0 2023-05-01 04:36:52,348 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=202891.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 04:37:11,276 INFO [train.py:904] (4/8) Epoch 20, batch 10050, loss[loss=0.1904, simple_loss=0.2892, pruned_loss=0.04584, over 16175.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2646, pruned_loss=0.03674, over 3042690.22 frames. ], batch size: 165, lr: 3.34e-03, grad_scale: 4.0 2023-05-01 04:38:05,392 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6401, 4.9549, 4.5405, 4.8675, 4.5488, 4.4438, 4.4052, 4.9897], device='cuda:4'), covar=tensor([0.2230, 0.1640, 0.2494, 0.1283, 0.1553, 0.1922, 0.2515, 0.1920], device='cuda:4'), in_proj_covar=tensor([0.0622, 0.0762, 0.0621, 0.0571, 0.0485, 0.0495, 0.0640, 0.0592], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:38:41,705 INFO [train.py:904] (4/8) Epoch 20, batch 10100, loss[loss=0.1601, simple_loss=0.2521, pruned_loss=0.03412, over 16114.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2651, pruned_loss=0.03698, over 3061220.78 frames. ], batch size: 165, lr: 3.34e-03, grad_scale: 4.0 2023-05-01 04:38:45,098 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2720, 1.5929, 1.8976, 2.0826, 2.2248, 2.3052, 1.7417, 2.2234], device='cuda:4'), covar=tensor([0.0231, 0.0493, 0.0289, 0.0347, 0.0315, 0.0231, 0.0504, 0.0145], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0183, 0.0170, 0.0173, 0.0186, 0.0142, 0.0187, 0.0138], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:38:48,267 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9715, 3.1584, 3.1783, 2.1541, 2.9377, 3.2550, 3.0763, 1.8306], device='cuda:4'), covar=tensor([0.0578, 0.0071, 0.0065, 0.0426, 0.0136, 0.0106, 0.0122, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0077, 0.0077, 0.0128, 0.0094, 0.0103, 0.0088, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 04:39:40,384 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.271e+02 2.682e+02 3.301e+02 6.522e+02, threshold=5.364e+02, percent-clipped=2.0 2023-05-01 04:40:23,119 INFO [train.py:904] (4/8) Epoch 21, batch 0, loss[loss=0.1703, simple_loss=0.2603, pruned_loss=0.04015, over 17237.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2603, pruned_loss=0.04015, over 17237.00 frames. ], batch size: 45, lr: 3.26e-03, grad_scale: 8.0 2023-05-01 04:40:23,119 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 04:40:30,890 INFO [train.py:938] (4/8) Epoch 21, validation: loss=0.1451, simple_loss=0.2491, pruned_loss=0.02058, over 944034.00 frames. 2023-05-01 04:40:30,890 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 04:41:07,246 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2883, 5.0638, 5.3200, 5.4673, 5.6152, 4.8841, 5.5789, 5.5824], device='cuda:4'), covar=tensor([0.1799, 0.1157, 0.1562, 0.0727, 0.0509, 0.0799, 0.0505, 0.0633], device='cuda:4'), in_proj_covar=tensor([0.0594, 0.0726, 0.0847, 0.0750, 0.0565, 0.0582, 0.0603, 0.0699], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:41:19,654 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=203038.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:41:36,911 INFO [train.py:904] (4/8) Epoch 21, batch 50, loss[loss=0.2005, simple_loss=0.2732, pruned_loss=0.0639, over 16765.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.2708, pruned_loss=0.04896, over 749911.70 frames. ], batch size: 124, lr: 3.26e-03, grad_scale: 2.0 2023-05-01 04:42:25,305 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=203085.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:42:26,043 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.693e+02 2.424e+02 2.967e+02 3.842e+02 7.470e+02, threshold=5.934e+02, percent-clipped=2.0 2023-05-01 04:42:28,849 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8946, 4.9017, 4.7928, 4.4002, 4.8249, 2.0812, 4.6125, 4.6079], device='cuda:4'), covar=tensor([0.0111, 0.0086, 0.0193, 0.0296, 0.0113, 0.2470, 0.0144, 0.0225], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0147, 0.0186, 0.0167, 0.0166, 0.0201, 0.0177, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:42:32,596 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-01 04:42:43,791 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=203099.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:42:48,524 INFO [train.py:904] (4/8) Epoch 21, batch 100, loss[loss=0.1751, simple_loss=0.2783, pruned_loss=0.03594, over 17076.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2682, pruned_loss=0.04735, over 1326551.92 frames. ], batch size: 49, lr: 3.26e-03, grad_scale: 2.0 2023-05-01 04:43:12,271 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-01 04:43:50,289 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=203146.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:43:57,339 INFO [train.py:904] (4/8) Epoch 21, batch 150, loss[loss=0.1792, simple_loss=0.2621, pruned_loss=0.04811, over 16654.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2657, pruned_loss=0.04596, over 1776561.84 frames. ], batch size: 89, lr: 3.26e-03, grad_scale: 2.0 2023-05-01 04:44:06,730 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6286, 2.2643, 2.3647, 4.4667, 2.2984, 2.6566, 2.3511, 2.4465], device='cuda:4'), covar=tensor([0.1117, 0.3756, 0.2948, 0.0430, 0.4083, 0.2640, 0.3603, 0.3576], device='cuda:4'), in_proj_covar=tensor([0.0390, 0.0434, 0.0359, 0.0317, 0.0427, 0.0495, 0.0403, 0.0505], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:44:44,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.604e+02 2.249e+02 2.595e+02 3.084e+02 6.012e+02, threshold=5.190e+02, percent-clipped=1.0 2023-05-01 04:44:44,942 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=203186.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 04:45:04,855 INFO [train.py:904] (4/8) Epoch 21, batch 200, loss[loss=0.1989, simple_loss=0.2727, pruned_loss=0.06257, over 16608.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2662, pruned_loss=0.04656, over 2117995.70 frames. ], batch size: 134, lr: 3.26e-03, grad_scale: 2.0 2023-05-01 04:46:05,949 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8897, 3.0630, 2.9241, 5.0993, 4.3481, 4.4828, 1.6642, 3.4720], device='cuda:4'), covar=tensor([0.1279, 0.0702, 0.1063, 0.0217, 0.0226, 0.0391, 0.1645, 0.0646], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0172, 0.0192, 0.0184, 0.0201, 0.0211, 0.0200, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:46:15,295 INFO [train.py:904] (4/8) Epoch 21, batch 250, loss[loss=0.1772, simple_loss=0.2584, pruned_loss=0.04799, over 15684.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.265, pruned_loss=0.04659, over 2385070.24 frames. ], batch size: 191, lr: 3.26e-03, grad_scale: 2.0 2023-05-01 04:46:45,985 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9983, 2.0277, 2.4651, 2.8678, 2.6752, 3.3728, 2.3283, 3.4021], device='cuda:4'), covar=tensor([0.0250, 0.0513, 0.0349, 0.0327, 0.0349, 0.0188, 0.0472, 0.0182], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0187, 0.0174, 0.0178, 0.0190, 0.0147, 0.0191, 0.0142], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:47:00,829 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.546e+02 2.163e+02 2.580e+02 3.064e+02 7.723e+02, threshold=5.160e+02, percent-clipped=4.0 2023-05-01 04:47:23,329 INFO [train.py:904] (4/8) Epoch 21, batch 300, loss[loss=0.1565, simple_loss=0.2362, pruned_loss=0.03846, over 15777.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2617, pruned_loss=0.04499, over 2586242.80 frames. ], batch size: 35, lr: 3.26e-03, grad_scale: 2.0 2023-05-01 04:48:04,126 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1454, 2.1748, 2.3805, 3.9263, 2.1979, 2.4943, 2.2648, 2.3586], device='cuda:4'), covar=tensor([0.1486, 0.3726, 0.2988, 0.0687, 0.3983, 0.2599, 0.3902, 0.3036], device='cuda:4'), in_proj_covar=tensor([0.0395, 0.0440, 0.0364, 0.0322, 0.0432, 0.0503, 0.0410, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:48:32,603 INFO [train.py:904] (4/8) Epoch 21, batch 350, loss[loss=0.1749, simple_loss=0.2734, pruned_loss=0.03824, over 17254.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2592, pruned_loss=0.04353, over 2754145.87 frames. ], batch size: 52, lr: 3.25e-03, grad_scale: 2.0 2023-05-01 04:48:55,243 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1890, 2.1699, 2.2768, 3.9469, 2.2345, 2.5215, 2.2250, 2.3585], device='cuda:4'), covar=tensor([0.1503, 0.3763, 0.3024, 0.0650, 0.3933, 0.2515, 0.3904, 0.3049], device='cuda:4'), in_proj_covar=tensor([0.0395, 0.0441, 0.0364, 0.0323, 0.0432, 0.0503, 0.0410, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 04:49:17,935 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 2.171e+02 2.523e+02 3.052e+02 8.873e+02, threshold=5.047e+02, percent-clipped=2.0 2023-05-01 04:49:30,126 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=203394.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:49:41,954 INFO [train.py:904] (4/8) Epoch 21, batch 400, loss[loss=0.1847, simple_loss=0.2749, pruned_loss=0.04727, over 16559.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2572, pruned_loss=0.0427, over 2875934.79 frames. ], batch size: 68, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 04:50:13,128 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.11 vs. limit=2.0 2023-05-01 04:50:19,382 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6953, 3.6053, 4.0322, 2.2668, 4.2090, 4.1980, 3.1256, 3.1570], device='cuda:4'), covar=tensor([0.0804, 0.0262, 0.0204, 0.1106, 0.0087, 0.0185, 0.0424, 0.0462], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0107, 0.0095, 0.0139, 0.0078, 0.0121, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:50:37,061 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=203441.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:50:48,762 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-05-01 04:50:50,734 INFO [train.py:904] (4/8) Epoch 21, batch 450, loss[loss=0.1588, simple_loss=0.2527, pruned_loss=0.03244, over 16710.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2556, pruned_loss=0.04198, over 2978185.64 frames. ], batch size: 57, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 04:50:55,502 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7255, 4.0012, 3.0949, 2.2992, 2.5757, 2.4929, 4.1269, 3.4236], device='cuda:4'), covar=tensor([0.2818, 0.0564, 0.1670, 0.2938, 0.2793, 0.1999, 0.0453, 0.1426], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0267, 0.0303, 0.0308, 0.0291, 0.0256, 0.0291, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 04:51:10,949 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-05-01 04:51:11,836 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=203467.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:51:37,694 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.511e+02 2.094e+02 2.458e+02 3.011e+02 6.519e+02, threshold=4.916e+02, percent-clipped=3.0 2023-05-01 04:51:38,087 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=203486.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 04:51:59,114 INFO [train.py:904] (4/8) Epoch 21, batch 500, loss[loss=0.1744, simple_loss=0.267, pruned_loss=0.04087, over 17027.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2544, pruned_loss=0.04142, over 3060832.86 frames. ], batch size: 50, lr: 3.25e-03, grad_scale: 2.0 2023-05-01 04:52:36,270 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=203528.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:52:43,061 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=203534.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 04:53:08,035 INFO [train.py:904] (4/8) Epoch 21, batch 550, loss[loss=0.2093, simple_loss=0.2819, pruned_loss=0.06839, over 16369.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2537, pruned_loss=0.04098, over 3122650.10 frames. ], batch size: 145, lr: 3.25e-03, grad_scale: 2.0 2023-05-01 04:53:52,780 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-05-01 04:53:56,912 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.204e+02 2.652e+02 3.296e+02 1.402e+03, threshold=5.303e+02, percent-clipped=5.0 2023-05-01 04:54:17,582 INFO [train.py:904] (4/8) Epoch 21, batch 600, loss[loss=0.1693, simple_loss=0.2613, pruned_loss=0.03866, over 16720.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2543, pruned_loss=0.04167, over 3165466.60 frames. ], batch size: 57, lr: 3.25e-03, grad_scale: 2.0 2023-05-01 04:55:27,189 INFO [train.py:904] (4/8) Epoch 21, batch 650, loss[loss=0.1665, simple_loss=0.2393, pruned_loss=0.04683, over 16834.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2528, pruned_loss=0.04169, over 3196276.92 frames. ], batch size: 116, lr: 3.25e-03, grad_scale: 2.0 2023-05-01 04:55:28,682 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5700, 3.4050, 3.7677, 2.0173, 3.9293, 3.9496, 3.0626, 2.9255], device='cuda:4'), covar=tensor([0.0750, 0.0261, 0.0216, 0.1162, 0.0101, 0.0183, 0.0429, 0.0450], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0096, 0.0140, 0.0079, 0.0123, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:55:39,391 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 04:56:14,236 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.434e+02 2.016e+02 2.395e+02 3.216e+02 7.414e+02, threshold=4.790e+02, percent-clipped=2.0 2023-05-01 04:56:24,928 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=203694.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:56:34,645 INFO [train.py:904] (4/8) Epoch 21, batch 700, loss[loss=0.1656, simple_loss=0.243, pruned_loss=0.04414, over 15593.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.252, pruned_loss=0.0413, over 3222909.47 frames. ], batch size: 190, lr: 3.25e-03, grad_scale: 2.0 2023-05-01 04:57:30,217 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=203741.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:57:31,298 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=203742.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:57:34,934 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3849, 3.6055, 3.9631, 2.1326, 3.2340, 2.5162, 3.8660, 3.7827], device='cuda:4'), covar=tensor([0.0254, 0.0896, 0.0478, 0.1965, 0.0747, 0.0942, 0.0550, 0.1022], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0160, 0.0165, 0.0151, 0.0143, 0.0128, 0.0143, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 04:57:38,344 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=203747.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:57:44,448 INFO [train.py:904] (4/8) Epoch 21, batch 750, loss[loss=0.1681, simple_loss=0.2597, pruned_loss=0.03827, over 17115.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2533, pruned_loss=0.04141, over 3249574.07 frames. ], batch size: 49, lr: 3.25e-03, grad_scale: 2.0 2023-05-01 04:58:03,803 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=203766.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:58:33,072 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.632e+02 2.172e+02 2.611e+02 3.083e+02 3.128e+03, threshold=5.221e+02, percent-clipped=7.0 2023-05-01 04:58:36,275 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=203789.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:58:54,751 INFO [train.py:904] (4/8) Epoch 21, batch 800, loss[loss=0.151, simple_loss=0.2417, pruned_loss=0.0302, over 17212.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2529, pruned_loss=0.04169, over 3265514.86 frames. ], batch size: 46, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 04:59:03,698 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=203808.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:59:20,580 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-01 04:59:22,980 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-05-01 04:59:24,700 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=203823.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:59:30,088 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=203827.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 04:59:58,194 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=203847.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:00:04,930 INFO [train.py:904] (4/8) Epoch 21, batch 850, loss[loss=0.1767, simple_loss=0.249, pruned_loss=0.05221, over 12132.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.252, pruned_loss=0.0414, over 3278018.22 frames. ], batch size: 247, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:00:53,307 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.368e+02 1.964e+02 2.295e+02 2.674e+02 5.967e+02, threshold=4.591e+02, percent-clipped=3.0 2023-05-01 05:01:11,193 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=203900.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 05:01:12,967 INFO [train.py:904] (4/8) Epoch 21, batch 900, loss[loss=0.1897, simple_loss=0.2636, pruned_loss=0.0579, over 16668.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2514, pruned_loss=0.04098, over 3277855.97 frames. ], batch size: 134, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:01:20,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2163, 4.2879, 4.3849, 4.1958, 4.2621, 4.8144, 4.3690, 4.0043], device='cuda:4'), covar=tensor([0.1964, 0.2304, 0.2522, 0.2378, 0.3067, 0.1412, 0.1952, 0.2928], device='cuda:4'), in_proj_covar=tensor([0.0403, 0.0591, 0.0652, 0.0489, 0.0652, 0.0682, 0.0512, 0.0657], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:01:21,581 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=203908.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 05:02:21,435 INFO [train.py:904] (4/8) Epoch 21, batch 950, loss[loss=0.1801, simple_loss=0.2502, pruned_loss=0.05501, over 16684.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2516, pruned_loss=0.04097, over 3292103.63 frames. ], batch size: 134, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:02:33,308 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-01 05:02:35,127 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=203961.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 05:02:37,025 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.07 vs. limit=5.0 2023-05-01 05:03:10,185 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.322e+02 2.138e+02 2.475e+02 2.994e+02 7.321e+02, threshold=4.951e+02, percent-clipped=5.0 2023-05-01 05:03:22,624 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0826, 4.0728, 2.4637, 4.6938, 3.1226, 4.6623, 2.5018, 3.3523], device='cuda:4'), covar=tensor([0.0254, 0.0378, 0.1669, 0.0223, 0.0812, 0.0483, 0.1584, 0.0663], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0178, 0.0195, 0.0162, 0.0178, 0.0215, 0.0204, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:03:34,206 INFO [train.py:904] (4/8) Epoch 21, batch 1000, loss[loss=0.1846, simple_loss=0.2479, pruned_loss=0.06062, over 16862.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2504, pruned_loss=0.04071, over 3300840.93 frames. ], batch size: 124, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:04:07,787 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7980, 5.1823, 4.9403, 4.9747, 4.6974, 4.6439, 4.5831, 5.2608], device='cuda:4'), covar=tensor([0.1432, 0.0968, 0.1062, 0.0890, 0.0867, 0.1223, 0.1285, 0.0907], device='cuda:4'), in_proj_covar=tensor([0.0672, 0.0818, 0.0674, 0.0617, 0.0522, 0.0529, 0.0690, 0.0636], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:04:18,853 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-01 05:04:43,737 INFO [train.py:904] (4/8) Epoch 21, batch 1050, loss[loss=0.1816, simple_loss=0.2516, pruned_loss=0.05579, over 16899.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2504, pruned_loss=0.0406, over 3300668.89 frames. ], batch size: 109, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:04:47,560 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1072, 3.8709, 4.0342, 4.3038, 4.3752, 3.9938, 4.1893, 4.3804], device='cuda:4'), covar=tensor([0.1734, 0.1445, 0.1797, 0.0850, 0.0798, 0.1488, 0.3401, 0.0927], device='cuda:4'), in_proj_covar=tensor([0.0645, 0.0788, 0.0920, 0.0810, 0.0604, 0.0629, 0.0651, 0.0751], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:05:19,238 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=204077.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:05:22,891 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5912, 3.5012, 3.8427, 1.9863, 4.0217, 4.0353, 3.0976, 2.9440], device='cuda:4'), covar=tensor([0.0809, 0.0236, 0.0200, 0.1180, 0.0097, 0.0179, 0.0417, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0108, 0.0096, 0.0139, 0.0079, 0.0122, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:05:34,276 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.391e+02 2.155e+02 2.486e+02 2.944e+02 8.933e+02, threshold=4.972e+02, percent-clipped=2.0 2023-05-01 05:05:35,487 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9788, 4.5521, 3.2894, 2.3607, 2.7530, 2.5680, 4.8918, 3.6172], device='cuda:4'), covar=tensor([0.2666, 0.0470, 0.1633, 0.3026, 0.3040, 0.2116, 0.0291, 0.1438], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0269, 0.0303, 0.0309, 0.0292, 0.0256, 0.0293, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:05:44,049 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7660, 4.1976, 3.0524, 2.2520, 2.6992, 2.5241, 4.6153, 3.5409], device='cuda:4'), covar=tensor([0.2766, 0.0612, 0.1721, 0.2811, 0.2780, 0.2042, 0.0346, 0.1332], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0269, 0.0303, 0.0309, 0.0292, 0.0256, 0.0293, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:05:47,462 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3686, 5.3298, 5.1874, 4.6875, 4.7978, 5.2292, 5.1393, 4.8224], device='cuda:4'), covar=tensor([0.0598, 0.0457, 0.0327, 0.0368, 0.1086, 0.0433, 0.0323, 0.0847], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0422, 0.0341, 0.0337, 0.0349, 0.0393, 0.0234, 0.0410], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:05:54,735 INFO [train.py:904] (4/8) Epoch 21, batch 1100, loss[loss=0.1655, simple_loss=0.2534, pruned_loss=0.03884, over 17125.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2505, pruned_loss=0.04035, over 3309962.11 frames. ], batch size: 48, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:05:56,315 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=204103.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:06:24,117 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=204122.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:06:25,344 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=204123.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:06:45,642 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=204138.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 05:07:05,258 INFO [train.py:904] (4/8) Epoch 21, batch 1150, loss[loss=0.1477, simple_loss=0.247, pruned_loss=0.02419, over 17086.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2491, pruned_loss=0.03942, over 3300373.37 frames. ], batch size: 49, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:07:20,970 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=204163.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:07:30,663 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3693, 3.3515, 2.1379, 3.5638, 2.6947, 3.6052, 2.2035, 2.7520], device='cuda:4'), covar=tensor([0.0285, 0.0432, 0.1496, 0.0351, 0.0779, 0.0636, 0.1461, 0.0684], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0177, 0.0194, 0.0162, 0.0177, 0.0214, 0.0203, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:07:31,657 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=204171.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:07:53,760 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 2.065e+02 2.500e+02 2.935e+02 5.015e+02, threshold=5.000e+02, percent-clipped=1.0 2023-05-01 05:08:07,147 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3264, 5.7002, 5.4374, 5.4970, 5.1368, 5.1797, 5.0789, 5.7972], device='cuda:4'), covar=tensor([0.1362, 0.1026, 0.1128, 0.1059, 0.0975, 0.0821, 0.1232, 0.1040], device='cuda:4'), in_proj_covar=tensor([0.0678, 0.0826, 0.0679, 0.0623, 0.0526, 0.0533, 0.0696, 0.0642], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:08:13,780 INFO [train.py:904] (4/8) Epoch 21, batch 1200, loss[loss=0.1588, simple_loss=0.252, pruned_loss=0.03279, over 16718.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2478, pruned_loss=0.03872, over 3307259.31 frames. ], batch size: 57, lr: 3.25e-03, grad_scale: 8.0 2023-05-01 05:08:15,242 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=204203.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 05:08:15,356 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8129, 4.9723, 4.7045, 4.3717, 4.0061, 4.9751, 4.9485, 4.5108], device='cuda:4'), covar=tensor([0.1072, 0.0833, 0.0619, 0.0549, 0.2010, 0.0638, 0.0406, 0.0910], device='cuda:4'), in_proj_covar=tensor([0.0298, 0.0426, 0.0344, 0.0341, 0.0352, 0.0397, 0.0237, 0.0414], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:08:28,552 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6517, 4.7535, 4.9032, 4.7676, 4.7574, 5.3403, 4.8194, 4.5317], device='cuda:4'), covar=tensor([0.1492, 0.2015, 0.2360, 0.2155, 0.2826, 0.1044, 0.1710, 0.2471], device='cuda:4'), in_proj_covar=tensor([0.0404, 0.0594, 0.0654, 0.0491, 0.0656, 0.0683, 0.0515, 0.0659], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:08:44,925 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=204224.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:09:06,280 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-01 05:09:24,353 INFO [train.py:904] (4/8) Epoch 21, batch 1250, loss[loss=0.1594, simple_loss=0.2385, pruned_loss=0.0401, over 12253.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2485, pruned_loss=0.03954, over 3299857.21 frames. ], batch size: 246, lr: 3.25e-03, grad_scale: 8.0 2023-05-01 05:09:30,706 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=204256.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 05:10:01,375 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1709, 2.1081, 1.6953, 1.8901, 2.3003, 2.0824, 2.1303, 2.4318], device='cuda:4'), covar=tensor([0.0311, 0.0447, 0.0610, 0.0489, 0.0298, 0.0389, 0.0291, 0.0318], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0238, 0.0228, 0.0228, 0.0237, 0.0236, 0.0238, 0.0234], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:10:04,696 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9783, 2.1682, 2.5347, 2.9267, 2.7663, 3.4447, 2.3689, 3.3874], device='cuda:4'), covar=tensor([0.0263, 0.0511, 0.0339, 0.0339, 0.0344, 0.0202, 0.0454, 0.0178], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0194, 0.0180, 0.0184, 0.0196, 0.0153, 0.0196, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:10:12,294 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.450e+02 2.352e+02 2.804e+02 3.486e+02 9.364e+02, threshold=5.607e+02, percent-clipped=5.0 2023-05-01 05:10:21,128 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6690, 1.7713, 1.5905, 1.4479, 1.8193, 1.5527, 1.5830, 1.8789], device='cuda:4'), covar=tensor([0.0248, 0.0327, 0.0456, 0.0418, 0.0251, 0.0328, 0.0240, 0.0259], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0237, 0.0227, 0.0227, 0.0237, 0.0236, 0.0238, 0.0233], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:10:33,931 INFO [train.py:904] (4/8) Epoch 21, batch 1300, loss[loss=0.1708, simple_loss=0.2404, pruned_loss=0.05058, over 16662.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2479, pruned_loss=0.03959, over 3295332.44 frames. ], batch size: 134, lr: 3.25e-03, grad_scale: 8.0 2023-05-01 05:11:17,941 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9766, 2.6382, 2.9007, 2.0997, 2.6781, 2.1368, 2.8218, 2.8681], device='cuda:4'), covar=tensor([0.0328, 0.0916, 0.0570, 0.1798, 0.0850, 0.0933, 0.0636, 0.0839], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0167, 0.0153, 0.0145, 0.0130, 0.0145, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:11:42,901 INFO [train.py:904] (4/8) Epoch 21, batch 1350, loss[loss=0.1414, simple_loss=0.2224, pruned_loss=0.03022, over 16790.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2475, pruned_loss=0.03899, over 3296859.55 frames. ], batch size: 39, lr: 3.25e-03, grad_scale: 8.0 2023-05-01 05:11:48,284 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1537, 4.8756, 5.1831, 5.3626, 5.6105, 4.8705, 5.5210, 5.5634], device='cuda:4'), covar=tensor([0.1854, 0.1305, 0.1672, 0.0708, 0.0502, 0.0833, 0.0473, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0650, 0.0795, 0.0928, 0.0818, 0.0611, 0.0635, 0.0657, 0.0759], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:12:17,604 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-01 05:12:31,606 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.473e+02 2.188e+02 2.581e+02 2.996e+02 9.160e+02, threshold=5.162e+02, percent-clipped=1.0 2023-05-01 05:12:52,518 INFO [train.py:904] (4/8) Epoch 21, batch 1400, loss[loss=0.1717, simple_loss=0.2559, pruned_loss=0.0437, over 16567.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2477, pruned_loss=0.03869, over 3312589.48 frames. ], batch size: 75, lr: 3.25e-03, grad_scale: 8.0 2023-05-01 05:12:54,521 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=204403.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:13:20,715 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=204422.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:13:35,304 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=204433.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 05:14:00,494 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=204451.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:14:01,405 INFO [train.py:904] (4/8) Epoch 21, batch 1450, loss[loss=0.1729, simple_loss=0.2435, pruned_loss=0.05115, over 16278.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2468, pruned_loss=0.03902, over 3296471.73 frames. ], batch size: 165, lr: 3.25e-03, grad_scale: 8.0 2023-05-01 05:14:26,854 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=204470.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:14:28,977 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9778, 2.8479, 2.6312, 4.4659, 3.7423, 4.2262, 1.7507, 3.1785], device='cuda:4'), covar=tensor([0.1304, 0.0722, 0.1172, 0.0226, 0.0181, 0.0411, 0.1539, 0.0737], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0195, 0.0189, 0.0205, 0.0216, 0.0202, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:14:50,905 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.523e+02 2.087e+02 2.507e+02 3.018e+02 6.138e+02, threshold=5.015e+02, percent-clipped=2.0 2023-05-01 05:14:57,712 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9290, 5.3127, 5.0528, 5.0888, 4.7789, 4.8347, 4.7307, 5.3963], device='cuda:4'), covar=tensor([0.1368, 0.1002, 0.1119, 0.0944, 0.0934, 0.1058, 0.1210, 0.0946], device='cuda:4'), in_proj_covar=tensor([0.0681, 0.0831, 0.0680, 0.0627, 0.0529, 0.0535, 0.0699, 0.0644], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:15:10,560 INFO [train.py:904] (4/8) Epoch 21, batch 1500, loss[loss=0.1764, simple_loss=0.2672, pruned_loss=0.04273, over 17192.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2467, pruned_loss=0.03896, over 3305351.09 frames. ], batch size: 46, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:15:12,817 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=204503.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:15:33,800 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=204519.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:16:17,197 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=204551.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:16:18,101 INFO [train.py:904] (4/8) Epoch 21, batch 1550, loss[loss=0.1364, simple_loss=0.2208, pruned_loss=0.02603, over 16779.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.248, pruned_loss=0.03969, over 3313426.52 frames. ], batch size: 39, lr: 3.25e-03, grad_scale: 4.0 2023-05-01 05:16:23,635 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=204556.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 05:16:26,442 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.78 vs. limit=5.0 2023-05-01 05:17:07,001 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.626e+02 2.391e+02 2.852e+02 3.354e+02 7.624e+02, threshold=5.704e+02, percent-clipped=5.0 2023-05-01 05:17:08,251 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-05-01 05:17:26,284 INFO [train.py:904] (4/8) Epoch 21, batch 1600, loss[loss=0.1946, simple_loss=0.2766, pruned_loss=0.05628, over 16420.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2504, pruned_loss=0.04086, over 3322437.72 frames. ], batch size: 75, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:17:28,784 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=204604.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 05:18:06,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7841, 3.8309, 3.8939, 3.6960, 3.8624, 4.2708, 3.8460, 3.4915], device='cuda:4'), covar=tensor([0.2202, 0.2442, 0.2395, 0.2499, 0.2741, 0.1880, 0.1744, 0.3060], device='cuda:4'), in_proj_covar=tensor([0.0403, 0.0591, 0.0651, 0.0487, 0.0651, 0.0680, 0.0513, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:18:24,665 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9026, 2.1216, 2.4028, 2.6691, 2.7638, 2.7989, 2.0362, 2.9798], device='cuda:4'), covar=tensor([0.0196, 0.0472, 0.0350, 0.0305, 0.0305, 0.0279, 0.0552, 0.0194], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0193, 0.0180, 0.0184, 0.0196, 0.0154, 0.0196, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:18:35,822 INFO [train.py:904] (4/8) Epoch 21, batch 1650, loss[loss=0.1705, simple_loss=0.2649, pruned_loss=0.03805, over 17122.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2521, pruned_loss=0.04169, over 3320883.79 frames. ], batch size: 48, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:19:25,764 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.742e+02 2.251e+02 2.625e+02 3.143e+02 5.667e+02, threshold=5.251e+02, percent-clipped=0.0 2023-05-01 05:19:38,405 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.86 vs. limit=5.0 2023-05-01 05:19:45,540 INFO [train.py:904] (4/8) Epoch 21, batch 1700, loss[loss=0.1639, simple_loss=0.2515, pruned_loss=0.03814, over 16863.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2534, pruned_loss=0.04144, over 3328857.13 frames. ], batch size: 42, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:20:29,899 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=204733.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 05:20:46,395 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0393, 3.1817, 3.2340, 2.1370, 2.8333, 2.2722, 3.5239, 3.5108], device='cuda:4'), covar=tensor([0.0246, 0.0878, 0.0639, 0.1873, 0.0855, 0.1023, 0.0540, 0.0882], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0167, 0.0152, 0.0144, 0.0129, 0.0144, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:20:55,679 INFO [train.py:904] (4/8) Epoch 21, batch 1750, loss[loss=0.2059, simple_loss=0.2882, pruned_loss=0.06178, over 15609.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2544, pruned_loss=0.04166, over 3330130.35 frames. ], batch size: 191, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:21:37,075 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=204781.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:21:46,293 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.620e+02 2.363e+02 2.773e+02 3.345e+02 7.836e+02, threshold=5.546e+02, percent-clipped=6.0 2023-05-01 05:22:06,996 INFO [train.py:904] (4/8) Epoch 21, batch 1800, loss[loss=0.1836, simple_loss=0.2633, pruned_loss=0.05195, over 16784.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.255, pruned_loss=0.04144, over 3334276.81 frames. ], batch size: 96, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:22:30,016 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=204819.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:22:32,913 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=204821.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:22:34,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7268, 4.9263, 5.0835, 4.8321, 4.9303, 5.5459, 5.0390, 4.6691], device='cuda:4'), covar=tensor([0.1422, 0.2191, 0.2434, 0.2375, 0.2814, 0.1049, 0.1653, 0.2767], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0599, 0.0660, 0.0495, 0.0661, 0.0689, 0.0521, 0.0664], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:23:03,707 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9358, 5.2700, 5.0364, 5.0376, 4.7982, 4.7550, 4.7452, 5.3447], device='cuda:4'), covar=tensor([0.1241, 0.0834, 0.1086, 0.0898, 0.0816, 0.1031, 0.1139, 0.0935], device='cuda:4'), in_proj_covar=tensor([0.0678, 0.0826, 0.0679, 0.0624, 0.0526, 0.0531, 0.0697, 0.0640], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:23:15,231 INFO [train.py:904] (4/8) Epoch 21, batch 1850, loss[loss=0.1585, simple_loss=0.2485, pruned_loss=0.03424, over 17198.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2563, pruned_loss=0.0419, over 3334083.89 frames. ], batch size: 46, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:23:37,468 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=204867.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:23:57,930 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=204882.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:24:05,812 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-05-01 05:24:06,221 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.348e+02 2.082e+02 2.497e+02 3.026e+02 6.100e+02, threshold=4.995e+02, percent-clipped=2.0 2023-05-01 05:24:26,176 INFO [train.py:904] (4/8) Epoch 21, batch 1900, loss[loss=0.164, simple_loss=0.2559, pruned_loss=0.036, over 17139.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2559, pruned_loss=0.04127, over 3330061.26 frames. ], batch size: 48, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:25:35,963 INFO [train.py:904] (4/8) Epoch 21, batch 1950, loss[loss=0.1472, simple_loss=0.2337, pruned_loss=0.03036, over 16963.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2556, pruned_loss=0.0412, over 3328886.66 frames. ], batch size: 41, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:26:26,411 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.621e+02 2.195e+02 2.614e+02 3.172e+02 4.787e+02, threshold=5.228e+02, percent-clipped=0.0 2023-05-01 05:26:36,808 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7755, 3.8414, 2.4892, 4.4607, 2.9953, 4.4087, 2.5652, 3.1162], device='cuda:4'), covar=tensor([0.0312, 0.0388, 0.1473, 0.0294, 0.0813, 0.0462, 0.1438, 0.0733], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0179, 0.0196, 0.0166, 0.0178, 0.0219, 0.0205, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:26:44,684 INFO [train.py:904] (4/8) Epoch 21, batch 2000, loss[loss=0.1798, simple_loss=0.2506, pruned_loss=0.05445, over 16846.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2551, pruned_loss=0.04112, over 3326306.76 frames. ], batch size: 83, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:27:55,522 INFO [train.py:904] (4/8) Epoch 21, batch 2050, loss[loss=0.1591, simple_loss=0.251, pruned_loss=0.03355, over 17140.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2548, pruned_loss=0.04108, over 3322121.99 frames. ], batch size: 47, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:28:18,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1077, 5.0864, 4.8869, 4.4474, 4.3810, 5.0601, 4.9787, 4.5071], device='cuda:4'), covar=tensor([0.0696, 0.0654, 0.0387, 0.0463, 0.1336, 0.0470, 0.0355, 0.0830], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0436, 0.0353, 0.0349, 0.0361, 0.0404, 0.0240, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:28:44,313 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.164e+02 2.478e+02 3.098e+02 5.896e+02, threshold=4.956e+02, percent-clipped=1.0 2023-05-01 05:28:50,297 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-05-01 05:29:04,151 INFO [train.py:904] (4/8) Epoch 21, batch 2100, loss[loss=0.1421, simple_loss=0.2243, pruned_loss=0.02997, over 16797.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2558, pruned_loss=0.04147, over 3326237.06 frames. ], batch size: 39, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:29:14,604 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=205109.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:30:14,923 INFO [train.py:904] (4/8) Epoch 21, batch 2150, loss[loss=0.1654, simple_loss=0.2536, pruned_loss=0.03858, over 16300.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.257, pruned_loss=0.04228, over 3314738.00 frames. ], batch size: 165, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:30:26,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7976, 2.6320, 2.4314, 4.1100, 3.4263, 4.0415, 1.6286, 2.9533], device='cuda:4'), covar=tensor([0.1343, 0.0707, 0.1225, 0.0176, 0.0172, 0.0389, 0.1540, 0.0836], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0175, 0.0194, 0.0190, 0.0205, 0.0216, 0.0201, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:30:39,968 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=205170.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 05:30:50,421 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=205177.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:31:04,862 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.289e+02 2.743e+02 3.268e+02 8.955e+02, threshold=5.486e+02, percent-clipped=2.0 2023-05-01 05:31:19,705 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=205198.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:31:24,882 INFO [train.py:904] (4/8) Epoch 21, batch 2200, loss[loss=0.1826, simple_loss=0.2581, pruned_loss=0.05359, over 16693.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.258, pruned_loss=0.04253, over 3319910.69 frames. ], batch size: 124, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:31:44,780 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-01 05:32:34,214 INFO [train.py:904] (4/8) Epoch 21, batch 2250, loss[loss=0.2261, simple_loss=0.3038, pruned_loss=0.07418, over 11839.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2589, pruned_loss=0.04258, over 3316077.37 frames. ], batch size: 246, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:32:45,036 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=205259.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:32:49,107 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2120, 5.6865, 5.8343, 5.4545, 5.5667, 6.1790, 5.6747, 5.4147], device='cuda:4'), covar=tensor([0.0841, 0.2005, 0.2821, 0.2246, 0.2802, 0.0915, 0.1536, 0.2360], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0601, 0.0662, 0.0499, 0.0663, 0.0692, 0.0521, 0.0668], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:33:19,318 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8793, 2.1002, 2.5627, 2.8434, 2.6490, 3.3279, 2.2517, 3.2798], device='cuda:4'), covar=tensor([0.0255, 0.0462, 0.0325, 0.0314, 0.0346, 0.0177, 0.0485, 0.0164], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0194, 0.0180, 0.0186, 0.0198, 0.0155, 0.0197, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:33:23,477 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.631e+02 2.149e+02 2.507e+02 2.983e+02 5.754e+02, threshold=5.014e+02, percent-clipped=1.0 2023-05-01 05:33:44,283 INFO [train.py:904] (4/8) Epoch 21, batch 2300, loss[loss=0.1562, simple_loss=0.2482, pruned_loss=0.03209, over 17234.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2594, pruned_loss=0.04257, over 3316073.30 frames. ], batch size: 44, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:34:53,179 INFO [train.py:904] (4/8) Epoch 21, batch 2350, loss[loss=0.2069, simple_loss=0.2792, pruned_loss=0.06726, over 16894.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2596, pruned_loss=0.04307, over 3313550.32 frames. ], batch size: 109, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:35:18,400 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9207, 4.4899, 2.9220, 2.2250, 2.8059, 2.4382, 4.7172, 3.6578], device='cuda:4'), covar=tensor([0.2901, 0.0538, 0.2027, 0.3114, 0.3066, 0.2225, 0.0410, 0.1391], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0270, 0.0303, 0.0309, 0.0296, 0.0257, 0.0294, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:35:42,801 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.569e+02 2.163e+02 2.457e+02 2.980e+02 4.846e+02, threshold=4.914e+02, percent-clipped=0.0 2023-05-01 05:35:55,184 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-05-01 05:36:02,959 INFO [train.py:904] (4/8) Epoch 21, batch 2400, loss[loss=0.1836, simple_loss=0.2647, pruned_loss=0.05122, over 16693.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2601, pruned_loss=0.04329, over 3323266.73 frames. ], batch size: 134, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:36:10,930 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-01 05:36:51,173 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-05-01 05:37:10,842 INFO [train.py:904] (4/8) Epoch 21, batch 2450, loss[loss=0.1818, simple_loss=0.269, pruned_loss=0.04734, over 15609.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.261, pruned_loss=0.04338, over 3318275.50 frames. ], batch size: 190, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:37:24,963 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1486, 3.0735, 3.1517, 2.2932, 2.9475, 3.2685, 2.9810, 1.8530], device='cuda:4'), covar=tensor([0.0466, 0.0128, 0.0070, 0.0372, 0.0137, 0.0102, 0.0121, 0.0479], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0084, 0.0083, 0.0134, 0.0099, 0.0109, 0.0094, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:37:29,570 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=205465.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 05:37:46,014 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=205477.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:38:00,842 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.425e+02 2.346e+02 2.648e+02 3.284e+02 5.369e+02, threshold=5.296e+02, percent-clipped=1.0 2023-05-01 05:38:16,902 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7990, 1.9544, 2.4331, 2.6691, 2.7507, 2.6930, 2.1000, 2.9174], device='cuda:4'), covar=tensor([0.0190, 0.0465, 0.0308, 0.0263, 0.0277, 0.0277, 0.0455, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0195, 0.0180, 0.0186, 0.0198, 0.0155, 0.0198, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:38:21,262 INFO [train.py:904] (4/8) Epoch 21, batch 2500, loss[loss=0.171, simple_loss=0.2654, pruned_loss=0.0383, over 17072.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2601, pruned_loss=0.04323, over 3322679.17 frames. ], batch size: 53, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:38:40,336 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9968, 4.5274, 4.5657, 3.3489, 3.7806, 4.4519, 3.9054, 2.6896], device='cuda:4'), covar=tensor([0.0447, 0.0054, 0.0037, 0.0326, 0.0128, 0.0098, 0.0089, 0.0426], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0084, 0.0083, 0.0134, 0.0099, 0.0109, 0.0094, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:38:53,422 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=205525.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:39:30,351 INFO [train.py:904] (4/8) Epoch 21, batch 2550, loss[loss=0.1678, simple_loss=0.2523, pruned_loss=0.04172, over 16044.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2598, pruned_loss=0.04304, over 3322367.68 frames. ], batch size: 35, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:39:33,578 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=205554.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:40:17,560 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9835, 4.7320, 5.0288, 5.2093, 5.4205, 4.7537, 5.3979, 5.3759], device='cuda:4'), covar=tensor([0.1847, 0.1309, 0.1726, 0.0758, 0.0537, 0.0968, 0.0541, 0.0611], device='cuda:4'), in_proj_covar=tensor([0.0660, 0.0810, 0.0949, 0.0832, 0.0619, 0.0649, 0.0670, 0.0776], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:40:19,332 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.360e+02 2.328e+02 2.679e+02 3.297e+02 1.189e+03, threshold=5.358e+02, percent-clipped=3.0 2023-05-01 05:40:20,940 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5058, 3.5541, 2.8191, 2.1378, 2.3184, 2.3048, 3.6868, 3.1234], device='cuda:4'), covar=tensor([0.2949, 0.0684, 0.1770, 0.3203, 0.2676, 0.2128, 0.0608, 0.1559], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0270, 0.0304, 0.0308, 0.0296, 0.0257, 0.0294, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:40:38,687 INFO [train.py:904] (4/8) Epoch 21, batch 2600, loss[loss=0.178, simple_loss=0.2756, pruned_loss=0.04018, over 16651.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2592, pruned_loss=0.04231, over 3326468.12 frames. ], batch size: 57, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:41:49,801 INFO [train.py:904] (4/8) Epoch 21, batch 2650, loss[loss=0.1541, simple_loss=0.2497, pruned_loss=0.02925, over 17091.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2599, pruned_loss=0.04221, over 3327755.45 frames. ], batch size: 47, lr: 3.24e-03, grad_scale: 4.0 2023-05-01 05:41:54,613 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=205655.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:42:40,190 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 2.048e+02 2.349e+02 2.910e+02 5.870e+02, threshold=4.699e+02, percent-clipped=1.0 2023-05-01 05:43:00,011 INFO [train.py:904] (4/8) Epoch 21, batch 2700, loss[loss=0.1759, simple_loss=0.2721, pruned_loss=0.0399, over 17261.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2609, pruned_loss=0.04203, over 3323410.30 frames. ], batch size: 52, lr: 3.24e-03, grad_scale: 4.0 2023-05-01 05:43:18,681 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=205716.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:43:59,570 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=205744.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:44:09,875 INFO [train.py:904] (4/8) Epoch 21, batch 2750, loss[loss=0.1523, simple_loss=0.2445, pruned_loss=0.03006, over 17185.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.261, pruned_loss=0.04165, over 3323901.10 frames. ], batch size: 44, lr: 3.24e-03, grad_scale: 4.0 2023-05-01 05:44:13,737 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7240, 3.8610, 3.0086, 2.2975, 2.5957, 2.4912, 4.1062, 3.3677], device='cuda:4'), covar=tensor([0.2773, 0.0634, 0.1682, 0.2733, 0.2527, 0.1955, 0.0532, 0.1376], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0270, 0.0304, 0.0309, 0.0297, 0.0257, 0.0294, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:44:29,225 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=205765.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:45:02,029 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.167e+02 2.431e+02 2.778e+02 3.742e+02, threshold=4.863e+02, percent-clipped=0.0 2023-05-01 05:45:19,232 INFO [train.py:904] (4/8) Epoch 21, batch 2800, loss[loss=0.1411, simple_loss=0.2306, pruned_loss=0.02575, over 16991.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2599, pruned_loss=0.04126, over 3329195.52 frames. ], batch size: 41, lr: 3.24e-03, grad_scale: 8.0 2023-05-01 05:45:22,635 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8844, 4.8568, 4.6894, 4.1350, 4.7865, 1.9806, 4.5365, 4.4975], device='cuda:4'), covar=tensor([0.0105, 0.0084, 0.0201, 0.0334, 0.0099, 0.2628, 0.0137, 0.0196], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0157, 0.0199, 0.0180, 0.0178, 0.0210, 0.0190, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:45:24,442 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=205805.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:45:35,278 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=205813.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:46:29,085 INFO [train.py:904] (4/8) Epoch 21, batch 2850, loss[loss=0.145, simple_loss=0.2234, pruned_loss=0.0333, over 16783.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2594, pruned_loss=0.04202, over 3325558.20 frames. ], batch size: 102, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:46:31,705 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=205854.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:47:17,820 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.397e+02 2.169e+02 2.532e+02 3.008e+02 5.585e+02, threshold=5.064e+02, percent-clipped=2.0 2023-05-01 05:47:19,189 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2262, 5.5844, 5.7367, 5.4413, 5.4864, 6.1213, 5.6439, 5.3364], device='cuda:4'), covar=tensor([0.0856, 0.2022, 0.2283, 0.2034, 0.2676, 0.0950, 0.1416, 0.2285], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0611, 0.0669, 0.0505, 0.0669, 0.0699, 0.0526, 0.0673], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:47:19,304 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=205890.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:47:36,338 INFO [train.py:904] (4/8) Epoch 21, batch 2900, loss[loss=0.1745, simple_loss=0.2605, pruned_loss=0.04423, over 16438.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2582, pruned_loss=0.0417, over 3330022.45 frames. ], batch size: 68, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:47:36,636 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=205902.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:48:44,799 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=205951.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:48:45,570 INFO [train.py:904] (4/8) Epoch 21, batch 2950, loss[loss=0.1755, simple_loss=0.2522, pruned_loss=0.04945, over 16855.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2582, pruned_loss=0.04257, over 3323869.31 frames. ], batch size: 102, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:49:12,727 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0420, 2.1112, 2.2390, 3.6197, 2.1127, 2.3946, 2.2460, 2.2561], device='cuda:4'), covar=tensor([0.1444, 0.3799, 0.3022, 0.0704, 0.3965, 0.2664, 0.3592, 0.3429], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0449, 0.0369, 0.0331, 0.0435, 0.0515, 0.0418, 0.0525], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:49:28,992 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8977, 4.8915, 4.7957, 4.4176, 4.4775, 4.8584, 4.7377, 4.5835], device='cuda:4'), covar=tensor([0.0734, 0.0795, 0.0514, 0.0353, 0.1082, 0.0569, 0.0428, 0.0727], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0437, 0.0355, 0.0350, 0.0363, 0.0407, 0.0243, 0.0427], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:49:36,948 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.585e+02 2.251e+02 2.659e+02 3.342e+02 9.466e+02, threshold=5.318e+02, percent-clipped=4.0 2023-05-01 05:49:58,057 INFO [train.py:904] (4/8) Epoch 21, batch 3000, loss[loss=0.1509, simple_loss=0.24, pruned_loss=0.03096, over 17179.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2583, pruned_loss=0.04268, over 3327875.41 frames. ], batch size: 46, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:49:58,057 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 05:50:06,473 INFO [train.py:938] (4/8) Epoch 21, validation: loss=0.1356, simple_loss=0.2408, pruned_loss=0.01521, over 944034.00 frames. 2023-05-01 05:50:06,474 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 05:50:18,435 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=206011.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:51:14,522 INFO [train.py:904] (4/8) Epoch 21, batch 3050, loss[loss=0.1786, simple_loss=0.2566, pruned_loss=0.05026, over 16716.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.257, pruned_loss=0.0426, over 3333117.75 frames. ], batch size: 124, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:51:21,064 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=206056.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:51:56,636 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8879, 5.2548, 4.9689, 5.0254, 4.7657, 4.7478, 4.6942, 5.3615], device='cuda:4'), covar=tensor([0.1220, 0.0917, 0.1098, 0.0887, 0.0860, 0.1041, 0.1147, 0.0865], device='cuda:4'), in_proj_covar=tensor([0.0681, 0.0841, 0.0692, 0.0631, 0.0535, 0.0537, 0.0703, 0.0650], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:52:02,023 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0824, 4.8646, 5.0617, 5.2844, 5.5352, 4.7944, 5.5108, 5.4927], device='cuda:4'), covar=tensor([0.1858, 0.1240, 0.1930, 0.0805, 0.0511, 0.0969, 0.0466, 0.0575], device='cuda:4'), in_proj_covar=tensor([0.0668, 0.0818, 0.0959, 0.0841, 0.0623, 0.0658, 0.0674, 0.0781], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:52:05,514 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.452e+02 2.000e+02 2.421e+02 3.043e+02 5.222e+02, threshold=4.843e+02, percent-clipped=0.0 2023-05-01 05:52:21,874 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=206100.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:52:24,074 INFO [train.py:904] (4/8) Epoch 21, batch 3100, loss[loss=0.1565, simple_loss=0.2381, pruned_loss=0.03745, over 16978.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2562, pruned_loss=0.04244, over 3323116.81 frames. ], batch size: 41, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:52:42,982 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-01 05:52:43,662 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=206117.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:53:11,698 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4677, 3.5135, 3.6722, 2.5793, 3.3626, 3.7692, 3.4863, 2.1888], device='cuda:4'), covar=tensor([0.0478, 0.0152, 0.0065, 0.0384, 0.0117, 0.0098, 0.0094, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0083, 0.0082, 0.0133, 0.0098, 0.0108, 0.0093, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:53:30,989 INFO [train.py:904] (4/8) Epoch 21, batch 3150, loss[loss=0.2016, simple_loss=0.2772, pruned_loss=0.06304, over 16687.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2557, pruned_loss=0.0425, over 3323708.77 frames. ], batch size: 134, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:53:39,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7686, 3.8759, 4.1261, 4.1097, 4.1404, 3.8942, 3.9407, 3.8966], device='cuda:4'), covar=tensor([0.0396, 0.0565, 0.0410, 0.0411, 0.0516, 0.0438, 0.0751, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0465, 0.0453, 0.0420, 0.0498, 0.0474, 0.0568, 0.0380], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 05:53:48,276 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2004, 3.9314, 4.4475, 2.2735, 4.6610, 4.7017, 3.4236, 3.5807], device='cuda:4'), covar=tensor([0.0650, 0.0243, 0.0196, 0.1050, 0.0069, 0.0137, 0.0404, 0.0386], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0098, 0.0139, 0.0080, 0.0125, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:53:51,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2938, 4.0976, 4.5099, 2.5274, 4.7942, 4.7638, 3.4888, 3.7731], device='cuda:4'), covar=tensor([0.0646, 0.0217, 0.0211, 0.0947, 0.0066, 0.0161, 0.0376, 0.0353], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0098, 0.0139, 0.0080, 0.0125, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 05:54:22,972 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.468e+02 2.110e+02 2.360e+02 2.759e+02 5.694e+02, threshold=4.721e+02, percent-clipped=1.0 2023-05-01 05:54:34,982 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=206197.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:54:41,549 INFO [train.py:904] (4/8) Epoch 21, batch 3200, loss[loss=0.1702, simple_loss=0.2623, pruned_loss=0.03905, over 17263.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2554, pruned_loss=0.04203, over 3325676.12 frames. ], batch size: 52, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:54:55,261 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0899, 2.0303, 2.2081, 3.7865, 2.0370, 2.3471, 2.1417, 2.1648], device='cuda:4'), covar=tensor([0.1575, 0.4246, 0.3023, 0.0712, 0.4503, 0.2921, 0.4035, 0.3585], device='cuda:4'), in_proj_covar=tensor([0.0404, 0.0451, 0.0370, 0.0332, 0.0437, 0.0517, 0.0420, 0.0527], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 05:55:43,200 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=206246.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:55:51,281 INFO [train.py:904] (4/8) Epoch 21, batch 3250, loss[loss=0.1633, simple_loss=0.2453, pruned_loss=0.04063, over 16789.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.256, pruned_loss=0.04156, over 3322034.08 frames. ], batch size: 39, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:55:59,429 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=206258.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:56:30,697 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2633, 5.8457, 5.9572, 5.6960, 5.8181, 6.2705, 5.8577, 5.5844], device='cuda:4'), covar=tensor([0.0847, 0.2013, 0.2301, 0.2032, 0.2275, 0.0915, 0.1422, 0.2316], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0608, 0.0667, 0.0503, 0.0667, 0.0698, 0.0523, 0.0670], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:56:42,634 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.332e+02 2.138e+02 2.458e+02 2.946e+02 5.794e+02, threshold=4.917e+02, percent-clipped=2.0 2023-05-01 05:56:47,608 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-01 05:57:00,581 INFO [train.py:904] (4/8) Epoch 21, batch 3300, loss[loss=0.1681, simple_loss=0.2564, pruned_loss=0.0399, over 16466.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2572, pruned_loss=0.04195, over 3323186.26 frames. ], batch size: 68, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:57:13,937 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=206311.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:58:09,885 INFO [train.py:904] (4/8) Epoch 21, batch 3350, loss[loss=0.1849, simple_loss=0.2695, pruned_loss=0.05015, over 15477.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2589, pruned_loss=0.04261, over 3314875.04 frames. ], batch size: 190, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:58:20,156 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=206359.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:58:30,642 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0527, 4.5340, 4.5153, 3.2475, 3.7770, 4.4911, 3.8986, 2.5587], device='cuda:4'), covar=tensor([0.0450, 0.0060, 0.0040, 0.0337, 0.0123, 0.0090, 0.0092, 0.0461], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0085, 0.0083, 0.0134, 0.0099, 0.0110, 0.0095, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 05:59:00,090 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.620e+02 2.208e+02 2.654e+02 3.136e+02 4.728e+02, threshold=5.309e+02, percent-clipped=0.0 2023-05-01 05:59:06,222 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 05:59:12,023 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6052, 3.7090, 2.9148, 2.2430, 2.4126, 2.4136, 3.8312, 3.3070], device='cuda:4'), covar=tensor([0.2840, 0.0681, 0.1685, 0.2934, 0.2609, 0.2002, 0.0539, 0.1515], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0272, 0.0306, 0.0311, 0.0299, 0.0260, 0.0296, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 05:59:16,447 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=206400.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 05:59:18,363 INFO [train.py:904] (4/8) Epoch 21, batch 3400, loss[loss=0.1657, simple_loss=0.2635, pruned_loss=0.03398, over 17146.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2574, pruned_loss=0.04181, over 3324262.15 frames. ], batch size: 48, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 05:59:32,318 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=206412.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:00:21,655 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=206448.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:00:26,881 INFO [train.py:904] (4/8) Epoch 21, batch 3450, loss[loss=0.1683, simple_loss=0.2561, pruned_loss=0.04025, over 17142.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2568, pruned_loss=0.04138, over 3321717.88 frames. ], batch size: 49, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:01:17,174 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.532e+02 2.080e+02 2.437e+02 2.944e+02 7.361e+02, threshold=4.875e+02, percent-clipped=3.0 2023-05-01 06:01:36,849 INFO [train.py:904] (4/8) Epoch 21, batch 3500, loss[loss=0.1631, simple_loss=0.251, pruned_loss=0.03755, over 17013.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2553, pruned_loss=0.04112, over 3325557.21 frames. ], batch size: 41, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:02:23,339 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2869, 2.2669, 2.4505, 4.1659, 2.2766, 2.6415, 2.3352, 2.4036], device='cuda:4'), covar=tensor([0.1465, 0.3789, 0.2828, 0.0550, 0.3890, 0.2638, 0.3936, 0.3086], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0451, 0.0370, 0.0334, 0.0439, 0.0519, 0.0420, 0.0528], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:02:29,431 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.70 vs. limit=5.0 2023-05-01 06:02:37,920 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=206546.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:02:44,994 INFO [train.py:904] (4/8) Epoch 21, batch 3550, loss[loss=0.1549, simple_loss=0.2497, pruned_loss=0.03, over 17258.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2538, pruned_loss=0.04065, over 3330228.90 frames. ], batch size: 52, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:02:47,115 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=206553.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:03:36,211 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1759, 5.7589, 5.8250, 5.5422, 5.5970, 6.1849, 5.7063, 5.3922], device='cuda:4'), covar=tensor([0.0884, 0.1969, 0.2479, 0.2025, 0.2923, 0.1000, 0.1566, 0.2645], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0607, 0.0667, 0.0503, 0.0669, 0.0697, 0.0521, 0.0668], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:03:37,043 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.453e+02 2.127e+02 2.401e+02 2.794e+02 4.680e+02, threshold=4.802e+02, percent-clipped=0.0 2023-05-01 06:03:44,189 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=206594.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:03:55,932 INFO [train.py:904] (4/8) Epoch 21, batch 3600, loss[loss=0.1492, simple_loss=0.2434, pruned_loss=0.02755, over 17111.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2527, pruned_loss=0.04019, over 3331180.49 frames. ], batch size: 48, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:04:27,545 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3579, 5.6609, 5.4342, 5.5124, 5.1038, 5.1086, 5.0853, 5.7868], device='cuda:4'), covar=tensor([0.1325, 0.0903, 0.1093, 0.1001, 0.0964, 0.0836, 0.1193, 0.0894], device='cuda:4'), in_proj_covar=tensor([0.0681, 0.0840, 0.0690, 0.0632, 0.0533, 0.0538, 0.0703, 0.0651], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:05:06,975 INFO [train.py:904] (4/8) Epoch 21, batch 3650, loss[loss=0.1556, simple_loss=0.236, pruned_loss=0.03766, over 15564.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.252, pruned_loss=0.04078, over 3319005.03 frames. ], batch size: 191, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:05:24,297 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.02 vs. limit=5.0 2023-05-01 06:05:59,785 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.607e+02 2.224e+02 2.566e+02 3.274e+02 6.534e+02, threshold=5.132e+02, percent-clipped=5.0 2023-05-01 06:06:18,969 INFO [train.py:904] (4/8) Epoch 21, batch 3700, loss[loss=0.1707, simple_loss=0.2451, pruned_loss=0.04814, over 16817.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2502, pruned_loss=0.04184, over 3298986.90 frames. ], batch size: 90, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:06:28,545 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0750, 2.1271, 2.3370, 3.7420, 2.2080, 2.4290, 2.2430, 2.2518], device='cuda:4'), covar=tensor([0.1520, 0.3764, 0.2853, 0.0627, 0.3833, 0.2653, 0.3815, 0.3226], device='cuda:4'), in_proj_covar=tensor([0.0403, 0.0450, 0.0369, 0.0333, 0.0437, 0.0518, 0.0419, 0.0526], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:06:34,594 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=206712.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:06:48,104 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-05-01 06:06:59,118 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8104, 4.0557, 3.0613, 2.4972, 2.8684, 2.6787, 4.4437, 3.7134], device='cuda:4'), covar=tensor([0.2671, 0.0559, 0.1673, 0.2348, 0.2339, 0.1761, 0.0370, 0.0947], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0271, 0.0306, 0.0311, 0.0299, 0.0260, 0.0296, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:07:32,563 INFO [train.py:904] (4/8) Epoch 21, batch 3750, loss[loss=0.1513, simple_loss=0.2294, pruned_loss=0.03655, over 16830.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2509, pruned_loss=0.04298, over 3290234.33 frames. ], batch size: 102, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:07:39,258 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5978, 4.7172, 4.8711, 4.7274, 4.6715, 5.2906, 4.8060, 4.4750], device='cuda:4'), covar=tensor([0.1437, 0.2051, 0.2233, 0.2018, 0.2804, 0.1063, 0.1670, 0.2633], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0605, 0.0661, 0.0501, 0.0666, 0.0694, 0.0519, 0.0666], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:07:45,681 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=206760.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:08:26,980 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.616e+02 2.267e+02 2.726e+02 3.357e+02 9.074e+02, threshold=5.452e+02, percent-clipped=3.0 2023-05-01 06:08:45,460 INFO [train.py:904] (4/8) Epoch 21, batch 3800, loss[loss=0.2, simple_loss=0.2838, pruned_loss=0.05808, over 12268.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2521, pruned_loss=0.04424, over 3279447.63 frames. ], batch size: 248, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:09:32,787 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=206834.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:09:58,314 INFO [train.py:904] (4/8) Epoch 21, batch 3850, loss[loss=0.1639, simple_loss=0.2374, pruned_loss=0.04521, over 16861.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2526, pruned_loss=0.04535, over 3288623.35 frames. ], batch size: 90, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:10:00,686 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=206853.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:10:11,178 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6964, 3.7620, 2.9148, 2.2636, 2.3668, 2.3787, 3.7901, 3.2413], device='cuda:4'), covar=tensor([0.2580, 0.0523, 0.1693, 0.2914, 0.2800, 0.2092, 0.0482, 0.1515], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0273, 0.0307, 0.0313, 0.0301, 0.0260, 0.0297, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:10:52,982 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.673e+02 2.267e+02 2.812e+02 3.321e+02 1.032e+03, threshold=5.623e+02, percent-clipped=3.0 2023-05-01 06:11:00,928 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=206895.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 06:11:09,938 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=206901.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:11:10,164 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9075, 3.0285, 2.6670, 4.5332, 3.7485, 4.1343, 1.8095, 3.1040], device='cuda:4'), covar=tensor([0.1266, 0.0707, 0.1156, 0.0181, 0.0260, 0.0407, 0.1507, 0.0812], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0175, 0.0194, 0.0192, 0.0207, 0.0217, 0.0201, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:11:10,713 INFO [train.py:904] (4/8) Epoch 21, batch 3900, loss[loss=0.1701, simple_loss=0.2441, pruned_loss=0.04809, over 16537.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2529, pruned_loss=0.04622, over 3285591.93 frames. ], batch size: 68, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:12:24,775 INFO [train.py:904] (4/8) Epoch 21, batch 3950, loss[loss=0.2061, simple_loss=0.2822, pruned_loss=0.065, over 15554.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2533, pruned_loss=0.04688, over 3278902.87 frames. ], batch size: 191, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:13:16,330 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.684e+02 2.243e+02 2.750e+02 3.370e+02 5.125e+02, threshold=5.500e+02, percent-clipped=0.0 2023-05-01 06:13:35,019 INFO [train.py:904] (4/8) Epoch 21, batch 4000, loss[loss=0.1964, simple_loss=0.2824, pruned_loss=0.05519, over 12431.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2538, pruned_loss=0.04787, over 3281416.30 frames. ], batch size: 247, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:13:42,708 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1655, 3.1664, 1.9391, 3.3715, 2.3808, 3.5220, 2.1571, 2.5849], device='cuda:4'), covar=tensor([0.0306, 0.0381, 0.1743, 0.0199, 0.0971, 0.0436, 0.1608, 0.0787], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0179, 0.0195, 0.0166, 0.0179, 0.0219, 0.0202, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:13:42,948 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 06:13:53,407 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8719, 2.0026, 2.5527, 2.8583, 2.8368, 3.0759, 2.0777, 3.0851], device='cuda:4'), covar=tensor([0.0227, 0.0498, 0.0312, 0.0291, 0.0302, 0.0196, 0.0544, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0195, 0.0181, 0.0186, 0.0199, 0.0156, 0.0199, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:14:03,357 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6652, 2.4842, 2.3301, 3.4507, 2.5776, 3.6985, 1.5747, 2.6797], device='cuda:4'), covar=tensor([0.1450, 0.0879, 0.1309, 0.0225, 0.0219, 0.0385, 0.1715, 0.0949], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0175, 0.0193, 0.0192, 0.0206, 0.0216, 0.0201, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:14:12,182 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4625, 3.5285, 2.4406, 2.2383, 2.2343, 2.1497, 3.4552, 3.0154], device='cuda:4'), covar=tensor([0.3041, 0.0724, 0.2299, 0.2868, 0.3051, 0.2442, 0.0712, 0.1350], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0273, 0.0307, 0.0313, 0.0301, 0.0261, 0.0297, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:14:16,910 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1987, 5.0757, 5.2396, 5.3986, 5.5690, 4.9709, 5.5408, 5.6110], device='cuda:4'), covar=tensor([0.1493, 0.1026, 0.1380, 0.0572, 0.0377, 0.0699, 0.0463, 0.0378], device='cuda:4'), in_proj_covar=tensor([0.0656, 0.0810, 0.0948, 0.0830, 0.0617, 0.0646, 0.0666, 0.0771], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:14:36,360 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.89 vs. limit=5.0 2023-05-01 06:14:45,530 INFO [train.py:904] (4/8) Epoch 21, batch 4050, loss[loss=0.1783, simple_loss=0.2654, pruned_loss=0.04562, over 17104.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2547, pruned_loss=0.0472, over 3286088.28 frames. ], batch size: 47, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:14:53,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2151, 5.8825, 6.0543, 5.7153, 5.8303, 6.3112, 5.8668, 5.6135], device='cuda:4'), covar=tensor([0.0851, 0.1642, 0.1671, 0.1650, 0.2113, 0.0847, 0.1318, 0.1934], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0602, 0.0657, 0.0498, 0.0662, 0.0688, 0.0515, 0.0663], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:15:36,733 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0495, 5.5134, 5.6924, 5.4490, 5.4799, 6.0237, 5.5284, 5.2347], device='cuda:4'), covar=tensor([0.0920, 0.1625, 0.1624, 0.1665, 0.2118, 0.0790, 0.1318, 0.2071], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0601, 0.0655, 0.0496, 0.0661, 0.0687, 0.0514, 0.0662], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:15:37,537 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.811e+02 2.022e+02 2.362e+02 4.277e+02, threshold=4.045e+02, percent-clipped=0.0 2023-05-01 06:15:55,999 INFO [train.py:904] (4/8) Epoch 21, batch 4100, loss[loss=0.1757, simple_loss=0.2574, pruned_loss=0.04704, over 12060.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.256, pruned_loss=0.0465, over 3285021.41 frames. ], batch size: 248, lr: 3.23e-03, grad_scale: 8.0 2023-05-01 06:16:33,216 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9531, 4.1093, 4.3027, 4.2886, 4.3015, 4.0601, 4.0502, 4.0157], device='cuda:4'), covar=tensor([0.0340, 0.0474, 0.0398, 0.0382, 0.0464, 0.0383, 0.0794, 0.0509], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0460, 0.0445, 0.0413, 0.0491, 0.0467, 0.0556, 0.0375], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 06:16:39,968 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2117, 2.5068, 2.0536, 2.1404, 2.7559, 2.3785, 2.8564, 2.8979], device='cuda:4'), covar=tensor([0.0116, 0.0372, 0.0514, 0.0455, 0.0247, 0.0352, 0.0198, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0237, 0.0226, 0.0227, 0.0238, 0.0236, 0.0241, 0.0235], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:16:48,144 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7419, 2.2862, 1.8596, 1.9954, 2.5612, 2.2366, 2.5635, 2.7065], device='cuda:4'), covar=tensor([0.0177, 0.0391, 0.0525, 0.0475, 0.0254, 0.0352, 0.0185, 0.0260], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0237, 0.0226, 0.0227, 0.0238, 0.0236, 0.0240, 0.0235], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:17:10,035 INFO [train.py:904] (4/8) Epoch 21, batch 4150, loss[loss=0.1813, simple_loss=0.2819, pruned_loss=0.04038, over 16861.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2629, pruned_loss=0.0484, over 3256971.32 frames. ], batch size: 96, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:17:45,666 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=207176.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:18:04,517 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.144e+02 2.646e+02 3.287e+02 6.191e+02, threshold=5.292e+02, percent-clipped=8.0 2023-05-01 06:18:06,781 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=207190.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 06:18:16,948 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-01 06:18:24,668 INFO [train.py:904] (4/8) Epoch 21, batch 4200, loss[loss=0.1958, simple_loss=0.2869, pruned_loss=0.05233, over 16759.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2699, pruned_loss=0.05017, over 3219502.15 frames. ], batch size: 39, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:19:18,488 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=207237.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:19:40,047 INFO [train.py:904] (4/8) Epoch 21, batch 4250, loss[loss=0.171, simple_loss=0.2598, pruned_loss=0.04106, over 15429.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2725, pruned_loss=0.04986, over 3189801.17 frames. ], batch size: 190, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:20:35,819 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 2.204e+02 2.470e+02 3.047e+02 4.735e+02, threshold=4.941e+02, percent-clipped=0.0 2023-05-01 06:20:55,796 INFO [train.py:904] (4/8) Epoch 21, batch 4300, loss[loss=0.1896, simple_loss=0.2877, pruned_loss=0.04574, over 16404.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2737, pruned_loss=0.0489, over 3192217.41 frames. ], batch size: 146, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:22:07,510 INFO [train.py:904] (4/8) Epoch 21, batch 4350, loss[loss=0.1751, simple_loss=0.2693, pruned_loss=0.04046, over 16867.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2768, pruned_loss=0.04994, over 3180682.58 frames. ], batch size: 96, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:22:21,848 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1808, 3.2932, 1.9640, 3.6119, 2.4627, 3.6193, 2.2846, 2.6555], device='cuda:4'), covar=tensor([0.0337, 0.0368, 0.1688, 0.0163, 0.0854, 0.0563, 0.1289, 0.0791], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0177, 0.0194, 0.0163, 0.0177, 0.0217, 0.0200, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:22:34,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5587, 3.2703, 3.7902, 1.7022, 4.0085, 4.0444, 2.9282, 2.7927], device='cuda:4'), covar=tensor([0.0805, 0.0302, 0.0199, 0.1312, 0.0067, 0.0126, 0.0489, 0.0503], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0097, 0.0139, 0.0080, 0.0125, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:23:02,764 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.651e+02 2.219e+02 2.598e+02 3.036e+02 7.307e+02, threshold=5.195e+02, percent-clipped=1.0 2023-05-01 06:23:22,052 INFO [train.py:904] (4/8) Epoch 21, batch 4400, loss[loss=0.1819, simple_loss=0.2765, pruned_loss=0.04368, over 16679.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2785, pruned_loss=0.05064, over 3197636.67 frames. ], batch size: 76, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:23:30,547 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1984, 5.2764, 5.5771, 5.5874, 5.6250, 5.2926, 5.1738, 4.8846], device='cuda:4'), covar=tensor([0.0247, 0.0355, 0.0286, 0.0337, 0.0412, 0.0277, 0.0892, 0.0449], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0454, 0.0440, 0.0408, 0.0484, 0.0461, 0.0549, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 06:24:02,539 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2745, 1.5665, 1.9798, 2.2156, 2.2834, 2.5567, 1.6563, 2.3611], device='cuda:4'), covar=tensor([0.0225, 0.0527, 0.0312, 0.0330, 0.0312, 0.0192, 0.0574, 0.0156], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0193, 0.0178, 0.0185, 0.0196, 0.0153, 0.0197, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:24:37,195 INFO [train.py:904] (4/8) Epoch 21, batch 4450, loss[loss=0.2088, simple_loss=0.3012, pruned_loss=0.05816, over 17012.00 frames. ], tot_loss[loss=0.1928, simple_loss=0.2819, pruned_loss=0.05188, over 3205817.05 frames. ], batch size: 53, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:25:31,509 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 2.026e+02 2.353e+02 2.743e+02 3.990e+02, threshold=4.706e+02, percent-clipped=0.0 2023-05-01 06:25:33,767 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=207490.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 06:25:35,167 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5229, 2.5637, 2.4526, 3.5755, 2.7995, 3.8318, 1.5164, 2.7657], device='cuda:4'), covar=tensor([0.1375, 0.0832, 0.1174, 0.0197, 0.0265, 0.0343, 0.1693, 0.0824], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0175, 0.0194, 0.0191, 0.0208, 0.0215, 0.0201, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:25:50,489 INFO [train.py:904] (4/8) Epoch 21, batch 4500, loss[loss=0.1813, simple_loss=0.2653, pruned_loss=0.04864, over 17062.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2823, pruned_loss=0.05254, over 3206688.77 frames. ], batch size: 53, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:26:35,638 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=207532.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:26:43,761 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=207538.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:27:03,811 INFO [train.py:904] (4/8) Epoch 21, batch 4550, loss[loss=0.2191, simple_loss=0.3031, pruned_loss=0.0676, over 16683.00 frames. ], tot_loss[loss=0.195, simple_loss=0.283, pruned_loss=0.05346, over 3223221.58 frames. ], batch size: 57, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:27:05,692 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5659, 3.6814, 2.1729, 4.2904, 2.7694, 4.1762, 2.3725, 2.8585], device='cuda:4'), covar=tensor([0.0316, 0.0339, 0.1712, 0.0145, 0.0938, 0.0510, 0.1546, 0.0861], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0177, 0.0194, 0.0163, 0.0177, 0.0216, 0.0200, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:27:41,042 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 06:27:57,547 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.324e+02 1.874e+02 2.187e+02 2.648e+02 6.639e+02, threshold=4.374e+02, percent-clipped=1.0 2023-05-01 06:28:16,259 INFO [train.py:904] (4/8) Epoch 21, batch 4600, loss[loss=0.1874, simple_loss=0.2752, pruned_loss=0.04984, over 16699.00 frames. ], tot_loss[loss=0.1963, simple_loss=0.2844, pruned_loss=0.05414, over 3224759.85 frames. ], batch size: 57, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:29:02,822 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.01 vs. limit=5.0 2023-05-01 06:29:07,197 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1746, 4.3062, 4.4433, 4.1872, 4.2813, 4.8163, 4.3821, 4.0804], device='cuda:4'), covar=tensor([0.1740, 0.1952, 0.2113, 0.2144, 0.2664, 0.1044, 0.1422, 0.2497], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0584, 0.0637, 0.0486, 0.0646, 0.0672, 0.0498, 0.0648], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:29:29,142 INFO [train.py:904] (4/8) Epoch 21, batch 4650, loss[loss=0.179, simple_loss=0.2647, pruned_loss=0.04662, over 16474.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2835, pruned_loss=0.05436, over 3218522.80 frames. ], batch size: 68, lr: 3.22e-03, grad_scale: 16.0 2023-05-01 06:30:23,480 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 1.925e+02 2.221e+02 2.572e+02 4.933e+02, threshold=4.441e+02, percent-clipped=1.0 2023-05-01 06:30:42,424 INFO [train.py:904] (4/8) Epoch 21, batch 4700, loss[loss=0.1746, simple_loss=0.2669, pruned_loss=0.04115, over 16402.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2808, pruned_loss=0.05319, over 3219583.91 frames. ], batch size: 146, lr: 3.22e-03, grad_scale: 16.0 2023-05-01 06:31:56,744 INFO [train.py:904] (4/8) Epoch 21, batch 4750, loss[loss=0.1533, simple_loss=0.2501, pruned_loss=0.02824, over 16891.00 frames. ], tot_loss[loss=0.1898, simple_loss=0.2769, pruned_loss=0.05137, over 3205672.16 frames. ], batch size: 96, lr: 3.22e-03, grad_scale: 16.0 2023-05-01 06:32:03,209 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.51 vs. limit=5.0 2023-05-01 06:32:50,130 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.318e+02 1.797e+02 2.146e+02 2.444e+02 4.649e+02, threshold=4.292e+02, percent-clipped=1.0 2023-05-01 06:33:00,146 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0664, 5.0682, 4.9124, 4.2054, 5.0088, 1.8461, 4.6913, 4.6616], device='cuda:4'), covar=tensor([0.0095, 0.0099, 0.0160, 0.0480, 0.0102, 0.2838, 0.0144, 0.0231], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0155, 0.0197, 0.0178, 0.0175, 0.0207, 0.0187, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:33:11,157 INFO [train.py:904] (4/8) Epoch 21, batch 4800, loss[loss=0.1593, simple_loss=0.2502, pruned_loss=0.03421, over 16699.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2735, pruned_loss=0.04974, over 3189259.91 frames. ], batch size: 76, lr: 3.22e-03, grad_scale: 16.0 2023-05-01 06:33:14,431 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 06:33:55,500 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=207832.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:34:24,552 INFO [train.py:904] (4/8) Epoch 21, batch 4850, loss[loss=0.182, simple_loss=0.2729, pruned_loss=0.04554, over 16754.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2742, pruned_loss=0.04924, over 3168434.26 frames. ], batch size: 124, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:34:32,031 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 06:35:08,160 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=207880.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:35:22,070 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 1.907e+02 2.222e+02 2.565e+02 4.310e+02, threshold=4.444e+02, percent-clipped=1.0 2023-05-01 06:35:40,341 INFO [train.py:904] (4/8) Epoch 21, batch 4900, loss[loss=0.1568, simple_loss=0.2458, pruned_loss=0.0339, over 16854.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2738, pruned_loss=0.04794, over 3166579.85 frames. ], batch size: 42, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:36:52,780 INFO [train.py:904] (4/8) Epoch 21, batch 4950, loss[loss=0.2001, simple_loss=0.2828, pruned_loss=0.05875, over 12226.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2731, pruned_loss=0.04697, over 3169200.12 frames. ], batch size: 246, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:37:05,247 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5090, 3.7131, 2.7442, 2.2181, 2.4429, 2.4706, 3.9409, 3.3238], device='cuda:4'), covar=tensor([0.3145, 0.0645, 0.1861, 0.2715, 0.2630, 0.1959, 0.0502, 0.1180], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0269, 0.0303, 0.0310, 0.0296, 0.0256, 0.0295, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 06:37:22,542 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5530, 4.6716, 4.4439, 4.1417, 4.1511, 4.5439, 4.3961, 4.2567], device='cuda:4'), covar=tensor([0.0622, 0.0445, 0.0291, 0.0289, 0.0883, 0.0450, 0.0485, 0.0584], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0413, 0.0335, 0.0330, 0.0344, 0.0384, 0.0229, 0.0402], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:37:47,818 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.452e+02 1.964e+02 2.322e+02 2.960e+02 4.980e+02, threshold=4.644e+02, percent-clipped=1.0 2023-05-01 06:37:57,047 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 06:38:08,406 INFO [train.py:904] (4/8) Epoch 21, batch 5000, loss[loss=0.1807, simple_loss=0.2795, pruned_loss=0.04089, over 16793.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2744, pruned_loss=0.04657, over 3188914.23 frames. ], batch size: 83, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:38:52,047 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=208032.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:39:21,547 INFO [train.py:904] (4/8) Epoch 21, batch 5050, loss[loss=0.2159, simple_loss=0.3028, pruned_loss=0.06448, over 15352.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2745, pruned_loss=0.0464, over 3189893.24 frames. ], batch size: 190, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:40:18,551 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.548e+02 1.998e+02 2.352e+02 2.773e+02 5.877e+02, threshold=4.704e+02, percent-clipped=4.0 2023-05-01 06:40:23,308 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=208093.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 06:40:35,285 INFO [train.py:904] (4/8) Epoch 21, batch 5100, loss[loss=0.1639, simple_loss=0.2481, pruned_loss=0.03987, over 16638.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.273, pruned_loss=0.04622, over 3170809.07 frames. ], batch size: 57, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:40:52,436 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7281, 2.6551, 2.5287, 4.1396, 3.0517, 3.9605, 1.4949, 2.9441], device='cuda:4'), covar=tensor([0.1289, 0.0761, 0.1177, 0.0141, 0.0227, 0.0364, 0.1606, 0.0798], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0175, 0.0194, 0.0189, 0.0208, 0.0214, 0.0201, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:40:59,938 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1715, 4.2562, 4.0682, 3.7764, 3.7382, 4.1257, 3.8929, 3.9138], device='cuda:4'), covar=tensor([0.0621, 0.0459, 0.0326, 0.0305, 0.0874, 0.0501, 0.0732, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0413, 0.0335, 0.0329, 0.0344, 0.0385, 0.0229, 0.0402], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:41:48,631 INFO [train.py:904] (4/8) Epoch 21, batch 5150, loss[loss=0.1944, simple_loss=0.2798, pruned_loss=0.05448, over 12038.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2733, pruned_loss=0.04544, over 3165920.43 frames. ], batch size: 246, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:42:27,983 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8077, 3.2238, 3.3166, 1.9631, 2.7512, 2.1957, 3.3189, 3.4540], device='cuda:4'), covar=tensor([0.0264, 0.0710, 0.0613, 0.1961, 0.0875, 0.0949, 0.0623, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0163, 0.0167, 0.0152, 0.0145, 0.0130, 0.0144, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:42:43,673 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.371e+02 1.920e+02 2.264e+02 2.595e+02 3.716e+02, threshold=4.527e+02, percent-clipped=0.0 2023-05-01 06:43:01,061 INFO [train.py:904] (4/8) Epoch 21, batch 5200, loss[loss=0.1638, simple_loss=0.2566, pruned_loss=0.03549, over 16247.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2717, pruned_loss=0.04486, over 3174034.21 frames. ], batch size: 165, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:43:49,632 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9921, 5.0354, 4.8704, 4.5038, 4.5356, 4.9151, 4.8249, 4.6304], device='cuda:4'), covar=tensor([0.0607, 0.0511, 0.0281, 0.0281, 0.0939, 0.0596, 0.0339, 0.0609], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0414, 0.0336, 0.0331, 0.0345, 0.0386, 0.0229, 0.0402], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:44:11,868 INFO [train.py:904] (4/8) Epoch 21, batch 5250, loss[loss=0.1833, simple_loss=0.2663, pruned_loss=0.05015, over 12187.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2693, pruned_loss=0.04447, over 3182632.32 frames. ], batch size: 248, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:44:31,790 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=208265.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:45:03,598 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 06:45:07,260 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 1.864e+02 2.120e+02 2.516e+02 5.355e+02, threshold=4.239e+02, percent-clipped=2.0 2023-05-01 06:45:25,337 INFO [train.py:904] (4/8) Epoch 21, batch 5300, loss[loss=0.167, simple_loss=0.2485, pruned_loss=0.04278, over 16693.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2656, pruned_loss=0.04303, over 3194045.82 frames. ], batch size: 62, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:46:00,439 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=208326.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:46:37,966 INFO [train.py:904] (4/8) Epoch 21, batch 5350, loss[loss=0.1711, simple_loss=0.2673, pruned_loss=0.03741, over 16492.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2641, pruned_loss=0.04265, over 3180846.74 frames. ], batch size: 75, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:46:58,530 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-01 06:47:21,653 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-01 06:47:32,334 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=208388.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 06:47:34,961 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.374e+02 1.963e+02 2.291e+02 2.599e+02 4.161e+02, threshold=4.581e+02, percent-clipped=0.0 2023-05-01 06:47:53,658 INFO [train.py:904] (4/8) Epoch 21, batch 5400, loss[loss=0.191, simple_loss=0.2824, pruned_loss=0.04977, over 16940.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2664, pruned_loss=0.04294, over 3180322.64 frames. ], batch size: 109, lr: 3.22e-03, grad_scale: 8.0 2023-05-01 06:49:10,975 INFO [train.py:904] (4/8) Epoch 21, batch 5450, loss[loss=0.2066, simple_loss=0.2937, pruned_loss=0.05977, over 16377.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2691, pruned_loss=0.04434, over 3178995.12 frames. ], batch size: 35, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 06:50:09,156 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.336e+02 2.337e+02 2.736e+02 3.594e+02 9.529e+02, threshold=5.472e+02, percent-clipped=14.0 2023-05-01 06:50:28,299 INFO [train.py:904] (4/8) Epoch 21, batch 5500, loss[loss=0.2029, simple_loss=0.3008, pruned_loss=0.05249, over 16840.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2767, pruned_loss=0.04877, over 3150851.34 frames. ], batch size: 102, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 06:51:47,653 INFO [train.py:904] (4/8) Epoch 21, batch 5550, loss[loss=0.2826, simple_loss=0.3402, pruned_loss=0.1124, over 11329.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2832, pruned_loss=0.05346, over 3126876.09 frames. ], batch size: 250, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 06:52:05,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3670, 3.3566, 3.3727, 3.4689, 3.4875, 3.2681, 3.4745, 3.5526], device='cuda:4'), covar=tensor([0.1172, 0.0878, 0.1069, 0.0603, 0.0704, 0.2163, 0.1038, 0.0789], device='cuda:4'), in_proj_covar=tensor([0.0626, 0.0771, 0.0907, 0.0789, 0.0593, 0.0624, 0.0636, 0.0740], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:52:10,282 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-05-01 06:52:27,844 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=208576.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:52:49,284 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.296e+02 3.115e+02 3.921e+02 4.748e+02 1.231e+03, threshold=7.841e+02, percent-clipped=12.0 2023-05-01 06:53:07,711 INFO [train.py:904] (4/8) Epoch 21, batch 5600, loss[loss=0.2456, simple_loss=0.3239, pruned_loss=0.08362, over 15369.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2885, pruned_loss=0.05813, over 3092871.94 frames. ], batch size: 190, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 06:53:39,435 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=208621.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:54:07,071 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=208637.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:54:29,857 INFO [train.py:904] (4/8) Epoch 21, batch 5650, loss[loss=0.2654, simple_loss=0.3336, pruned_loss=0.09866, over 11301.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2937, pruned_loss=0.06256, over 3052749.20 frames. ], batch size: 246, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 06:54:42,794 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6579, 1.7571, 2.2863, 2.5828, 2.6162, 2.8963, 1.9464, 2.8321], device='cuda:4'), covar=tensor([0.0202, 0.0505, 0.0282, 0.0316, 0.0266, 0.0176, 0.0508, 0.0123], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0192, 0.0178, 0.0184, 0.0195, 0.0151, 0.0195, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:54:48,300 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0246, 2.4445, 2.5880, 1.9666, 2.6670, 2.7698, 2.4169, 2.3605], device='cuda:4'), covar=tensor([0.0674, 0.0243, 0.0221, 0.0870, 0.0115, 0.0289, 0.0446, 0.0434], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0096, 0.0138, 0.0080, 0.0124, 0.0128, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:54:55,530 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0973, 2.1640, 2.2671, 3.6713, 2.0331, 2.4826, 2.2907, 2.3464], device='cuda:4'), covar=tensor([0.1319, 0.3226, 0.2705, 0.0564, 0.3932, 0.2340, 0.3161, 0.3160], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0442, 0.0362, 0.0324, 0.0430, 0.0509, 0.0412, 0.0516], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 06:55:28,191 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=208688.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:55:30,664 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.148e+02 3.160e+02 3.845e+02 4.644e+02 8.553e+02, threshold=7.690e+02, percent-clipped=3.0 2023-05-01 06:55:50,856 INFO [train.py:904] (4/8) Epoch 21, batch 5700, loss[loss=0.1947, simple_loss=0.2977, pruned_loss=0.04586, over 16723.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.2958, pruned_loss=0.06442, over 3044381.89 frames. ], batch size: 89, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 06:55:52,982 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-05-01 06:56:45,711 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=208736.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 06:57:10,969 INFO [train.py:904] (4/8) Epoch 21, batch 5750, loss[loss=0.2181, simple_loss=0.3087, pruned_loss=0.06382, over 16393.00 frames. ], tot_loss[loss=0.2156, simple_loss=0.2984, pruned_loss=0.06643, over 2993817.23 frames. ], batch size: 68, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 06:57:18,235 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5416, 3.5868, 2.0483, 4.0464, 2.6806, 4.0042, 2.2598, 2.8076], device='cuda:4'), covar=tensor([0.0287, 0.0394, 0.1816, 0.0186, 0.0864, 0.0542, 0.1559, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0192, 0.0160, 0.0175, 0.0214, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:57:49,368 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9195, 3.9924, 2.3478, 4.5135, 2.9970, 4.3972, 2.4646, 3.0860], device='cuda:4'), covar=tensor([0.0239, 0.0328, 0.1639, 0.0206, 0.0790, 0.0539, 0.1581, 0.0769], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0193, 0.0160, 0.0176, 0.0214, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 06:58:13,920 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.275e+02 2.983e+02 3.733e+02 4.569e+02 7.128e+02, threshold=7.466e+02, percent-clipped=0.0 2023-05-01 06:58:33,852 INFO [train.py:904] (4/8) Epoch 21, batch 5800, loss[loss=0.1898, simple_loss=0.2819, pruned_loss=0.04879, over 16924.00 frames. ], tot_loss[loss=0.2137, simple_loss=0.2974, pruned_loss=0.06504, over 3000711.74 frames. ], batch size: 109, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 06:59:52,593 INFO [train.py:904] (4/8) Epoch 21, batch 5850, loss[loss=0.1903, simple_loss=0.2852, pruned_loss=0.04774, over 16773.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.2945, pruned_loss=0.06256, over 3018175.59 frames. ], batch size: 124, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:00:53,525 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.655e+02 2.740e+02 3.068e+02 3.692e+02 6.527e+02, threshold=6.136e+02, percent-clipped=0.0 2023-05-01 07:01:12,711 INFO [train.py:904] (4/8) Epoch 21, batch 5900, loss[loss=0.2375, simple_loss=0.3001, pruned_loss=0.0875, over 11610.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2946, pruned_loss=0.06213, over 3025337.16 frames. ], batch size: 246, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:01:48,389 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=208921.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:02:04,493 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=208932.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:02:21,470 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.61 vs. limit=2.0 2023-05-01 07:02:36,224 INFO [train.py:904] (4/8) Epoch 21, batch 5950, loss[loss=0.205, simple_loss=0.2947, pruned_loss=0.05762, over 17055.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2954, pruned_loss=0.06042, over 3054052.97 frames. ], batch size: 55, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:03:02,963 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=208969.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:03:36,545 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.874e+02 2.651e+02 3.328e+02 3.727e+02 9.744e+02, threshold=6.656e+02, percent-clipped=4.0 2023-05-01 07:03:56,984 INFO [train.py:904] (4/8) Epoch 21, batch 6000, loss[loss=0.2227, simple_loss=0.3061, pruned_loss=0.06968, over 16734.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2947, pruned_loss=0.05997, over 3070613.33 frames. ], batch size: 124, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:03:56,985 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 07:04:08,270 INFO [train.py:938] (4/8) Epoch 21, validation: loss=0.1512, simple_loss=0.2639, pruned_loss=0.01924, over 944034.00 frames. 2023-05-01 07:04:08,271 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 07:05:25,680 INFO [train.py:904] (4/8) Epoch 21, batch 6050, loss[loss=0.2028, simple_loss=0.2935, pruned_loss=0.05602, over 16686.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.2928, pruned_loss=0.05942, over 3076980.82 frames. ], batch size: 62, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:05:26,951 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-05-01 07:06:26,989 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.899e+02 2.646e+02 3.170e+02 4.163e+02 9.783e+02, threshold=6.340e+02, percent-clipped=1.0 2023-05-01 07:06:37,553 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4523, 4.5604, 4.3458, 4.0375, 4.0434, 4.4546, 4.2691, 4.1353], device='cuda:4'), covar=tensor([0.0675, 0.0515, 0.0306, 0.0339, 0.0858, 0.0520, 0.0558, 0.0674], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0415, 0.0335, 0.0332, 0.0344, 0.0386, 0.0231, 0.0403], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:06:45,918 INFO [train.py:904] (4/8) Epoch 21, batch 6100, loss[loss=0.2025, simple_loss=0.2957, pruned_loss=0.05467, over 16296.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2924, pruned_loss=0.05817, over 3095023.93 frames. ], batch size: 165, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:07:40,336 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5840, 4.8032, 5.0069, 4.7867, 4.9027, 5.4130, 4.8883, 4.6401], device='cuda:4'), covar=tensor([0.1321, 0.2050, 0.2169, 0.2028, 0.2419, 0.0953, 0.1594, 0.2440], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0584, 0.0639, 0.0487, 0.0645, 0.0673, 0.0503, 0.0651], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 07:08:04,715 INFO [train.py:904] (4/8) Epoch 21, batch 6150, loss[loss=0.1747, simple_loss=0.2619, pruned_loss=0.04378, over 16447.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2908, pruned_loss=0.05791, over 3103301.14 frames. ], batch size: 68, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:08:42,466 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 07:09:02,671 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-05-01 07:09:04,297 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.750e+02 2.685e+02 3.365e+02 4.144e+02 7.779e+02, threshold=6.731e+02, percent-clipped=2.0 2023-05-01 07:09:16,794 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0159, 3.9986, 3.9967, 3.2279, 3.9894, 1.8161, 3.7781, 3.5912], device='cuda:4'), covar=tensor([0.0147, 0.0115, 0.0172, 0.0318, 0.0105, 0.2657, 0.0151, 0.0243], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0150, 0.0192, 0.0173, 0.0170, 0.0200, 0.0181, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:09:23,301 INFO [train.py:904] (4/8) Epoch 21, batch 6200, loss[loss=0.1821, simple_loss=0.2748, pruned_loss=0.04471, over 16836.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2888, pruned_loss=0.05743, over 3117322.82 frames. ], batch size: 102, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:09:23,864 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=209202.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:10:10,049 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=209232.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:10:11,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8693, 4.7102, 4.8819, 5.0619, 5.2455, 4.6652, 5.2767, 5.2527], device='cuda:4'), covar=tensor([0.1961, 0.1256, 0.1722, 0.0777, 0.0630, 0.0879, 0.0636, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0623, 0.0766, 0.0896, 0.0780, 0.0586, 0.0616, 0.0635, 0.0736], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:10:39,910 INFO [train.py:904] (4/8) Epoch 21, batch 6250, loss[loss=0.2058, simple_loss=0.2827, pruned_loss=0.06442, over 11793.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2876, pruned_loss=0.05683, over 3120272.28 frames. ], batch size: 246, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:10:46,242 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=209255.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:10:57,587 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=209263.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:11:22,154 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=209280.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:11:35,896 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.929e+02 2.681e+02 3.167e+02 3.896e+02 8.805e+02, threshold=6.333e+02, percent-clipped=2.0 2023-05-01 07:11:50,452 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=209298.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:11:54,987 INFO [train.py:904] (4/8) Epoch 21, batch 6300, loss[loss=0.1965, simple_loss=0.2886, pruned_loss=0.05216, over 16864.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2877, pruned_loss=0.05655, over 3128180.23 frames. ], batch size: 96, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:12:07,943 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6015, 4.6617, 4.4716, 4.1653, 4.1786, 4.5766, 4.3432, 4.2637], device='cuda:4'), covar=tensor([0.0668, 0.0599, 0.0302, 0.0336, 0.0839, 0.0528, 0.0647, 0.0739], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0412, 0.0333, 0.0330, 0.0341, 0.0382, 0.0229, 0.0400], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:12:08,103 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1217, 2.2039, 2.0939, 3.8292, 2.1182, 2.5731, 2.3070, 2.3572], device='cuda:4'), covar=tensor([0.1429, 0.3554, 0.3134, 0.0527, 0.4029, 0.2422, 0.3455, 0.3281], device='cuda:4'), in_proj_covar=tensor([0.0399, 0.0446, 0.0365, 0.0326, 0.0435, 0.0512, 0.0414, 0.0520], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:12:17,952 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=209316.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:13:13,583 INFO [train.py:904] (4/8) Epoch 21, batch 6350, loss[loss=0.1672, simple_loss=0.2593, pruned_loss=0.03761, over 17254.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2876, pruned_loss=0.05678, over 3136103.11 frames. ], batch size: 52, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:13:24,864 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=209359.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:14:13,986 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.652e+02 3.041e+02 3.551e+02 4.346e+02 9.300e+02, threshold=7.102e+02, percent-clipped=4.0 2023-05-01 07:14:31,817 INFO [train.py:904] (4/8) Epoch 21, batch 6400, loss[loss=0.2041, simple_loss=0.2847, pruned_loss=0.06178, over 16647.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.288, pruned_loss=0.05818, over 3125681.05 frames. ], batch size: 134, lr: 3.21e-03, grad_scale: 8.0 2023-05-01 07:15:48,009 INFO [train.py:904] (4/8) Epoch 21, batch 6450, loss[loss=0.2173, simple_loss=0.3118, pruned_loss=0.06145, over 16729.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.2888, pruned_loss=0.05824, over 3097373.51 frames. ], batch size: 76, lr: 3.21e-03, grad_scale: 4.0 2023-05-01 07:16:10,822 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=209466.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:16:42,857 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0936, 2.3612, 2.3757, 2.6865, 1.8581, 3.1816, 1.8719, 2.7115], device='cuda:4'), covar=tensor([0.1044, 0.0631, 0.1108, 0.0180, 0.0142, 0.0489, 0.1408, 0.0732], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0173, 0.0194, 0.0188, 0.0206, 0.0214, 0.0201, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 07:16:52,760 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.025e+02 2.825e+02 3.422e+02 4.079e+02 7.372e+02, threshold=6.844e+02, percent-clipped=3.0 2023-05-01 07:16:59,409 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3703, 2.5422, 2.4441, 4.4825, 2.4327, 3.0424, 2.4925, 2.7625], device='cuda:4'), covar=tensor([0.1329, 0.3243, 0.2793, 0.0413, 0.3734, 0.2190, 0.3392, 0.2813], device='cuda:4'), in_proj_covar=tensor([0.0399, 0.0444, 0.0363, 0.0325, 0.0433, 0.0511, 0.0413, 0.0519], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:17:08,454 INFO [train.py:904] (4/8) Epoch 21, batch 6500, loss[loss=0.2052, simple_loss=0.2874, pruned_loss=0.06147, over 16931.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2869, pruned_loss=0.05763, over 3091729.46 frames. ], batch size: 109, lr: 3.21e-03, grad_scale: 4.0 2023-05-01 07:17:20,805 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=209510.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:17:48,029 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=209527.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:18:29,349 INFO [train.py:904] (4/8) Epoch 21, batch 6550, loss[loss=0.2181, simple_loss=0.3161, pruned_loss=0.06003, over 16436.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2896, pruned_loss=0.0585, over 3093709.63 frames. ], batch size: 146, lr: 3.21e-03, grad_scale: 4.0 2023-05-01 07:18:40,049 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=209558.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:19:01,117 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=209571.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:19:33,743 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.848e+02 2.646e+02 3.318e+02 3.875e+02 1.016e+03, threshold=6.636e+02, percent-clipped=2.0 2023-05-01 07:19:49,515 INFO [train.py:904] (4/8) Epoch 21, batch 6600, loss[loss=0.2217, simple_loss=0.3116, pruned_loss=0.06585, over 16891.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2915, pruned_loss=0.05858, over 3095022.60 frames. ], batch size: 116, lr: 3.21e-03, grad_scale: 4.0 2023-05-01 07:19:55,990 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=209606.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:20:03,741 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=209611.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:21:07,997 INFO [train.py:904] (4/8) Epoch 21, batch 6650, loss[loss=0.1811, simple_loss=0.2651, pruned_loss=0.04859, over 17028.00 frames. ], tot_loss[loss=0.205, simple_loss=0.2917, pruned_loss=0.05916, over 3104092.89 frames. ], batch size: 53, lr: 3.21e-03, grad_scale: 2.0 2023-05-01 07:21:08,417 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4073, 4.5392, 4.7098, 4.5004, 4.5545, 5.0414, 4.5855, 4.3218], device='cuda:4'), covar=tensor([0.1447, 0.1911, 0.1913, 0.1869, 0.2383, 0.0978, 0.1664, 0.2522], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0589, 0.0646, 0.0491, 0.0651, 0.0680, 0.0507, 0.0659], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 07:21:12,194 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=209654.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:21:32,344 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=209667.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:22:05,004 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=209688.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:22:11,939 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.980e+02 3.003e+02 3.834e+02 5.149e+02 1.142e+03, threshold=7.669e+02, percent-clipped=10.0 2023-05-01 07:22:25,654 INFO [train.py:904] (4/8) Epoch 21, batch 6700, loss[loss=0.213, simple_loss=0.2977, pruned_loss=0.06418, over 15109.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.291, pruned_loss=0.05982, over 3081086.89 frames. ], batch size: 190, lr: 3.21e-03, grad_scale: 2.0 2023-05-01 07:23:40,228 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=209749.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:23:43,724 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-05-01 07:23:44,112 INFO [train.py:904] (4/8) Epoch 21, batch 6750, loss[loss=0.2221, simple_loss=0.3018, pruned_loss=0.07119, over 16949.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2907, pruned_loss=0.0602, over 3090072.89 frames. ], batch size: 116, lr: 3.20e-03, grad_scale: 2.0 2023-05-01 07:24:47,245 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.958e+02 2.606e+02 3.292e+02 3.909e+02 7.681e+02, threshold=6.584e+02, percent-clipped=1.0 2023-05-01 07:25:01,547 INFO [train.py:904] (4/8) Epoch 21, batch 6800, loss[loss=0.2155, simple_loss=0.3021, pruned_loss=0.06444, over 16661.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.2908, pruned_loss=0.06047, over 3073158.79 frames. ], batch size: 62, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:25:33,963 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=209822.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:25:47,278 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6193, 1.7701, 1.5932, 1.5157, 1.9475, 1.6404, 1.6175, 1.9236], device='cuda:4'), covar=tensor([0.0202, 0.0295, 0.0412, 0.0350, 0.0202, 0.0251, 0.0178, 0.0211], device='cuda:4'), in_proj_covar=tensor([0.0203, 0.0230, 0.0222, 0.0222, 0.0231, 0.0229, 0.0230, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:26:20,405 INFO [train.py:904] (4/8) Epoch 21, batch 6850, loss[loss=0.1978, simple_loss=0.2984, pruned_loss=0.04864, over 16425.00 frames. ], tot_loss[loss=0.2084, simple_loss=0.2932, pruned_loss=0.06178, over 3059072.47 frames. ], batch size: 68, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:26:29,620 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=209858.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:26:42,250 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=209866.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:26:46,465 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8373, 4.8081, 4.7067, 3.8370, 4.7024, 1.6804, 4.4153, 4.3039], device='cuda:4'), covar=tensor([0.0129, 0.0134, 0.0196, 0.0413, 0.0127, 0.2956, 0.0232, 0.0257], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0151, 0.0194, 0.0175, 0.0172, 0.0204, 0.0183, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:27:21,614 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.924e+02 2.648e+02 3.249e+02 3.781e+02 8.526e+02, threshold=6.499e+02, percent-clipped=1.0 2023-05-01 07:27:34,652 INFO [train.py:904] (4/8) Epoch 21, batch 6900, loss[loss=0.1954, simple_loss=0.291, pruned_loss=0.04996, over 16788.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.2945, pruned_loss=0.0605, over 3071591.42 frames. ], batch size: 83, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:27:41,304 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=209906.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:27:48,363 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=209911.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:27:50,245 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6882, 4.7607, 4.5400, 4.2224, 4.1973, 4.6551, 4.4979, 4.3552], device='cuda:4'), covar=tensor([0.0700, 0.0800, 0.0389, 0.0386, 0.1079, 0.0630, 0.0536, 0.0819], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0409, 0.0330, 0.0327, 0.0341, 0.0380, 0.0228, 0.0398], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:28:51,082 INFO [train.py:904] (4/8) Epoch 21, batch 6950, loss[loss=0.2668, simple_loss=0.3293, pruned_loss=0.1021, over 11201.00 frames. ], tot_loss[loss=0.2111, simple_loss=0.2969, pruned_loss=0.06261, over 3062062.72 frames. ], batch size: 248, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:28:54,453 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=209954.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:29:02,106 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=209959.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:29:06,368 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=209962.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:29:51,930 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.035e+02 3.251e+02 3.766e+02 4.564e+02 9.580e+02, threshold=7.532e+02, percent-clipped=9.0 2023-05-01 07:30:09,077 INFO [train.py:904] (4/8) Epoch 21, batch 7000, loss[loss=0.1837, simple_loss=0.2865, pruned_loss=0.04046, over 17132.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2962, pruned_loss=0.06137, over 3068490.50 frames. ], batch size: 47, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:30:09,333 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=210002.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:31:02,837 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6029, 4.6966, 4.4776, 4.1976, 4.1865, 4.5801, 4.3593, 4.2691], device='cuda:4'), covar=tensor([0.0606, 0.0542, 0.0279, 0.0285, 0.0829, 0.0480, 0.0527, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0407, 0.0329, 0.0325, 0.0339, 0.0379, 0.0227, 0.0396], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:31:11,189 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=210044.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:31:18,464 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 07:31:22,808 INFO [train.py:904] (4/8) Epoch 21, batch 7050, loss[loss=0.2298, simple_loss=0.2967, pruned_loss=0.08143, over 11505.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.2975, pruned_loss=0.06158, over 3072327.29 frames. ], batch size: 247, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:31:30,587 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=210057.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:32:24,157 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.868e+02 2.885e+02 3.523e+02 4.096e+02 8.338e+02, threshold=7.047e+02, percent-clipped=1.0 2023-05-01 07:32:37,544 INFO [train.py:904] (4/8) Epoch 21, batch 7100, loss[loss=0.1981, simple_loss=0.2917, pruned_loss=0.05225, over 16835.00 frames. ], tot_loss[loss=0.2091, simple_loss=0.2956, pruned_loss=0.06131, over 3073734.64 frames. ], batch size: 102, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:33:03,215 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=210118.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:33:09,183 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=210122.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:33:21,093 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9503, 1.9436, 2.3715, 2.8279, 2.7819, 3.3301, 2.1056, 3.3765], device='cuda:4'), covar=tensor([0.0240, 0.0542, 0.0361, 0.0325, 0.0311, 0.0174, 0.0548, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0192, 0.0176, 0.0181, 0.0194, 0.0151, 0.0193, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:33:55,429 INFO [train.py:904] (4/8) Epoch 21, batch 7150, loss[loss=0.233, simple_loss=0.3008, pruned_loss=0.08261, over 11627.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2934, pruned_loss=0.06095, over 3071069.81 frames. ], batch size: 247, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:33:55,947 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0544, 4.1215, 4.4263, 4.3781, 4.3911, 4.1224, 4.1292, 4.0133], device='cuda:4'), covar=tensor([0.0356, 0.0615, 0.0368, 0.0446, 0.0521, 0.0431, 0.0964, 0.0611], device='cuda:4'), in_proj_covar=tensor([0.0403, 0.0448, 0.0436, 0.0405, 0.0484, 0.0458, 0.0544, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 07:34:16,466 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=210166.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:34:21,258 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=210170.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:34:27,111 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3927, 3.4368, 2.0769, 3.8470, 2.6028, 3.8350, 2.2329, 2.7090], device='cuda:4'), covar=tensor([0.0308, 0.0385, 0.1667, 0.0173, 0.0813, 0.0541, 0.1575, 0.0768], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0194, 0.0159, 0.0176, 0.0214, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 07:34:53,944 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.076e+02 2.584e+02 3.186e+02 4.207e+02 8.537e+02, threshold=6.371e+02, percent-clipped=2.0 2023-05-01 07:35:07,924 INFO [train.py:904] (4/8) Epoch 21, batch 7200, loss[loss=0.2114, simple_loss=0.2923, pruned_loss=0.06528, over 12020.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2905, pruned_loss=0.05921, over 3063184.73 frames. ], batch size: 248, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:35:26,244 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=210214.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:35:58,441 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8494, 2.7311, 2.6856, 1.9579, 2.5615, 2.6904, 2.5659, 1.9234], device='cuda:4'), covar=tensor([0.0445, 0.0083, 0.0085, 0.0359, 0.0134, 0.0120, 0.0123, 0.0401], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0081, 0.0082, 0.0132, 0.0096, 0.0107, 0.0092, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 07:36:28,283 INFO [train.py:904] (4/8) Epoch 21, batch 7250, loss[loss=0.1948, simple_loss=0.2804, pruned_loss=0.05455, over 16751.00 frames. ], tot_loss[loss=0.2013, simple_loss=0.2879, pruned_loss=0.05739, over 3078528.00 frames. ], batch size: 83, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:36:43,724 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=210262.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:36:57,951 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-05-01 07:37:24,328 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1360, 3.1800, 1.9413, 3.4469, 2.4293, 3.4966, 2.0713, 2.5388], device='cuda:4'), covar=tensor([0.0326, 0.0405, 0.1726, 0.0248, 0.0874, 0.0674, 0.1629, 0.0876], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0194, 0.0160, 0.0176, 0.0215, 0.0201, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 07:37:31,771 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.805e+02 2.676e+02 3.072e+02 3.848e+02 7.717e+02, threshold=6.145e+02, percent-clipped=2.0 2023-05-01 07:37:44,653 INFO [train.py:904] (4/8) Epoch 21, batch 7300, loss[loss=0.2221, simple_loss=0.2944, pruned_loss=0.07487, over 11754.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2875, pruned_loss=0.05705, over 3092612.39 frames. ], batch size: 247, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:37:58,790 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=210310.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:38:19,733 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-05-01 07:38:48,742 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=210344.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:38:59,602 INFO [train.py:904] (4/8) Epoch 21, batch 7350, loss[loss=0.1952, simple_loss=0.2896, pruned_loss=0.0504, over 16761.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.289, pruned_loss=0.05823, over 3086636.35 frames. ], batch size: 124, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:40:00,127 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=210392.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:40:02,082 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.675e+02 2.648e+02 3.204e+02 3.841e+02 6.410e+02, threshold=6.409e+02, percent-clipped=2.0 2023-05-01 07:40:14,853 INFO [train.py:904] (4/8) Epoch 21, batch 7400, loss[loss=0.228, simple_loss=0.3103, pruned_loss=0.07284, over 16276.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.29, pruned_loss=0.05889, over 3074042.85 frames. ], batch size: 165, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:40:16,050 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1041, 2.4667, 2.5354, 1.9597, 2.7282, 2.7911, 2.4013, 2.3828], device='cuda:4'), covar=tensor([0.0647, 0.0248, 0.0213, 0.0875, 0.0106, 0.0245, 0.0424, 0.0403], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0106, 0.0095, 0.0136, 0.0078, 0.0121, 0.0126, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 07:40:32,266 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=210413.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:41:32,223 INFO [train.py:904] (4/8) Epoch 21, batch 7450, loss[loss=0.2078, simple_loss=0.2972, pruned_loss=0.05922, over 15395.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2903, pruned_loss=0.05919, over 3095563.62 frames. ], batch size: 190, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:41:46,449 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 07:42:17,291 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.92 vs. limit=5.0 2023-05-01 07:42:42,640 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.331e+02 3.072e+02 3.553e+02 4.443e+02 7.195e+02, threshold=7.106e+02, percent-clipped=1.0 2023-05-01 07:42:53,253 INFO [train.py:904] (4/8) Epoch 21, batch 7500, loss[loss=0.2398, simple_loss=0.3135, pruned_loss=0.08305, over 16474.00 frames. ], tot_loss[loss=0.2042, simple_loss=0.2907, pruned_loss=0.05886, over 3089053.96 frames. ], batch size: 146, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:44:09,157 INFO [train.py:904] (4/8) Epoch 21, batch 7550, loss[loss=0.1956, simple_loss=0.2889, pruned_loss=0.05119, over 16390.00 frames. ], tot_loss[loss=0.205, simple_loss=0.2907, pruned_loss=0.05966, over 3074570.32 frames. ], batch size: 146, lr: 3.20e-03, grad_scale: 4.0 2023-05-01 07:45:11,347 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.786e+02 2.775e+02 3.392e+02 4.108e+02 6.854e+02, threshold=6.785e+02, percent-clipped=0.0 2023-05-01 07:45:23,227 INFO [train.py:904] (4/8) Epoch 21, batch 7600, loss[loss=0.1966, simple_loss=0.2876, pruned_loss=0.05281, over 17114.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2897, pruned_loss=0.0592, over 3089533.11 frames. ], batch size: 49, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:46:04,186 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-05-01 07:46:08,241 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=210632.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:46:37,229 INFO [train.py:904] (4/8) Epoch 21, batch 7650, loss[loss=0.2532, simple_loss=0.3213, pruned_loss=0.0925, over 11402.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2906, pruned_loss=0.06028, over 3078934.85 frames. ], batch size: 248, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:46:59,444 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=210666.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:47:26,055 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=210682.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:47:42,154 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=210693.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:47:42,756 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.936e+02 3.052e+02 3.513e+02 4.198e+02 7.732e+02, threshold=7.025e+02, percent-clipped=2.0 2023-05-01 07:47:55,351 INFO [train.py:904] (4/8) Epoch 21, batch 7700, loss[loss=0.2202, simple_loss=0.2965, pruned_loss=0.07189, over 11591.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2908, pruned_loss=0.06099, over 3076129.78 frames. ], batch size: 246, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:48:12,259 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=210713.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:48:23,315 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5238, 2.9809, 3.0753, 1.9399, 2.6810, 2.0703, 3.0786, 3.2093], device='cuda:4'), covar=tensor([0.0279, 0.0814, 0.0616, 0.2085, 0.0930, 0.1078, 0.0668, 0.0860], device='cuda:4'), in_proj_covar=tensor([0.0153, 0.0163, 0.0166, 0.0151, 0.0143, 0.0128, 0.0142, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 07:48:29,583 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2915, 5.6235, 5.3410, 5.3977, 5.0343, 5.0673, 5.0324, 5.7326], device='cuda:4'), covar=tensor([0.1207, 0.0828, 0.1016, 0.0922, 0.0834, 0.0766, 0.1210, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0658, 0.0798, 0.0666, 0.0607, 0.0506, 0.0518, 0.0670, 0.0626], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:48:34,061 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=210727.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:48:54,782 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4111, 3.4554, 3.7364, 2.1833, 3.1106, 2.3953, 3.7446, 3.7179], device='cuda:4'), covar=tensor([0.0214, 0.0811, 0.0533, 0.1925, 0.0827, 0.0954, 0.0573, 0.1002], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0163, 0.0167, 0.0152, 0.0144, 0.0129, 0.0143, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 07:48:59,591 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=210743.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:49:12,532 INFO [train.py:904] (4/8) Epoch 21, batch 7750, loss[loss=0.2003, simple_loss=0.2904, pruned_loss=0.05507, over 16632.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.291, pruned_loss=0.06085, over 3067130.34 frames. ], batch size: 57, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:49:27,321 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=210761.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:49:33,072 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=210764.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:50:05,925 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0191, 2.0942, 2.2216, 3.4489, 2.0994, 2.3733, 2.2131, 2.2563], device='cuda:4'), covar=tensor([0.1396, 0.3746, 0.2906, 0.0620, 0.4330, 0.2561, 0.3484, 0.3422], device='cuda:4'), in_proj_covar=tensor([0.0396, 0.0445, 0.0364, 0.0324, 0.0435, 0.0511, 0.0414, 0.0518], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 07:50:16,860 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8007, 2.8761, 2.6054, 5.0935, 3.9246, 4.3609, 1.7996, 3.1800], device='cuda:4'), covar=tensor([0.1337, 0.0792, 0.1285, 0.0174, 0.0391, 0.0420, 0.1575, 0.0837], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0196, 0.0190, 0.0207, 0.0216, 0.0202, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 07:50:18,619 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.950e+02 2.795e+02 3.327e+02 4.093e+02 7.547e+02, threshold=6.654e+02, percent-clipped=2.0 2023-05-01 07:50:31,004 INFO [train.py:904] (4/8) Epoch 21, batch 7800, loss[loss=0.2142, simple_loss=0.2927, pruned_loss=0.06789, over 15331.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2909, pruned_loss=0.06067, over 3078335.84 frames. ], batch size: 190, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:51:07,562 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=210825.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:51:31,579 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2535, 3.1178, 3.3984, 1.7587, 3.5657, 3.6423, 2.7938, 2.6923], device='cuda:4'), covar=tensor([0.0857, 0.0277, 0.0198, 0.1232, 0.0082, 0.0189, 0.0465, 0.0500], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0107, 0.0096, 0.0137, 0.0079, 0.0123, 0.0128, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 07:51:48,910 INFO [train.py:904] (4/8) Epoch 21, batch 7850, loss[loss=0.2054, simple_loss=0.2902, pruned_loss=0.06028, over 15335.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2917, pruned_loss=0.06013, over 3092672.20 frames. ], batch size: 190, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:52:54,099 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.040e+02 2.722e+02 3.190e+02 3.938e+02 6.813e+02, threshold=6.379e+02, percent-clipped=1.0 2023-05-01 07:53:05,678 INFO [train.py:904] (4/8) Epoch 21, batch 7900, loss[loss=0.2207, simple_loss=0.3069, pruned_loss=0.06728, over 15204.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2911, pruned_loss=0.05982, over 3094306.49 frames. ], batch size: 190, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:53:34,933 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.98 vs. limit=5.0 2023-05-01 07:54:24,386 INFO [train.py:904] (4/8) Epoch 21, batch 7950, loss[loss=0.2132, simple_loss=0.2919, pruned_loss=0.06728, over 16604.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2916, pruned_loss=0.06036, over 3083523.66 frames. ], batch size: 57, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:55:14,453 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=210984.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:55:20,868 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=210988.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:55:29,110 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.637e+02 2.961e+02 3.312e+02 4.005e+02 9.237e+02, threshold=6.624e+02, percent-clipped=4.0 2023-05-01 07:55:41,322 INFO [train.py:904] (4/8) Epoch 21, batch 8000, loss[loss=0.2312, simple_loss=0.2989, pruned_loss=0.08174, over 11372.00 frames. ], tot_loss[loss=0.2075, simple_loss=0.2924, pruned_loss=0.06135, over 3064384.45 frames. ], batch size: 248, lr: 3.20e-03, grad_scale: 8.0 2023-05-01 07:56:12,602 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=211022.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:56:35,931 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=211038.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:56:36,318 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-05-01 07:56:45,452 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=211045.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:56:56,538 INFO [train.py:904] (4/8) Epoch 21, batch 8050, loss[loss=0.1868, simple_loss=0.2861, pruned_loss=0.04373, over 16865.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2918, pruned_loss=0.06036, over 3079459.59 frames. ], batch size: 96, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 07:57:59,280 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.685e+02 2.757e+02 3.257e+02 3.943e+02 6.625e+02, threshold=6.515e+02, percent-clipped=2.0 2023-05-01 07:58:08,454 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-05-01 07:58:10,496 INFO [train.py:904] (4/8) Epoch 21, batch 8100, loss[loss=0.2315, simple_loss=0.3238, pruned_loss=0.06953, over 16744.00 frames. ], tot_loss[loss=0.2066, simple_loss=0.2919, pruned_loss=0.06065, over 3053435.61 frames. ], batch size: 134, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 07:58:38,299 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=211120.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 07:59:22,860 INFO [train.py:904] (4/8) Epoch 21, batch 8150, loss[loss=0.181, simple_loss=0.2595, pruned_loss=0.05122, over 16307.00 frames. ], tot_loss[loss=0.2035, simple_loss=0.2888, pruned_loss=0.0591, over 3079769.60 frames. ], batch size: 146, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:00:27,467 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.932e+02 2.753e+02 3.326e+02 4.060e+02 8.278e+02, threshold=6.652e+02, percent-clipped=2.0 2023-05-01 08:00:40,743 INFO [train.py:904] (4/8) Epoch 21, batch 8200, loss[loss=0.2011, simple_loss=0.2823, pruned_loss=0.05997, over 16694.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2861, pruned_loss=0.05804, over 3088706.90 frames. ], batch size: 62, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:00:45,114 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6106, 2.2571, 1.8769, 1.9921, 2.5492, 2.2851, 2.3293, 2.6957], device='cuda:4'), covar=tensor([0.0261, 0.0419, 0.0575, 0.0503, 0.0275, 0.0383, 0.0264, 0.0289], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0229, 0.0223, 0.0223, 0.0230, 0.0229, 0.0230, 0.0226], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:01:59,672 INFO [train.py:904] (4/8) Epoch 21, batch 8250, loss[loss=0.176, simple_loss=0.2726, pruned_loss=0.03972, over 16243.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2851, pruned_loss=0.05561, over 3073266.50 frames. ], batch size: 165, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:02:56,784 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=211288.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:03:04,841 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-05-01 08:03:06,691 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.662e+02 2.346e+02 2.848e+02 3.670e+02 8.296e+02, threshold=5.695e+02, percent-clipped=3.0 2023-05-01 08:03:18,672 INFO [train.py:904] (4/8) Epoch 21, batch 8300, loss[loss=0.1736, simple_loss=0.2797, pruned_loss=0.03374, over 16887.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2819, pruned_loss=0.05275, over 3047329.27 frames. ], batch size: 102, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:03:41,574 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-01 08:03:51,768 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=211322.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:04:11,194 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7282, 3.2333, 3.4866, 2.0398, 3.0376, 2.2131, 3.3409, 3.3892], device='cuda:4'), covar=tensor([0.0286, 0.0790, 0.0448, 0.2055, 0.0725, 0.1025, 0.0547, 0.0764], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0160, 0.0164, 0.0150, 0.0142, 0.0127, 0.0141, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 08:04:12,349 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=211336.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:04:14,375 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0132, 2.3753, 2.0455, 2.1989, 2.6601, 2.4173, 2.5861, 2.8918], device='cuda:4'), covar=tensor([0.0177, 0.0407, 0.0489, 0.0417, 0.0284, 0.0330, 0.0244, 0.0244], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0225, 0.0219, 0.0219, 0.0227, 0.0225, 0.0226, 0.0222], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:04:16,003 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=211338.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:04:18,456 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=211340.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:04:20,494 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6942, 2.6143, 2.3673, 3.8614, 2.2903, 3.9198, 1.4804, 3.0146], device='cuda:4'), covar=tensor([0.1376, 0.0741, 0.1187, 0.0201, 0.0115, 0.0363, 0.1694, 0.0656], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0174, 0.0195, 0.0189, 0.0206, 0.0215, 0.0201, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 08:04:36,926 INFO [train.py:904] (4/8) Epoch 21, batch 8350, loss[loss=0.1859, simple_loss=0.2908, pruned_loss=0.04051, over 16843.00 frames. ], tot_loss[loss=0.1914, simple_loss=0.2809, pruned_loss=0.05094, over 3032342.05 frames. ], batch size: 116, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:04:56,764 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1332, 2.0805, 2.1576, 3.6551, 2.0625, 2.3931, 2.2232, 2.2355], device='cuda:4'), covar=tensor([0.1250, 0.3732, 0.3196, 0.0605, 0.4688, 0.2828, 0.3730, 0.3613], device='cuda:4'), in_proj_covar=tensor([0.0389, 0.0436, 0.0357, 0.0317, 0.0428, 0.0500, 0.0407, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:04:59,825 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7650, 3.7483, 3.9462, 3.7182, 3.9010, 4.2906, 3.9499, 3.6294], device='cuda:4'), covar=tensor([0.2086, 0.2434, 0.2256, 0.2523, 0.2803, 0.1753, 0.1587, 0.2656], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0575, 0.0634, 0.0481, 0.0638, 0.0667, 0.0502, 0.0646], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 08:05:05,458 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=211370.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:05:30,340 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=211386.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:05:30,436 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=211386.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:05:43,486 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 2.192e+02 2.619e+02 3.085e+02 8.186e+02, threshold=5.238e+02, percent-clipped=3.0 2023-05-01 08:05:55,506 INFO [train.py:904] (4/8) Epoch 21, batch 8400, loss[loss=0.1665, simple_loss=0.2666, pruned_loss=0.03322, over 16524.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2783, pruned_loss=0.04894, over 3031973.79 frames. ], batch size: 75, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:06:02,433 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-01 08:06:10,874 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=211412.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:06:22,507 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=211420.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:07:03,754 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=211447.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:07:10,964 INFO [train.py:904] (4/8) Epoch 21, batch 8450, loss[loss=0.1637, simple_loss=0.2627, pruned_loss=0.03239, over 16832.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2767, pruned_loss=0.04697, over 3054515.68 frames. ], batch size: 102, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:07:13,165 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-01 08:07:36,101 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=211468.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:07:44,866 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=211473.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:08:18,590 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.603e+02 2.143e+02 2.501e+02 3.093e+02 5.428e+02, threshold=5.001e+02, percent-clipped=1.0 2023-05-01 08:08:29,988 INFO [train.py:904] (4/8) Epoch 21, batch 8500, loss[loss=0.1723, simple_loss=0.2533, pruned_loss=0.04567, over 12149.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2729, pruned_loss=0.04488, over 3034555.57 frames. ], batch size: 247, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:08:41,882 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1163, 2.4186, 2.0699, 2.1364, 2.6677, 2.4258, 2.7290, 2.9315], device='cuda:4'), covar=tensor([0.0180, 0.0406, 0.0500, 0.0485, 0.0283, 0.0383, 0.0217, 0.0271], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0224, 0.0217, 0.0218, 0.0226, 0.0224, 0.0225, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:09:54,385 INFO [train.py:904] (4/8) Epoch 21, batch 8550, loss[loss=0.181, simple_loss=0.281, pruned_loss=0.04048, over 16755.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2703, pruned_loss=0.04396, over 3022490.53 frames. ], batch size: 124, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:10:03,672 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 08:11:18,418 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.555e+02 2.272e+02 2.648e+02 3.114e+02 4.500e+02, threshold=5.297e+02, percent-clipped=0.0 2023-05-01 08:11:32,286 INFO [train.py:904] (4/8) Epoch 21, batch 8600, loss[loss=0.1877, simple_loss=0.2901, pruned_loss=0.04267, over 16919.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2708, pruned_loss=0.04291, over 3044315.22 frames. ], batch size: 116, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:12:10,366 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2636, 3.5372, 3.7737, 2.5180, 3.4063, 3.7512, 3.5620, 2.1715], device='cuda:4'), covar=tensor([0.0506, 0.0062, 0.0047, 0.0404, 0.0099, 0.0084, 0.0069, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0133, 0.0080, 0.0081, 0.0132, 0.0095, 0.0106, 0.0092, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 08:12:17,632 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-01 08:12:26,265 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8387, 4.8315, 4.6082, 3.5062, 4.6831, 1.6229, 4.2592, 4.3649], device='cuda:4'), covar=tensor([0.0128, 0.0102, 0.0236, 0.0648, 0.0139, 0.3540, 0.0188, 0.0334], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0148, 0.0190, 0.0170, 0.0168, 0.0200, 0.0178, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:12:48,711 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=211640.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:13:09,685 INFO [train.py:904] (4/8) Epoch 21, batch 8650, loss[loss=0.164, simple_loss=0.2779, pruned_loss=0.02499, over 16836.00 frames. ], tot_loss[loss=0.1757, simple_loss=0.2694, pruned_loss=0.04101, over 3068854.66 frames. ], batch size: 102, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:14:30,914 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=211688.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:14:43,483 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.495e+02 2.069e+02 2.535e+02 3.155e+02 6.979e+02, threshold=5.069e+02, percent-clipped=2.0 2023-05-01 08:14:57,322 INFO [train.py:904] (4/8) Epoch 21, batch 8700, loss[loss=0.1802, simple_loss=0.2719, pruned_loss=0.04426, over 16800.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2674, pruned_loss=0.04022, over 3055388.39 frames. ], batch size: 124, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:15:34,129 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4157, 4.5177, 4.3049, 3.9963, 4.0300, 4.4232, 4.1527, 4.1405], device='cuda:4'), covar=tensor([0.0609, 0.0533, 0.0318, 0.0341, 0.0825, 0.0507, 0.0636, 0.0659], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0407, 0.0326, 0.0323, 0.0335, 0.0377, 0.0226, 0.0392], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:16:10,366 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=211742.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:16:17,981 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6054, 3.5618, 3.5276, 2.8045, 3.4416, 1.9937, 3.1513, 2.9182], device='cuda:4'), covar=tensor([0.0118, 0.0097, 0.0165, 0.0202, 0.0096, 0.2554, 0.0121, 0.0234], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0147, 0.0188, 0.0168, 0.0167, 0.0199, 0.0176, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:16:26,090 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5069, 4.6946, 4.8168, 4.6014, 4.6184, 5.1811, 4.6844, 4.4304], device='cuda:4'), covar=tensor([0.1205, 0.1837, 0.1997, 0.2094, 0.2534, 0.0899, 0.1599, 0.2345], device='cuda:4'), in_proj_covar=tensor([0.0385, 0.0561, 0.0617, 0.0468, 0.0620, 0.0651, 0.0491, 0.0629], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 08:16:31,418 INFO [train.py:904] (4/8) Epoch 21, batch 8750, loss[loss=0.1798, simple_loss=0.2784, pruned_loss=0.04057, over 15431.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2672, pruned_loss=0.03978, over 3052776.12 frames. ], batch size: 191, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:16:45,102 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9739, 2.2276, 2.3254, 3.0481, 1.8894, 3.2126, 1.7485, 2.6847], device='cuda:4'), covar=tensor([0.1181, 0.0711, 0.1041, 0.0166, 0.0080, 0.0326, 0.1454, 0.0732], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0171, 0.0191, 0.0184, 0.0201, 0.0210, 0.0198, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 08:17:13,747 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=211768.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:18:09,821 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.610e+02 2.163e+02 2.564e+02 3.220e+02 5.959e+02, threshold=5.128e+02, percent-clipped=1.0 2023-05-01 08:18:23,762 INFO [train.py:904] (4/8) Epoch 21, batch 8800, loss[loss=0.1685, simple_loss=0.2713, pruned_loss=0.03282, over 16858.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2648, pruned_loss=0.03845, over 3034236.36 frames. ], batch size: 96, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:20:07,833 INFO [train.py:904] (4/8) Epoch 21, batch 8850, loss[loss=0.1874, simple_loss=0.2917, pruned_loss=0.04159, over 16692.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2674, pruned_loss=0.0378, over 3037555.33 frames. ], batch size: 134, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:21:38,910 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.506e+02 2.129e+02 2.519e+02 2.912e+02 9.426e+02, threshold=5.038e+02, percent-clipped=1.0 2023-05-01 08:21:54,320 INFO [train.py:904] (4/8) Epoch 21, batch 8900, loss[loss=0.1647, simple_loss=0.2584, pruned_loss=0.03554, over 16494.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2675, pruned_loss=0.03721, over 3041937.65 frames. ], batch size: 68, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:22:56,708 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-05-01 08:23:57,812 INFO [train.py:904] (4/8) Epoch 21, batch 8950, loss[loss=0.1459, simple_loss=0.2387, pruned_loss=0.02655, over 15335.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2675, pruned_loss=0.03794, over 3058094.58 frames. ], batch size: 191, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:24:06,458 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1475, 1.6645, 1.9274, 2.0831, 2.2405, 2.3478, 1.7868, 2.2546], device='cuda:4'), covar=tensor([0.0250, 0.0512, 0.0313, 0.0329, 0.0337, 0.0216, 0.0475, 0.0156], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0188, 0.0172, 0.0176, 0.0190, 0.0147, 0.0189, 0.0143], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:25:11,937 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6530, 4.9799, 4.7707, 4.7634, 4.4916, 4.5313, 4.4251, 5.0334], device='cuda:4'), covar=tensor([0.1212, 0.0825, 0.1016, 0.0822, 0.0836, 0.1122, 0.1164, 0.0906], device='cuda:4'), in_proj_covar=tensor([0.0643, 0.0779, 0.0649, 0.0593, 0.0495, 0.0506, 0.0655, 0.0612], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:25:29,292 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.253e+02 2.092e+02 2.341e+02 2.927e+02 4.703e+02, threshold=4.682e+02, percent-clipped=0.0 2023-05-01 08:25:46,908 INFO [train.py:904] (4/8) Epoch 21, batch 9000, loss[loss=0.1423, simple_loss=0.234, pruned_loss=0.02534, over 16424.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.264, pruned_loss=0.03646, over 3057147.83 frames. ], batch size: 68, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:25:46,909 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 08:25:57,431 INFO [train.py:938] (4/8) Epoch 21, validation: loss=0.1457, simple_loss=0.2498, pruned_loss=0.02077, over 944034.00 frames. 2023-05-01 08:25:57,432 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 08:27:14,182 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2176, 3.5643, 3.7154, 1.9234, 3.1411, 2.3480, 3.5676, 3.6233], device='cuda:4'), covar=tensor([0.0214, 0.0764, 0.0476, 0.2188, 0.0762, 0.0978, 0.0605, 0.0943], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0157, 0.0162, 0.0149, 0.0140, 0.0126, 0.0138, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 08:27:20,366 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=212042.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:27:39,212 INFO [train.py:904] (4/8) Epoch 21, batch 9050, loss[loss=0.149, simple_loss=0.243, pruned_loss=0.02749, over 16908.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2651, pruned_loss=0.03663, over 3077240.62 frames. ], batch size: 96, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:28:14,802 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=212068.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:28:58,993 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=212090.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:29:10,423 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.640e+02 2.192e+02 2.458e+02 2.901e+02 5.041e+02, threshold=4.916e+02, percent-clipped=1.0 2023-05-01 08:29:26,256 INFO [train.py:904] (4/8) Epoch 21, batch 9100, loss[loss=0.1922, simple_loss=0.2829, pruned_loss=0.05076, over 16960.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2649, pruned_loss=0.03739, over 3095439.99 frames. ], batch size: 109, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:29:30,928 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=212104.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 08:29:53,443 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=212116.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:31:22,738 INFO [train.py:904] (4/8) Epoch 21, batch 9150, loss[loss=0.1591, simple_loss=0.258, pruned_loss=0.03008, over 15427.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2654, pruned_loss=0.03714, over 3089710.84 frames. ], batch size: 192, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:31:52,675 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=212165.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 08:32:57,277 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.621e+02 2.079e+02 2.542e+02 3.144e+02 5.396e+02, threshold=5.084e+02, percent-clipped=1.0 2023-05-01 08:33:09,949 INFO [train.py:904] (4/8) Epoch 21, batch 9200, loss[loss=0.1592, simple_loss=0.2538, pruned_loss=0.0323, over 16205.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2615, pruned_loss=0.03657, over 3078813.81 frames. ], batch size: 165, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:33:30,563 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4159, 3.3455, 3.4671, 3.5386, 3.5593, 3.3196, 3.5460, 3.6123], device='cuda:4'), covar=tensor([0.1140, 0.0944, 0.0913, 0.0542, 0.0617, 0.1987, 0.0819, 0.0753], device='cuda:4'), in_proj_covar=tensor([0.0594, 0.0734, 0.0853, 0.0747, 0.0565, 0.0594, 0.0611, 0.0707], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:34:28,488 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8486, 1.9385, 2.2846, 3.1091, 2.0966, 2.0916, 2.1624, 2.0409], device='cuda:4'), covar=tensor([0.1415, 0.4444, 0.2747, 0.0851, 0.5344, 0.3409, 0.3828, 0.4448], device='cuda:4'), in_proj_covar=tensor([0.0387, 0.0434, 0.0358, 0.0315, 0.0426, 0.0496, 0.0406, 0.0506], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:34:48,817 INFO [train.py:904] (4/8) Epoch 21, batch 9250, loss[loss=0.157, simple_loss=0.2563, pruned_loss=0.02884, over 15221.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.261, pruned_loss=0.03669, over 3066339.15 frames. ], batch size: 191, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:34:51,604 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=212253.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:36:05,937 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3847, 4.4296, 4.7516, 4.7164, 4.7240, 4.4735, 4.4104, 4.4337], device='cuda:4'), covar=tensor([0.0420, 0.1133, 0.0532, 0.0537, 0.0605, 0.0587, 0.0962, 0.0502], device='cuda:4'), in_proj_covar=tensor([0.0386, 0.0430, 0.0419, 0.0388, 0.0466, 0.0438, 0.0519, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 08:36:21,620 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4137, 3.3987, 3.5113, 3.5847, 3.6135, 3.3245, 3.5921, 3.6746], device='cuda:4'), covar=tensor([0.1448, 0.1071, 0.1056, 0.0610, 0.0607, 0.2362, 0.0787, 0.0696], device='cuda:4'), in_proj_covar=tensor([0.0595, 0.0734, 0.0853, 0.0747, 0.0566, 0.0594, 0.0610, 0.0707], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:36:25,634 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.441e+02 2.162e+02 2.598e+02 3.159e+02 6.088e+02, threshold=5.195e+02, percent-clipped=3.0 2023-05-01 08:36:39,742 INFO [train.py:904] (4/8) Epoch 21, batch 9300, loss[loss=0.1563, simple_loss=0.2395, pruned_loss=0.03656, over 12334.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2599, pruned_loss=0.03635, over 3062453.35 frames. ], batch size: 250, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:37:07,753 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=212314.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 08:37:30,170 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6041, 1.9910, 1.6544, 1.8498, 2.3073, 2.0478, 2.0738, 2.5018], device='cuda:4'), covar=tensor([0.0222, 0.0503, 0.0647, 0.0541, 0.0299, 0.0449, 0.0233, 0.0285], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0225, 0.0218, 0.0218, 0.0225, 0.0224, 0.0222, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:37:32,748 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 08:38:11,316 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3040, 4.3483, 2.9167, 4.9346, 3.4466, 4.8224, 3.0637, 3.7784], device='cuda:4'), covar=tensor([0.0227, 0.0270, 0.1260, 0.0207, 0.0616, 0.0383, 0.1289, 0.0495], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0169, 0.0188, 0.0153, 0.0171, 0.0207, 0.0195, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 08:38:22,830 INFO [train.py:904] (4/8) Epoch 21, batch 9350, loss[loss=0.1821, simple_loss=0.273, pruned_loss=0.04557, over 16639.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2591, pruned_loss=0.03606, over 3049889.86 frames. ], batch size: 134, lr: 3.19e-03, grad_scale: 8.0 2023-05-01 08:39:27,228 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-01 08:39:41,368 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8678, 2.7912, 2.7373, 1.9824, 2.6036, 2.8230, 2.6768, 1.8297], device='cuda:4'), covar=tensor([0.0469, 0.0078, 0.0077, 0.0400, 0.0156, 0.0105, 0.0111, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0131, 0.0079, 0.0080, 0.0130, 0.0094, 0.0104, 0.0091, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 08:39:47,852 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.322e+02 2.212e+02 2.571e+02 3.078e+02 5.703e+02, threshold=5.142e+02, percent-clipped=1.0 2023-05-01 08:40:02,993 INFO [train.py:904] (4/8) Epoch 21, batch 9400, loss[loss=0.162, simple_loss=0.2633, pruned_loss=0.03032, over 15307.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2593, pruned_loss=0.03621, over 3029125.94 frames. ], batch size: 191, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:40:07,534 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-05-01 08:41:16,811 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-05-01 08:41:18,575 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=212439.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:41:42,917 INFO [train.py:904] (4/8) Epoch 21, batch 9450, loss[loss=0.1659, simple_loss=0.2541, pruned_loss=0.03887, over 12236.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2612, pruned_loss=0.03655, over 3036233.17 frames. ], batch size: 250, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:41:57,283 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0219, 5.5059, 5.6807, 5.3919, 5.5204, 6.0182, 5.5558, 5.2812], device='cuda:4'), covar=tensor([0.0822, 0.1592, 0.1781, 0.1928, 0.2066, 0.0782, 0.1428, 0.2131], device='cuda:4'), in_proj_covar=tensor([0.0380, 0.0557, 0.0614, 0.0464, 0.0613, 0.0648, 0.0487, 0.0622], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 08:41:58,722 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=212460.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 08:41:59,184 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-05-01 08:42:26,890 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4205, 3.3776, 3.4873, 3.5773, 3.5870, 3.3205, 3.5779, 3.6394], device='cuda:4'), covar=tensor([0.1436, 0.1237, 0.1155, 0.0671, 0.0775, 0.2485, 0.0999, 0.0859], device='cuda:4'), in_proj_covar=tensor([0.0594, 0.0731, 0.0851, 0.0745, 0.0564, 0.0593, 0.0609, 0.0706], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:43:11,245 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.570e+02 1.974e+02 2.463e+02 3.006e+02 5.675e+02, threshold=4.927e+02, percent-clipped=1.0 2023-05-01 08:43:21,174 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=212500.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:43:25,063 INFO [train.py:904] (4/8) Epoch 21, batch 9500, loss[loss=0.1687, simple_loss=0.26, pruned_loss=0.03866, over 16638.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2606, pruned_loss=0.03639, over 3035959.28 frames. ], batch size: 62, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:45:12,071 INFO [train.py:904] (4/8) Epoch 21, batch 9550, loss[loss=0.1703, simple_loss=0.2748, pruned_loss=0.03286, over 15320.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2599, pruned_loss=0.03614, over 3041336.16 frames. ], batch size: 192, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:46:39,906 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.364e+02 2.109e+02 2.597e+02 3.170e+02 5.717e+02, threshold=5.194e+02, percent-clipped=4.0 2023-05-01 08:46:50,960 INFO [train.py:904] (4/8) Epoch 21, batch 9600, loss[loss=0.1907, simple_loss=0.2931, pruned_loss=0.04409, over 16122.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2616, pruned_loss=0.0371, over 3034799.02 frames. ], batch size: 165, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:47:06,244 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=212609.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 08:48:37,651 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0632, 3.1036, 1.8850, 3.3261, 2.2565, 3.3204, 2.1241, 2.5566], device='cuda:4'), covar=tensor([0.0316, 0.0363, 0.1633, 0.0262, 0.0907, 0.0544, 0.1517, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0169, 0.0189, 0.0153, 0.0172, 0.0207, 0.0196, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 08:48:38,603 INFO [train.py:904] (4/8) Epoch 21, batch 9650, loss[loss=0.1607, simple_loss=0.2633, pruned_loss=0.0291, over 15443.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2631, pruned_loss=0.03744, over 3024854.91 frames. ], batch size: 192, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:50:11,920 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.547e+02 2.239e+02 2.609e+02 3.201e+02 8.250e+02, threshold=5.217e+02, percent-clipped=3.0 2023-05-01 08:50:20,713 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=212698.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:50:25,300 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4549, 3.4279, 3.5217, 3.5810, 3.6194, 3.3260, 3.6050, 3.6636], device='cuda:4'), covar=tensor([0.1333, 0.0890, 0.0974, 0.0630, 0.0576, 0.2439, 0.0791, 0.0808], device='cuda:4'), in_proj_covar=tensor([0.0594, 0.0730, 0.0851, 0.0745, 0.0565, 0.0593, 0.0610, 0.0707], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 08:50:28,029 INFO [train.py:904] (4/8) Epoch 21, batch 9700, loss[loss=0.1819, simple_loss=0.2704, pruned_loss=0.0467, over 16934.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2621, pruned_loss=0.03712, over 3023883.31 frames. ], batch size: 109, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:50:46,012 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=212711.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:52:10,270 INFO [train.py:904] (4/8) Epoch 21, batch 9750, loss[loss=0.1688, simple_loss=0.2664, pruned_loss=0.03562, over 16719.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2613, pruned_loss=0.03736, over 3025970.38 frames. ], batch size: 124, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:52:25,405 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=212759.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:52:27,046 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=212760.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 08:52:29,605 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3223, 3.0646, 3.3030, 1.7055, 3.4697, 3.5740, 2.8791, 2.7560], device='cuda:4'), covar=tensor([0.0728, 0.0277, 0.0201, 0.1239, 0.0088, 0.0175, 0.0401, 0.0436], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0103, 0.0091, 0.0133, 0.0076, 0.0117, 0.0123, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 08:52:49,260 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=212772.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:52:53,758 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=212774.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:53:36,936 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.235e+02 2.010e+02 2.340e+02 2.880e+02 4.703e+02, threshold=4.680e+02, percent-clipped=0.0 2023-05-01 08:53:37,559 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=212795.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:53:48,982 INFO [train.py:904] (4/8) Epoch 21, batch 9800, loss[loss=0.1544, simple_loss=0.2408, pruned_loss=0.03395, over 12204.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.262, pruned_loss=0.03674, over 3040106.76 frames. ], batch size: 250, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:54:01,428 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=212808.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 08:54:10,477 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1231, 2.5303, 2.6431, 1.8622, 2.8096, 2.8926, 2.4891, 2.4778], device='cuda:4'), covar=tensor([0.0659, 0.0243, 0.0207, 0.1015, 0.0103, 0.0200, 0.0462, 0.0411], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0103, 0.0091, 0.0133, 0.0076, 0.0117, 0.0123, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 08:54:43,123 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 08:54:50,070 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.84 vs. limit=2.0 2023-05-01 08:54:51,570 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=212835.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:55:31,572 INFO [train.py:904] (4/8) Epoch 21, batch 9850, loss[loss=0.1586, simple_loss=0.2584, pruned_loss=0.02938, over 16312.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2632, pruned_loss=0.03632, over 3060570.09 frames. ], batch size: 146, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:55:35,580 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-05-01 08:57:07,911 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.603e+02 2.176e+02 2.812e+02 3.349e+02 5.684e+02, threshold=5.624e+02, percent-clipped=7.0 2023-05-01 08:57:21,369 INFO [train.py:904] (4/8) Epoch 21, batch 9900, loss[loss=0.1875, simple_loss=0.2902, pruned_loss=0.04241, over 16258.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.263, pruned_loss=0.03612, over 3046117.95 frames. ], batch size: 165, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:57:38,616 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=212909.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 08:59:17,573 INFO [train.py:904] (4/8) Epoch 21, batch 9950, loss[loss=0.1667, simple_loss=0.2668, pruned_loss=0.03325, over 16875.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2651, pruned_loss=0.03673, over 3034052.46 frames. ], batch size: 102, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 08:59:29,578 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=212957.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:00:58,646 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.360e+02 2.203e+02 2.698e+02 3.794e+02 1.048e+03, threshold=5.396e+02, percent-clipped=2.0 2023-05-01 09:01:17,069 INFO [train.py:904] (4/8) Epoch 21, batch 10000, loss[loss=0.1703, simple_loss=0.2716, pruned_loss=0.03449, over 16516.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2636, pruned_loss=0.03611, over 3045134.86 frames. ], batch size: 148, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 09:02:58,011 INFO [train.py:904] (4/8) Epoch 21, batch 10050, loss[loss=0.1741, simple_loss=0.2759, pruned_loss=0.0361, over 16122.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2637, pruned_loss=0.03615, over 3051536.93 frames. ], batch size: 165, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 09:03:02,065 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=213054.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:03:28,753 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=213067.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:04:21,048 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.335e+02 2.063e+02 2.495e+02 3.076e+02 5.792e+02, threshold=4.990e+02, percent-clipped=2.0 2023-05-01 09:04:21,678 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=213095.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:04:32,157 INFO [train.py:904] (4/8) Epoch 21, batch 10100, loss[loss=0.1791, simple_loss=0.2646, pruned_loss=0.04679, over 16689.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2643, pruned_loss=0.03661, over 3050096.10 frames. ], batch size: 134, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 09:04:59,262 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9864, 4.2693, 4.0971, 4.1090, 3.8159, 3.8526, 3.8539, 4.2448], device='cuda:4'), covar=tensor([0.1119, 0.0900, 0.0931, 0.0821, 0.0776, 0.1717, 0.0977, 0.1100], device='cuda:4'), in_proj_covar=tensor([0.0635, 0.0772, 0.0640, 0.0585, 0.0491, 0.0502, 0.0646, 0.0604], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:05:27,315 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=213130.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:05:42,672 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=213143.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:05:51,218 INFO [train.py:904] (4/8) Epoch 21, batch 10150, loss[loss=0.1716, simple_loss=0.2547, pruned_loss=0.04425, over 12193.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.263, pruned_loss=0.03671, over 3037553.74 frames. ], batch size: 246, lr: 3.18e-03, grad_scale: 8.0 2023-05-01 09:06:16,172 INFO [train.py:904] (4/8) Epoch 22, batch 0, loss[loss=0.2039, simple_loss=0.2992, pruned_loss=0.0543, over 17117.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2992, pruned_loss=0.0543, over 17117.00 frames. ], batch size: 49, lr: 3.10e-03, grad_scale: 8.0 2023-05-01 09:06:16,172 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 09:06:23,632 INFO [train.py:938] (4/8) Epoch 22, validation: loss=0.1457, simple_loss=0.2489, pruned_loss=0.0212, over 944034.00 frames. 2023-05-01 09:06:23,633 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 09:06:52,733 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8334, 4.7477, 4.7306, 4.4124, 4.4431, 4.7798, 4.6952, 4.4623], device='cuda:4'), covar=tensor([0.0654, 0.0828, 0.0395, 0.0329, 0.1045, 0.0591, 0.0360, 0.0773], device='cuda:4'), in_proj_covar=tensor([0.0273, 0.0395, 0.0319, 0.0314, 0.0325, 0.0366, 0.0220, 0.0381], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-05-01 09:07:26,338 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.831e+02 2.320e+02 2.870e+02 3.515e+02 6.456e+02, threshold=5.740e+02, percent-clipped=5.0 2023-05-01 09:07:34,020 INFO [train.py:904] (4/8) Epoch 22, batch 50, loss[loss=0.187, simple_loss=0.2841, pruned_loss=0.04492, over 17034.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.2727, pruned_loss=0.04811, over 757919.48 frames. ], batch size: 50, lr: 3.10e-03, grad_scale: 1.0 2023-05-01 09:08:21,602 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.41 vs. limit=5.0 2023-05-01 09:08:41,847 INFO [train.py:904] (4/8) Epoch 22, batch 100, loss[loss=0.1979, simple_loss=0.2986, pruned_loss=0.04858, over 17266.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2668, pruned_loss=0.04576, over 1324310.31 frames. ], batch size: 52, lr: 3.10e-03, grad_scale: 1.0 2023-05-01 09:09:44,707 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.515e+02 2.194e+02 2.643e+02 3.119e+02 6.629e+02, threshold=5.286e+02, percent-clipped=1.0 2023-05-01 09:09:51,952 INFO [train.py:904] (4/8) Epoch 22, batch 150, loss[loss=0.1743, simple_loss=0.2467, pruned_loss=0.05097, over 16918.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2651, pruned_loss=0.0455, over 1772598.88 frames. ], batch size: 90, lr: 3.10e-03, grad_scale: 1.0 2023-05-01 09:10:26,766 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0230, 5.0165, 5.4473, 5.4422, 5.4638, 5.0991, 5.0496, 4.8280], device='cuda:4'), covar=tensor([0.0338, 0.0610, 0.0416, 0.0417, 0.0430, 0.0399, 0.0939, 0.0463], device='cuda:4'), in_proj_covar=tensor([0.0387, 0.0432, 0.0419, 0.0393, 0.0466, 0.0441, 0.0518, 0.0354], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 09:10:45,673 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6752, 3.6096, 3.9228, 2.1077, 4.0786, 4.0724, 3.1800, 3.0510], device='cuda:4'), covar=tensor([0.0764, 0.0221, 0.0165, 0.1129, 0.0073, 0.0176, 0.0363, 0.0425], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0106, 0.0093, 0.0136, 0.0077, 0.0121, 0.0126, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 09:11:00,681 INFO [train.py:904] (4/8) Epoch 22, batch 200, loss[loss=0.1605, simple_loss=0.2499, pruned_loss=0.03556, over 17244.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2628, pruned_loss=0.04483, over 2118880.54 frames. ], batch size: 45, lr: 3.10e-03, grad_scale: 1.0 2023-05-01 09:11:02,370 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=213354.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:11:20,317 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=213367.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:11:27,585 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2912, 2.4693, 2.8556, 3.1821, 3.0840, 3.7202, 2.6891, 3.6889], device='cuda:4'), covar=tensor([0.0229, 0.0435, 0.0308, 0.0285, 0.0289, 0.0151, 0.0452, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0189, 0.0175, 0.0179, 0.0193, 0.0149, 0.0192, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:12:00,918 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.471e+02 2.291e+02 2.770e+02 3.341e+02 7.349e+02, threshold=5.539e+02, percent-clipped=1.0 2023-05-01 09:12:07,194 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=213402.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:12:08,108 INFO [train.py:904] (4/8) Epoch 22, batch 250, loss[loss=0.1959, simple_loss=0.2768, pruned_loss=0.05748, over 11922.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2617, pruned_loss=0.04599, over 2380062.11 frames. ], batch size: 247, lr: 3.10e-03, grad_scale: 1.0 2023-05-01 09:12:25,569 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=213415.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:12:43,308 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4463, 4.5810, 4.6608, 4.5662, 4.5893, 5.1518, 4.6742, 4.3532], device='cuda:4'), covar=tensor([0.1737, 0.2082, 0.2514, 0.2243, 0.2905, 0.1173, 0.1712, 0.2548], device='cuda:4'), in_proj_covar=tensor([0.0400, 0.0583, 0.0646, 0.0486, 0.0644, 0.0678, 0.0507, 0.0648], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 09:12:46,433 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=213430.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:13:11,248 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.64 vs. limit=5.0 2023-05-01 09:13:17,487 INFO [train.py:904] (4/8) Epoch 22, batch 300, loss[loss=0.1836, simple_loss=0.2768, pruned_loss=0.04525, over 17094.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2592, pruned_loss=0.04512, over 2584302.88 frames. ], batch size: 53, lr: 3.10e-03, grad_scale: 1.0 2023-05-01 09:13:45,410 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-05-01 09:13:52,063 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=213478.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:14:19,525 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.474e+02 2.218e+02 2.593e+02 2.976e+02 4.668e+02, threshold=5.187e+02, percent-clipped=0.0 2023-05-01 09:14:25,241 INFO [train.py:904] (4/8) Epoch 22, batch 350, loss[loss=0.1563, simple_loss=0.2392, pruned_loss=0.0367, over 16832.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2566, pruned_loss=0.04383, over 2747031.76 frames. ], batch size: 42, lr: 3.10e-03, grad_scale: 1.0 2023-05-01 09:14:30,294 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9616, 5.3727, 5.0837, 5.0694, 4.7940, 4.7006, 4.7634, 5.4695], device='cuda:4'), covar=tensor([0.1451, 0.0911, 0.1162, 0.0891, 0.0967, 0.1175, 0.1264, 0.0986], device='cuda:4'), in_proj_covar=tensor([0.0658, 0.0804, 0.0663, 0.0610, 0.0511, 0.0520, 0.0672, 0.0628], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:14:43,943 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-05-01 09:15:16,962 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7740, 2.9427, 2.6376, 5.0612, 4.0663, 4.4210, 1.7113, 3.3587], device='cuda:4'), covar=tensor([0.1369, 0.0764, 0.1275, 0.0181, 0.0203, 0.0371, 0.1571, 0.0699], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0172, 0.0192, 0.0185, 0.0198, 0.0211, 0.0200, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 09:15:34,172 INFO [train.py:904] (4/8) Epoch 22, batch 400, loss[loss=0.1974, simple_loss=0.2676, pruned_loss=0.06354, over 16814.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2557, pruned_loss=0.04429, over 2870734.64 frames. ], batch size: 116, lr: 3.10e-03, grad_scale: 2.0 2023-05-01 09:15:40,956 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2163, 4.3087, 4.4038, 4.2800, 4.3096, 4.8380, 4.3784, 4.0908], device='cuda:4'), covar=tensor([0.1980, 0.2369, 0.2673, 0.2584, 0.3246, 0.1389, 0.1890, 0.2845], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0591, 0.0654, 0.0493, 0.0652, 0.0688, 0.0514, 0.0658], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 09:15:52,121 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=213566.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:16:15,102 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1885, 5.2383, 5.6318, 5.6254, 5.6490, 5.3134, 5.2284, 5.0640], device='cuda:4'), covar=tensor([0.0339, 0.0518, 0.0353, 0.0382, 0.0456, 0.0359, 0.0889, 0.0445], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0443, 0.0427, 0.0401, 0.0475, 0.0450, 0.0531, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 09:16:36,629 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.747e+02 2.265e+02 2.651e+02 3.318e+02 1.993e+03, threshold=5.302e+02, percent-clipped=3.0 2023-05-01 09:16:38,896 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9753, 2.2723, 2.3501, 2.7228, 2.1009, 3.1566, 1.7799, 2.7103], device='cuda:4'), covar=tensor([0.1143, 0.0712, 0.1105, 0.0195, 0.0124, 0.0359, 0.1395, 0.0719], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0172, 0.0192, 0.0185, 0.0199, 0.0212, 0.0200, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 09:16:43,116 INFO [train.py:904] (4/8) Epoch 22, batch 450, loss[loss=0.1386, simple_loss=0.216, pruned_loss=0.03059, over 16775.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2543, pruned_loss=0.04307, over 2962922.26 frames. ], batch size: 83, lr: 3.10e-03, grad_scale: 2.0 2023-05-01 09:17:16,858 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=213627.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:17:53,028 INFO [train.py:904] (4/8) Epoch 22, batch 500, loss[loss=0.1679, simple_loss=0.2535, pruned_loss=0.04119, over 16879.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2535, pruned_loss=0.04214, over 3037434.98 frames. ], batch size: 116, lr: 3.10e-03, grad_scale: 2.0 2023-05-01 09:18:54,885 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.471e+02 2.187e+02 2.448e+02 2.877e+02 6.308e+02, threshold=4.896e+02, percent-clipped=1.0 2023-05-01 09:19:01,577 INFO [train.py:904] (4/8) Epoch 22, batch 550, loss[loss=0.1555, simple_loss=0.2478, pruned_loss=0.03157, over 17213.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2531, pruned_loss=0.0417, over 3098751.06 frames. ], batch size: 45, lr: 3.10e-03, grad_scale: 2.0 2023-05-01 09:20:10,672 INFO [train.py:904] (4/8) Epoch 22, batch 600, loss[loss=0.1661, simple_loss=0.2441, pruned_loss=0.04408, over 16535.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2519, pruned_loss=0.04147, over 3148848.64 frames. ], batch size: 68, lr: 3.10e-03, grad_scale: 2.0 2023-05-01 09:20:47,780 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-05-01 09:21:13,517 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.232e+02 2.565e+02 3.000e+02 5.415e+02, threshold=5.129e+02, percent-clipped=2.0 2023-05-01 09:21:21,578 INFO [train.py:904] (4/8) Epoch 22, batch 650, loss[loss=0.15, simple_loss=0.2326, pruned_loss=0.03369, over 16877.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2503, pruned_loss=0.04063, over 3183978.56 frames. ], batch size: 42, lr: 3.10e-03, grad_scale: 2.0 2023-05-01 09:21:54,527 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-05-01 09:22:30,248 INFO [train.py:904] (4/8) Epoch 22, batch 700, loss[loss=0.1919, simple_loss=0.2619, pruned_loss=0.06096, over 16884.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2499, pruned_loss=0.04032, over 3208472.31 frames. ], batch size: 109, lr: 3.10e-03, grad_scale: 2.0 2023-05-01 09:23:15,774 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3848, 4.4380, 4.5407, 4.4254, 4.4215, 5.0218, 4.5317, 4.2457], device='cuda:4'), covar=tensor([0.1697, 0.2110, 0.2918, 0.2322, 0.3042, 0.1269, 0.1989, 0.2630], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0595, 0.0658, 0.0495, 0.0655, 0.0691, 0.0517, 0.0658], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 09:23:35,483 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.408e+02 2.030e+02 2.447e+02 3.057e+02 6.587e+02, threshold=4.894e+02, percent-clipped=3.0 2023-05-01 09:23:41,789 INFO [train.py:904] (4/8) Epoch 22, batch 750, loss[loss=0.18, simple_loss=0.2581, pruned_loss=0.05094, over 16529.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2503, pruned_loss=0.04074, over 3238019.22 frames. ], batch size: 146, lr: 3.10e-03, grad_scale: 2.0 2023-05-01 09:23:55,189 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3339, 3.5235, 3.9120, 2.2099, 3.1621, 2.4511, 3.8237, 3.7314], device='cuda:4'), covar=tensor([0.0278, 0.0933, 0.0461, 0.1937, 0.0818, 0.0959, 0.0561, 0.1036], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0161, 0.0166, 0.0153, 0.0144, 0.0129, 0.0143, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 09:23:58,456 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5852, 4.7100, 4.8189, 4.6714, 4.6773, 5.3227, 4.8279, 4.5386], device='cuda:4'), covar=tensor([0.1495, 0.2149, 0.2704, 0.2224, 0.2970, 0.1210, 0.1911, 0.2437], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0596, 0.0658, 0.0496, 0.0656, 0.0692, 0.0518, 0.0658], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 09:24:08,388 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=213922.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:24:26,190 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6982, 3.8481, 2.5011, 4.5146, 3.0412, 4.4149, 2.5454, 3.1115], device='cuda:4'), covar=tensor([0.0347, 0.0444, 0.1635, 0.0278, 0.0893, 0.0594, 0.1602, 0.0802], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0177, 0.0196, 0.0163, 0.0178, 0.0217, 0.0203, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 09:24:53,399 INFO [train.py:904] (4/8) Epoch 22, batch 800, loss[loss=0.1707, simple_loss=0.2463, pruned_loss=0.04753, over 16838.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2504, pruned_loss=0.04044, over 3261577.94 frames. ], batch size: 83, lr: 3.10e-03, grad_scale: 4.0 2023-05-01 09:25:56,892 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.478e+02 2.094e+02 2.441e+02 2.760e+02 5.264e+02, threshold=4.882e+02, percent-clipped=1.0 2023-05-01 09:26:06,619 INFO [train.py:904] (4/8) Epoch 22, batch 850, loss[loss=0.1777, simple_loss=0.26, pruned_loss=0.04767, over 16722.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2498, pruned_loss=0.04048, over 3278659.66 frames. ], batch size: 134, lr: 3.10e-03, grad_scale: 4.0 2023-05-01 09:26:29,078 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6776, 3.7494, 2.3960, 4.2664, 2.9359, 4.2466, 2.5039, 3.1064], device='cuda:4'), covar=tensor([0.0310, 0.0402, 0.1605, 0.0338, 0.0827, 0.0554, 0.1564, 0.0729], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0178, 0.0197, 0.0164, 0.0179, 0.0219, 0.0204, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 09:27:17,240 INFO [train.py:904] (4/8) Epoch 22, batch 900, loss[loss=0.1982, simple_loss=0.2786, pruned_loss=0.0589, over 16683.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2504, pruned_loss=0.03991, over 3290512.67 frames. ], batch size: 134, lr: 3.10e-03, grad_scale: 4.0 2023-05-01 09:28:19,798 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.353e+02 1.995e+02 2.382e+02 2.694e+02 6.707e+02, threshold=4.763e+02, percent-clipped=1.0 2023-05-01 09:28:24,211 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7424, 5.1158, 4.9027, 4.9162, 4.6622, 4.6298, 4.6164, 5.1992], device='cuda:4'), covar=tensor([0.1399, 0.0932, 0.1014, 0.0793, 0.0848, 0.1178, 0.1105, 0.0925], device='cuda:4'), in_proj_covar=tensor([0.0676, 0.0821, 0.0679, 0.0626, 0.0525, 0.0532, 0.0690, 0.0646], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:28:27,527 INFO [train.py:904] (4/8) Epoch 22, batch 950, loss[loss=0.1718, simple_loss=0.2698, pruned_loss=0.03688, over 17060.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2504, pruned_loss=0.0398, over 3300508.53 frames. ], batch size: 50, lr: 3.10e-03, grad_scale: 4.0 2023-05-01 09:29:03,321 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=214129.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:29:35,585 INFO [train.py:904] (4/8) Epoch 22, batch 1000, loss[loss=0.1397, simple_loss=0.2234, pruned_loss=0.028, over 16657.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2494, pruned_loss=0.03948, over 3300340.89 frames. ], batch size: 68, lr: 3.10e-03, grad_scale: 4.0 2023-05-01 09:30:15,722 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1554, 2.7020, 2.1805, 2.4028, 3.0265, 2.7551, 3.0558, 3.0957], device='cuda:4'), covar=tensor([0.0258, 0.0444, 0.0615, 0.0499, 0.0300, 0.0382, 0.0313, 0.0317], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0238, 0.0228, 0.0229, 0.0239, 0.0237, 0.0238, 0.0233], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:30:20,601 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1885, 5.1958, 4.9665, 4.4307, 5.0485, 1.9801, 4.7732, 4.8188], device='cuda:4'), covar=tensor([0.0089, 0.0079, 0.0204, 0.0403, 0.0109, 0.2761, 0.0149, 0.0225], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0155, 0.0197, 0.0174, 0.0175, 0.0208, 0.0186, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:30:29,093 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=214190.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:30:39,708 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.418e+02 2.141e+02 2.430e+02 2.890e+02 5.407e+02, threshold=4.860e+02, percent-clipped=2.0 2023-05-01 09:30:46,635 INFO [train.py:904] (4/8) Epoch 22, batch 1050, loss[loss=0.1641, simple_loss=0.2547, pruned_loss=0.0368, over 17207.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2496, pruned_loss=0.03981, over 3308672.70 frames. ], batch size: 45, lr: 3.10e-03, grad_scale: 4.0 2023-05-01 09:30:51,616 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3935, 3.6151, 3.8597, 2.0840, 3.0112, 2.4762, 3.8392, 3.7758], device='cuda:4'), covar=tensor([0.0308, 0.0899, 0.0550, 0.2074, 0.0919, 0.0994, 0.0626, 0.1026], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0167, 0.0154, 0.0145, 0.0130, 0.0144, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 09:31:07,763 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=214218.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:31:13,762 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=214222.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:31:56,094 INFO [train.py:904] (4/8) Epoch 22, batch 1100, loss[loss=0.1771, simple_loss=0.2442, pruned_loss=0.05502, over 16869.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2488, pruned_loss=0.03964, over 3310816.48 frames. ], batch size: 109, lr: 3.10e-03, grad_scale: 4.0 2023-05-01 09:32:19,668 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=214270.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:32:31,197 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=214279.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:32:57,752 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.499e+02 2.112e+02 2.466e+02 2.921e+02 8.908e+02, threshold=4.932e+02, percent-clipped=7.0 2023-05-01 09:33:03,837 INFO [train.py:904] (4/8) Epoch 22, batch 1150, loss[loss=0.1455, simple_loss=0.2309, pruned_loss=0.03004, over 16990.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2477, pruned_loss=0.03891, over 3304257.95 frames. ], batch size: 41, lr: 3.10e-03, grad_scale: 4.0 2023-05-01 09:34:15,635 INFO [train.py:904] (4/8) Epoch 22, batch 1200, loss[loss=0.1824, simple_loss=0.2546, pruned_loss=0.05515, over 16756.00 frames. ], tot_loss[loss=0.1617, simple_loss=0.2465, pruned_loss=0.03842, over 3298355.98 frames. ], batch size: 134, lr: 3.10e-03, grad_scale: 8.0 2023-05-01 09:34:54,174 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3805, 4.1731, 4.4260, 4.5639, 4.6504, 4.2176, 4.4448, 4.6354], device='cuda:4'), covar=tensor([0.1606, 0.1293, 0.1265, 0.0694, 0.0656, 0.1144, 0.3156, 0.0807], device='cuda:4'), in_proj_covar=tensor([0.0648, 0.0796, 0.0927, 0.0809, 0.0611, 0.0641, 0.0665, 0.0771], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:35:18,107 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.558e+02 2.085e+02 2.414e+02 2.922e+02 4.636e+02, threshold=4.828e+02, percent-clipped=0.0 2023-05-01 09:35:25,090 INFO [train.py:904] (4/8) Epoch 22, batch 1250, loss[loss=0.1669, simple_loss=0.2519, pruned_loss=0.04095, over 16540.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2471, pruned_loss=0.03885, over 3305902.23 frames. ], batch size: 68, lr: 3.10e-03, grad_scale: 8.0 2023-05-01 09:36:35,040 INFO [train.py:904] (4/8) Epoch 22, batch 1300, loss[loss=0.16, simple_loss=0.2495, pruned_loss=0.03527, over 17201.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2468, pruned_loss=0.03874, over 3313115.62 frames. ], batch size: 45, lr: 3.10e-03, grad_scale: 8.0 2023-05-01 09:37:18,456 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=214485.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:37:34,614 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.423e+02 2.092e+02 2.483e+02 2.893e+02 4.900e+02, threshold=4.965e+02, percent-clipped=1.0 2023-05-01 09:37:42,347 INFO [train.py:904] (4/8) Epoch 22, batch 1350, loss[loss=0.1462, simple_loss=0.2308, pruned_loss=0.0308, over 16966.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2474, pruned_loss=0.03855, over 3319806.17 frames. ], batch size: 41, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:37:55,051 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8480, 4.4420, 3.2084, 2.3207, 2.7955, 2.7101, 4.8062, 3.6598], device='cuda:4'), covar=tensor([0.2754, 0.0540, 0.1643, 0.2871, 0.2884, 0.2091, 0.0301, 0.1434], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0271, 0.0307, 0.0313, 0.0297, 0.0261, 0.0295, 0.0338], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 09:38:48,216 INFO [train.py:904] (4/8) Epoch 22, batch 1400, loss[loss=0.1524, simple_loss=0.2401, pruned_loss=0.03234, over 17208.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2478, pruned_loss=0.03863, over 3327457.18 frames. ], batch size: 44, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:39:17,566 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=214574.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:39:50,087 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.375e+02 2.086e+02 2.334e+02 2.820e+02 4.887e+02, threshold=4.669e+02, percent-clipped=0.0 2023-05-01 09:39:57,056 INFO [train.py:904] (4/8) Epoch 22, batch 1450, loss[loss=0.1692, simple_loss=0.248, pruned_loss=0.04524, over 16870.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2472, pruned_loss=0.03847, over 3329543.65 frames. ], batch size: 116, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:41:07,194 INFO [train.py:904] (4/8) Epoch 22, batch 1500, loss[loss=0.1449, simple_loss=0.2429, pruned_loss=0.02351, over 16832.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2465, pruned_loss=0.03826, over 3332501.61 frames. ], batch size: 42, lr: 3.09e-03, grad_scale: 4.0 2023-05-01 09:42:07,406 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-05-01 09:42:10,190 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.429e+02 2.149e+02 2.472e+02 3.063e+02 5.864e+02, threshold=4.945e+02, percent-clipped=3.0 2023-05-01 09:42:16,354 INFO [train.py:904] (4/8) Epoch 22, batch 1550, loss[loss=0.201, simple_loss=0.2724, pruned_loss=0.06485, over 16739.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2485, pruned_loss=0.03965, over 3330118.79 frames. ], batch size: 89, lr: 3.09e-03, grad_scale: 4.0 2023-05-01 09:43:04,913 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8491, 2.4334, 2.0183, 2.3037, 2.9152, 2.6575, 2.9067, 2.9925], device='cuda:4'), covar=tensor([0.0188, 0.0428, 0.0551, 0.0446, 0.0210, 0.0314, 0.0228, 0.0284], device='cuda:4'), in_proj_covar=tensor([0.0216, 0.0240, 0.0230, 0.0230, 0.0240, 0.0239, 0.0242, 0.0236], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:43:27,996 INFO [train.py:904] (4/8) Epoch 22, batch 1600, loss[loss=0.1907, simple_loss=0.2729, pruned_loss=0.05428, over 16201.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2513, pruned_loss=0.04095, over 3316380.00 frames. ], batch size: 165, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:43:57,802 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9544, 3.1437, 3.1661, 2.0599, 2.9181, 3.2352, 3.0380, 1.9574], device='cuda:4'), covar=tensor([0.0571, 0.0114, 0.0078, 0.0476, 0.0141, 0.0118, 0.0115, 0.0487], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0084, 0.0084, 0.0134, 0.0098, 0.0109, 0.0095, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 09:44:12,493 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=214785.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:44:32,064 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.512e+02 2.192e+02 2.616e+02 3.002e+02 5.148e+02, threshold=5.232e+02, percent-clipped=2.0 2023-05-01 09:44:37,871 INFO [train.py:904] (4/8) Epoch 22, batch 1650, loss[loss=0.1584, simple_loss=0.2441, pruned_loss=0.03635, over 17221.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2529, pruned_loss=0.04122, over 3319015.12 frames. ], batch size: 44, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:44:50,439 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9422, 3.6701, 4.1071, 2.2050, 4.3324, 4.4250, 3.2181, 3.4016], device='cuda:4'), covar=tensor([0.0681, 0.0249, 0.0244, 0.1087, 0.0073, 0.0193, 0.0434, 0.0374], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0108, 0.0097, 0.0138, 0.0080, 0.0126, 0.0129, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 09:45:20,795 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=214833.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:45:39,508 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-05-01 09:45:47,643 INFO [train.py:904] (4/8) Epoch 22, batch 1700, loss[loss=0.1848, simple_loss=0.2658, pruned_loss=0.05189, over 16786.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2547, pruned_loss=0.04158, over 3326125.92 frames. ], batch size: 83, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:46:18,478 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=214874.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:46:53,102 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.696e+02 2.472e+02 2.965e+02 3.692e+02 1.718e+03, threshold=5.930e+02, percent-clipped=4.0 2023-05-01 09:46:58,548 INFO [train.py:904] (4/8) Epoch 22, batch 1750, loss[loss=0.1718, simple_loss=0.2617, pruned_loss=0.04092, over 17074.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2547, pruned_loss=0.04123, over 3321979.89 frames. ], batch size: 53, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:47:06,243 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5911, 3.7218, 2.2617, 4.0500, 2.8496, 3.9575, 2.4366, 2.9922], device='cuda:4'), covar=tensor([0.0317, 0.0385, 0.1651, 0.0330, 0.0859, 0.0824, 0.1408, 0.0743], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0181, 0.0198, 0.0168, 0.0181, 0.0223, 0.0206, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 09:47:25,706 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=214922.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:48:06,727 INFO [train.py:904] (4/8) Epoch 22, batch 1800, loss[loss=0.1816, simple_loss=0.2664, pruned_loss=0.04839, over 16532.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2558, pruned_loss=0.04139, over 3321237.73 frames. ], batch size: 75, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:48:31,714 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0909, 4.4772, 4.5149, 3.2777, 3.7668, 4.4174, 4.0131, 2.8122], device='cuda:4'), covar=tensor([0.0386, 0.0058, 0.0039, 0.0324, 0.0123, 0.0091, 0.0095, 0.0389], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0083, 0.0084, 0.0134, 0.0098, 0.0109, 0.0095, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 09:49:03,175 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-01 09:49:13,316 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.476e+02 2.061e+02 2.528e+02 2.863e+02 4.992e+02, threshold=5.055e+02, percent-clipped=0.0 2023-05-01 09:49:18,204 INFO [train.py:904] (4/8) Epoch 22, batch 1850, loss[loss=0.1883, simple_loss=0.2712, pruned_loss=0.05276, over 11995.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2565, pruned_loss=0.04146, over 3324071.92 frames. ], batch size: 246, lr: 3.09e-03, grad_scale: 4.0 2023-05-01 09:49:34,414 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7051, 2.3132, 1.8406, 2.1633, 2.7323, 2.5202, 2.7353, 2.8516], device='cuda:4'), covar=tensor([0.0207, 0.0443, 0.0603, 0.0500, 0.0253, 0.0328, 0.0220, 0.0299], device='cuda:4'), in_proj_covar=tensor([0.0218, 0.0241, 0.0232, 0.0232, 0.0242, 0.0241, 0.0244, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:49:54,708 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 09:50:27,231 INFO [train.py:904] (4/8) Epoch 22, batch 1900, loss[loss=0.152, simple_loss=0.2488, pruned_loss=0.02757, over 17138.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2557, pruned_loss=0.04114, over 3308926.54 frames. ], batch size: 47, lr: 3.09e-03, grad_scale: 4.0 2023-05-01 09:50:41,466 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-01 09:51:31,792 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.538e+02 2.086e+02 2.418e+02 2.993e+02 8.338e+02, threshold=4.836e+02, percent-clipped=2.0 2023-05-01 09:51:36,123 INFO [train.py:904] (4/8) Epoch 22, batch 1950, loss[loss=0.1324, simple_loss=0.2216, pruned_loss=0.02158, over 16759.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2557, pruned_loss=0.04057, over 3318478.47 frames. ], batch size: 39, lr: 3.09e-03, grad_scale: 4.0 2023-05-01 09:52:44,789 INFO [train.py:904] (4/8) Epoch 22, batch 2000, loss[loss=0.1789, simple_loss=0.2706, pruned_loss=0.04366, over 16733.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2552, pruned_loss=0.0403, over 3323401.16 frames. ], batch size: 57, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:53:48,736 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.559e+02 2.173e+02 2.667e+02 3.241e+02 6.066e+02, threshold=5.334e+02, percent-clipped=3.0 2023-05-01 09:53:53,918 INFO [train.py:904] (4/8) Epoch 22, batch 2050, loss[loss=0.1928, simple_loss=0.2612, pruned_loss=0.06215, over 16737.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2557, pruned_loss=0.04097, over 3318403.55 frames. ], batch size: 124, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:54:31,415 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=215230.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 09:55:01,197 INFO [train.py:904] (4/8) Epoch 22, batch 2100, loss[loss=0.2092, simple_loss=0.2855, pruned_loss=0.06649, over 16876.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2567, pruned_loss=0.0417, over 3321764.41 frames. ], batch size: 109, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:55:54,373 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=215291.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 09:56:04,628 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.468e+02 2.033e+02 2.346e+02 2.893e+02 7.245e+02, threshold=4.692e+02, percent-clipped=2.0 2023-05-01 09:56:09,028 INFO [train.py:904] (4/8) Epoch 22, batch 2150, loss[loss=0.1699, simple_loss=0.2664, pruned_loss=0.03666, over 16804.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2568, pruned_loss=0.04158, over 3327145.20 frames. ], batch size: 62, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:56:28,645 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7207, 4.5402, 4.8101, 4.9704, 5.1405, 4.5532, 5.1158, 5.1511], device='cuda:4'), covar=tensor([0.1935, 0.1346, 0.1726, 0.0746, 0.0597, 0.1032, 0.0759, 0.0624], device='cuda:4'), in_proj_covar=tensor([0.0656, 0.0810, 0.0947, 0.0819, 0.0618, 0.0651, 0.0672, 0.0779], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:56:39,624 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=215325.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 09:56:50,999 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=215333.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 09:57:18,610 INFO [train.py:904] (4/8) Epoch 22, batch 2200, loss[loss=0.1744, simple_loss=0.2559, pruned_loss=0.04645, over 16856.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2573, pruned_loss=0.04192, over 3325640.43 frames. ], batch size: 42, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:58:02,283 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=215386.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 09:58:14,366 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=215394.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 09:58:20,772 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.171e+02 2.606e+02 3.111e+02 9.974e+02, threshold=5.211e+02, percent-clipped=4.0 2023-05-01 09:58:24,659 INFO [train.py:904] (4/8) Epoch 22, batch 2250, loss[loss=0.1639, simple_loss=0.2565, pruned_loss=0.03569, over 17074.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2574, pruned_loss=0.04219, over 3330042.98 frames. ], batch size: 53, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:58:33,302 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4837, 5.4715, 5.1825, 4.7024, 5.2833, 2.2432, 5.0293, 5.1387], device='cuda:4'), covar=tensor([0.0080, 0.0065, 0.0195, 0.0356, 0.0097, 0.2436, 0.0128, 0.0180], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0158, 0.0200, 0.0179, 0.0179, 0.0209, 0.0190, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 09:59:36,632 INFO [train.py:904] (4/8) Epoch 22, batch 2300, loss[loss=0.1663, simple_loss=0.2652, pruned_loss=0.0337, over 17117.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2573, pruned_loss=0.04233, over 3327739.95 frames. ], batch size: 49, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 09:59:59,948 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8398, 4.1750, 3.1549, 2.4188, 2.7075, 2.6665, 4.5398, 3.5623], device='cuda:4'), covar=tensor([0.2709, 0.0565, 0.1712, 0.2852, 0.2731, 0.1921, 0.0384, 0.1293], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0270, 0.0306, 0.0313, 0.0297, 0.0261, 0.0295, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:00:42,879 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.468e+02 2.161e+02 2.569e+02 3.134e+02 6.590e+02, threshold=5.137e+02, percent-clipped=2.0 2023-05-01 10:00:46,496 INFO [train.py:904] (4/8) Epoch 22, batch 2350, loss[loss=0.1617, simple_loss=0.2543, pruned_loss=0.03456, over 17073.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.257, pruned_loss=0.0417, over 3332990.24 frames. ], batch size: 53, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 10:00:46,898 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=215503.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:01:31,151 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-05-01 10:01:33,961 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-01 10:01:39,943 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-05-01 10:01:48,825 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=215548.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:01:55,001 INFO [train.py:904] (4/8) Epoch 22, batch 2400, loss[loss=0.1849, simple_loss=0.2608, pruned_loss=0.05449, over 16867.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2583, pruned_loss=0.04204, over 3337946.67 frames. ], batch size: 116, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 10:02:11,580 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=215564.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:02:12,658 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=215565.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:02:27,817 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9063, 3.0480, 2.7559, 5.0455, 3.9424, 4.4393, 1.8737, 3.2639], device='cuda:4'), covar=tensor([0.1420, 0.0810, 0.1250, 0.0218, 0.0314, 0.0466, 0.1696, 0.0859], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0194, 0.0192, 0.0205, 0.0218, 0.0203, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 10:02:34,339 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-01 10:02:41,691 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=215586.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:02:59,708 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.274e+02 2.111e+02 2.344e+02 3.035e+02 7.498e+02, threshold=4.688e+02, percent-clipped=4.0 2023-05-01 10:03:04,615 INFO [train.py:904] (4/8) Epoch 22, batch 2450, loss[loss=0.1585, simple_loss=0.2528, pruned_loss=0.03206, over 17218.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2588, pruned_loss=0.04177, over 3338981.37 frames. ], batch size: 44, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 10:03:12,054 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=215609.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 10:03:30,365 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-01 10:03:35,698 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=215626.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:04:11,712 INFO [train.py:904] (4/8) Epoch 22, batch 2500, loss[loss=0.1831, simple_loss=0.2661, pruned_loss=0.05004, over 16073.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2581, pruned_loss=0.04138, over 3340803.32 frames. ], batch size: 35, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 10:04:13,190 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7966, 4.8747, 5.0316, 4.8591, 4.8574, 5.5057, 4.9516, 4.5994], device='cuda:4'), covar=tensor([0.1430, 0.2142, 0.2246, 0.2208, 0.2700, 0.1095, 0.1850, 0.2714], device='cuda:4'), in_proj_covar=tensor([0.0424, 0.0620, 0.0683, 0.0516, 0.0687, 0.0717, 0.0537, 0.0687], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:04:38,366 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9043, 4.2881, 4.2303, 3.0600, 3.7297, 4.2108, 3.9367, 2.0226], device='cuda:4'), covar=tensor([0.0563, 0.0108, 0.0074, 0.0453, 0.0191, 0.0156, 0.0126, 0.0743], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0084, 0.0084, 0.0134, 0.0099, 0.0109, 0.0095, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:04:51,429 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=215681.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 10:05:02,669 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=215689.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 10:05:11,784 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5288, 3.5181, 3.4453, 2.7783, 3.2497, 2.1007, 3.0884, 2.6744], device='cuda:4'), covar=tensor([0.0153, 0.0130, 0.0180, 0.0227, 0.0111, 0.2389, 0.0136, 0.0261], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0158, 0.0201, 0.0179, 0.0179, 0.0210, 0.0191, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:05:15,511 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.601e+02 2.254e+02 2.666e+02 2.993e+02 6.188e+02, threshold=5.333e+02, percent-clipped=1.0 2023-05-01 10:05:20,561 INFO [train.py:904] (4/8) Epoch 22, batch 2550, loss[loss=0.1566, simple_loss=0.2504, pruned_loss=0.03137, over 17174.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2581, pruned_loss=0.042, over 3333575.19 frames. ], batch size: 46, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 10:05:23,499 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9433, 4.5299, 3.3704, 2.4642, 2.8275, 2.7970, 4.9382, 3.7541], device='cuda:4'), covar=tensor([0.2947, 0.0576, 0.1677, 0.2927, 0.3094, 0.2074, 0.0343, 0.1500], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0272, 0.0308, 0.0316, 0.0299, 0.0262, 0.0296, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:06:25,143 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3494, 3.7192, 4.3746, 2.1112, 4.5546, 4.7260, 3.3574, 3.6429], device='cuda:4'), covar=tensor([0.0582, 0.0288, 0.0207, 0.1184, 0.0070, 0.0155, 0.0428, 0.0357], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0099, 0.0140, 0.0081, 0.0128, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 10:06:30,382 INFO [train.py:904] (4/8) Epoch 22, batch 2600, loss[loss=0.1597, simple_loss=0.2396, pruned_loss=0.03989, over 16822.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2579, pruned_loss=0.04131, over 3333567.41 frames. ], batch size: 83, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 10:07:01,345 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 10:07:36,118 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.534e+02 2.223e+02 2.543e+02 2.941e+02 8.287e+02, threshold=5.085e+02, percent-clipped=5.0 2023-05-01 10:07:39,997 INFO [train.py:904] (4/8) Epoch 22, batch 2650, loss[loss=0.1974, simple_loss=0.2795, pruned_loss=0.05765, over 16682.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2591, pruned_loss=0.04142, over 3336914.29 frames. ], batch size: 124, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 10:08:16,623 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=215829.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:08:48,042 INFO [train.py:904] (4/8) Epoch 22, batch 2700, loss[loss=0.1648, simple_loss=0.2517, pruned_loss=0.03898, over 17189.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2584, pruned_loss=0.04071, over 3335061.56 frames. ], batch size: 44, lr: 3.09e-03, grad_scale: 8.0 2023-05-01 10:08:57,510 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=215859.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:09:22,095 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-01 10:09:34,773 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=215886.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 10:09:39,275 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=215890.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:09:53,922 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.480e+02 2.142e+02 2.465e+02 2.972e+02 5.160e+02, threshold=4.929e+02, percent-clipped=1.0 2023-05-01 10:09:57,373 INFO [train.py:904] (4/8) Epoch 22, batch 2750, loss[loss=0.1779, simple_loss=0.2607, pruned_loss=0.04755, over 16884.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2583, pruned_loss=0.04051, over 3336328.40 frames. ], batch size: 90, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:09:59,453 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=215904.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:10:20,127 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6725, 2.3155, 1.9432, 2.2501, 2.7228, 2.5191, 2.6516, 2.8442], device='cuda:4'), covar=tensor([0.0225, 0.0430, 0.0535, 0.0449, 0.0226, 0.0334, 0.0208, 0.0264], device='cuda:4'), in_proj_covar=tensor([0.0218, 0.0240, 0.0230, 0.0231, 0.0241, 0.0240, 0.0243, 0.0238], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:10:21,930 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=215921.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:10:38,795 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=215934.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:11:05,009 INFO [train.py:904] (4/8) Epoch 22, batch 2800, loss[loss=0.1842, simple_loss=0.2741, pruned_loss=0.04716, over 16660.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2581, pruned_loss=0.04062, over 3337786.51 frames. ], batch size: 57, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:11:42,292 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=215981.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 10:11:53,839 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=215989.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 10:12:07,418 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.162e+02 2.449e+02 3.092e+02 5.987e+02, threshold=4.899e+02, percent-clipped=3.0 2023-05-01 10:12:14,945 INFO [train.py:904] (4/8) Epoch 22, batch 2850, loss[loss=0.1736, simple_loss=0.2488, pruned_loss=0.04918, over 16760.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2578, pruned_loss=0.04046, over 3338823.53 frames. ], batch size: 124, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:12:51,478 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=216028.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:12:52,493 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=216029.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:13:04,677 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=216037.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:13:20,002 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 10:13:26,222 INFO [train.py:904] (4/8) Epoch 22, batch 2900, loss[loss=0.1734, simple_loss=0.2688, pruned_loss=0.03905, over 17065.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2565, pruned_loss=0.0406, over 3338837.41 frames. ], batch size: 55, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:13:33,727 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6948, 2.6637, 2.3953, 2.6690, 2.9845, 2.8518, 3.2826, 3.2153], device='cuda:4'), covar=tensor([0.0157, 0.0454, 0.0521, 0.0443, 0.0289, 0.0380, 0.0235, 0.0301], device='cuda:4'), in_proj_covar=tensor([0.0218, 0.0240, 0.0230, 0.0230, 0.0241, 0.0240, 0.0243, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:13:38,342 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7141, 4.7833, 4.9679, 4.7833, 4.7597, 5.4427, 4.9548, 4.6691], device='cuda:4'), covar=tensor([0.1441, 0.2332, 0.2602, 0.2159, 0.2861, 0.1057, 0.1758, 0.2719], device='cuda:4'), in_proj_covar=tensor([0.0423, 0.0615, 0.0676, 0.0513, 0.0682, 0.0712, 0.0533, 0.0681], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:14:17,795 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=216089.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:14:31,569 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.538e+02 2.264e+02 2.773e+02 3.437e+02 6.159e+02, threshold=5.547e+02, percent-clipped=1.0 2023-05-01 10:14:35,745 INFO [train.py:904] (4/8) Epoch 22, batch 2950, loss[loss=0.1603, simple_loss=0.2533, pruned_loss=0.03371, over 17145.00 frames. ], tot_loss[loss=0.17, simple_loss=0.257, pruned_loss=0.04152, over 3336582.99 frames. ], batch size: 49, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:15:45,385 INFO [train.py:904] (4/8) Epoch 22, batch 3000, loss[loss=0.1648, simple_loss=0.2619, pruned_loss=0.03383, over 17052.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2573, pruned_loss=0.04159, over 3337129.47 frames. ], batch size: 50, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:15:45,385 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 10:15:50,207 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4849, 5.4130, 5.1111, 4.6048, 5.2379, 2.3972, 5.0865, 4.9498], device='cuda:4'), covar=tensor([0.0059, 0.0060, 0.0211, 0.0388, 0.0092, 0.2823, 0.0120, 0.0261], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0159, 0.0202, 0.0181, 0.0181, 0.0211, 0.0192, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:15:54,107 INFO [train.py:938] (4/8) Epoch 22, validation: loss=0.1347, simple_loss=0.2399, pruned_loss=0.0148, over 944034.00 frames. 2023-05-01 10:15:54,107 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 10:16:02,783 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=216159.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:16:38,785 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=216185.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:16:50,633 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9297, 3.0287, 2.9361, 5.1388, 4.2530, 4.4564, 1.6661, 3.2950], device='cuda:4'), covar=tensor([0.1328, 0.0720, 0.1106, 0.0193, 0.0255, 0.0418, 0.1639, 0.0755], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0174, 0.0194, 0.0193, 0.0204, 0.0218, 0.0202, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 10:16:59,839 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.653e+02 2.146e+02 2.550e+02 3.086e+02 1.158e+03, threshold=5.100e+02, percent-clipped=1.0 2023-05-01 10:17:03,727 INFO [train.py:904] (4/8) Epoch 22, batch 3050, loss[loss=0.141, simple_loss=0.2313, pruned_loss=0.02532, over 16814.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2566, pruned_loss=0.04146, over 3332982.53 frames. ], batch size: 39, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:17:05,182 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=216204.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:17:09,668 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=216207.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:17:29,262 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=216221.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:18:10,763 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=216252.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:18:12,222 INFO [train.py:904] (4/8) Epoch 22, batch 3100, loss[loss=0.172, simple_loss=0.2628, pruned_loss=0.04059, over 17119.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2563, pruned_loss=0.04196, over 3333324.75 frames. ], batch size: 47, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:18:33,891 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=216269.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:18:36,969 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1442, 3.3891, 3.5607, 2.3676, 3.2767, 3.6598, 3.3567, 2.0291], device='cuda:4'), covar=tensor([0.0545, 0.0118, 0.0064, 0.0412, 0.0117, 0.0099, 0.0096, 0.0474], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0084, 0.0084, 0.0133, 0.0099, 0.0109, 0.0095, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:19:16,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.355e+02 2.131e+02 2.523e+02 3.440e+02 6.792e+02, threshold=5.047e+02, percent-clipped=4.0 2023-05-01 10:19:21,072 INFO [train.py:904] (4/8) Epoch 22, batch 3150, loss[loss=0.186, simple_loss=0.2771, pruned_loss=0.04749, over 17053.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.256, pruned_loss=0.04196, over 3335248.05 frames. ], batch size: 53, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:19:43,787 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0262, 3.1080, 2.6778, 2.9231, 3.4329, 3.1377, 3.6774, 3.5527], device='cuda:4'), covar=tensor([0.0148, 0.0346, 0.0473, 0.0406, 0.0221, 0.0324, 0.0235, 0.0222], device='cuda:4'), in_proj_covar=tensor([0.0220, 0.0241, 0.0231, 0.0232, 0.0242, 0.0242, 0.0246, 0.0239], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:20:29,459 INFO [train.py:904] (4/8) Epoch 22, batch 3200, loss[loss=0.1573, simple_loss=0.2407, pruned_loss=0.03688, over 15509.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2551, pruned_loss=0.04151, over 3330804.18 frames. ], batch size: 190, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:21:13,818 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=216384.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:21:36,409 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.593e+02 2.182e+02 2.419e+02 2.918e+02 5.281e+02, threshold=4.837e+02, percent-clipped=1.0 2023-05-01 10:21:40,317 INFO [train.py:904] (4/8) Epoch 22, batch 3250, loss[loss=0.1761, simple_loss=0.2571, pruned_loss=0.04752, over 16806.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2549, pruned_loss=0.04162, over 3330661.04 frames. ], batch size: 102, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:22:11,953 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-05-01 10:22:52,219 INFO [train.py:904] (4/8) Epoch 22, batch 3300, loss[loss=0.1753, simple_loss=0.252, pruned_loss=0.0493, over 16729.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2562, pruned_loss=0.04187, over 3328000.63 frames. ], batch size: 83, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:23:01,882 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5018, 2.3491, 2.3279, 4.4154, 2.3270, 2.7195, 2.3940, 2.5567], device='cuda:4'), covar=tensor([0.1271, 0.3845, 0.3250, 0.0519, 0.4298, 0.2870, 0.3868, 0.3715], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0453, 0.0373, 0.0333, 0.0441, 0.0522, 0.0425, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:23:36,594 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=216485.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:23:56,628 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.388e+02 2.110e+02 2.633e+02 3.149e+02 8.538e+02, threshold=5.267e+02, percent-clipped=2.0 2023-05-01 10:24:00,075 INFO [train.py:904] (4/8) Epoch 22, batch 3350, loss[loss=0.1496, simple_loss=0.2405, pruned_loss=0.02933, over 17222.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2571, pruned_loss=0.04195, over 3326407.92 frames. ], batch size: 44, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:24:42,589 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=216533.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:25:10,643 INFO [train.py:904] (4/8) Epoch 22, batch 3400, loss[loss=0.1819, simple_loss=0.2582, pruned_loss=0.05279, over 16866.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2569, pruned_loss=0.0412, over 3333266.96 frames. ], batch size: 116, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:25:20,888 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7801, 2.4709, 2.3072, 3.3476, 2.3794, 3.5382, 1.6472, 2.6605], device='cuda:4'), covar=tensor([0.1419, 0.0796, 0.1247, 0.0228, 0.0179, 0.0430, 0.1692, 0.0910], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0195, 0.0194, 0.0206, 0.0219, 0.0203, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 10:26:02,841 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=216591.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:26:15,739 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.115e+02 2.508e+02 2.941e+02 4.033e+02, threshold=5.017e+02, percent-clipped=0.0 2023-05-01 10:26:19,664 INFO [train.py:904] (4/8) Epoch 22, batch 3450, loss[loss=0.1562, simple_loss=0.2375, pruned_loss=0.03744, over 16884.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2556, pruned_loss=0.04102, over 3330635.42 frames. ], batch size: 90, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:27:29,278 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=216652.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:27:29,991 INFO [train.py:904] (4/8) Epoch 22, batch 3500, loss[loss=0.1751, simple_loss=0.2494, pruned_loss=0.05038, over 16894.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2542, pruned_loss=0.04081, over 3336839.43 frames. ], batch size: 116, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:27:32,745 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=216655.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:27:44,532 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-05-01 10:28:12,580 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=216684.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:28:35,764 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.548e+02 1.992e+02 2.337e+02 2.852e+02 6.760e+02, threshold=4.673e+02, percent-clipped=2.0 2023-05-01 10:28:39,276 INFO [train.py:904] (4/8) Epoch 22, batch 3550, loss[loss=0.1559, simple_loss=0.2373, pruned_loss=0.03722, over 16917.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2531, pruned_loss=0.04054, over 3336835.22 frames. ], batch size: 96, lr: 3.08e-03, grad_scale: 4.0 2023-05-01 10:28:42,525 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=216705.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:28:57,312 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=216716.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 10:29:20,498 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=216732.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:29:49,493 INFO [train.py:904] (4/8) Epoch 22, batch 3600, loss[loss=0.1699, simple_loss=0.2609, pruned_loss=0.03949, over 17019.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2521, pruned_loss=0.04002, over 3335494.69 frames. ], batch size: 55, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:30:08,399 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=216766.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:30:51,316 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 10:31:00,613 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.582e+02 2.119e+02 2.478e+02 2.922e+02 5.214e+02, threshold=4.955e+02, percent-clipped=3.0 2023-05-01 10:31:03,547 INFO [train.py:904] (4/8) Epoch 22, batch 3650, loss[loss=0.1753, simple_loss=0.2726, pruned_loss=0.03894, over 16993.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2517, pruned_loss=0.04034, over 3319248.41 frames. ], batch size: 55, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:32:13,473 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-05-01 10:32:18,423 INFO [train.py:904] (4/8) Epoch 22, batch 3700, loss[loss=0.1676, simple_loss=0.2426, pruned_loss=0.04634, over 16895.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.251, pruned_loss=0.04198, over 3296436.45 frames. ], batch size: 116, lr: 3.08e-03, grad_scale: 4.0 2023-05-01 10:32:52,389 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-01 10:32:54,288 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0194, 5.0048, 4.8588, 4.3596, 4.9682, 2.0856, 4.7472, 4.5564], device='cuda:4'), covar=tensor([0.0095, 0.0093, 0.0191, 0.0333, 0.0086, 0.2593, 0.0116, 0.0209], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0161, 0.0205, 0.0182, 0.0183, 0.0212, 0.0194, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:33:31,443 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.634e+02 2.238e+02 2.459e+02 3.071e+02 7.166e+02, threshold=4.918e+02, percent-clipped=1.0 2023-05-01 10:33:32,645 INFO [train.py:904] (4/8) Epoch 22, batch 3750, loss[loss=0.1603, simple_loss=0.2404, pruned_loss=0.04005, over 16794.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2522, pruned_loss=0.04361, over 3277076.96 frames. ], batch size: 83, lr: 3.08e-03, grad_scale: 4.0 2023-05-01 10:34:11,424 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3638, 3.5441, 3.6479, 3.6290, 3.6370, 3.4797, 3.5027, 3.5166], device='cuda:4'), covar=tensor([0.0410, 0.0676, 0.0460, 0.0453, 0.0602, 0.0495, 0.0787, 0.0528], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0467, 0.0453, 0.0420, 0.0502, 0.0477, 0.0562, 0.0382], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 10:34:31,493 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6722, 2.7636, 2.6768, 4.7908, 2.5093, 3.0234, 2.8185, 2.9488], device='cuda:4'), covar=tensor([0.1174, 0.2919, 0.2567, 0.0311, 0.3609, 0.2117, 0.2916, 0.2719], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0455, 0.0374, 0.0334, 0.0442, 0.0525, 0.0426, 0.0534], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:34:36,639 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=216947.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:34:44,481 INFO [train.py:904] (4/8) Epoch 22, batch 3800, loss[loss=0.1628, simple_loss=0.2393, pruned_loss=0.04317, over 16682.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2536, pruned_loss=0.04503, over 3271969.98 frames. ], batch size: 76, lr: 3.08e-03, grad_scale: 4.0 2023-05-01 10:35:03,065 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=216965.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:35:55,032 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.604e+02 2.325e+02 2.685e+02 3.161e+02 5.857e+02, threshold=5.370e+02, percent-clipped=3.0 2023-05-01 10:35:56,826 INFO [train.py:904] (4/8) Epoch 22, batch 3850, loss[loss=0.1968, simple_loss=0.2708, pruned_loss=0.06136, over 15663.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2539, pruned_loss=0.0458, over 3266733.70 frames. ], batch size: 190, lr: 3.08e-03, grad_scale: 4.0 2023-05-01 10:36:08,427 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=217011.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 10:36:29,882 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=217026.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:37:09,496 INFO [train.py:904] (4/8) Epoch 22, batch 3900, loss[loss=0.1747, simple_loss=0.2503, pruned_loss=0.04958, over 16916.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2531, pruned_loss=0.04612, over 3269091.72 frames. ], batch size: 116, lr: 3.08e-03, grad_scale: 4.0 2023-05-01 10:37:22,071 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=217061.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:37:40,291 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8406, 5.1334, 4.9132, 4.9302, 4.7116, 4.5786, 4.6015, 5.2181], device='cuda:4'), covar=tensor([0.1160, 0.0789, 0.0933, 0.0852, 0.0781, 0.1093, 0.1158, 0.0844], device='cuda:4'), in_proj_covar=tensor([0.0696, 0.0853, 0.0701, 0.0647, 0.0540, 0.0548, 0.0714, 0.0667], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:38:21,560 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.630e+02 2.294e+02 2.626e+02 3.091e+02 8.707e+02, threshold=5.252e+02, percent-clipped=1.0 2023-05-01 10:38:22,850 INFO [train.py:904] (4/8) Epoch 22, batch 3950, loss[loss=0.1551, simple_loss=0.2436, pruned_loss=0.03327, over 17061.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2526, pruned_loss=0.0467, over 3274302.59 frames. ], batch size: 50, lr: 3.08e-03, grad_scale: 4.0 2023-05-01 10:38:41,452 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9408, 5.3366, 5.5552, 5.2480, 5.3660, 5.9209, 5.3751, 5.0220], device='cuda:4'), covar=tensor([0.1008, 0.1771, 0.1779, 0.1899, 0.2318, 0.0910, 0.1363, 0.2341], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0607, 0.0665, 0.0504, 0.0672, 0.0699, 0.0522, 0.0671], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:38:49,579 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.85 vs. limit=2.0 2023-05-01 10:39:35,757 INFO [train.py:904] (4/8) Epoch 22, batch 4000, loss[loss=0.1906, simple_loss=0.2699, pruned_loss=0.05568, over 16880.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2517, pruned_loss=0.04639, over 3286074.06 frames. ], batch size: 116, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:39:38,892 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-05-01 10:40:48,053 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 1.995e+02 2.386e+02 2.991e+02 7.237e+02, threshold=4.771e+02, percent-clipped=2.0 2023-05-01 10:40:49,969 INFO [train.py:904] (4/8) Epoch 22, batch 4050, loss[loss=0.1758, simple_loss=0.2558, pruned_loss=0.04794, over 16632.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2525, pruned_loss=0.04578, over 3281588.69 frames. ], batch size: 57, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:41:55,388 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=217247.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:42:04,909 INFO [train.py:904] (4/8) Epoch 22, batch 4100, loss[loss=0.1731, simple_loss=0.2592, pruned_loss=0.04353, over 16591.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2541, pruned_loss=0.04545, over 3268588.15 frames. ], batch size: 62, lr: 3.08e-03, grad_scale: 8.0 2023-05-01 10:43:11,027 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=217295.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:43:21,265 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.414e+02 2.082e+02 2.531e+02 2.908e+02 7.878e+02, threshold=5.062e+02, percent-clipped=6.0 2023-05-01 10:43:23,202 INFO [train.py:904] (4/8) Epoch 22, batch 4150, loss[loss=0.1759, simple_loss=0.2761, pruned_loss=0.03789, over 16786.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2611, pruned_loss=0.04772, over 3221415.85 frames. ], batch size: 102, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:43:36,121 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=217311.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 10:43:48,708 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5979, 4.5626, 4.6879, 4.8028, 4.9929, 4.5046, 4.9709, 5.0076], device='cuda:4'), covar=tensor([0.1962, 0.1119, 0.1485, 0.0763, 0.0562, 0.1021, 0.0714, 0.0563], device='cuda:4'), in_proj_covar=tensor([0.0653, 0.0808, 0.0938, 0.0823, 0.0621, 0.0649, 0.0672, 0.0777], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:43:52,134 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=217321.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:44:30,933 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9102, 4.1624, 4.0374, 4.0643, 3.7315, 3.7339, 3.8264, 4.1692], device='cuda:4'), covar=tensor([0.0996, 0.0835, 0.0883, 0.0784, 0.0717, 0.1746, 0.0859, 0.1003], device='cuda:4'), in_proj_covar=tensor([0.0685, 0.0842, 0.0692, 0.0640, 0.0533, 0.0542, 0.0704, 0.0658], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:44:39,665 INFO [train.py:904] (4/8) Epoch 22, batch 4200, loss[loss=0.2021, simple_loss=0.2964, pruned_loss=0.05392, over 16675.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2683, pruned_loss=0.0497, over 3195112.18 frames. ], batch size: 89, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:44:50,069 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=217359.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 10:44:52,709 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=217361.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:45:32,834 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=217388.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:45:53,920 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.404e+02 2.187e+02 2.459e+02 3.018e+02 5.838e+02, threshold=4.919e+02, percent-clipped=1.0 2023-05-01 10:45:55,239 INFO [train.py:904] (4/8) Epoch 22, batch 4250, loss[loss=0.1675, simple_loss=0.2672, pruned_loss=0.03393, over 16640.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2724, pruned_loss=0.05024, over 3160552.58 frames. ], batch size: 62, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:46:04,665 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=217409.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:47:04,209 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=217449.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:47:09,152 INFO [train.py:904] (4/8) Epoch 22, batch 4300, loss[loss=0.1868, simple_loss=0.2855, pruned_loss=0.04407, over 16435.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2737, pruned_loss=0.04928, over 3165364.86 frames. ], batch size: 75, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:47:40,700 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7575, 4.7404, 4.5106, 3.8178, 4.6653, 1.7006, 4.4370, 4.1834], device='cuda:4'), covar=tensor([0.0075, 0.0071, 0.0186, 0.0344, 0.0080, 0.3012, 0.0110, 0.0270], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0158, 0.0202, 0.0182, 0.0181, 0.0210, 0.0192, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:48:07,879 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=217492.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:48:15,681 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.57 vs. limit=5.0 2023-05-01 10:48:23,055 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 2.190e+02 2.513e+02 2.861e+02 5.953e+02, threshold=5.026e+02, percent-clipped=1.0 2023-05-01 10:48:24,299 INFO [train.py:904] (4/8) Epoch 22, batch 4350, loss[loss=0.2178, simple_loss=0.2923, pruned_loss=0.07164, over 11569.00 frames. ], tot_loss[loss=0.1878, simple_loss=0.2762, pruned_loss=0.04973, over 3171930.16 frames. ], batch size: 247, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:49:04,504 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.81 vs. limit=5.0 2023-05-01 10:49:31,315 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-05-01 10:49:39,978 INFO [train.py:904] (4/8) Epoch 22, batch 4400, loss[loss=0.1952, simple_loss=0.2819, pruned_loss=0.05426, over 16226.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.2781, pruned_loss=0.05102, over 3139214.25 frames. ], batch size: 165, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:49:41,130 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=217553.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:49:49,382 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.95 vs. limit=5.0 2023-05-01 10:50:04,069 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-01 10:50:52,392 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.401e+02 2.074e+02 2.354e+02 2.786e+02 4.691e+02, threshold=4.709e+02, percent-clipped=0.0 2023-05-01 10:50:53,484 INFO [train.py:904] (4/8) Epoch 22, batch 4450, loss[loss=0.1931, simple_loss=0.2893, pruned_loss=0.04846, over 17010.00 frames. ], tot_loss[loss=0.1927, simple_loss=0.2815, pruned_loss=0.05199, over 3164815.17 frames. ], batch size: 41, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:50:54,213 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-01 10:51:04,120 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7079, 2.9972, 3.2364, 2.0022, 2.7772, 2.0723, 3.2357, 3.2984], device='cuda:4'), covar=tensor([0.0239, 0.0894, 0.0590, 0.2115, 0.0874, 0.1165, 0.0643, 0.0895], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0164, 0.0166, 0.0152, 0.0145, 0.0130, 0.0143, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 10:51:16,910 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9047, 5.4244, 5.6508, 5.3322, 5.4643, 6.0116, 5.4982, 5.1861], device='cuda:4'), covar=tensor([0.1001, 0.1647, 0.1550, 0.1708, 0.2207, 0.0857, 0.1184, 0.2148], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0594, 0.0647, 0.0491, 0.0655, 0.0684, 0.0510, 0.0659], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:51:20,384 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=217621.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:51:35,682 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3665, 5.6088, 5.3624, 5.4491, 5.1286, 4.9671, 5.0510, 5.7180], device='cuda:4'), covar=tensor([0.1073, 0.0703, 0.0938, 0.0711, 0.0680, 0.0706, 0.1061, 0.0736], device='cuda:4'), in_proj_covar=tensor([0.0678, 0.0828, 0.0685, 0.0631, 0.0525, 0.0535, 0.0694, 0.0651], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:51:53,006 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1241, 4.0915, 3.9839, 3.2643, 3.9423, 1.8039, 3.7320, 3.4200], device='cuda:4'), covar=tensor([0.0084, 0.0072, 0.0161, 0.0247, 0.0069, 0.2956, 0.0099, 0.0251], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0159, 0.0203, 0.0182, 0.0180, 0.0211, 0.0192, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:52:08,354 INFO [train.py:904] (4/8) Epoch 22, batch 4500, loss[loss=0.1826, simple_loss=0.2734, pruned_loss=0.04596, over 16736.00 frames. ], tot_loss[loss=0.1938, simple_loss=0.2822, pruned_loss=0.05267, over 3177050.03 frames. ], batch size: 76, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:52:24,004 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=217664.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:52:32,190 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=217669.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:53:11,375 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 10:53:18,212 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.413e+02 1.861e+02 2.083e+02 2.379e+02 4.476e+02, threshold=4.167e+02, percent-clipped=0.0 2023-05-01 10:53:19,294 INFO [train.py:904] (4/8) Epoch 22, batch 4550, loss[loss=0.2279, simple_loss=0.2938, pruned_loss=0.08104, over 11872.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2835, pruned_loss=0.05374, over 3177564.53 frames. ], batch size: 246, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:53:44,458 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0808, 2.1917, 2.1690, 3.7765, 2.1706, 2.5011, 2.2871, 2.3038], device='cuda:4'), covar=tensor([0.1450, 0.3522, 0.2992, 0.0572, 0.4278, 0.2554, 0.3492, 0.3582], device='cuda:4'), in_proj_covar=tensor([0.0404, 0.0451, 0.0368, 0.0329, 0.0439, 0.0520, 0.0422, 0.0528], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 10:53:53,304 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=217725.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:54:13,154 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9430, 3.2313, 3.4925, 1.9913, 2.9867, 2.1982, 3.3012, 3.4067], device='cuda:4'), covar=tensor([0.0249, 0.0777, 0.0569, 0.2125, 0.0846, 0.1089, 0.0612, 0.0869], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0164, 0.0166, 0.0153, 0.0145, 0.0130, 0.0143, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 10:54:19,826 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=217744.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:54:31,797 INFO [train.py:904] (4/8) Epoch 22, batch 4600, loss[loss=0.1951, simple_loss=0.2818, pruned_loss=0.05418, over 16661.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2842, pruned_loss=0.05372, over 3193614.12 frames. ], batch size: 134, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:55:06,166 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-05-01 10:55:41,527 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.434e+02 1.780e+02 2.045e+02 2.386e+02 4.197e+02, threshold=4.091e+02, percent-clipped=1.0 2023-05-01 10:55:42,804 INFO [train.py:904] (4/8) Epoch 22, batch 4650, loss[loss=0.1738, simple_loss=0.2601, pruned_loss=0.04375, over 16852.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2837, pruned_loss=0.05399, over 3203476.07 frames. ], batch size: 96, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:56:46,088 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=217848.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 10:56:52,991 INFO [train.py:904] (4/8) Epoch 22, batch 4700, loss[loss=0.1693, simple_loss=0.2581, pruned_loss=0.04022, over 16745.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2808, pruned_loss=0.05294, over 3192986.54 frames. ], batch size: 124, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:57:01,297 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1458, 3.7406, 3.7378, 2.3496, 3.3411, 3.7630, 3.3600, 2.1783], device='cuda:4'), covar=tensor([0.0603, 0.0053, 0.0053, 0.0445, 0.0110, 0.0094, 0.0135, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0083, 0.0084, 0.0133, 0.0099, 0.0110, 0.0095, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 10:57:16,754 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4093, 2.8810, 2.6387, 2.2675, 2.2042, 2.1997, 2.9551, 2.7514], device='cuda:4'), covar=tensor([0.2646, 0.0841, 0.1711, 0.2766, 0.2432, 0.2276, 0.0540, 0.1250], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0270, 0.0306, 0.0316, 0.0299, 0.0261, 0.0296, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 10:57:38,754 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3459, 3.3834, 2.0571, 3.7752, 2.6157, 3.7901, 2.1825, 2.7593], device='cuda:4'), covar=tensor([0.0329, 0.0400, 0.1817, 0.0157, 0.0879, 0.0578, 0.1694, 0.0824], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0178, 0.0195, 0.0164, 0.0178, 0.0220, 0.0202, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 10:57:59,150 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=217898.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 10:58:04,665 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.316e+02 1.895e+02 2.114e+02 2.418e+02 3.922e+02, threshold=4.227e+02, percent-clipped=0.0 2023-05-01 10:58:05,922 INFO [train.py:904] (4/8) Epoch 22, batch 4750, loss[loss=0.1651, simple_loss=0.2477, pruned_loss=0.04119, over 17126.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2766, pruned_loss=0.05099, over 3194583.69 frames. ], batch size: 49, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:58:49,678 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-05-01 10:59:17,330 INFO [train.py:904] (4/8) Epoch 22, batch 4800, loss[loss=0.1891, simple_loss=0.2824, pruned_loss=0.04794, over 15454.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2723, pruned_loss=0.04852, over 3208724.47 frames. ], batch size: 191, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 10:59:27,595 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=217959.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 11:00:36,339 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.772e+02 2.165e+02 2.536e+02 4.360e+02, threshold=4.330e+02, percent-clipped=2.0 2023-05-01 11:00:36,354 INFO [train.py:904] (4/8) Epoch 22, batch 4850, loss[loss=0.1796, simple_loss=0.277, pruned_loss=0.04114, over 16779.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2738, pruned_loss=0.048, over 3179974.09 frames. ], batch size: 89, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:01:01,608 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=218020.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:01:38,581 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=218044.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:01:51,558 INFO [train.py:904] (4/8) Epoch 22, batch 4900, loss[loss=0.1772, simple_loss=0.2643, pruned_loss=0.04505, over 16441.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.273, pruned_loss=0.04706, over 3161057.45 frames. ], batch size: 75, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:02:49,274 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=218092.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:03:01,562 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=218100.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:03:05,447 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.536e+02 2.042e+02 2.302e+02 2.921e+02 6.655e+02, threshold=4.605e+02, percent-clipped=4.0 2023-05-01 11:03:05,462 INFO [train.py:904] (4/8) Epoch 22, batch 4950, loss[loss=0.176, simple_loss=0.272, pruned_loss=0.03998, over 17152.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2729, pruned_loss=0.04665, over 3158458.01 frames. ], batch size: 47, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:03:20,835 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-05-01 11:04:11,472 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=218148.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:04:12,859 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7946, 1.8560, 2.4792, 2.7540, 2.7236, 3.2613, 2.0237, 3.2340], device='cuda:4'), covar=tensor([0.0291, 0.0572, 0.0367, 0.0349, 0.0349, 0.0176, 0.0603, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0197, 0.0182, 0.0188, 0.0202, 0.0156, 0.0200, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:04:18,892 INFO [train.py:904] (4/8) Epoch 22, batch 5000, loss[loss=0.1833, simple_loss=0.2739, pruned_loss=0.04638, over 16634.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2745, pruned_loss=0.04637, over 3185379.04 frames. ], batch size: 68, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:04:30,134 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=218161.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:04:30,293 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-01 11:04:56,632 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1748, 4.1732, 4.0630, 2.6449, 3.6193, 4.1967, 3.5729, 2.1569], device='cuda:4'), covar=tensor([0.0596, 0.0039, 0.0044, 0.0402, 0.0100, 0.0068, 0.0106, 0.0480], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0082, 0.0084, 0.0132, 0.0098, 0.0109, 0.0094, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:05:01,527 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=218182.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:05:21,357 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=218196.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:05:32,358 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.421e+02 2.007e+02 2.287e+02 2.605e+02 5.161e+02, threshold=4.575e+02, percent-clipped=1.0 2023-05-01 11:05:32,373 INFO [train.py:904] (4/8) Epoch 22, batch 5050, loss[loss=0.1903, simple_loss=0.2853, pruned_loss=0.04762, over 16277.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2751, pruned_loss=0.04615, over 3186182.33 frames. ], batch size: 165, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:05:35,946 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3769, 4.3353, 4.3123, 3.4529, 4.3323, 1.7144, 4.0352, 3.9668], device='cuda:4'), covar=tensor([0.0111, 0.0121, 0.0163, 0.0454, 0.0105, 0.2830, 0.0148, 0.0262], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0155, 0.0198, 0.0178, 0.0176, 0.0207, 0.0188, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:05:42,344 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=218210.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:06:30,553 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=218243.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:06:44,148 INFO [train.py:904] (4/8) Epoch 22, batch 5100, loss[loss=0.1848, simple_loss=0.2738, pruned_loss=0.04792, over 16859.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2729, pruned_loss=0.04528, over 3200662.78 frames. ], batch size: 116, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:06:45,785 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=218254.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 11:07:10,596 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=218271.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 11:07:22,632 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2818, 4.1239, 4.3423, 4.4672, 4.6209, 4.1826, 4.5555, 4.6392], device='cuda:4'), covar=tensor([0.1573, 0.1261, 0.1375, 0.0715, 0.0467, 0.1119, 0.0744, 0.0634], device='cuda:4'), in_proj_covar=tensor([0.0631, 0.0784, 0.0907, 0.0795, 0.0598, 0.0630, 0.0649, 0.0752], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:07:24,309 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6535, 3.1845, 3.1206, 1.9005, 2.7627, 2.2599, 3.1836, 3.3737], device='cuda:4'), covar=tensor([0.0355, 0.0719, 0.0729, 0.2126, 0.0923, 0.0979, 0.0715, 0.0788], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0164, 0.0168, 0.0154, 0.0146, 0.0131, 0.0144, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:07:47,298 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1379, 1.9861, 1.7058, 1.7651, 2.2519, 1.9907, 1.8439, 2.3857], device='cuda:4'), covar=tensor([0.0167, 0.0439, 0.0552, 0.0447, 0.0256, 0.0347, 0.0200, 0.0263], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0233, 0.0225, 0.0225, 0.0234, 0.0234, 0.0236, 0.0231], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:07:57,331 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 1.925e+02 2.150e+02 2.531e+02 3.604e+02, threshold=4.300e+02, percent-clipped=0.0 2023-05-01 11:07:57,346 INFO [train.py:904] (4/8) Epoch 22, batch 5150, loss[loss=0.1697, simple_loss=0.2736, pruned_loss=0.03286, over 16238.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2734, pruned_loss=0.04484, over 3199671.48 frames. ], batch size: 165, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:08:02,940 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0926, 2.4718, 2.6976, 1.8775, 2.7344, 2.8711, 2.4674, 2.3415], device='cuda:4'), covar=tensor([0.0704, 0.0264, 0.0200, 0.1034, 0.0112, 0.0212, 0.0421, 0.0501], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0099, 0.0140, 0.0081, 0.0126, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:08:23,087 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=218320.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:09:11,243 INFO [train.py:904] (4/8) Epoch 22, batch 5200, loss[loss=0.177, simple_loss=0.2669, pruned_loss=0.04355, over 16692.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2724, pruned_loss=0.04461, over 3189649.46 frames. ], batch size: 89, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:09:33,041 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=218368.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:09:52,923 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5893, 3.6235, 2.7742, 2.2324, 2.3277, 2.3665, 3.7798, 3.1585], device='cuda:4'), covar=tensor([0.2740, 0.0625, 0.1779, 0.2810, 0.2612, 0.1992, 0.0537, 0.1225], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0269, 0.0304, 0.0313, 0.0297, 0.0259, 0.0295, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:10:23,968 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.335e+02 1.955e+02 2.305e+02 2.711e+02 5.108e+02, threshold=4.610e+02, percent-clipped=2.0 2023-05-01 11:10:23,983 INFO [train.py:904] (4/8) Epoch 22, batch 5250, loss[loss=0.1642, simple_loss=0.2497, pruned_loss=0.0393, over 16465.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.269, pruned_loss=0.04367, over 3202914.66 frames. ], batch size: 68, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:11:37,152 INFO [train.py:904] (4/8) Epoch 22, batch 5300, loss[loss=0.1747, simple_loss=0.2603, pruned_loss=0.04451, over 16365.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2654, pruned_loss=0.04271, over 3213582.62 frames. ], batch size: 35, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:11:41,559 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=218456.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:12:20,704 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1565, 4.9398, 5.1578, 5.3256, 5.5687, 4.8705, 5.5480, 5.5477], device='cuda:4'), covar=tensor([0.1673, 0.1287, 0.1596, 0.0802, 0.0471, 0.0876, 0.0467, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0631, 0.0784, 0.0910, 0.0795, 0.0598, 0.0630, 0.0650, 0.0750], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:12:35,822 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8518, 3.2637, 3.3458, 2.0738, 2.8824, 2.2636, 3.2661, 3.5442], device='cuda:4'), covar=tensor([0.0336, 0.0780, 0.0658, 0.2014, 0.0935, 0.0992, 0.0797, 0.0896], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0164, 0.0168, 0.0154, 0.0146, 0.0131, 0.0144, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:12:51,254 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.460e+02 1.955e+02 2.262e+02 2.726e+02 6.747e+02, threshold=4.525e+02, percent-clipped=3.0 2023-05-01 11:12:51,269 INFO [train.py:904] (4/8) Epoch 22, batch 5350, loss[loss=0.1891, simple_loss=0.2776, pruned_loss=0.05032, over 12139.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2637, pruned_loss=0.042, over 3212819.28 frames. ], batch size: 247, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:13:00,562 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=218509.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:13:43,244 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=218538.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:14:03,262 INFO [train.py:904] (4/8) Epoch 22, batch 5400, loss[loss=0.1823, simple_loss=0.2795, pruned_loss=0.04252, over 16377.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2661, pruned_loss=0.04244, over 3202860.21 frames. ], batch size: 146, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:14:06,205 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=218554.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 11:14:23,139 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=218566.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 11:14:28,950 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=218570.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:14:34,753 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=218574.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:14:50,851 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4156, 3.3853, 3.4385, 3.5249, 3.5843, 3.3382, 3.5413, 3.6336], device='cuda:4'), covar=tensor([0.1198, 0.0989, 0.0952, 0.0630, 0.0591, 0.2384, 0.1050, 0.0714], device='cuda:4'), in_proj_covar=tensor([0.0633, 0.0786, 0.0913, 0.0797, 0.0599, 0.0631, 0.0652, 0.0753], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:15:00,791 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.90 vs. limit=5.0 2023-05-01 11:15:18,765 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=218602.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 11:15:19,491 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.606e+02 2.152e+02 2.360e+02 2.669e+02 5.554e+02, threshold=4.720e+02, percent-clipped=2.0 2023-05-01 11:15:19,508 INFO [train.py:904] (4/8) Epoch 22, batch 5450, loss[loss=0.1926, simple_loss=0.2802, pruned_loss=0.05255, over 16459.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2678, pruned_loss=0.04314, over 3201637.46 frames. ], batch size: 68, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:16:10,928 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=218635.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:16:37,991 INFO [train.py:904] (4/8) Epoch 22, batch 5500, loss[loss=0.2017, simple_loss=0.2988, pruned_loss=0.05227, over 16897.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2754, pruned_loss=0.04763, over 3183988.39 frames. ], batch size: 96, lr: 3.07e-03, grad_scale: 8.0 2023-05-01 11:17:57,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.055e+02 2.950e+02 3.517e+02 4.298e+02 7.348e+02, threshold=7.033e+02, percent-clipped=15.0 2023-05-01 11:17:57,952 INFO [train.py:904] (4/8) Epoch 22, batch 5550, loss[loss=0.2179, simple_loss=0.2959, pruned_loss=0.06994, over 17264.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2824, pruned_loss=0.0522, over 3154655.07 frames. ], batch size: 52, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:18:44,664 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5892, 2.6497, 2.3114, 3.8804, 2.6453, 3.8513, 1.4231, 2.8496], device='cuda:4'), covar=tensor([0.1540, 0.0829, 0.1410, 0.0187, 0.0233, 0.0422, 0.1899, 0.0886], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0174, 0.0194, 0.0190, 0.0205, 0.0215, 0.0201, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:18:47,532 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5805, 4.7816, 4.9475, 4.7299, 4.8333, 5.3619, 4.8190, 4.5688], device='cuda:4'), covar=tensor([0.1206, 0.1840, 0.1903, 0.1834, 0.2206, 0.0879, 0.1582, 0.2343], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0581, 0.0640, 0.0487, 0.0647, 0.0675, 0.0507, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:19:13,548 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0616, 4.1173, 4.4122, 4.3702, 4.4057, 4.1691, 4.1733, 4.1100], device='cuda:4'), covar=tensor([0.0357, 0.0682, 0.0429, 0.0471, 0.0484, 0.0475, 0.0870, 0.0537], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0446, 0.0432, 0.0399, 0.0478, 0.0452, 0.0537, 0.0362], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 11:19:20,438 INFO [train.py:904] (4/8) Epoch 22, batch 5600, loss[loss=0.1877, simple_loss=0.2789, pruned_loss=0.04824, over 16843.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.286, pruned_loss=0.05517, over 3143160.66 frames. ], batch size: 90, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:19:25,769 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=218756.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:20:41,871 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.037e+02 3.163e+02 3.914e+02 4.827e+02 8.104e+02, threshold=7.827e+02, percent-clipped=4.0 2023-05-01 11:20:41,886 INFO [train.py:904] (4/8) Epoch 22, batch 5650, loss[loss=0.2683, simple_loss=0.3344, pruned_loss=0.1011, over 11393.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.2918, pruned_loss=0.05998, over 3110072.02 frames. ], batch size: 248, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:20:44,084 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=218804.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:20:55,291 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 11:21:35,393 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=218838.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:21:56,891 INFO [train.py:904] (4/8) Epoch 22, batch 5700, loss[loss=0.2153, simple_loss=0.3089, pruned_loss=0.0609, over 16907.00 frames. ], tot_loss[loss=0.2079, simple_loss=0.2932, pruned_loss=0.06132, over 3103222.41 frames. ], batch size: 109, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:22:12,542 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9031, 2.1591, 2.4304, 3.1481, 2.2254, 2.3482, 2.3427, 2.2864], device='cuda:4'), covar=tensor([0.1300, 0.3006, 0.2285, 0.0678, 0.3681, 0.2220, 0.2857, 0.2876], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0447, 0.0367, 0.0326, 0.0434, 0.0516, 0.0419, 0.0522], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:22:16,161 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=218865.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:22:17,561 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=218866.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:22:32,498 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-05-01 11:22:48,658 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=218886.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:23:13,875 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.920e+02 2.926e+02 3.484e+02 4.272e+02 7.645e+02, threshold=6.969e+02, percent-clipped=0.0 2023-05-01 11:23:13,890 INFO [train.py:904] (4/8) Epoch 22, batch 5750, loss[loss=0.223, simple_loss=0.2891, pruned_loss=0.07841, over 11225.00 frames. ], tot_loss[loss=0.2113, simple_loss=0.2961, pruned_loss=0.06325, over 3069880.51 frames. ], batch size: 247, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:23:25,640 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7352, 2.4968, 2.2700, 3.2143, 2.2570, 3.5331, 1.5599, 2.7357], device='cuda:4'), covar=tensor([0.1401, 0.0801, 0.1375, 0.0229, 0.0221, 0.0480, 0.1743, 0.0852], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0173, 0.0194, 0.0190, 0.0204, 0.0214, 0.0201, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:23:29,744 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1359, 2.4172, 2.3577, 2.7128, 1.9674, 3.1355, 1.9545, 2.7872], device='cuda:4'), covar=tensor([0.1182, 0.0577, 0.1161, 0.0193, 0.0133, 0.0429, 0.1453, 0.0724], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0173, 0.0194, 0.0190, 0.0204, 0.0214, 0.0201, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:23:32,117 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=218914.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:23:57,588 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=218930.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:24:34,379 INFO [train.py:904] (4/8) Epoch 22, batch 5800, loss[loss=0.1715, simple_loss=0.2653, pruned_loss=0.03886, over 16684.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.296, pruned_loss=0.06257, over 3063820.38 frames. ], batch size: 76, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:24:54,971 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=218965.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:25:24,042 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=218984.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:25:52,934 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4364, 3.3998, 2.6830, 2.1786, 2.2821, 2.3059, 3.5720, 3.0812], device='cuda:4'), covar=tensor([0.3116, 0.0661, 0.1815, 0.2752, 0.2549, 0.2145, 0.0514, 0.1293], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0269, 0.0305, 0.0314, 0.0297, 0.0259, 0.0296, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:25:53,554 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.200e+02 2.954e+02 3.399e+02 3.875e+02 5.875e+02, threshold=6.799e+02, percent-clipped=0.0 2023-05-01 11:25:53,575 INFO [train.py:904] (4/8) Epoch 22, batch 5850, loss[loss=0.1893, simple_loss=0.2776, pruned_loss=0.05047, over 16289.00 frames. ], tot_loss[loss=0.2077, simple_loss=0.2935, pruned_loss=0.0609, over 3076448.56 frames. ], batch size: 165, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:26:03,864 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9193, 4.1751, 4.0029, 4.0532, 3.7428, 3.8249, 3.8421, 4.1860], device='cuda:4'), covar=tensor([0.1114, 0.0902, 0.1082, 0.0823, 0.0752, 0.1472, 0.0917, 0.1002], device='cuda:4'), in_proj_covar=tensor([0.0666, 0.0812, 0.0674, 0.0618, 0.0514, 0.0522, 0.0677, 0.0635], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:26:21,534 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=219021.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:26:30,010 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219026.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:27:02,414 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219045.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:27:14,927 INFO [train.py:904] (4/8) Epoch 22, batch 5900, loss[loss=0.2259, simple_loss=0.2989, pruned_loss=0.07644, over 15511.00 frames. ], tot_loss[loss=0.207, simple_loss=0.2928, pruned_loss=0.0606, over 3083062.92 frames. ], batch size: 191, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:27:49,373 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=219072.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:28:04,739 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219082.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:28:10,338 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6410, 2.2004, 1.8752, 1.9202, 2.5943, 2.1889, 2.3311, 2.7277], device='cuda:4'), covar=tensor([0.0236, 0.0467, 0.0585, 0.0544, 0.0278, 0.0441, 0.0259, 0.0287], device='cuda:4'), in_proj_covar=tensor([0.0209, 0.0231, 0.0224, 0.0223, 0.0232, 0.0232, 0.0232, 0.0228], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:28:30,024 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-01 11:28:35,897 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.790e+02 2.529e+02 2.968e+02 3.874e+02 6.567e+02, threshold=5.937e+02, percent-clipped=0.0 2023-05-01 11:28:35,912 INFO [train.py:904] (4/8) Epoch 22, batch 5950, loss[loss=0.2361, simple_loss=0.3165, pruned_loss=0.07787, over 15285.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2935, pruned_loss=0.05961, over 3082018.91 frames. ], batch size: 190, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:28:55,264 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2021, 5.9188, 6.1438, 5.8174, 5.8393, 6.3818, 5.8897, 5.5737], device='cuda:4'), covar=tensor([0.0889, 0.1774, 0.2144, 0.1869, 0.2335, 0.0875, 0.1514, 0.2388], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0585, 0.0647, 0.0490, 0.0652, 0.0679, 0.0511, 0.0658], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:29:24,451 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219133.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:29:48,604 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-01 11:29:54,903 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0793, 4.1048, 4.4385, 4.3926, 4.4096, 4.1488, 4.1486, 4.1131], device='cuda:4'), covar=tensor([0.0344, 0.0574, 0.0357, 0.0411, 0.0465, 0.0438, 0.0884, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0400, 0.0445, 0.0431, 0.0399, 0.0479, 0.0452, 0.0536, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 11:29:57,662 INFO [train.py:904] (4/8) Epoch 22, batch 6000, loss[loss=0.1887, simple_loss=0.2779, pruned_loss=0.04982, over 16871.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2925, pruned_loss=0.05906, over 3087764.75 frames. ], batch size: 109, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:29:57,662 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 11:30:07,617 INFO [train.py:938] (4/8) Epoch 22, validation: loss=0.1507, simple_loss=0.2632, pruned_loss=0.01907, over 944034.00 frames. 2023-05-01 11:30:07,617 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 11:30:21,232 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-05-01 11:30:27,708 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=219165.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:30:33,307 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0272, 3.0765, 1.9431, 3.2671, 2.3610, 3.3402, 2.1787, 2.6157], device='cuda:4'), covar=tensor([0.0311, 0.0399, 0.1623, 0.0217, 0.0879, 0.0604, 0.1440, 0.0736], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0193, 0.0161, 0.0175, 0.0215, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:30:58,005 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0644, 4.0170, 3.9655, 3.1563, 3.9971, 1.7978, 3.7633, 3.4180], device='cuda:4'), covar=tensor([0.0107, 0.0113, 0.0180, 0.0301, 0.0095, 0.2785, 0.0133, 0.0305], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0155, 0.0198, 0.0178, 0.0175, 0.0206, 0.0187, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:31:28,922 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.979e+02 2.852e+02 3.411e+02 4.220e+02 7.320e+02, threshold=6.821e+02, percent-clipped=6.0 2023-05-01 11:31:28,937 INFO [train.py:904] (4/8) Epoch 22, batch 6050, loss[loss=0.1759, simple_loss=0.2757, pruned_loss=0.03805, over 16668.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2915, pruned_loss=0.0587, over 3079927.98 frames. ], batch size: 68, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:31:29,590 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=219203.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:31:44,909 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=219213.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:32:11,180 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=219230.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:32:46,221 INFO [train.py:904] (4/8) Epoch 22, batch 6100, loss[loss=0.1946, simple_loss=0.2847, pruned_loss=0.05231, over 15436.00 frames. ], tot_loss[loss=0.2032, simple_loss=0.2911, pruned_loss=0.05771, over 3091854.82 frames. ], batch size: 191, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:33:05,407 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219264.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:33:26,347 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=219278.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:33:38,937 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6292, 5.9838, 5.6102, 5.7570, 5.3867, 5.3409, 5.3166, 6.0873], device='cuda:4'), covar=tensor([0.1258, 0.0798, 0.1059, 0.0820, 0.0858, 0.0654, 0.1178, 0.0876], device='cuda:4'), in_proj_covar=tensor([0.0668, 0.0816, 0.0678, 0.0622, 0.0518, 0.0526, 0.0681, 0.0641], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:34:04,107 INFO [train.py:904] (4/8) Epoch 22, batch 6150, loss[loss=0.1995, simple_loss=0.2883, pruned_loss=0.05532, over 15512.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.29, pruned_loss=0.05762, over 3086469.93 frames. ], batch size: 191, lr: 3.06e-03, grad_scale: 4.0 2023-05-01 11:34:05,864 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.700e+02 2.531e+02 2.938e+02 3.735e+02 5.885e+02, threshold=5.877e+02, percent-clipped=0.0 2023-05-01 11:34:14,698 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4007, 4.5263, 4.7084, 4.4793, 4.5600, 5.0537, 4.5519, 4.3021], device='cuda:4'), covar=tensor([0.1468, 0.1890, 0.2074, 0.1984, 0.2438, 0.0982, 0.1689, 0.2560], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0587, 0.0647, 0.0491, 0.0653, 0.0679, 0.0514, 0.0659], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:34:33,692 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=219321.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:34:50,231 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8990, 5.2444, 5.4549, 5.2115, 5.2583, 5.8000, 5.2757, 5.0457], device='cuda:4'), covar=tensor([0.1003, 0.1862, 0.2047, 0.1779, 0.2364, 0.0885, 0.1586, 0.2358], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0586, 0.0645, 0.0490, 0.0653, 0.0677, 0.0513, 0.0657], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:35:03,737 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=219340.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:35:06,798 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=219342.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 11:35:22,806 INFO [train.py:904] (4/8) Epoch 22, batch 6200, loss[loss=0.1965, simple_loss=0.281, pruned_loss=0.05598, over 16917.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2878, pruned_loss=0.05716, over 3087981.59 frames. ], batch size: 109, lr: 3.06e-03, grad_scale: 4.0 2023-05-01 11:35:51,385 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0762, 3.9777, 4.1429, 4.2773, 4.3881, 4.0159, 4.3281, 4.3977], device='cuda:4'), covar=tensor([0.1733, 0.1226, 0.1463, 0.0703, 0.0580, 0.1283, 0.0865, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0628, 0.0779, 0.0901, 0.0788, 0.0597, 0.0623, 0.0649, 0.0749], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:36:02,464 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=219377.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:36:37,366 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2817, 2.2307, 2.8271, 3.2175, 3.0735, 3.6541, 2.2041, 3.6102], device='cuda:4'), covar=tensor([0.0183, 0.0483, 0.0290, 0.0239, 0.0295, 0.0146, 0.0572, 0.0135], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0192, 0.0179, 0.0183, 0.0197, 0.0152, 0.0196, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:36:42,472 INFO [train.py:904] (4/8) Epoch 22, batch 6250, loss[loss=0.1875, simple_loss=0.2753, pruned_loss=0.04986, over 16248.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2871, pruned_loss=0.05652, over 3101128.52 frames. ], batch size: 165, lr: 3.06e-03, grad_scale: 4.0 2023-05-01 11:36:43,079 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9085, 3.3286, 3.3259, 2.1205, 3.1511, 3.4585, 3.2017, 1.8637], device='cuda:4'), covar=tensor([0.0604, 0.0073, 0.0069, 0.0461, 0.0107, 0.0106, 0.0099, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0082, 0.0083, 0.0132, 0.0097, 0.0109, 0.0094, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:36:43,110 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219403.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 11:36:43,733 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.827e+02 2.722e+02 3.215e+02 3.900e+02 7.497e+02, threshold=6.430e+02, percent-clipped=6.0 2023-05-01 11:37:20,998 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=219428.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:37:57,565 INFO [train.py:904] (4/8) Epoch 22, batch 6300, loss[loss=0.2086, simple_loss=0.2847, pruned_loss=0.0663, over 16685.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.287, pruned_loss=0.0563, over 3098437.18 frames. ], batch size: 134, lr: 3.06e-03, grad_scale: 4.0 2023-05-01 11:38:55,995 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=219490.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:39:15,298 INFO [train.py:904] (4/8) Epoch 22, batch 6350, loss[loss=0.2044, simple_loss=0.29, pruned_loss=0.05942, over 16764.00 frames. ], tot_loss[loss=0.2013, simple_loss=0.2877, pruned_loss=0.05746, over 3092087.45 frames. ], batch size: 134, lr: 3.06e-03, grad_scale: 4.0 2023-05-01 11:39:16,413 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.781e+02 2.840e+02 3.669e+02 4.695e+02 9.321e+02, threshold=7.339e+02, percent-clipped=9.0 2023-05-01 11:39:25,120 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5983, 3.6320, 2.6825, 2.2280, 2.4600, 2.4186, 3.8720, 3.3083], device='cuda:4'), covar=tensor([0.2949, 0.0765, 0.2023, 0.2888, 0.2529, 0.2063, 0.0568, 0.1379], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0268, 0.0303, 0.0313, 0.0296, 0.0258, 0.0294, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:40:28,619 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219551.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:40:31,086 INFO [train.py:904] (4/8) Epoch 22, batch 6400, loss[loss=0.174, simple_loss=0.2643, pruned_loss=0.0419, over 16706.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2888, pruned_loss=0.05892, over 3067943.80 frames. ], batch size: 83, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:40:40,478 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=219559.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:40:41,752 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0164, 5.4770, 5.6348, 5.3608, 5.4776, 5.9622, 5.4811, 5.2107], device='cuda:4'), covar=tensor([0.0918, 0.1702, 0.2121, 0.1777, 0.2177, 0.0931, 0.1532, 0.2309], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0591, 0.0653, 0.0495, 0.0657, 0.0683, 0.0513, 0.0662], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:41:08,369 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-01 11:41:45,413 INFO [train.py:904] (4/8) Epoch 22, batch 6450, loss[loss=0.2055, simple_loss=0.2951, pruned_loss=0.05795, over 16736.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2889, pruned_loss=0.05803, over 3074098.10 frames. ], batch size: 124, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:41:47,189 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.783e+02 2.795e+02 3.528e+02 4.334e+02 7.123e+02, threshold=7.056e+02, percent-clipped=0.0 2023-05-01 11:42:13,293 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=219621.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:42:16,362 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2953, 4.1706, 4.3436, 4.4891, 4.6198, 4.1900, 4.5669, 4.6193], device='cuda:4'), covar=tensor([0.1668, 0.1227, 0.1437, 0.0666, 0.0547, 0.1236, 0.0817, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0627, 0.0778, 0.0903, 0.0786, 0.0596, 0.0622, 0.0648, 0.0748], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:42:31,376 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9672, 4.9933, 4.8392, 4.5149, 4.4877, 4.9551, 4.8259, 4.5913], device='cuda:4'), covar=tensor([0.0616, 0.0511, 0.0327, 0.0321, 0.1044, 0.0411, 0.0321, 0.0697], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0429, 0.0341, 0.0338, 0.0347, 0.0393, 0.0235, 0.0407], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:42:45,379 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=219640.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:43:04,572 INFO [train.py:904] (4/8) Epoch 22, batch 6500, loss[loss=0.1739, simple_loss=0.2718, pruned_loss=0.03801, over 16826.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2867, pruned_loss=0.05683, over 3096797.73 frames. ], batch size: 102, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:43:29,506 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3481, 3.2431, 2.6302, 2.1543, 2.2371, 2.2849, 3.4364, 2.9820], device='cuda:4'), covar=tensor([0.3023, 0.0813, 0.1827, 0.2687, 0.2626, 0.2178, 0.0493, 0.1382], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0267, 0.0302, 0.0311, 0.0295, 0.0257, 0.0293, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:43:30,548 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=219669.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:43:43,538 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=219677.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:44:01,338 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=219688.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:44:19,768 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=219698.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 11:44:26,651 INFO [train.py:904] (4/8) Epoch 22, batch 6550, loss[loss=0.2203, simple_loss=0.3142, pruned_loss=0.06317, over 16885.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2896, pruned_loss=0.05801, over 3104334.78 frames. ], batch size: 116, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:44:28,428 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.870e+02 2.583e+02 3.105e+02 3.738e+02 8.268e+02, threshold=6.210e+02, percent-clipped=2.0 2023-05-01 11:45:03,110 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=219725.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:45:05,871 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5954, 2.5432, 1.8615, 2.6877, 2.0975, 2.7828, 2.1608, 2.3939], device='cuda:4'), covar=tensor([0.0313, 0.0382, 0.1289, 0.0326, 0.0671, 0.0534, 0.1113, 0.0562], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0174, 0.0192, 0.0160, 0.0175, 0.0214, 0.0199, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:45:08,809 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=219728.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:45:46,181 INFO [train.py:904] (4/8) Epoch 22, batch 6600, loss[loss=0.2037, simple_loss=0.2925, pruned_loss=0.05745, over 15312.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2918, pruned_loss=0.05879, over 3096958.43 frames. ], batch size: 190, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:45:57,889 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9859, 3.4485, 3.4613, 2.2551, 3.1762, 3.5108, 3.2736, 2.0218], device='cuda:4'), covar=tensor([0.0621, 0.0067, 0.0064, 0.0458, 0.0121, 0.0126, 0.0114, 0.0488], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0083, 0.0083, 0.0132, 0.0097, 0.0109, 0.0094, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:46:02,117 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=219763.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:46:23,295 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=219776.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:47:06,507 INFO [train.py:904] (4/8) Epoch 22, batch 6650, loss[loss=0.1714, simple_loss=0.2661, pruned_loss=0.0383, over 16715.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2919, pruned_loss=0.05961, over 3087484.95 frames. ], batch size: 89, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:47:07,640 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.837e+02 2.943e+02 3.397e+02 4.500e+02 8.184e+02, threshold=6.793e+02, percent-clipped=7.0 2023-05-01 11:47:39,850 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219824.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:47:42,894 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6649, 4.6786, 5.0256, 5.0033, 5.0399, 4.7260, 4.7196, 4.5074], device='cuda:4'), covar=tensor([0.0327, 0.0522, 0.0384, 0.0402, 0.0473, 0.0376, 0.0902, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0453, 0.0438, 0.0407, 0.0487, 0.0461, 0.0546, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 11:48:10,733 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=219846.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:48:21,239 INFO [train.py:904] (4/8) Epoch 22, batch 6700, loss[loss=0.2132, simple_loss=0.3019, pruned_loss=0.06227, over 16836.00 frames. ], tot_loss[loss=0.205, simple_loss=0.291, pruned_loss=0.05953, over 3096747.25 frames. ], batch size: 116, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:48:27,598 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=219857.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:48:30,120 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=219859.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:48:45,596 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-05-01 11:48:51,947 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 11:49:30,048 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0910, 3.2030, 1.9811, 3.4426, 2.4759, 3.4853, 2.0660, 2.6251], device='cuda:4'), covar=tensor([0.0363, 0.0397, 0.1693, 0.0196, 0.0835, 0.0623, 0.1695, 0.0784], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0194, 0.0161, 0.0175, 0.0216, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:49:36,042 INFO [train.py:904] (4/8) Epoch 22, batch 6750, loss[loss=0.1952, simple_loss=0.2768, pruned_loss=0.05679, over 16589.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2906, pruned_loss=0.06003, over 3097479.58 frames. ], batch size: 68, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:49:37,872 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.993e+02 2.924e+02 3.493e+02 4.502e+02 8.889e+02, threshold=6.985e+02, percent-clipped=2.0 2023-05-01 11:49:43,852 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=219907.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:50:00,686 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=219918.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:50:54,126 INFO [train.py:904] (4/8) Epoch 22, batch 6800, loss[loss=0.1999, simple_loss=0.2923, pruned_loss=0.05375, over 17111.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2899, pruned_loss=0.05933, over 3120073.65 frames. ], batch size: 49, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:51:22,367 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5792, 3.0105, 3.2122, 1.9898, 2.8012, 2.1123, 3.1775, 3.2144], device='cuda:4'), covar=tensor([0.0283, 0.0748, 0.0579, 0.2049, 0.0852, 0.1024, 0.0656, 0.0858], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0165, 0.0168, 0.0155, 0.0146, 0.0131, 0.0144, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:51:53,899 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-05-01 11:52:05,018 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=219998.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 11:52:15,003 INFO [train.py:904] (4/8) Epoch 22, batch 6850, loss[loss=0.2059, simple_loss=0.3245, pruned_loss=0.04366, over 16797.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2916, pruned_loss=0.05952, over 3119599.08 frames. ], batch size: 102, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:52:16,799 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.941e+02 2.883e+02 3.357e+02 3.930e+02 6.765e+02, threshold=6.713e+02, percent-clipped=0.0 2023-05-01 11:53:20,648 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=220046.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 11:53:31,115 INFO [train.py:904] (4/8) Epoch 22, batch 6900, loss[loss=0.1959, simple_loss=0.2882, pruned_loss=0.05183, over 16860.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2934, pruned_loss=0.05861, over 3138785.02 frames. ], batch size: 116, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:54:24,697 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=220087.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:54:41,927 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=220098.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:54:49,991 INFO [train.py:904] (4/8) Epoch 22, batch 6950, loss[loss=0.2042, simple_loss=0.2942, pruned_loss=0.05711, over 16889.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2949, pruned_loss=0.06049, over 3113690.51 frames. ], batch size: 109, lr: 3.06e-03, grad_scale: 8.0 2023-05-01 11:54:51,079 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.857e+02 2.861e+02 3.409e+02 4.251e+02 1.328e+03, threshold=6.818e+02, percent-clipped=1.0 2023-05-01 11:55:14,107 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=220119.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:55:54,693 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=220146.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:55:57,599 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=220148.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:56:03,772 INFO [train.py:904] (4/8) Epoch 22, batch 7000, loss[loss=0.2012, simple_loss=0.3053, pruned_loss=0.04851, over 16441.00 frames. ], tot_loss[loss=0.2066, simple_loss=0.2947, pruned_loss=0.05927, over 3125494.70 frames. ], batch size: 146, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 11:56:06,179 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6886, 1.7212, 2.2012, 2.5573, 2.5326, 2.8612, 1.8867, 2.8742], device='cuda:4'), covar=tensor([0.0202, 0.0577, 0.0366, 0.0353, 0.0346, 0.0227, 0.0604, 0.0165], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0192, 0.0178, 0.0182, 0.0196, 0.0152, 0.0195, 0.0150], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 11:56:11,262 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6409, 2.5299, 1.8895, 2.6980, 2.1105, 2.7835, 2.1866, 2.4000], device='cuda:4'), covar=tensor([0.0317, 0.0352, 0.1284, 0.0242, 0.0677, 0.0556, 0.1076, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0194, 0.0161, 0.0175, 0.0216, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 11:56:14,302 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=220159.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:57:05,650 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=220194.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:57:18,841 INFO [train.py:904] (4/8) Epoch 22, batch 7050, loss[loss=0.2375, simple_loss=0.3103, pruned_loss=0.08238, over 15428.00 frames. ], tot_loss[loss=0.2067, simple_loss=0.295, pruned_loss=0.05923, over 3116297.29 frames. ], batch size: 190, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 11:57:21,955 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.068e+02 2.577e+02 3.204e+02 3.781e+02 6.030e+02, threshold=6.408e+02, percent-clipped=0.0 2023-05-01 11:57:23,846 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3813, 2.9429, 2.6970, 2.3117, 2.2931, 2.3459, 2.9724, 2.8273], device='cuda:4'), covar=tensor([0.2372, 0.0717, 0.1615, 0.2272, 0.2274, 0.2079, 0.0433, 0.1316], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0267, 0.0303, 0.0312, 0.0296, 0.0258, 0.0293, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 11:57:35,244 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=220213.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:57:38,231 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-05-01 11:57:59,252 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=220229.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:58:19,432 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-05-01 11:58:36,504 INFO [train.py:904] (4/8) Epoch 22, batch 7100, loss[loss=0.24, simple_loss=0.2972, pruned_loss=0.09136, over 11316.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2928, pruned_loss=0.05871, over 3104176.58 frames. ], batch size: 250, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 11:58:45,426 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=220258.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:59:35,717 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=220290.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 11:59:54,481 INFO [train.py:904] (4/8) Epoch 22, batch 7150, loss[loss=0.1827, simple_loss=0.2781, pruned_loss=0.04369, over 16835.00 frames. ], tot_loss[loss=0.2049, simple_loss=0.2918, pruned_loss=0.05896, over 3102968.79 frames. ], batch size: 102, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 11:59:58,138 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.530e+02 2.701e+02 3.416e+02 4.420e+02 1.024e+03, threshold=6.833e+02, percent-clipped=4.0 2023-05-01 12:00:15,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9741, 3.9141, 4.0432, 4.1574, 4.2688, 3.8789, 4.1940, 4.2814], device='cuda:4'), covar=tensor([0.1697, 0.1160, 0.1329, 0.0700, 0.0569, 0.1648, 0.0851, 0.0675], device='cuda:4'), in_proj_covar=tensor([0.0629, 0.0778, 0.0900, 0.0787, 0.0597, 0.0620, 0.0650, 0.0749], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:00:20,907 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=220319.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:01:07,949 INFO [train.py:904] (4/8) Epoch 22, batch 7200, loss[loss=0.1918, simple_loss=0.2734, pruned_loss=0.05508, over 11624.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2892, pruned_loss=0.05707, over 3099540.10 frames. ], batch size: 247, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:02:28,580 INFO [train.py:904] (4/8) Epoch 22, batch 7250, loss[loss=0.1625, simple_loss=0.2521, pruned_loss=0.03645, over 16775.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2867, pruned_loss=0.05636, over 3083086.80 frames. ], batch size: 83, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:02:30,895 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.713e+02 2.548e+02 3.054e+02 3.716e+02 9.083e+02, threshold=6.108e+02, percent-clipped=2.0 2023-05-01 12:02:53,035 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=220419.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:03:30,263 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=220443.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:03:45,047 INFO [train.py:904] (4/8) Epoch 22, batch 7300, loss[loss=0.2447, simple_loss=0.3089, pruned_loss=0.09022, over 11452.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.286, pruned_loss=0.05629, over 3077667.69 frames. ], batch size: 247, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:03:46,865 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=220454.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:04:07,576 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=220467.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:04:18,752 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.91 vs. limit=5.0 2023-05-01 12:04:55,534 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0888, 2.1616, 2.2200, 3.6139, 2.1549, 2.5109, 2.2794, 2.3065], device='cuda:4'), covar=tensor([0.1365, 0.3353, 0.2878, 0.0611, 0.4125, 0.2362, 0.3230, 0.3612], device='cuda:4'), in_proj_covar=tensor([0.0399, 0.0447, 0.0365, 0.0325, 0.0436, 0.0515, 0.0418, 0.0522], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:05:02,409 INFO [train.py:904] (4/8) Epoch 22, batch 7350, loss[loss=0.2228, simple_loss=0.2982, pruned_loss=0.07373, over 10965.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2878, pruned_loss=0.05787, over 3053637.04 frames. ], batch size: 246, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:05:05,570 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.460e+02 2.713e+02 3.260e+02 3.989e+02 1.125e+03, threshold=6.520e+02, percent-clipped=5.0 2023-05-01 12:05:18,432 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=220513.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:06:20,489 INFO [train.py:904] (4/8) Epoch 22, batch 7400, loss[loss=0.1818, simple_loss=0.27, pruned_loss=0.04676, over 16874.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.289, pruned_loss=0.05864, over 3047365.79 frames. ], batch size: 109, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:06:25,377 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=220555.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:06:35,403 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=220561.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:07:05,330 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8989, 4.8791, 4.6569, 3.9524, 4.7789, 1.6773, 4.5026, 4.4347], device='cuda:4'), covar=tensor([0.0081, 0.0077, 0.0190, 0.0356, 0.0088, 0.2849, 0.0128, 0.0216], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0154, 0.0196, 0.0177, 0.0173, 0.0206, 0.0185, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:07:13,115 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=220585.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:07:41,479 INFO [train.py:904] (4/8) Epoch 22, batch 7450, loss[loss=0.1978, simple_loss=0.294, pruned_loss=0.05078, over 16674.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2891, pruned_loss=0.05847, over 3078768.87 frames. ], batch size: 62, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:07:42,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7217, 1.7869, 1.5842, 1.5288, 1.9561, 1.6382, 1.6625, 1.9619], device='cuda:4'), covar=tensor([0.0206, 0.0315, 0.0436, 0.0371, 0.0218, 0.0284, 0.0175, 0.0225], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0232, 0.0225, 0.0223, 0.0234, 0.0232, 0.0232, 0.0227], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:07:43,936 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.847e+02 2.745e+02 3.256e+02 3.859e+02 7.938e+02, threshold=6.512e+02, percent-clipped=1.0 2023-05-01 12:08:01,209 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=220614.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:08:04,711 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=220616.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 12:09:00,841 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=220651.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:09:03,151 INFO [train.py:904] (4/8) Epoch 22, batch 7500, loss[loss=0.1989, simple_loss=0.2848, pruned_loss=0.05652, over 16716.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.2895, pruned_loss=0.05794, over 3082266.26 frames. ], batch size: 89, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:09:33,652 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.22 vs. limit=5.0 2023-05-01 12:09:34,654 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9371, 4.9315, 4.6763, 4.0400, 4.8201, 1.8461, 4.5683, 4.5071], device='cuda:4'), covar=tensor([0.0081, 0.0083, 0.0192, 0.0364, 0.0091, 0.2746, 0.0115, 0.0210], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0154, 0.0197, 0.0177, 0.0173, 0.0206, 0.0185, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:09:59,232 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-01 12:10:21,148 INFO [train.py:904] (4/8) Epoch 22, batch 7550, loss[loss=0.1829, simple_loss=0.2776, pruned_loss=0.04405, over 16748.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2891, pruned_loss=0.05845, over 3072753.72 frames. ], batch size: 124, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:10:24,492 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.979e+02 2.918e+02 3.506e+02 4.599e+02 8.060e+02, threshold=7.013e+02, percent-clipped=6.0 2023-05-01 12:10:36,081 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=220712.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:11:23,170 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=220743.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:11:38,534 INFO [train.py:904] (4/8) Epoch 22, batch 7600, loss[loss=0.266, simple_loss=0.3147, pruned_loss=0.1086, over 11393.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2884, pruned_loss=0.05915, over 3062977.19 frames. ], batch size: 248, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:11:40,748 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=220754.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:11:42,525 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4343, 3.5216, 2.1834, 3.9362, 2.6179, 3.9240, 2.1609, 2.7654], device='cuda:4'), covar=tensor([0.0314, 0.0426, 0.1650, 0.0261, 0.0931, 0.0657, 0.1737, 0.0932], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0174, 0.0193, 0.0160, 0.0174, 0.0214, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 12:12:15,330 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-05-01 12:12:37,570 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=220791.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:12:55,413 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=220802.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:12:56,365 INFO [train.py:904] (4/8) Epoch 22, batch 7650, loss[loss=0.2306, simple_loss=0.3051, pruned_loss=0.07809, over 11252.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2892, pruned_loss=0.06009, over 3050711.12 frames. ], batch size: 248, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:12:59,168 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.956e+02 2.818e+02 3.454e+02 4.152e+02 9.180e+02, threshold=6.907e+02, percent-clipped=1.0 2023-05-01 12:13:27,107 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=220822.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:14:11,869 INFO [train.py:904] (4/8) Epoch 22, batch 7700, loss[loss=0.2293, simple_loss=0.2998, pruned_loss=0.07942, over 11630.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2888, pruned_loss=0.06011, over 3057919.25 frames. ], batch size: 248, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:14:27,345 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7800, 4.0050, 2.8240, 2.3619, 2.8670, 2.4871, 4.1886, 3.5704], device='cuda:4'), covar=tensor([0.3082, 0.0705, 0.2147, 0.2795, 0.2757, 0.2164, 0.0585, 0.1260], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0269, 0.0307, 0.0315, 0.0299, 0.0261, 0.0297, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 12:14:58,202 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=220883.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:15:01,861 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9755, 5.2879, 5.0830, 5.0614, 4.8211, 4.7240, 4.7220, 5.3981], device='cuda:4'), covar=tensor([0.1391, 0.0881, 0.0986, 0.0894, 0.0838, 0.0973, 0.1211, 0.0901], device='cuda:4'), in_proj_covar=tensor([0.0668, 0.0814, 0.0674, 0.0618, 0.0511, 0.0527, 0.0681, 0.0636], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:15:01,913 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=220885.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:15:29,148 INFO [train.py:904] (4/8) Epoch 22, batch 7750, loss[loss=0.2193, simple_loss=0.2955, pruned_loss=0.07154, over 11600.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2893, pruned_loss=0.06004, over 3063104.64 frames. ], batch size: 246, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:15:30,784 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=220904.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:15:32,046 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.793e+02 2.706e+02 3.407e+02 4.136e+02 9.876e+02, threshold=6.815e+02, percent-clipped=5.0 2023-05-01 12:15:40,075 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=220911.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 12:15:45,921 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=220914.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:16:09,524 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-01 12:16:15,122 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=220933.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:16:27,997 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.57 vs. limit=5.0 2023-05-01 12:16:40,646 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-05-01 12:16:42,221 INFO [train.py:904] (4/8) Epoch 22, batch 7800, loss[loss=0.1937, simple_loss=0.2806, pruned_loss=0.05336, over 16490.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2905, pruned_loss=0.06116, over 3055439.56 frames. ], batch size: 68, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:16:56,569 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=220962.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:17:00,294 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=220965.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:17:55,724 INFO [train.py:904] (4/8) Epoch 22, batch 7850, loss[loss=0.1867, simple_loss=0.2798, pruned_loss=0.04684, over 16266.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2913, pruned_loss=0.06048, over 3075236.45 frames. ], batch size: 165, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:17:58,013 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.636e+02 2.766e+02 3.554e+02 4.297e+02 6.446e+02, threshold=7.108e+02, percent-clipped=0.0 2023-05-01 12:18:02,987 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=221007.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:18:17,895 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 12:18:20,574 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7303, 4.4763, 4.4333, 3.0541, 3.8728, 4.4392, 3.7978, 2.7538], device='cuda:4'), covar=tensor([0.0486, 0.0038, 0.0039, 0.0354, 0.0098, 0.0087, 0.0093, 0.0371], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0083, 0.0084, 0.0133, 0.0098, 0.0110, 0.0095, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 12:19:09,446 INFO [train.py:904] (4/8) Epoch 22, batch 7900, loss[loss=0.2272, simple_loss=0.3075, pruned_loss=0.07347, over 15237.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2897, pruned_loss=0.05981, over 3060205.77 frames. ], batch size: 190, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:19:24,394 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1562, 3.7757, 3.7687, 2.4468, 3.4625, 3.8110, 3.4124, 2.2691], device='cuda:4'), covar=tensor([0.0602, 0.0064, 0.0066, 0.0438, 0.0109, 0.0121, 0.0113, 0.0441], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0083, 0.0084, 0.0133, 0.0097, 0.0110, 0.0094, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 12:20:27,100 INFO [train.py:904] (4/8) Epoch 22, batch 7950, loss[loss=0.2147, simple_loss=0.2966, pruned_loss=0.0664, over 16674.00 frames. ], tot_loss[loss=0.2049, simple_loss=0.29, pruned_loss=0.05992, over 3069888.83 frames. ], batch size: 134, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:20:32,013 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.765e+02 2.674e+02 3.039e+02 3.526e+02 5.687e+02, threshold=6.078e+02, percent-clipped=0.0 2023-05-01 12:21:10,490 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7692, 2.6297, 2.3393, 3.6713, 2.4693, 3.7647, 1.5075, 2.6681], device='cuda:4'), covar=tensor([0.1306, 0.0762, 0.1256, 0.0214, 0.0204, 0.0388, 0.1724, 0.0884], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0195, 0.0190, 0.0207, 0.0215, 0.0203, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 12:21:41,818 INFO [train.py:904] (4/8) Epoch 22, batch 8000, loss[loss=0.2102, simple_loss=0.2889, pruned_loss=0.06577, over 16722.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2906, pruned_loss=0.0607, over 3069745.35 frames. ], batch size: 57, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:22:19,111 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=221178.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:22:43,891 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5234, 2.4712, 2.2468, 4.0275, 2.5701, 3.8473, 1.3310, 2.6800], device='cuda:4'), covar=tensor([0.1674, 0.1010, 0.1526, 0.0270, 0.0288, 0.0426, 0.2141, 0.0992], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0174, 0.0194, 0.0189, 0.0207, 0.0214, 0.0202, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 12:22:55,165 INFO [train.py:904] (4/8) Epoch 22, batch 8050, loss[loss=0.1898, simple_loss=0.2813, pruned_loss=0.04915, over 16792.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2905, pruned_loss=0.06004, over 3077682.24 frames. ], batch size: 83, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:23:01,251 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.760e+02 2.816e+02 3.381e+02 4.131e+02 1.220e+03, threshold=6.763e+02, percent-clipped=7.0 2023-05-01 12:23:07,206 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=221211.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:24:08,783 INFO [train.py:904] (4/8) Epoch 22, batch 8100, loss[loss=0.1958, simple_loss=0.2845, pruned_loss=0.05359, over 16390.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.29, pruned_loss=0.05937, over 3095328.87 frames. ], batch size: 146, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:24:17,302 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=221259.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:24:19,980 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=221260.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:25:24,335 INFO [train.py:904] (4/8) Epoch 22, batch 8150, loss[loss=0.201, simple_loss=0.2716, pruned_loss=0.06518, over 11638.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.2874, pruned_loss=0.05841, over 3094856.41 frames. ], batch size: 248, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:25:31,011 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.995e+02 2.655e+02 3.146e+02 3.866e+02 6.459e+02, threshold=6.292e+02, percent-clipped=0.0 2023-05-01 12:25:31,341 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=221307.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:26:41,082 INFO [train.py:904] (4/8) Epoch 22, batch 8200, loss[loss=0.194, simple_loss=0.2822, pruned_loss=0.05292, over 16476.00 frames. ], tot_loss[loss=0.1994, simple_loss=0.2843, pruned_loss=0.05727, over 3091162.42 frames. ], batch size: 68, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:26:44,114 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=221355.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:27:58,679 INFO [train.py:904] (4/8) Epoch 22, batch 8250, loss[loss=0.1985, simple_loss=0.2848, pruned_loss=0.05606, over 16879.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2834, pruned_loss=0.05474, over 3082299.76 frames. ], batch size: 42, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:28:05,605 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.564e+02 2.638e+02 3.060e+02 3.657e+02 7.056e+02, threshold=6.120e+02, percent-clipped=1.0 2023-05-01 12:29:17,813 INFO [train.py:904] (4/8) Epoch 22, batch 8300, loss[loss=0.1935, simple_loss=0.2872, pruned_loss=0.04996, over 16173.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.281, pruned_loss=0.05187, over 3069778.29 frames. ], batch size: 165, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:29:57,540 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=221478.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:30:38,299 INFO [train.py:904] (4/8) Epoch 22, batch 8350, loss[loss=0.2166, simple_loss=0.3073, pruned_loss=0.06294, over 16734.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2803, pruned_loss=0.05021, over 3053469.87 frames. ], batch size: 124, lr: 3.05e-03, grad_scale: 4.0 2023-05-01 12:30:43,699 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.406e+02 2.123e+02 2.436e+02 3.235e+02 5.612e+02, threshold=4.873e+02, percent-clipped=0.0 2023-05-01 12:30:44,280 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=221507.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:31:15,660 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=221526.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:31:29,401 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5023, 4.5794, 4.3996, 4.0969, 4.0667, 4.4948, 4.2556, 4.1883], device='cuda:4'), covar=tensor([0.0587, 0.0652, 0.0317, 0.0316, 0.0946, 0.0548, 0.0567, 0.0743], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0422, 0.0336, 0.0331, 0.0341, 0.0387, 0.0231, 0.0401], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:31:56,139 INFO [train.py:904] (4/8) Epoch 22, batch 8400, loss[loss=0.1734, simple_loss=0.2665, pruned_loss=0.04016, over 16858.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2781, pruned_loss=0.04871, over 3030113.98 frames. ], batch size: 116, lr: 3.05e-03, grad_scale: 8.0 2023-05-01 12:32:08,948 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=221560.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:32:19,746 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9706, 4.2494, 4.1006, 4.1247, 3.7834, 3.8883, 3.8786, 4.2500], device='cuda:4'), covar=tensor([0.1112, 0.0888, 0.0932, 0.0861, 0.0846, 0.1660, 0.0943, 0.0953], device='cuda:4'), in_proj_covar=tensor([0.0662, 0.0803, 0.0666, 0.0610, 0.0505, 0.0521, 0.0672, 0.0629], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:32:21,008 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=221568.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:33:17,378 INFO [train.py:904] (4/8) Epoch 22, batch 8450, loss[loss=0.1625, simple_loss=0.2584, pruned_loss=0.03333, over 16153.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2764, pruned_loss=0.0469, over 3043301.58 frames. ], batch size: 165, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:33:24,323 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.610e+02 2.296e+02 2.720e+02 3.439e+02 5.542e+02, threshold=5.440e+02, percent-clipped=2.0 2023-05-01 12:33:26,152 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=221608.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:34:38,832 INFO [train.py:904] (4/8) Epoch 22, batch 8500, loss[loss=0.1653, simple_loss=0.2574, pruned_loss=0.03663, over 16407.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2728, pruned_loss=0.04459, over 3050157.91 frames. ], batch size: 68, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:35:39,619 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8227, 3.7880, 4.1148, 4.1124, 4.1039, 3.9237, 3.8730, 3.9143], device='cuda:4'), covar=tensor([0.0461, 0.0855, 0.0559, 0.0537, 0.0599, 0.0541, 0.1061, 0.0567], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0457, 0.0440, 0.0408, 0.0487, 0.0463, 0.0547, 0.0371], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 12:35:44,622 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-05-01 12:35:59,815 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 12:36:02,539 INFO [train.py:904] (4/8) Epoch 22, batch 8550, loss[loss=0.1692, simple_loss=0.2667, pruned_loss=0.0358, over 16880.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2704, pruned_loss=0.04362, over 3034477.94 frames. ], batch size: 96, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:36:10,060 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.538e+02 2.073e+02 2.527e+02 2.950e+02 5.682e+02, threshold=5.053e+02, percent-clipped=1.0 2023-05-01 12:37:41,421 INFO [train.py:904] (4/8) Epoch 22, batch 8600, loss[loss=0.1761, simple_loss=0.283, pruned_loss=0.03455, over 16887.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2705, pruned_loss=0.04275, over 3020476.60 frames. ], batch size: 90, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:37:53,223 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=221758.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:39:21,622 INFO [train.py:904] (4/8) Epoch 22, batch 8650, loss[loss=0.1674, simple_loss=0.2641, pruned_loss=0.03536, over 16736.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2691, pruned_loss=0.04135, over 3015905.45 frames. ], batch size: 134, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:39:30,557 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.618e+02 2.228e+02 2.506e+02 3.024e+02 6.242e+02, threshold=5.012e+02, percent-clipped=2.0 2023-05-01 12:39:58,940 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=221819.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:40:39,736 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 12:41:06,645 INFO [train.py:904] (4/8) Epoch 22, batch 8700, loss[loss=0.1592, simple_loss=0.2462, pruned_loss=0.03608, over 17041.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2668, pruned_loss=0.04032, over 3022980.55 frames. ], batch size: 55, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:41:27,672 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=221863.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:41:42,965 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3944, 3.5074, 3.7172, 3.7004, 3.7078, 3.5387, 3.5707, 3.6014], device='cuda:4'), covar=tensor([0.0518, 0.1058, 0.0686, 0.0800, 0.0749, 0.1010, 0.0911, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0449, 0.0434, 0.0400, 0.0480, 0.0454, 0.0538, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 12:42:16,878 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=221890.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:42:24,233 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9927, 3.0491, 1.7498, 3.3023, 2.2848, 3.2567, 2.0088, 2.5558], device='cuda:4'), covar=tensor([0.0318, 0.0391, 0.1794, 0.0256, 0.0835, 0.0610, 0.1628, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0171, 0.0189, 0.0157, 0.0171, 0.0209, 0.0198, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 12:42:42,866 INFO [train.py:904] (4/8) Epoch 22, batch 8750, loss[loss=0.156, simple_loss=0.2673, pruned_loss=0.02237, over 16856.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2662, pruned_loss=0.03969, over 3025912.33 frames. ], batch size: 96, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:42:47,144 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2537, 1.5031, 1.9508, 2.2397, 2.2780, 2.5683, 1.7122, 2.4996], device='cuda:4'), covar=tensor([0.0256, 0.0634, 0.0363, 0.0360, 0.0387, 0.0222, 0.0621, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0189, 0.0175, 0.0180, 0.0192, 0.0149, 0.0192, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:42:53,174 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.058e+02 2.561e+02 3.126e+02 5.404e+02, threshold=5.122e+02, percent-clipped=2.0 2023-05-01 12:43:31,312 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-05-01 12:44:31,200 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=221951.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:44:34,617 INFO [train.py:904] (4/8) Epoch 22, batch 8800, loss[loss=0.177, simple_loss=0.2672, pruned_loss=0.04336, over 12433.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2645, pruned_loss=0.03861, over 3028756.95 frames. ], batch size: 247, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:45:18,501 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1162, 4.0962, 4.4319, 4.4168, 4.4526, 4.2087, 4.2102, 4.1361], device='cuda:4'), covar=tensor([0.0292, 0.0561, 0.0404, 0.0416, 0.0366, 0.0355, 0.0733, 0.0419], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0444, 0.0430, 0.0397, 0.0475, 0.0449, 0.0532, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 12:45:46,745 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8326, 3.1261, 3.5654, 1.9091, 2.8878, 2.2060, 3.3891, 3.2649], device='cuda:4'), covar=tensor([0.0271, 0.0945, 0.0553, 0.2240, 0.0906, 0.1087, 0.0649, 0.1084], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0159, 0.0163, 0.0151, 0.0142, 0.0127, 0.0140, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 12:45:58,220 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7032, 3.4020, 3.8884, 1.7471, 3.9747, 4.1096, 3.0663, 2.9783], device='cuda:4'), covar=tensor([0.0712, 0.0258, 0.0177, 0.1310, 0.0081, 0.0128, 0.0427, 0.0443], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0104, 0.0093, 0.0133, 0.0077, 0.0119, 0.0124, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 12:46:22,303 INFO [train.py:904] (4/8) Epoch 22, batch 8850, loss[loss=0.156, simple_loss=0.2464, pruned_loss=0.03284, over 12074.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2676, pruned_loss=0.03795, over 3035771.25 frames. ], batch size: 249, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:46:28,904 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.701e+02 2.177e+02 2.540e+02 3.054e+02 9.241e+02, threshold=5.079e+02, percent-clipped=4.0 2023-05-01 12:46:35,332 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=222009.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:48:07,874 INFO [train.py:904] (4/8) Epoch 22, batch 8900, loss[loss=0.1608, simple_loss=0.2563, pruned_loss=0.03261, over 16528.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2676, pruned_loss=0.03714, over 3037922.68 frames. ], batch size: 62, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 12:48:42,790 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=222070.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:49:14,099 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0830, 4.0457, 3.9470, 3.2517, 4.0151, 1.8006, 3.7798, 3.6220], device='cuda:4'), covar=tensor([0.0094, 0.0092, 0.0169, 0.0265, 0.0098, 0.2582, 0.0134, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0152, 0.0193, 0.0172, 0.0171, 0.0204, 0.0182, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:50:11,677 INFO [train.py:904] (4/8) Epoch 22, batch 8950, loss[loss=0.1721, simple_loss=0.2618, pruned_loss=0.04125, over 12699.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2675, pruned_loss=0.0374, over 3056770.04 frames. ], batch size: 250, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 12:50:23,606 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.486e+02 2.055e+02 2.537e+02 3.242e+02 5.018e+02, threshold=5.074e+02, percent-clipped=0.0 2023-05-01 12:50:37,674 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=222114.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:50:54,572 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-05-01 12:52:03,798 INFO [train.py:904] (4/8) Epoch 22, batch 9000, loss[loss=0.1597, simple_loss=0.2546, pruned_loss=0.03243, over 16333.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2641, pruned_loss=0.0362, over 3064621.45 frames. ], batch size: 146, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 12:52:03,799 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 12:52:14,709 INFO [train.py:938] (4/8) Epoch 22, validation: loss=0.1453, simple_loss=0.2492, pruned_loss=0.0207, over 944034.00 frames. 2023-05-01 12:52:14,709 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 12:52:31,537 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5735, 2.5404, 2.1632, 4.0338, 2.4282, 3.9098, 1.4062, 2.8329], device='cuda:4'), covar=tensor([0.1451, 0.0845, 0.1412, 0.0147, 0.0140, 0.0353, 0.1820, 0.0766], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0173, 0.0193, 0.0184, 0.0201, 0.0212, 0.0201, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 12:52:36,382 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=222163.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:53:22,206 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=222185.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:53:50,887 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-05-01 12:53:58,628 INFO [train.py:904] (4/8) Epoch 22, batch 9050, loss[loss=0.1583, simple_loss=0.2506, pruned_loss=0.03295, over 12837.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2653, pruned_loss=0.03683, over 3078757.26 frames. ], batch size: 248, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 12:54:04,295 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=222205.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:54:09,041 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.330e+02 2.762e+02 3.325e+02 5.542e+02, threshold=5.524e+02, percent-clipped=4.0 2023-05-01 12:54:17,175 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=222211.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:55:31,424 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=222246.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:55:31,560 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=222246.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 12:55:44,664 INFO [train.py:904] (4/8) Epoch 22, batch 9100, loss[loss=0.1755, simple_loss=0.2787, pruned_loss=0.03611, over 15329.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2648, pruned_loss=0.03736, over 3081045.50 frames. ], batch size: 191, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 12:55:50,482 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 12:56:06,304 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3642, 1.6385, 2.0619, 2.3373, 2.3044, 2.6291, 1.9087, 2.5088], device='cuda:4'), covar=tensor([0.0269, 0.0576, 0.0358, 0.0357, 0.0388, 0.0196, 0.0546, 0.0162], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0187, 0.0173, 0.0178, 0.0191, 0.0147, 0.0190, 0.0146], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 12:56:11,299 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=222266.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 12:57:42,765 INFO [train.py:904] (4/8) Epoch 22, batch 9150, loss[loss=0.1419, simple_loss=0.2412, pruned_loss=0.02131, over 16815.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2651, pruned_loss=0.03689, over 3089202.59 frames. ], batch size: 83, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 12:57:53,763 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.620e+02 2.177e+02 2.613e+02 3.280e+02 4.905e+02, threshold=5.227e+02, percent-clipped=0.0 2023-05-01 12:59:12,874 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.84 vs. limit=5.0 2023-05-01 12:59:28,938 INFO [train.py:904] (4/8) Epoch 22, batch 9200, loss[loss=0.1657, simple_loss=0.2505, pruned_loss=0.04051, over 12459.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2606, pruned_loss=0.03593, over 3081196.76 frames. ], batch size: 250, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 12:59:52,285 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=222365.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:01:05,903 INFO [train.py:904] (4/8) Epoch 22, batch 9250, loss[loss=0.1473, simple_loss=0.2485, pruned_loss=0.02304, over 16887.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2607, pruned_loss=0.03623, over 3079257.65 frames. ], batch size: 96, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 13:01:16,243 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.499e+02 2.130e+02 2.525e+02 3.050e+02 5.911e+02, threshold=5.049e+02, percent-clipped=2.0 2023-05-01 13:01:28,861 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=222414.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:02:56,973 INFO [train.py:904] (4/8) Epoch 22, batch 9300, loss[loss=0.1464, simple_loss=0.2358, pruned_loss=0.02853, over 16615.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.259, pruned_loss=0.03574, over 3074952.91 frames. ], batch size: 57, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 13:03:17,564 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=222462.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:03:38,699 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6239, 3.6967, 2.1178, 4.2515, 2.7972, 4.1302, 2.2355, 3.0224], device='cuda:4'), covar=tensor([0.0289, 0.0366, 0.1805, 0.0258, 0.0826, 0.0494, 0.1700, 0.0744], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0171, 0.0188, 0.0156, 0.0172, 0.0208, 0.0199, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:04:36,512 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6749, 2.7206, 2.3158, 4.1379, 2.5359, 4.0035, 1.5128, 2.9255], device='cuda:4'), covar=tensor([0.1416, 0.0788, 0.1300, 0.0169, 0.0139, 0.0359, 0.1760, 0.0737], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0173, 0.0193, 0.0184, 0.0200, 0.0212, 0.0201, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:04:40,775 INFO [train.py:904] (4/8) Epoch 22, batch 9350, loss[loss=0.1608, simple_loss=0.2625, pruned_loss=0.02957, over 16894.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2588, pruned_loss=0.03569, over 3073789.35 frames. ], batch size: 102, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 13:04:49,913 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.318e+02 1.994e+02 2.436e+02 3.069e+02 5.740e+02, threshold=4.871e+02, percent-clipped=2.0 2023-05-01 13:05:01,605 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4902, 3.3946, 2.7202, 2.1455, 2.1662, 2.3180, 3.5263, 3.0164], device='cuda:4'), covar=tensor([0.2787, 0.0608, 0.1698, 0.2996, 0.2895, 0.2253, 0.0389, 0.1360], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0261, 0.0298, 0.0307, 0.0286, 0.0255, 0.0288, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 13:05:21,183 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1733, 2.3669, 1.9894, 2.2584, 2.7533, 2.4088, 2.7086, 2.8848], device='cuda:4'), covar=tensor([0.0148, 0.0442, 0.0569, 0.0442, 0.0294, 0.0437, 0.0223, 0.0258], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0228, 0.0221, 0.0220, 0.0228, 0.0227, 0.0224, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:05:23,310 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=222523.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:05:23,402 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6337, 2.0975, 1.7645, 1.9677, 2.4127, 2.1061, 2.0839, 2.5231], device='cuda:4'), covar=tensor([0.0163, 0.0446, 0.0569, 0.0447, 0.0283, 0.0406, 0.0173, 0.0253], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0228, 0.0221, 0.0220, 0.0229, 0.0227, 0.0224, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:05:24,982 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6410, 2.0705, 1.6933, 1.8994, 2.4115, 2.0930, 2.1075, 2.4937], device='cuda:4'), covar=tensor([0.0157, 0.0458, 0.0606, 0.0515, 0.0277, 0.0430, 0.0203, 0.0280], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0228, 0.0221, 0.0220, 0.0229, 0.0227, 0.0224, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:05:57,779 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=222541.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 13:06:08,602 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=222546.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:06:20,190 INFO [train.py:904] (4/8) Epoch 22, batch 9400, loss[loss=0.184, simple_loss=0.2826, pruned_loss=0.04273, over 16861.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2597, pruned_loss=0.0354, over 3081731.59 frames. ], batch size: 90, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 13:06:38,147 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=222561.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:06:53,458 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2953, 4.1260, 4.3733, 4.4875, 4.6365, 4.2026, 4.6183, 4.6562], device='cuda:4'), covar=tensor([0.1761, 0.1152, 0.1522, 0.0742, 0.0602, 0.1242, 0.0669, 0.0623], device='cuda:4'), in_proj_covar=tensor([0.0605, 0.0751, 0.0865, 0.0757, 0.0578, 0.0602, 0.0629, 0.0725], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:06:55,322 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=222570.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:06:55,431 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8238, 3.1360, 3.2499, 1.7074, 2.6328, 1.8129, 3.2664, 3.3663], device='cuda:4'), covar=tensor([0.0271, 0.0928, 0.0633, 0.2753, 0.1100, 0.1505, 0.0670, 0.0986], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0157, 0.0161, 0.0149, 0.0140, 0.0126, 0.0138, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:07:24,491 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=222584.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 13:07:43,444 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=222594.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:08:00,581 INFO [train.py:904] (4/8) Epoch 22, batch 9450, loss[loss=0.1632, simple_loss=0.261, pruned_loss=0.03267, over 15244.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2606, pruned_loss=0.03579, over 3046911.31 frames. ], batch size: 190, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 13:08:08,412 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.477e+02 2.238e+02 2.730e+02 3.189e+02 9.782e+02, threshold=5.460e+02, percent-clipped=5.0 2023-05-01 13:08:23,938 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6931, 3.8235, 2.1364, 4.3801, 2.8837, 4.2175, 2.2496, 3.0337], device='cuda:4'), covar=tensor([0.0327, 0.0414, 0.2022, 0.0274, 0.0841, 0.0689, 0.1859, 0.0860], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0171, 0.0188, 0.0156, 0.0172, 0.0208, 0.0198, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:08:56,865 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=222631.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:09:40,773 INFO [train.py:904] (4/8) Epoch 22, batch 9500, loss[loss=0.1681, simple_loss=0.2683, pruned_loss=0.03396, over 16329.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2599, pruned_loss=0.03557, over 3061257.86 frames. ], batch size: 146, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 13:10:07,034 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=222665.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:10:21,030 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9898, 4.2563, 4.0955, 4.1131, 3.7706, 3.8555, 3.8569, 4.2523], device='cuda:4'), covar=tensor([0.1128, 0.1002, 0.1005, 0.0813, 0.0835, 0.1834, 0.1018, 0.0920], device='cuda:4'), in_proj_covar=tensor([0.0641, 0.0779, 0.0642, 0.0592, 0.0492, 0.0506, 0.0652, 0.0610], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:11:22,394 INFO [train.py:904] (4/8) Epoch 22, batch 9550, loss[loss=0.1788, simple_loss=0.2812, pruned_loss=0.03814, over 15258.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2589, pruned_loss=0.03528, over 3071248.17 frames. ], batch size: 191, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 13:11:34,507 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.364e+02 2.117e+02 2.369e+02 3.001e+02 4.954e+02, threshold=4.737e+02, percent-clipped=0.0 2023-05-01 13:11:44,580 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=222713.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:11:46,914 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-01 13:11:49,019 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3826, 4.4092, 4.7463, 4.7336, 4.7445, 4.5052, 4.4776, 4.3792], device='cuda:4'), covar=tensor([0.0342, 0.0732, 0.0470, 0.0444, 0.0484, 0.0391, 0.0812, 0.0438], device='cuda:4'), in_proj_covar=tensor([0.0393, 0.0437, 0.0425, 0.0393, 0.0469, 0.0445, 0.0525, 0.0358], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 13:13:00,660 INFO [train.py:904] (4/8) Epoch 22, batch 9600, loss[loss=0.1587, simple_loss=0.2564, pruned_loss=0.03056, over 16604.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2608, pruned_loss=0.03636, over 3066881.99 frames. ], batch size: 68, lr: 3.04e-03, grad_scale: 8.0 2023-05-01 13:13:31,687 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6547, 1.9997, 1.6994, 1.7660, 2.3366, 2.0454, 2.0531, 2.4866], device='cuda:4'), covar=tensor([0.0165, 0.0480, 0.0597, 0.0591, 0.0310, 0.0454, 0.0212, 0.0275], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0228, 0.0220, 0.0220, 0.0228, 0.0227, 0.0223, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:14:27,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2082, 3.6270, 3.6315, 2.3981, 3.2840, 3.6229, 3.3953, 2.1436], device='cuda:4'), covar=tensor([0.0556, 0.0051, 0.0052, 0.0415, 0.0107, 0.0088, 0.0092, 0.0491], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0081, 0.0082, 0.0132, 0.0096, 0.0107, 0.0092, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 13:14:44,327 INFO [train.py:904] (4/8) Epoch 22, batch 9650, loss[loss=0.1701, simple_loss=0.2801, pruned_loss=0.03001, over 16932.00 frames. ], tot_loss[loss=0.168, simple_loss=0.263, pruned_loss=0.0365, over 3068992.01 frames. ], batch size: 102, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 13:14:58,745 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.544e+02 2.257e+02 2.597e+02 3.466e+02 1.012e+03, threshold=5.195e+02, percent-clipped=6.0 2023-05-01 13:15:50,854 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0045, 2.2306, 2.3135, 3.0812, 1.8167, 3.2510, 1.7886, 2.8418], device='cuda:4'), covar=tensor([0.1171, 0.0701, 0.1068, 0.0158, 0.0076, 0.0331, 0.1471, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0170, 0.0190, 0.0181, 0.0196, 0.0209, 0.0198, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:16:03,729 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=222841.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 13:16:27,395 INFO [train.py:904] (4/8) Epoch 22, batch 9700, loss[loss=0.183, simple_loss=0.2716, pruned_loss=0.04725, over 16786.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2621, pruned_loss=0.03618, over 3072941.61 frames. ], batch size: 124, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 13:16:43,295 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=222861.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:17:18,472 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.04 vs. limit=2.0 2023-05-01 13:17:22,888 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=222879.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 13:17:42,992 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=222889.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:18:08,547 INFO [train.py:904] (4/8) Epoch 22, batch 9750, loss[loss=0.1496, simple_loss=0.2375, pruned_loss=0.0308, over 12557.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2607, pruned_loss=0.03637, over 3061421.41 frames. ], batch size: 248, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 13:18:18,158 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 2.037e+02 2.441e+02 3.104e+02 5.362e+02, threshold=4.883e+02, percent-clipped=3.0 2023-05-01 13:18:19,216 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=222909.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:18:52,780 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=222926.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:19:08,722 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0702, 5.3799, 5.1737, 5.2002, 4.8565, 4.8448, 4.7283, 5.4696], device='cuda:4'), covar=tensor([0.1260, 0.0887, 0.1031, 0.0837, 0.0862, 0.0891, 0.1285, 0.0824], device='cuda:4'), in_proj_covar=tensor([0.0645, 0.0783, 0.0642, 0.0594, 0.0494, 0.0508, 0.0654, 0.0611], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:19:24,040 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=222940.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:19:42,566 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4682, 3.3785, 3.7546, 1.9014, 3.8918, 4.0167, 3.0079, 2.9774], device='cuda:4'), covar=tensor([0.0804, 0.0276, 0.0182, 0.1301, 0.0081, 0.0121, 0.0463, 0.0463], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0104, 0.0093, 0.0135, 0.0077, 0.0119, 0.0124, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 13:19:45,665 INFO [train.py:904] (4/8) Epoch 22, batch 9800, loss[loss=0.172, simple_loss=0.277, pruned_loss=0.03349, over 16722.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2608, pruned_loss=0.03559, over 3071860.49 frames. ], batch size: 89, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 13:20:15,211 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=222969.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:20:40,777 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=222984.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:21:22,974 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=223001.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:21:26,486 INFO [train.py:904] (4/8) Epoch 22, batch 9850, loss[loss=0.1588, simple_loss=0.2559, pruned_loss=0.03087, over 16393.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2614, pruned_loss=0.03514, over 3068668.95 frames. ], batch size: 146, lr: 3.04e-03, grad_scale: 4.0 2023-05-01 13:21:37,440 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.461e+02 2.181e+02 2.565e+02 3.075e+02 7.070e+02, threshold=5.130e+02, percent-clipped=2.0 2023-05-01 13:21:56,454 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0463, 1.7700, 1.5799, 1.3998, 1.9499, 1.5830, 1.5081, 1.9162], device='cuda:4'), covar=tensor([0.0212, 0.0422, 0.0582, 0.0508, 0.0298, 0.0384, 0.0185, 0.0275], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0229, 0.0222, 0.0221, 0.0230, 0.0228, 0.0224, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:22:19,895 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=223030.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:23:01,413 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=223045.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:23:10,085 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5715, 3.4975, 3.5111, 2.8351, 3.3974, 2.0515, 3.1618, 2.8348], device='cuda:4'), covar=tensor([0.0136, 0.0144, 0.0167, 0.0190, 0.0119, 0.2341, 0.0138, 0.0259], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0150, 0.0190, 0.0167, 0.0169, 0.0203, 0.0180, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:23:15,937 INFO [train.py:904] (4/8) Epoch 22, batch 9900, loss[loss=0.1734, simple_loss=0.2769, pruned_loss=0.03494, over 16301.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2617, pruned_loss=0.03516, over 3067560.98 frames. ], batch size: 165, lr: 3.03e-03, grad_scale: 4.0 2023-05-01 13:23:43,528 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=223064.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:25:13,509 INFO [train.py:904] (4/8) Epoch 22, batch 9950, loss[loss=0.1711, simple_loss=0.2719, pruned_loss=0.03508, over 16349.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2635, pruned_loss=0.03523, over 3061816.77 frames. ], batch size: 146, lr: 3.03e-03, grad_scale: 4.0 2023-05-01 13:25:27,890 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.410e+02 2.209e+02 2.588e+02 3.009e+02 4.332e+02, threshold=5.177e+02, percent-clipped=0.0 2023-05-01 13:26:07,059 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=223125.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:26:35,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3133, 2.1046, 2.6366, 3.2815, 3.0089, 3.6008, 2.3816, 3.6164], device='cuda:4'), covar=tensor([0.0170, 0.0520, 0.0390, 0.0214, 0.0286, 0.0133, 0.0495, 0.0140], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0187, 0.0175, 0.0177, 0.0191, 0.0147, 0.0191, 0.0145], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:27:13,207 INFO [train.py:904] (4/8) Epoch 22, batch 10000, loss[loss=0.1736, simple_loss=0.2627, pruned_loss=0.04222, over 13104.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2624, pruned_loss=0.0351, over 3075168.46 frames. ], batch size: 248, lr: 3.03e-03, grad_scale: 8.0 2023-05-01 13:27:51,685 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=223173.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 13:28:04,772 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=223179.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 13:28:54,091 INFO [train.py:904] (4/8) Epoch 22, batch 10050, loss[loss=0.1749, simple_loss=0.2742, pruned_loss=0.03778, over 16353.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.263, pruned_loss=0.03545, over 3062306.00 frames. ], batch size: 146, lr: 3.03e-03, grad_scale: 8.0 2023-05-01 13:29:04,328 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.317e+02 2.078e+02 2.413e+02 2.769e+02 4.519e+02, threshold=4.827e+02, percent-clipped=0.0 2023-05-01 13:29:15,986 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9203, 2.2030, 1.7785, 2.0231, 2.5222, 2.2518, 2.3866, 2.7322], device='cuda:4'), covar=tensor([0.0181, 0.0532, 0.0689, 0.0581, 0.0359, 0.0464, 0.0281, 0.0303], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0228, 0.0221, 0.0220, 0.0229, 0.0227, 0.0223, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:29:32,732 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-01 13:29:40,703 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=223226.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:29:42,097 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=223227.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:29:54,674 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=223234.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 13:30:27,326 INFO [train.py:904] (4/8) Epoch 22, batch 10100, loss[loss=0.1544, simple_loss=0.2467, pruned_loss=0.03105, over 16215.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2632, pruned_loss=0.03563, over 3071165.76 frames. ], batch size: 165, lr: 3.03e-03, grad_scale: 8.0 2023-05-01 13:30:51,218 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5634, 2.9310, 3.2722, 1.9770, 2.7969, 2.0410, 3.0929, 3.1277], device='cuda:4'), covar=tensor([0.0309, 0.0896, 0.0500, 0.2089, 0.0838, 0.1073, 0.0680, 0.1022], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0156, 0.0162, 0.0150, 0.0141, 0.0126, 0.0138, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:31:10,810 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=223274.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:31:11,258 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.94 vs. limit=5.0 2023-05-01 13:31:39,312 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=223296.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:32:13,429 INFO [train.py:904] (4/8) Epoch 23, batch 0, loss[loss=0.167, simple_loss=0.2629, pruned_loss=0.03551, over 17055.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2629, pruned_loss=0.03551, over 17055.00 frames. ], batch size: 50, lr: 2.97e-03, grad_scale: 8.0 2023-05-01 13:32:13,430 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 13:32:20,848 INFO [train.py:938] (4/8) Epoch 23, validation: loss=0.1454, simple_loss=0.2487, pruned_loss=0.02111, over 944034.00 frames. 2023-05-01 13:32:20,849 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 13:32:28,407 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.331e+02 2.627e+02 3.041e+02 3.739e+02 7.225e+02, threshold=6.083e+02, percent-clipped=7.0 2023-05-01 13:32:40,767 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6910, 3.9506, 4.1504, 4.1052, 4.1529, 3.9334, 3.6428, 3.9308], device='cuda:4'), covar=tensor([0.0665, 0.0916, 0.0674, 0.0807, 0.0881, 0.0786, 0.1660, 0.0672], device='cuda:4'), in_proj_covar=tensor([0.0392, 0.0436, 0.0426, 0.0393, 0.0467, 0.0445, 0.0524, 0.0357], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 13:32:49,520 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=223325.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:33:09,274 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=223340.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:33:26,112 INFO [train.py:904] (4/8) Epoch 23, batch 50, loss[loss=0.1805, simple_loss=0.2697, pruned_loss=0.04561, over 16505.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2689, pruned_loss=0.04783, over 758646.64 frames. ], batch size: 68, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:34:32,069 INFO [train.py:904] (4/8) Epoch 23, batch 100, loss[loss=0.1465, simple_loss=0.236, pruned_loss=0.02844, over 16857.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2645, pruned_loss=0.046, over 1327165.33 frames. ], batch size: 39, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:34:35,402 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6620, 4.5644, 4.5520, 4.1967, 4.2486, 4.5874, 4.3979, 4.3184], device='cuda:4'), covar=tensor([0.0676, 0.0917, 0.0384, 0.0375, 0.1009, 0.0529, 0.0525, 0.0732], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0411, 0.0328, 0.0323, 0.0333, 0.0379, 0.0223, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-05-01 13:34:42,058 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.587e+02 2.242e+02 2.764e+02 3.330e+02 6.883e+02, threshold=5.527e+02, percent-clipped=1.0 2023-05-01 13:34:55,418 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=223420.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:35:38,698 INFO [train.py:904] (4/8) Epoch 23, batch 150, loss[loss=0.1504, simple_loss=0.2481, pruned_loss=0.0264, over 17028.00 frames. ], tot_loss[loss=0.1764, simple_loss=0.2628, pruned_loss=0.04497, over 1771613.87 frames. ], batch size: 50, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:36:47,606 INFO [train.py:904] (4/8) Epoch 23, batch 200, loss[loss=0.1516, simple_loss=0.2376, pruned_loss=0.03282, over 16803.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2636, pruned_loss=0.04543, over 2095331.79 frames. ], batch size: 42, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:36:57,903 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.401e+02 2.279e+02 2.532e+02 3.033e+02 5.752e+02, threshold=5.065e+02, percent-clipped=1.0 2023-05-01 13:37:21,176 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=223529.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 13:37:21,345 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9452, 1.9766, 2.5555, 2.8594, 2.7284, 3.2779, 2.2046, 3.2600], device='cuda:4'), covar=tensor([0.0269, 0.0543, 0.0355, 0.0331, 0.0385, 0.0216, 0.0571, 0.0174], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0190, 0.0177, 0.0180, 0.0194, 0.0151, 0.0194, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:37:42,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9166, 5.4454, 5.5556, 5.2616, 5.3636, 5.9619, 5.3897, 5.1073], device='cuda:4'), covar=tensor([0.1126, 0.1869, 0.2429, 0.2041, 0.2582, 0.0923, 0.1484, 0.2361], device='cuda:4'), in_proj_covar=tensor([0.0396, 0.0575, 0.0636, 0.0474, 0.0638, 0.0662, 0.0500, 0.0636], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 13:37:52,716 INFO [train.py:904] (4/8) Epoch 23, batch 250, loss[loss=0.1927, simple_loss=0.2637, pruned_loss=0.06083, over 12124.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2625, pruned_loss=0.04563, over 2357002.76 frames. ], batch size: 246, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:37:53,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3520, 4.0870, 4.5463, 2.4868, 4.6857, 4.8197, 3.6800, 3.8572], device='cuda:4'), covar=tensor([0.0654, 0.0241, 0.0209, 0.1157, 0.0093, 0.0135, 0.0384, 0.0361], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0107, 0.0096, 0.0139, 0.0080, 0.0123, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:38:29,090 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=223578.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:38:41,780 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6857, 1.8548, 2.2739, 2.4811, 2.6516, 2.5070, 2.0426, 2.7220], device='cuda:4'), covar=tensor([0.0187, 0.0495, 0.0332, 0.0311, 0.0276, 0.0342, 0.0500, 0.0179], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0190, 0.0177, 0.0181, 0.0195, 0.0152, 0.0195, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:38:54,796 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=223596.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:39:04,044 INFO [train.py:904] (4/8) Epoch 23, batch 300, loss[loss=0.173, simple_loss=0.2586, pruned_loss=0.04373, over 16328.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2595, pruned_loss=0.04435, over 2563894.33 frames. ], batch size: 145, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:39:14,755 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.539e+02 2.219e+02 2.672e+02 3.069e+02 6.824e+02, threshold=5.344e+02, percent-clipped=3.0 2023-05-01 13:39:35,515 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=223625.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:39:47,991 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2801, 5.6155, 5.4242, 5.4062, 5.1174, 5.0626, 5.0320, 5.7441], device='cuda:4'), covar=tensor([0.1304, 0.0947, 0.1041, 0.0861, 0.0869, 0.0865, 0.1233, 0.0902], device='cuda:4'), in_proj_covar=tensor([0.0665, 0.0808, 0.0665, 0.0613, 0.0510, 0.0522, 0.0679, 0.0631], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:39:54,108 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=223639.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:39:55,225 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=223640.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:39:57,319 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2216, 3.3308, 3.5750, 2.2814, 3.0345, 2.4130, 3.6296, 3.6426], device='cuda:4'), covar=tensor([0.0243, 0.0893, 0.0599, 0.1897, 0.0870, 0.1004, 0.0536, 0.1029], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0161, 0.0166, 0.0153, 0.0144, 0.0129, 0.0142, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:40:01,694 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=223644.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:40:01,868 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3365, 5.3050, 5.0823, 4.3008, 5.1297, 2.0040, 4.8093, 4.9114], device='cuda:4'), covar=tensor([0.0106, 0.0111, 0.0241, 0.0518, 0.0137, 0.2972, 0.0176, 0.0265], device='cuda:4'), in_proj_covar=tensor([0.0165, 0.0155, 0.0195, 0.0172, 0.0173, 0.0208, 0.0185, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:40:13,647 INFO [train.py:904] (4/8) Epoch 23, batch 350, loss[loss=0.1833, simple_loss=0.258, pruned_loss=0.05428, over 16267.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2572, pruned_loss=0.04349, over 2732869.99 frames. ], batch size: 165, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:40:42,278 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=223673.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:40:45,853 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1562, 4.5624, 4.5743, 3.5785, 3.8810, 4.5577, 4.0957, 2.8930], device='cuda:4'), covar=tensor([0.0438, 0.0060, 0.0039, 0.0273, 0.0125, 0.0079, 0.0084, 0.0415], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0084, 0.0084, 0.0133, 0.0098, 0.0108, 0.0094, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 13:41:03,081 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=223688.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:41:22,666 INFO [train.py:904] (4/8) Epoch 23, batch 400, loss[loss=0.1369, simple_loss=0.2307, pruned_loss=0.0216, over 16966.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2556, pruned_loss=0.04239, over 2865461.81 frames. ], batch size: 41, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:41:27,366 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5039, 5.8518, 5.5769, 5.6552, 5.2550, 5.2826, 5.2039, 5.9757], device='cuda:4'), covar=tensor([0.1446, 0.0997, 0.1133, 0.0892, 0.0937, 0.0693, 0.1386, 0.1038], device='cuda:4'), in_proj_covar=tensor([0.0671, 0.0815, 0.0671, 0.0618, 0.0514, 0.0526, 0.0685, 0.0636], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:41:34,906 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.389e+02 2.103e+02 2.400e+02 2.968e+02 1.328e+03, threshold=4.800e+02, percent-clipped=3.0 2023-05-01 13:41:46,573 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=223720.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:41:52,283 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-05-01 13:41:54,675 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.45 vs. limit=5.0 2023-05-01 13:42:32,054 INFO [train.py:904] (4/8) Epoch 23, batch 450, loss[loss=0.1437, simple_loss=0.2414, pruned_loss=0.02296, over 17202.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2548, pruned_loss=0.04137, over 2973424.43 frames. ], batch size: 46, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:42:52,142 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=223768.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:43:03,034 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5110, 5.4888, 5.3273, 4.7814, 4.9443, 5.3832, 5.2579, 5.0165], device='cuda:4'), covar=tensor([0.0609, 0.0401, 0.0358, 0.0372, 0.1200, 0.0517, 0.0265, 0.0786], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0427, 0.0341, 0.0336, 0.0347, 0.0394, 0.0232, 0.0407], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:43:40,975 INFO [train.py:904] (4/8) Epoch 23, batch 500, loss[loss=0.1647, simple_loss=0.2474, pruned_loss=0.04105, over 16754.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2524, pruned_loss=0.04042, over 3049715.56 frames. ], batch size: 57, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:43:52,984 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.410e+02 2.054e+02 2.358e+02 2.805e+02 6.007e+02, threshold=4.715e+02, percent-clipped=4.0 2023-05-01 13:43:56,033 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6574, 4.9424, 5.2592, 5.2614, 5.3053, 4.9549, 4.6795, 4.7899], device='cuda:4'), covar=tensor([0.0706, 0.0845, 0.0729, 0.0826, 0.0705, 0.0769, 0.1574, 0.0574], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0459, 0.0446, 0.0413, 0.0489, 0.0468, 0.0550, 0.0374], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 13:44:16,862 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=223829.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 13:44:25,962 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3915, 2.3555, 2.4649, 4.1007, 2.3673, 2.7531, 2.4585, 2.5772], device='cuda:4'), covar=tensor([0.1316, 0.3620, 0.3070, 0.0627, 0.4109, 0.2514, 0.3826, 0.3384], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0451, 0.0370, 0.0328, 0.0439, 0.0517, 0.0422, 0.0526], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:44:49,713 INFO [train.py:904] (4/8) Epoch 23, batch 550, loss[loss=0.1747, simple_loss=0.2629, pruned_loss=0.04331, over 16541.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2521, pruned_loss=0.03992, over 3111601.33 frames. ], batch size: 68, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:45:13,040 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.62 vs. limit=5.0 2023-05-01 13:45:16,350 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4193, 5.3748, 5.2448, 4.6982, 4.8902, 5.3092, 5.1846, 4.9137], device='cuda:4'), covar=tensor([0.0541, 0.0519, 0.0331, 0.0337, 0.1086, 0.0509, 0.0283, 0.0763], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0431, 0.0344, 0.0340, 0.0350, 0.0397, 0.0235, 0.0410], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:45:24,265 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=223877.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 13:45:50,296 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9195, 2.0549, 2.3401, 2.6794, 1.9988, 3.1341, 1.7918, 2.6243], device='cuda:4'), covar=tensor([0.1446, 0.0931, 0.1166, 0.0238, 0.0158, 0.0389, 0.1778, 0.0869], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0174, 0.0193, 0.0188, 0.0201, 0.0215, 0.0203, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 13:45:59,107 INFO [train.py:904] (4/8) Epoch 23, batch 600, loss[loss=0.151, simple_loss=0.2453, pruned_loss=0.02834, over 17141.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2523, pruned_loss=0.04017, over 3154046.55 frames. ], batch size: 47, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:46:10,991 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.619e+02 2.055e+02 2.460e+02 3.168e+02 5.428e+02, threshold=4.920e+02, percent-clipped=4.0 2023-05-01 13:46:20,890 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8806, 3.9455, 2.8137, 2.4233, 2.4898, 2.3860, 3.9716, 3.2585], device='cuda:4'), covar=tensor([0.2799, 0.0779, 0.2003, 0.2807, 0.2820, 0.2348, 0.0649, 0.1752], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0267, 0.0304, 0.0313, 0.0294, 0.0260, 0.0295, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 13:46:41,948 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=223934.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:47:08,244 INFO [train.py:904] (4/8) Epoch 23, batch 650, loss[loss=0.1815, simple_loss=0.2584, pruned_loss=0.05232, over 16290.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.25, pruned_loss=0.03933, over 3181856.22 frames. ], batch size: 145, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:48:22,017 INFO [train.py:904] (4/8) Epoch 23, batch 700, loss[loss=0.1927, simple_loss=0.273, pruned_loss=0.05617, over 16282.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2497, pruned_loss=0.03908, over 3210416.45 frames. ], batch size: 165, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:48:35,477 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.336e+02 2.741e+02 3.391e+02 8.377e+02, threshold=5.482e+02, percent-clipped=4.0 2023-05-01 13:48:38,634 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-05-01 13:48:40,798 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-05-01 13:49:24,341 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6786, 6.1005, 5.8048, 5.8485, 5.3916, 5.4662, 5.4650, 6.1890], device='cuda:4'), covar=tensor([0.1434, 0.0916, 0.1123, 0.0856, 0.1009, 0.0636, 0.1326, 0.0995], device='cuda:4'), in_proj_covar=tensor([0.0683, 0.0830, 0.0683, 0.0630, 0.0523, 0.0535, 0.0699, 0.0649], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:49:33,339 INFO [train.py:904] (4/8) Epoch 23, batch 750, loss[loss=0.1578, simple_loss=0.2523, pruned_loss=0.03167, over 17055.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2499, pruned_loss=0.03924, over 3241141.86 frames. ], batch size: 53, lr: 2.96e-03, grad_scale: 2.0 2023-05-01 13:50:41,984 INFO [train.py:904] (4/8) Epoch 23, batch 800, loss[loss=0.1492, simple_loss=0.2366, pruned_loss=0.03094, over 16760.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2497, pruned_loss=0.03979, over 3262203.62 frames. ], batch size: 39, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:50:54,817 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.342e+02 2.213e+02 2.619e+02 3.114e+02 4.585e+02, threshold=5.238e+02, percent-clipped=0.0 2023-05-01 13:51:51,763 INFO [train.py:904] (4/8) Epoch 23, batch 850, loss[loss=0.1506, simple_loss=0.2322, pruned_loss=0.03452, over 16184.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2498, pruned_loss=0.03945, over 3262514.13 frames. ], batch size: 165, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:52:36,196 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9819, 3.0852, 2.7337, 3.0070, 3.3367, 3.1125, 3.5715, 3.5229], device='cuda:4'), covar=tensor([0.0143, 0.0386, 0.0497, 0.0404, 0.0294, 0.0380, 0.0316, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0218, 0.0241, 0.0232, 0.0232, 0.0242, 0.0241, 0.0241, 0.0235], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 13:53:00,735 INFO [train.py:904] (4/8) Epoch 23, batch 900, loss[loss=0.152, simple_loss=0.2424, pruned_loss=0.03087, over 17222.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2492, pruned_loss=0.03885, over 3281338.33 frames. ], batch size: 44, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:53:14,895 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.103e+02 2.438e+02 2.948e+02 5.610e+02, threshold=4.876e+02, percent-clipped=1.0 2023-05-01 13:53:15,253 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=224212.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:53:25,964 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-01 13:53:46,030 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=224234.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:54:10,198 INFO [train.py:904] (4/8) Epoch 23, batch 950, loss[loss=0.149, simple_loss=0.2413, pruned_loss=0.02837, over 17113.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2487, pruned_loss=0.0389, over 3288140.49 frames. ], batch size: 47, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:54:38,544 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=224273.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:54:50,502 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=224282.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 13:55:20,379 INFO [train.py:904] (4/8) Epoch 23, batch 1000, loss[loss=0.1377, simple_loss=0.2237, pruned_loss=0.02585, over 16724.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2479, pruned_loss=0.03893, over 3289971.17 frames. ], batch size: 39, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:55:33,046 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.55 vs. limit=5.0 2023-05-01 13:55:33,530 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.322e+02 2.242e+02 2.591e+02 2.992e+02 5.709e+02, threshold=5.183e+02, percent-clipped=2.0 2023-05-01 13:56:17,326 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-05-01 13:56:31,385 INFO [train.py:904] (4/8) Epoch 23, batch 1050, loss[loss=0.1623, simple_loss=0.2433, pruned_loss=0.04064, over 16827.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2474, pruned_loss=0.03907, over 3292129.39 frames. ], batch size: 96, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:57:42,155 INFO [train.py:904] (4/8) Epoch 23, batch 1100, loss[loss=0.1711, simple_loss=0.259, pruned_loss=0.0416, over 16529.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2473, pruned_loss=0.03888, over 3292497.54 frames. ], batch size: 68, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:57:54,066 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.340e+02 2.008e+02 2.334e+02 2.683e+02 4.688e+02, threshold=4.667e+02, percent-clipped=0.0 2023-05-01 13:58:51,565 INFO [train.py:904] (4/8) Epoch 23, batch 1150, loss[loss=0.1638, simple_loss=0.2615, pruned_loss=0.03307, over 17209.00 frames. ], tot_loss[loss=0.162, simple_loss=0.2471, pruned_loss=0.03843, over 3294235.81 frames. ], batch size: 45, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 13:59:32,708 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.02 vs. limit=5.0 2023-05-01 14:00:00,970 INFO [train.py:904] (4/8) Epoch 23, batch 1200, loss[loss=0.155, simple_loss=0.2317, pruned_loss=0.03914, over 16704.00 frames. ], tot_loss[loss=0.1614, simple_loss=0.2465, pruned_loss=0.03811, over 3298386.88 frames. ], batch size: 89, lr: 2.96e-03, grad_scale: 8.0 2023-05-01 14:00:14,522 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.524e+02 2.279e+02 2.643e+02 3.361e+02 1.197e+03, threshold=5.285e+02, percent-clipped=8.0 2023-05-01 14:00:53,765 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0107, 5.5296, 5.6093, 5.3403, 5.4567, 6.0245, 5.5031, 5.2744], device='cuda:4'), covar=tensor([0.1046, 0.1942, 0.2490, 0.2107, 0.2790, 0.0993, 0.1616, 0.2550], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0603, 0.0670, 0.0496, 0.0669, 0.0694, 0.0523, 0.0664], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 14:01:10,499 INFO [train.py:904] (4/8) Epoch 23, batch 1250, loss[loss=0.1552, simple_loss=0.2533, pruned_loss=0.02857, over 17127.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2468, pruned_loss=0.03902, over 3303736.49 frames. ], batch size: 49, lr: 2.96e-03, grad_scale: 8.0 2023-05-01 14:01:22,910 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=224562.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:01:31,866 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=224568.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:01:52,195 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-01 14:02:07,456 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-01 14:02:20,416 INFO [train.py:904] (4/8) Epoch 23, batch 1300, loss[loss=0.1423, simple_loss=0.2312, pruned_loss=0.0267, over 16839.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.247, pruned_loss=0.03856, over 3320055.04 frames. ], batch size: 42, lr: 2.96e-03, grad_scale: 8.0 2023-05-01 14:02:33,670 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.337e+02 2.195e+02 2.523e+02 3.004e+02 8.579e+02, threshold=5.045e+02, percent-clipped=3.0 2023-05-01 14:02:49,079 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=224623.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:02:53,175 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0257, 3.9612, 4.4219, 2.3156, 4.6698, 4.6939, 3.4544, 3.6909], device='cuda:4'), covar=tensor([0.0708, 0.0210, 0.0200, 0.1097, 0.0067, 0.0159, 0.0380, 0.0361], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0098, 0.0140, 0.0081, 0.0126, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 14:03:29,364 INFO [train.py:904] (4/8) Epoch 23, batch 1350, loss[loss=0.1743, simple_loss=0.2652, pruned_loss=0.0417, over 17054.00 frames. ], tot_loss[loss=0.1618, simple_loss=0.2474, pruned_loss=0.0381, over 3328484.44 frames. ], batch size: 53, lr: 2.96e-03, grad_scale: 8.0 2023-05-01 14:03:48,973 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-01 14:04:40,083 INFO [train.py:904] (4/8) Epoch 23, batch 1400, loss[loss=0.1675, simple_loss=0.2583, pruned_loss=0.03836, over 17119.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2481, pruned_loss=0.03841, over 3331163.49 frames. ], batch size: 48, lr: 2.96e-03, grad_scale: 8.0 2023-05-01 14:04:52,698 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.431e+02 2.135e+02 2.533e+02 2.912e+02 4.513e+02, threshold=5.066e+02, percent-clipped=0.0 2023-05-01 14:05:13,965 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=224726.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:05:33,059 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=224740.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:05:50,555 INFO [train.py:904] (4/8) Epoch 23, batch 1450, loss[loss=0.1702, simple_loss=0.2459, pruned_loss=0.04727, over 16832.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2472, pruned_loss=0.0386, over 3333630.28 frames. ], batch size: 90, lr: 2.96e-03, grad_scale: 8.0 2023-05-01 14:06:39,725 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1324, 2.1888, 2.6910, 3.0719, 3.0137, 3.5756, 2.4461, 3.5430], device='cuda:4'), covar=tensor([0.0295, 0.0512, 0.0377, 0.0351, 0.0331, 0.0199, 0.0501, 0.0181], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0193, 0.0181, 0.0186, 0.0198, 0.0155, 0.0197, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:06:39,734 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=224787.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:06:59,994 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=224801.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:07:01,744 INFO [train.py:904] (4/8) Epoch 23, batch 1500, loss[loss=0.1931, simple_loss=0.2602, pruned_loss=0.06296, over 16794.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2474, pruned_loss=0.03923, over 3330183.56 frames. ], batch size: 102, lr: 2.96e-03, grad_scale: 8.0 2023-05-01 14:07:02,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7979, 3.7812, 2.2734, 4.1477, 2.9764, 4.0949, 2.4623, 3.0874], device='cuda:4'), covar=tensor([0.0270, 0.0434, 0.1535, 0.0344, 0.0736, 0.0631, 0.1353, 0.0646], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0181, 0.0197, 0.0170, 0.0179, 0.0221, 0.0207, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 14:07:12,852 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3016, 5.2700, 5.0705, 4.4990, 5.0860, 1.9494, 4.8411, 4.9853], device='cuda:4'), covar=tensor([0.0097, 0.0088, 0.0205, 0.0463, 0.0129, 0.2757, 0.0161, 0.0228], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0163, 0.0205, 0.0181, 0.0183, 0.0215, 0.0195, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:07:16,528 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.459e+02 2.087e+02 2.500e+02 3.086e+02 8.479e+02, threshold=5.000e+02, percent-clipped=3.0 2023-05-01 14:07:38,558 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5373, 1.8451, 2.1957, 2.3737, 2.5657, 2.5644, 1.9485, 2.6391], device='cuda:4'), covar=tensor([0.0216, 0.0468, 0.0382, 0.0323, 0.0309, 0.0278, 0.0507, 0.0186], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0193, 0.0181, 0.0185, 0.0198, 0.0155, 0.0197, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:08:14,249 INFO [train.py:904] (4/8) Epoch 23, batch 1550, loss[loss=0.1746, simple_loss=0.2578, pruned_loss=0.04567, over 15274.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2484, pruned_loss=0.0402, over 3319930.92 frames. ], batch size: 190, lr: 2.96e-03, grad_scale: 4.0 2023-05-01 14:08:34,627 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=224868.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:08:53,917 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8199, 3.7561, 3.8915, 3.9807, 4.0760, 3.6729, 3.9353, 4.1182], device='cuda:4'), covar=tensor([0.1588, 0.1124, 0.1231, 0.0698, 0.0587, 0.1829, 0.1751, 0.0704], device='cuda:4'), in_proj_covar=tensor([0.0673, 0.0830, 0.0962, 0.0839, 0.0633, 0.0664, 0.0690, 0.0801], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:09:23,918 INFO [train.py:904] (4/8) Epoch 23, batch 1600, loss[loss=0.1409, simple_loss=0.2247, pruned_loss=0.02857, over 16796.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2496, pruned_loss=0.04079, over 3322270.78 frames. ], batch size: 39, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:09:36,820 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.525e+02 2.147e+02 2.697e+02 3.461e+02 9.919e+02, threshold=5.394e+02, percent-clipped=4.0 2023-05-01 14:09:41,271 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=224916.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:09:43,665 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=224918.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:09:54,864 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 14:10:32,924 INFO [train.py:904] (4/8) Epoch 23, batch 1650, loss[loss=0.1762, simple_loss=0.2721, pruned_loss=0.04017, over 17097.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2515, pruned_loss=0.04064, over 3327387.36 frames. ], batch size: 55, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:11:41,679 INFO [train.py:904] (4/8) Epoch 23, batch 1700, loss[loss=0.1809, simple_loss=0.2604, pruned_loss=0.05068, over 16745.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2532, pruned_loss=0.04143, over 3324596.71 frames. ], batch size: 83, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:11:56,162 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.378e+02 2.319e+02 2.651e+02 3.431e+02 9.066e+02, threshold=5.301e+02, percent-clipped=1.0 2023-05-01 14:12:13,269 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3626, 2.1932, 1.7738, 2.0523, 2.4878, 2.2854, 2.3502, 2.5814], device='cuda:4'), covar=tensor([0.0278, 0.0404, 0.0564, 0.0502, 0.0250, 0.0332, 0.0258, 0.0287], device='cuda:4'), in_proj_covar=tensor([0.0220, 0.0242, 0.0232, 0.0232, 0.0243, 0.0242, 0.0244, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:12:52,528 INFO [train.py:904] (4/8) Epoch 23, batch 1750, loss[loss=0.1754, simple_loss=0.2716, pruned_loss=0.03957, over 16731.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2541, pruned_loss=0.04164, over 3320045.36 frames. ], batch size: 57, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:13:33,035 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225082.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:13:34,923 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4315, 4.3822, 4.3419, 4.0203, 4.0883, 4.3968, 4.1116, 4.1583], device='cuda:4'), covar=tensor([0.0610, 0.0693, 0.0339, 0.0320, 0.0806, 0.0516, 0.0679, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0448, 0.0357, 0.0354, 0.0362, 0.0414, 0.0242, 0.0426], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:13:52,603 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225096.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:14:01,699 INFO [train.py:904] (4/8) Epoch 23, batch 1800, loss[loss=0.1862, simple_loss=0.2739, pruned_loss=0.04926, over 15460.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2558, pruned_loss=0.04148, over 3325803.83 frames. ], batch size: 191, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:14:02,791 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225103.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:14:15,808 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.405e+02 2.184e+02 2.489e+02 3.045e+02 6.303e+02, threshold=4.978e+02, percent-clipped=2.0 2023-05-01 14:14:35,763 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2131, 4.0796, 4.3034, 4.4201, 4.5400, 4.1127, 4.3445, 4.5444], device='cuda:4'), covar=tensor([0.1736, 0.1298, 0.1477, 0.0780, 0.0650, 0.1509, 0.2919, 0.0833], device='cuda:4'), in_proj_covar=tensor([0.0674, 0.0832, 0.0964, 0.0842, 0.0636, 0.0668, 0.0692, 0.0803], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:14:57,253 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225142.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:15:12,129 INFO [train.py:904] (4/8) Epoch 23, batch 1850, loss[loss=0.1856, simple_loss=0.262, pruned_loss=0.05457, over 16265.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2569, pruned_loss=0.04208, over 3330659.11 frames. ], batch size: 165, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:15:27,772 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225164.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:16:22,251 INFO [train.py:904] (4/8) Epoch 23, batch 1900, loss[loss=0.1553, simple_loss=0.2564, pruned_loss=0.02711, over 17110.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2563, pruned_loss=0.04137, over 3325451.33 frames. ], batch size: 49, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:16:22,769 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225203.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:16:36,681 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.361e+02 2.136e+02 2.472e+02 2.864e+02 6.129e+02, threshold=4.945e+02, percent-clipped=2.0 2023-05-01 14:16:43,956 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=225218.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:17:11,955 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-01 14:17:31,819 INFO [train.py:904] (4/8) Epoch 23, batch 1950, loss[loss=0.1916, simple_loss=0.2691, pruned_loss=0.05706, over 16441.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2558, pruned_loss=0.04069, over 3327369.58 frames. ], batch size: 146, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:17:49,190 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9549, 2.0437, 2.2673, 3.4018, 2.0826, 2.3118, 2.1763, 2.1914], device='cuda:4'), covar=tensor([0.1536, 0.3905, 0.3121, 0.0777, 0.4215, 0.2802, 0.3961, 0.3454], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0456, 0.0374, 0.0332, 0.0441, 0.0526, 0.0428, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:17:50,056 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=225266.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:18:24,641 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2313, 4.9917, 5.2616, 5.4202, 5.6521, 4.8522, 5.5570, 5.5915], device='cuda:4'), covar=tensor([0.1957, 0.1360, 0.1898, 0.0853, 0.0533, 0.0948, 0.0617, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0677, 0.0838, 0.0971, 0.0844, 0.0639, 0.0672, 0.0694, 0.0805], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:18:37,329 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8010, 2.6093, 2.2077, 2.6049, 2.9792, 2.7599, 3.3614, 3.2160], device='cuda:4'), covar=tensor([0.0163, 0.0506, 0.0610, 0.0477, 0.0360, 0.0448, 0.0270, 0.0313], device='cuda:4'), in_proj_covar=tensor([0.0221, 0.0242, 0.0232, 0.0232, 0.0243, 0.0242, 0.0243, 0.0238], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:18:42,383 INFO [train.py:904] (4/8) Epoch 23, batch 2000, loss[loss=0.1512, simple_loss=0.2397, pruned_loss=0.03136, over 17205.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2552, pruned_loss=0.04029, over 3331211.16 frames. ], batch size: 46, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:18:56,051 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.456e+02 2.089e+02 2.408e+02 2.851e+02 8.100e+02, threshold=4.816e+02, percent-clipped=1.0 2023-05-01 14:19:52,854 INFO [train.py:904] (4/8) Epoch 23, batch 2050, loss[loss=0.161, simple_loss=0.2589, pruned_loss=0.03154, over 17191.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2542, pruned_loss=0.03995, over 3331195.34 frames. ], batch size: 46, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:20:02,796 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-01 14:20:35,691 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=225382.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:20:54,897 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=225396.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:21:04,272 INFO [train.py:904] (4/8) Epoch 23, batch 2100, loss[loss=0.1823, simple_loss=0.2657, pruned_loss=0.04944, over 16354.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.255, pruned_loss=0.04101, over 3328776.28 frames. ], batch size: 146, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:21:09,858 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0792, 2.0659, 2.6767, 3.0043, 3.0215, 3.5209, 2.0631, 3.5575], device='cuda:4'), covar=tensor([0.0254, 0.0638, 0.0360, 0.0348, 0.0318, 0.0219, 0.0711, 0.0167], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0195, 0.0183, 0.0188, 0.0200, 0.0158, 0.0199, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:21:18,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.459e+02 2.181e+02 2.818e+02 3.430e+02 7.396e+02, threshold=5.635e+02, percent-clipped=7.0 2023-05-01 14:21:29,420 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 14:21:44,080 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=225430.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:21:57,160 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225439.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:22:03,499 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=225444.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:22:15,762 INFO [train.py:904] (4/8) Epoch 23, batch 2150, loss[loss=0.1728, simple_loss=0.2671, pruned_loss=0.0392, over 17077.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2551, pruned_loss=0.04096, over 3336821.13 frames. ], batch size: 53, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:22:25,279 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225459.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:22:25,346 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1899, 5.1634, 5.0636, 4.5566, 4.7015, 5.0870, 5.0502, 4.7412], device='cuda:4'), covar=tensor([0.0639, 0.0614, 0.0338, 0.0401, 0.1102, 0.0540, 0.0331, 0.0793], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0456, 0.0363, 0.0361, 0.0367, 0.0421, 0.0246, 0.0433], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 14:22:29,353 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225462.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:22:57,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3132, 3.6093, 3.9304, 2.1861, 3.1298, 2.5123, 3.8681, 3.7888], device='cuda:4'), covar=tensor([0.0314, 0.0844, 0.0518, 0.2084, 0.0828, 0.1003, 0.0588, 0.1036], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0166, 0.0168, 0.0154, 0.0146, 0.0130, 0.0144, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 14:23:19,350 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225498.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:23:22,046 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225500.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:23:26,035 INFO [train.py:904] (4/8) Epoch 23, batch 2200, loss[loss=0.1621, simple_loss=0.2416, pruned_loss=0.04133, over 16648.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2573, pruned_loss=0.04214, over 3324207.21 frames. ], batch size: 89, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:23:29,387 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225505.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:23:40,047 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.652e+02 2.156e+02 2.513e+02 2.761e+02 5.563e+02, threshold=5.025e+02, percent-clipped=0.0 2023-05-01 14:23:53,549 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225522.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:23:55,302 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225523.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:24:36,376 INFO [train.py:904] (4/8) Epoch 23, batch 2250, loss[loss=0.2257, simple_loss=0.3007, pruned_loss=0.07532, over 11634.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2582, pruned_loss=0.04238, over 3321439.69 frames. ], batch size: 246, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:24:39,310 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1534, 2.2563, 2.7336, 3.0290, 2.9760, 3.5895, 2.6203, 3.5389], device='cuda:4'), covar=tensor([0.0253, 0.0516, 0.0337, 0.0342, 0.0332, 0.0176, 0.0441, 0.0172], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0196, 0.0183, 0.0188, 0.0200, 0.0157, 0.0199, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:24:45,101 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225559.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 14:24:54,964 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225566.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:25:18,800 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225583.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:25:46,976 INFO [train.py:904] (4/8) Epoch 23, batch 2300, loss[loss=0.1722, simple_loss=0.252, pruned_loss=0.04618, over 16757.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2582, pruned_loss=0.04228, over 3320924.64 frames. ], batch size: 124, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:25:56,970 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-05-01 14:26:01,555 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.669e+02 2.061e+02 2.466e+02 3.026e+02 4.196e+02, threshold=4.932e+02, percent-clipped=0.0 2023-05-01 14:26:11,623 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225620.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 14:26:58,883 INFO [train.py:904] (4/8) Epoch 23, batch 2350, loss[loss=0.1781, simple_loss=0.2585, pruned_loss=0.04885, over 16703.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2584, pruned_loss=0.04251, over 3308406.82 frames. ], batch size: 134, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:27:34,762 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9616, 2.0019, 2.1550, 3.4992, 2.0693, 2.2593, 2.1460, 2.1586], device='cuda:4'), covar=tensor([0.1651, 0.3986, 0.3155, 0.0793, 0.4297, 0.2865, 0.3935, 0.3593], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0459, 0.0376, 0.0335, 0.0443, 0.0530, 0.0431, 0.0537], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:28:10,320 INFO [train.py:904] (4/8) Epoch 23, batch 2400, loss[loss=0.1672, simple_loss=0.2528, pruned_loss=0.04077, over 16449.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2591, pruned_loss=0.04268, over 3307879.98 frames. ], batch size: 68, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:28:23,216 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.701e+02 2.227e+02 2.621e+02 3.138e+02 6.803e+02, threshold=5.242e+02, percent-clipped=3.0 2023-05-01 14:29:17,848 INFO [train.py:904] (4/8) Epoch 23, batch 2450, loss[loss=0.1658, simple_loss=0.2471, pruned_loss=0.04219, over 16764.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2593, pruned_loss=0.0422, over 3315695.37 frames. ], batch size: 89, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:29:26,302 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=225759.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:30:10,504 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.22 vs. limit=5.0 2023-05-01 14:30:17,193 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225795.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:30:22,200 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=225798.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:30:29,832 INFO [train.py:904] (4/8) Epoch 23, batch 2500, loss[loss=0.2199, simple_loss=0.2959, pruned_loss=0.07195, over 12125.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2593, pruned_loss=0.04233, over 3312450.04 frames. ], batch size: 246, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:30:36,108 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=225807.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:30:43,415 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.456e+02 2.105e+02 2.544e+02 3.075e+02 7.140e+02, threshold=5.088e+02, percent-clipped=2.0 2023-05-01 14:30:51,423 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225818.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:31:28,974 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=225846.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:31:38,355 INFO [train.py:904] (4/8) Epoch 23, batch 2550, loss[loss=0.1474, simple_loss=0.2314, pruned_loss=0.03171, over 16996.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2587, pruned_loss=0.04193, over 3315392.40 frames. ], batch size: 41, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:31:39,530 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6541, 1.8990, 2.2770, 2.4655, 2.6387, 2.5757, 1.9905, 2.7965], device='cuda:4'), covar=tensor([0.0171, 0.0456, 0.0306, 0.0295, 0.0294, 0.0337, 0.0482, 0.0147], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0195, 0.0183, 0.0188, 0.0201, 0.0158, 0.0198, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:31:49,563 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225861.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:31:50,898 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225862.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:32:14,110 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225878.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:32:38,023 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=225895.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:32:49,282 INFO [train.py:904] (4/8) Epoch 23, batch 2600, loss[loss=0.1413, simple_loss=0.2358, pruned_loss=0.02344, over 17204.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2583, pruned_loss=0.04138, over 3313147.09 frames. ], batch size: 44, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:32:51,054 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-01 14:33:03,045 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.463e+02 2.265e+02 2.588e+02 2.921e+02 5.309e+02, threshold=5.176e+02, percent-clipped=1.0 2023-05-01 14:33:06,084 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=225915.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 14:33:07,229 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0279, 5.0473, 5.4422, 5.4526, 5.4993, 5.1066, 5.0641, 4.8359], device='cuda:4'), covar=tensor([0.0357, 0.0601, 0.0409, 0.0435, 0.0562, 0.0426, 0.1082, 0.0502], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0476, 0.0462, 0.0427, 0.0509, 0.0485, 0.0568, 0.0387], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 14:33:16,601 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225923.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:33:58,792 INFO [train.py:904] (4/8) Epoch 23, batch 2650, loss[loss=0.1489, simple_loss=0.2395, pruned_loss=0.02915, over 17202.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2588, pruned_loss=0.04079, over 3324737.95 frames. ], batch size: 44, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:34:04,262 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=225956.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:34:51,897 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.54 vs. limit=5.0 2023-05-01 14:35:12,270 INFO [train.py:904] (4/8) Epoch 23, batch 2700, loss[loss=0.157, simple_loss=0.243, pruned_loss=0.03548, over 16979.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2597, pruned_loss=0.04074, over 3330702.15 frames. ], batch size: 41, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:35:24,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8109, 4.6152, 4.8693, 5.0195, 5.2002, 4.5970, 5.2001, 5.2229], device='cuda:4'), covar=tensor([0.1956, 0.1389, 0.1711, 0.0792, 0.0578, 0.1096, 0.0716, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0672, 0.0835, 0.0965, 0.0842, 0.0640, 0.0668, 0.0692, 0.0803], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:35:25,728 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.491e+02 2.127e+02 2.428e+02 2.889e+02 7.591e+02, threshold=4.856e+02, percent-clipped=2.0 2023-05-01 14:36:06,268 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=226041.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:36:13,633 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2924, 5.2535, 4.9776, 4.4727, 5.1044, 1.7686, 4.8849, 4.8874], device='cuda:4'), covar=tensor([0.0086, 0.0089, 0.0213, 0.0395, 0.0104, 0.3111, 0.0131, 0.0224], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0165, 0.0207, 0.0184, 0.0185, 0.0216, 0.0197, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:36:23,069 INFO [train.py:904] (4/8) Epoch 23, batch 2750, loss[loss=0.1737, simple_loss=0.2695, pruned_loss=0.03898, over 16669.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2596, pruned_loss=0.04066, over 3333811.26 frames. ], batch size: 57, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:37:21,895 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226095.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:37:25,087 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8033, 2.6605, 2.2708, 2.4207, 2.9488, 2.7536, 3.4521, 3.2886], device='cuda:4'), covar=tensor([0.0175, 0.0504, 0.0654, 0.0599, 0.0360, 0.0460, 0.0272, 0.0302], device='cuda:4'), in_proj_covar=tensor([0.0223, 0.0243, 0.0231, 0.0233, 0.0244, 0.0242, 0.0244, 0.0239], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:37:31,980 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=226102.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:37:32,648 INFO [train.py:904] (4/8) Epoch 23, batch 2800, loss[loss=0.1696, simple_loss=0.2712, pruned_loss=0.034, over 16730.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2589, pruned_loss=0.04007, over 3333196.56 frames. ], batch size: 57, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:37:47,345 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 2.068e+02 2.362e+02 2.950e+02 8.233e+02, threshold=4.725e+02, percent-clipped=4.0 2023-05-01 14:37:54,722 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226118.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:37:59,207 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9878, 1.9362, 2.5345, 2.8664, 2.8128, 3.4216, 2.1549, 3.4044], device='cuda:4'), covar=tensor([0.0288, 0.0646, 0.0405, 0.0374, 0.0408, 0.0224, 0.0662, 0.0226], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0197, 0.0184, 0.0190, 0.0202, 0.0160, 0.0201, 0.0156], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:38:29,329 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226143.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:38:42,389 INFO [train.py:904] (4/8) Epoch 23, batch 2850, loss[loss=0.1727, simple_loss=0.2632, pruned_loss=0.04113, over 17163.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2579, pruned_loss=0.03966, over 3340529.80 frames. ], batch size: 46, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:38:54,372 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226161.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:39:00,759 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226166.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:39:18,075 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226178.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:39:51,837 INFO [train.py:904] (4/8) Epoch 23, batch 2900, loss[loss=0.1852, simple_loss=0.2537, pruned_loss=0.05838, over 16302.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2565, pruned_loss=0.03987, over 3337072.70 frames. ], batch size: 165, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:39:59,559 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-01 14:40:00,216 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226209.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:40:05,815 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.378e+02 2.036e+02 2.506e+02 2.935e+02 4.421e+02, threshold=5.013e+02, percent-clipped=0.0 2023-05-01 14:40:09,300 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226215.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 14:40:13,513 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=226218.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:40:24,106 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226226.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:40:34,174 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=226233.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:40:58,437 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=226251.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:41:00,393 INFO [train.py:904] (4/8) Epoch 23, batch 2950, loss[loss=0.1784, simple_loss=0.2582, pruned_loss=0.04929, over 16422.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2559, pruned_loss=0.04036, over 3342046.65 frames. ], batch size: 146, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:41:14,089 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226263.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 14:41:52,922 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.16 vs. limit=5.0 2023-05-01 14:41:56,390 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=226294.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:42:08,778 INFO [train.py:904] (4/8) Epoch 23, batch 3000, loss[loss=0.1594, simple_loss=0.2587, pruned_loss=0.02998, over 17163.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2558, pruned_loss=0.04116, over 3330381.48 frames. ], batch size: 49, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:42:08,778 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 14:42:17,866 INFO [train.py:938] (4/8) Epoch 23, validation: loss=0.1344, simple_loss=0.2397, pruned_loss=0.01456, over 944034.00 frames. 2023-05-01 14:42:17,867 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 14:42:30,994 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.698e+02 2.313e+02 2.743e+02 3.282e+02 6.136e+02, threshold=5.486e+02, percent-clipped=4.0 2023-05-01 14:42:46,472 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5665, 3.3603, 3.7374, 1.8962, 3.7789, 3.8161, 3.0794, 2.8000], device='cuda:4'), covar=tensor([0.0774, 0.0246, 0.0202, 0.1179, 0.0115, 0.0217, 0.0412, 0.0478], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0100, 0.0140, 0.0083, 0.0128, 0.0129, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 14:43:13,621 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-05-01 14:43:26,883 INFO [train.py:904] (4/8) Epoch 23, batch 3050, loss[loss=0.1663, simple_loss=0.2502, pruned_loss=0.04121, over 16778.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2568, pruned_loss=0.04206, over 3318428.92 frames. ], batch size: 102, lr: 2.95e-03, grad_scale: 8.0 2023-05-01 14:44:29,374 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=226397.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:44:37,979 INFO [train.py:904] (4/8) Epoch 23, batch 3100, loss[loss=0.1618, simple_loss=0.2385, pruned_loss=0.0425, over 16739.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2558, pruned_loss=0.04198, over 3323277.65 frames. ], batch size: 89, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:44:43,824 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3977, 4.2337, 4.4792, 4.6062, 4.7311, 4.3138, 4.5579, 4.7164], device='cuda:4'), covar=tensor([0.1913, 0.1383, 0.1486, 0.0788, 0.0707, 0.1201, 0.2496, 0.0981], device='cuda:4'), in_proj_covar=tensor([0.0679, 0.0845, 0.0975, 0.0850, 0.0646, 0.0675, 0.0699, 0.0812], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:44:47,319 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2891, 5.8010, 5.9683, 5.6667, 5.7654, 6.2448, 5.8621, 5.5877], device='cuda:4'), covar=tensor([0.0875, 0.1850, 0.2841, 0.1885, 0.2603, 0.1004, 0.1371, 0.2190], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0616, 0.0681, 0.0506, 0.0681, 0.0703, 0.0531, 0.0678], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 14:44:51,600 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 2.140e+02 2.727e+02 3.303e+02 6.006e+02, threshold=5.454e+02, percent-clipped=2.0 2023-05-01 14:45:26,236 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-05-01 14:45:47,148 INFO [train.py:904] (4/8) Epoch 23, batch 3150, loss[loss=0.1512, simple_loss=0.2371, pruned_loss=0.0327, over 16998.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2553, pruned_loss=0.04207, over 3320106.54 frames. ], batch size: 41, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:46:11,138 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-05-01 14:46:36,309 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6872, 4.6772, 4.6217, 4.2870, 4.3401, 4.6772, 4.4081, 4.4431], device='cuda:4'), covar=tensor([0.0649, 0.0674, 0.0279, 0.0318, 0.0834, 0.0450, 0.0529, 0.0646], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0462, 0.0368, 0.0366, 0.0373, 0.0427, 0.0249, 0.0440], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 14:46:54,851 INFO [train.py:904] (4/8) Epoch 23, batch 3200, loss[loss=0.1543, simple_loss=0.2601, pruned_loss=0.02423, over 17209.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2547, pruned_loss=0.04146, over 3319827.31 frames. ], batch size: 46, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:47:09,850 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.604e+02 2.052e+02 2.422e+02 2.820e+02 4.460e+02, threshold=4.844e+02, percent-clipped=0.0 2023-05-01 14:47:15,793 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226518.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:47:45,886 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4681, 2.2727, 2.3280, 4.2332, 2.2465, 2.6723, 2.3945, 2.4917], device='cuda:4'), covar=tensor([0.1252, 0.3765, 0.3222, 0.0541, 0.4283, 0.2663, 0.3612, 0.3755], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0457, 0.0375, 0.0336, 0.0441, 0.0527, 0.0428, 0.0535], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:48:01,538 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226551.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:48:04,026 INFO [train.py:904] (4/8) Epoch 23, batch 3250, loss[loss=0.1612, simple_loss=0.2471, pruned_loss=0.03762, over 17247.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.255, pruned_loss=0.04189, over 3303632.11 frames. ], batch size: 44, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:48:22,283 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226566.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:48:54,455 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=226589.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:49:08,847 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226599.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:49:14,395 INFO [train.py:904] (4/8) Epoch 23, batch 3300, loss[loss=0.1603, simple_loss=0.2445, pruned_loss=0.03806, over 16917.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2565, pruned_loss=0.04184, over 3308033.70 frames. ], batch size: 109, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:49:27,925 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.422e+02 2.105e+02 2.619e+02 3.003e+02 4.974e+02, threshold=5.238e+02, percent-clipped=1.0 2023-05-01 14:50:23,128 INFO [train.py:904] (4/8) Epoch 23, batch 3350, loss[loss=0.2052, simple_loss=0.2893, pruned_loss=0.06058, over 15637.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2573, pruned_loss=0.042, over 3300585.58 frames. ], batch size: 191, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:50:24,784 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8109, 4.6069, 4.8436, 5.0285, 5.2350, 4.6329, 5.2211, 5.2271], device='cuda:4'), covar=tensor([0.2028, 0.1331, 0.1861, 0.0807, 0.0586, 0.1029, 0.0636, 0.0658], device='cuda:4'), in_proj_covar=tensor([0.0684, 0.0850, 0.0983, 0.0857, 0.0649, 0.0681, 0.0704, 0.0819], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 14:51:25,714 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226697.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:51:34,603 INFO [train.py:904] (4/8) Epoch 23, batch 3400, loss[loss=0.1674, simple_loss=0.2501, pruned_loss=0.04233, over 16837.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2575, pruned_loss=0.04156, over 3299764.59 frames. ], batch size: 102, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:51:47,766 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.365e+02 2.087e+02 2.446e+02 3.167e+02 4.781e+02, threshold=4.893e+02, percent-clipped=0.0 2023-05-01 14:52:33,238 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226745.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:52:43,440 INFO [train.py:904] (4/8) Epoch 23, batch 3450, loss[loss=0.1602, simple_loss=0.2406, pruned_loss=0.03992, over 16317.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2555, pruned_loss=0.04111, over 3300417.57 frames. ], batch size: 165, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:53:01,628 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7759, 3.8496, 2.4345, 4.3189, 3.0164, 4.2445, 2.6725, 3.1528], device='cuda:4'), covar=tensor([0.0292, 0.0412, 0.1511, 0.0325, 0.0723, 0.0616, 0.1321, 0.0704], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0182, 0.0197, 0.0172, 0.0180, 0.0224, 0.0207, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 14:53:09,944 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8307, 3.9485, 4.1309, 4.1156, 4.1326, 3.9479, 3.9388, 3.9099], device='cuda:4'), covar=tensor([0.0371, 0.0619, 0.0453, 0.0440, 0.0559, 0.0454, 0.0792, 0.0579], device='cuda:4'), in_proj_covar=tensor([0.0435, 0.0486, 0.0471, 0.0436, 0.0516, 0.0493, 0.0581, 0.0394], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 14:53:52,946 INFO [train.py:904] (4/8) Epoch 23, batch 3500, loss[loss=0.1838, simple_loss=0.2869, pruned_loss=0.04038, over 17130.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2546, pruned_loss=0.04083, over 3301061.00 frames. ], batch size: 48, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:54:07,205 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.312e+02 2.101e+02 2.445e+02 3.055e+02 4.723e+02, threshold=4.890e+02, percent-clipped=0.0 2023-05-01 14:55:03,892 INFO [train.py:904] (4/8) Epoch 23, batch 3550, loss[loss=0.1831, simple_loss=0.268, pruned_loss=0.04915, over 16692.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2533, pruned_loss=0.04051, over 3304093.88 frames. ], batch size: 134, lr: 2.94e-03, grad_scale: 16.0 2023-05-01 14:55:52,971 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=226889.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:56:12,760 INFO [train.py:904] (4/8) Epoch 23, batch 3600, loss[loss=0.18, simple_loss=0.2565, pruned_loss=0.05173, over 16776.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2518, pruned_loss=0.03991, over 3301439.96 frames. ], batch size: 124, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:56:28,214 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.070e+02 2.545e+02 2.955e+02 4.911e+02, threshold=5.089e+02, percent-clipped=1.0 2023-05-01 14:56:55,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7736, 2.9331, 3.2635, 2.1014, 2.8088, 2.1623, 3.4144, 3.2484], device='cuda:4'), covar=tensor([0.0237, 0.0997, 0.0554, 0.1946, 0.0887, 0.1084, 0.0540, 0.0924], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0171, 0.0171, 0.0157, 0.0149, 0.0133, 0.0147, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-01 14:57:02,339 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=226937.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:57:24,672 INFO [train.py:904] (4/8) Epoch 23, batch 3650, loss[loss=0.145, simple_loss=0.2226, pruned_loss=0.03366, over 16772.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2505, pruned_loss=0.0402, over 3302185.46 frames. ], batch size: 102, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:58:03,799 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.54 vs. limit=5.0 2023-05-01 14:58:19,609 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=226989.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:58:40,563 INFO [train.py:904] (4/8) Epoch 23, batch 3700, loss[loss=0.1449, simple_loss=0.2245, pruned_loss=0.03261, over 16798.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2492, pruned_loss=0.04139, over 3277007.72 frames. ], batch size: 96, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 14:58:56,577 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.602e+02 2.184e+02 2.522e+02 2.938e+02 6.329e+02, threshold=5.043e+02, percent-clipped=1.0 2023-05-01 14:59:50,735 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=227050.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 14:59:53,687 INFO [train.py:904] (4/8) Epoch 23, batch 3750, loss[loss=0.1875, simple_loss=0.2648, pruned_loss=0.05515, over 16733.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2493, pruned_loss=0.04265, over 3271191.60 frames. ], batch size: 124, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:00:07,861 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-01 15:00:42,445 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7128, 4.8119, 4.9929, 4.8071, 4.8478, 5.4072, 4.9007, 4.6152], device='cuda:4'), covar=tensor([0.1506, 0.1790, 0.2061, 0.2174, 0.2740, 0.1033, 0.1657, 0.2622], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0611, 0.0674, 0.0505, 0.0674, 0.0698, 0.0529, 0.0676], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 15:01:07,855 INFO [train.py:904] (4/8) Epoch 23, batch 3800, loss[loss=0.1704, simple_loss=0.2518, pruned_loss=0.04446, over 15587.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2509, pruned_loss=0.04406, over 3280524.00 frames. ], batch size: 190, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:01:25,259 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.697e+02 2.146e+02 2.630e+02 3.388e+02 6.503e+02, threshold=5.260e+02, percent-clipped=3.0 2023-05-01 15:01:55,375 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7470, 1.8193, 1.5999, 1.5035, 1.9597, 1.6714, 1.7105, 1.9745], device='cuda:4'), covar=tensor([0.0182, 0.0283, 0.0418, 0.0367, 0.0199, 0.0282, 0.0157, 0.0193], device='cuda:4'), in_proj_covar=tensor([0.0223, 0.0244, 0.0231, 0.0233, 0.0242, 0.0242, 0.0245, 0.0240], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:02:21,618 INFO [train.py:904] (4/8) Epoch 23, batch 3850, loss[loss=0.1581, simple_loss=0.2393, pruned_loss=0.03844, over 16887.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2509, pruned_loss=0.04463, over 3278953.80 frames. ], batch size: 96, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:03:34,968 INFO [train.py:904] (4/8) Epoch 23, batch 3900, loss[loss=0.1837, simple_loss=0.2623, pruned_loss=0.05262, over 16776.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2508, pruned_loss=0.04495, over 3274416.16 frames. ], batch size: 134, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:03:51,651 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.638e+02 2.230e+02 2.530e+02 3.049e+02 6.179e+02, threshold=5.060e+02, percent-clipped=1.0 2023-05-01 15:04:38,775 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.65 vs. limit=5.0 2023-05-01 15:04:47,725 INFO [train.py:904] (4/8) Epoch 23, batch 3950, loss[loss=0.1631, simple_loss=0.2421, pruned_loss=0.04204, over 16480.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2503, pruned_loss=0.04527, over 3280999.64 frames. ], batch size: 146, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:05:03,370 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-05-01 15:06:00,626 INFO [train.py:904] (4/8) Epoch 23, batch 4000, loss[loss=0.1589, simple_loss=0.243, pruned_loss=0.03738, over 16773.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2508, pruned_loss=0.04575, over 3286456.50 frames. ], batch size: 102, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:06:17,147 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.633e+02 2.172e+02 2.478e+02 2.977e+02 5.073e+02, threshold=4.957e+02, percent-clipped=1.0 2023-05-01 15:06:22,526 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2407, 3.3016, 3.6351, 2.2045, 3.0629, 2.3817, 3.7298, 3.6382], device='cuda:4'), covar=tensor([0.0208, 0.0846, 0.0560, 0.2012, 0.0860, 0.0973, 0.0458, 0.0834], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0170, 0.0170, 0.0156, 0.0147, 0.0132, 0.0145, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-01 15:07:01,811 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=227345.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:07:13,394 INFO [train.py:904] (4/8) Epoch 23, batch 4050, loss[loss=0.1623, simple_loss=0.2509, pruned_loss=0.03687, over 16691.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2515, pruned_loss=0.04513, over 3291997.62 frames. ], batch size: 62, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:07:30,281 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.19 vs. limit=5.0 2023-05-01 15:08:27,045 INFO [train.py:904] (4/8) Epoch 23, batch 4100, loss[loss=0.186, simple_loss=0.2714, pruned_loss=0.0503, over 16420.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2538, pruned_loss=0.0451, over 3277090.82 frames. ], batch size: 146, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:08:39,167 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=227411.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:08:42,769 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.244e+02 1.752e+02 1.921e+02 2.258e+02 5.560e+02, threshold=3.841e+02, percent-clipped=1.0 2023-05-01 15:08:57,013 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5639, 2.2507, 1.8156, 2.0397, 2.5292, 2.2028, 2.4576, 2.7150], device='cuda:4'), covar=tensor([0.0208, 0.0463, 0.0585, 0.0493, 0.0283, 0.0404, 0.0231, 0.0265], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0242, 0.0231, 0.0232, 0.0241, 0.0241, 0.0244, 0.0239], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:09:45,381 INFO [train.py:904] (4/8) Epoch 23, batch 4150, loss[loss=0.2006, simple_loss=0.2982, pruned_loss=0.05151, over 16231.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2613, pruned_loss=0.04787, over 3231339.47 frames. ], batch size: 165, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:09:54,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3734, 3.0491, 3.4083, 1.8346, 3.5517, 3.5529, 2.8529, 2.7498], device='cuda:4'), covar=tensor([0.0832, 0.0323, 0.0223, 0.1268, 0.0094, 0.0173, 0.0441, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0099, 0.0139, 0.0082, 0.0128, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:10:04,852 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9113, 4.1687, 4.0090, 4.0672, 3.7841, 3.8148, 3.8788, 4.1848], device='cuda:4'), covar=tensor([0.1165, 0.0934, 0.1051, 0.0767, 0.0716, 0.1659, 0.0866, 0.0978], device='cuda:4'), in_proj_covar=tensor([0.0697, 0.0849, 0.0699, 0.0645, 0.0536, 0.0542, 0.0713, 0.0664], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:10:16,255 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=227472.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:11:01,320 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5636, 4.5315, 4.3563, 3.5825, 4.4549, 1.5920, 4.2876, 3.9168], device='cuda:4'), covar=tensor([0.0075, 0.0068, 0.0187, 0.0318, 0.0074, 0.3196, 0.0111, 0.0275], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0165, 0.0206, 0.0184, 0.0184, 0.0213, 0.0195, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:11:03,933 INFO [train.py:904] (4/8) Epoch 23, batch 4200, loss[loss=0.1999, simple_loss=0.2929, pruned_loss=0.05348, over 16660.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2676, pruned_loss=0.04926, over 3206254.77 frames. ], batch size: 134, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:11:12,004 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9348, 4.9484, 4.8344, 4.4873, 4.5000, 4.8910, 4.6741, 4.6006], device='cuda:4'), covar=tensor([0.0537, 0.0468, 0.0261, 0.0270, 0.0888, 0.0405, 0.0403, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0454, 0.0361, 0.0359, 0.0366, 0.0417, 0.0245, 0.0430], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:11:20,462 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 2.273e+02 2.688e+02 3.076e+02 6.689e+02, threshold=5.376e+02, percent-clipped=9.0 2023-05-01 15:11:41,283 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-05-01 15:12:19,944 INFO [train.py:904] (4/8) Epoch 23, batch 4250, loss[loss=0.1753, simple_loss=0.2758, pruned_loss=0.03737, over 16889.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.272, pruned_loss=0.04958, over 3190470.96 frames. ], batch size: 96, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:12:39,181 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5584, 2.3811, 2.3235, 3.2662, 2.1832, 3.5575, 1.4522, 2.7395], device='cuda:4'), covar=tensor([0.1375, 0.0843, 0.1265, 0.0183, 0.0146, 0.0402, 0.1679, 0.0839], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0194, 0.0192, 0.0205, 0.0215, 0.0202, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:13:27,273 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-05-01 15:13:36,401 INFO [train.py:904] (4/8) Epoch 23, batch 4300, loss[loss=0.1864, simple_loss=0.2807, pruned_loss=0.04601, over 16790.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2728, pruned_loss=0.0482, over 3183817.33 frames. ], batch size: 83, lr: 2.94e-03, grad_scale: 4.0 2023-05-01 15:13:55,082 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.568e+02 2.112e+02 2.458e+02 3.022e+02 6.068e+02, threshold=4.917e+02, percent-clipped=1.0 2023-05-01 15:14:41,964 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=227645.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:14:53,544 INFO [train.py:904] (4/8) Epoch 23, batch 4350, loss[loss=0.1892, simple_loss=0.2751, pruned_loss=0.05164, over 16566.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2756, pruned_loss=0.04903, over 3180250.20 frames. ], batch size: 62, lr: 2.94e-03, grad_scale: 4.0 2023-05-01 15:15:00,413 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9271, 3.5189, 4.1868, 1.8077, 4.4391, 4.4114, 3.1187, 3.3668], device='cuda:4'), covar=tensor([0.0708, 0.0313, 0.0203, 0.1246, 0.0057, 0.0102, 0.0446, 0.0403], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0108, 0.0098, 0.0138, 0.0081, 0.0126, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:15:55,221 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=227693.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:15:55,351 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=227693.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:16:09,816 INFO [train.py:904] (4/8) Epoch 23, batch 4400, loss[loss=0.195, simple_loss=0.2834, pruned_loss=0.05327, over 16306.00 frames. ], tot_loss[loss=0.1885, simple_loss=0.2773, pruned_loss=0.04987, over 3176728.76 frames. ], batch size: 165, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:16:27,158 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.647e+02 2.099e+02 2.601e+02 2.882e+02 6.298e+02, threshold=5.202e+02, percent-clipped=2.0 2023-05-01 15:17:22,035 INFO [train.py:904] (4/8) Epoch 23, batch 4450, loss[loss=0.1847, simple_loss=0.2738, pruned_loss=0.04785, over 16692.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2804, pruned_loss=0.05074, over 3193106.72 frames. ], batch size: 62, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:17:23,606 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=227754.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:17:41,687 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=227767.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:18:35,088 INFO [train.py:904] (4/8) Epoch 23, batch 4500, loss[loss=0.2028, simple_loss=0.2872, pruned_loss=0.05924, over 17239.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2811, pruned_loss=0.05152, over 3208704.10 frames. ], batch size: 44, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:18:52,359 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.372e+02 1.777e+02 2.085e+02 2.523e+02 4.139e+02, threshold=4.170e+02, percent-clipped=0.0 2023-05-01 15:18:56,429 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-05-01 15:19:02,399 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0074, 2.1716, 2.2121, 3.5771, 2.0854, 2.4868, 2.3201, 2.2849], device='cuda:4'), covar=tensor([0.1425, 0.3168, 0.2854, 0.0669, 0.4185, 0.2457, 0.3007, 0.3435], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0457, 0.0372, 0.0332, 0.0439, 0.0527, 0.0426, 0.0533], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:19:11,661 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=227827.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:19:48,151 INFO [train.py:904] (4/8) Epoch 23, batch 4550, loss[loss=0.226, simple_loss=0.3105, pruned_loss=0.07074, over 16710.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2816, pruned_loss=0.05259, over 3207678.22 frames. ], batch size: 57, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:20:39,376 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=227888.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:21:00,790 INFO [train.py:904] (4/8) Epoch 23, batch 4600, loss[loss=0.1939, simple_loss=0.2861, pruned_loss=0.05086, over 16530.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2827, pruned_loss=0.05279, over 3224554.45 frames. ], batch size: 75, lr: 2.94e-03, grad_scale: 8.0 2023-05-01 15:21:18,784 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.806e+02 2.082e+02 2.426e+02 3.731e+02, threshold=4.163e+02, percent-clipped=0.0 2023-05-01 15:22:14,192 INFO [train.py:904] (4/8) Epoch 23, batch 4650, loss[loss=0.1706, simple_loss=0.2578, pruned_loss=0.04167, over 16528.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.2819, pruned_loss=0.05294, over 3225013.70 frames. ], batch size: 68, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:22:25,395 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1490, 3.0802, 1.9175, 3.4368, 2.4024, 3.4976, 2.1279, 2.5421], device='cuda:4'), covar=tensor([0.0350, 0.0443, 0.1743, 0.0178, 0.0931, 0.0485, 0.1581, 0.0919], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0177, 0.0194, 0.0166, 0.0177, 0.0218, 0.0201, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:23:28,866 INFO [train.py:904] (4/8) Epoch 23, batch 4700, loss[loss=0.1939, simple_loss=0.286, pruned_loss=0.0509, over 15318.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2793, pruned_loss=0.05211, over 3224523.22 frames. ], batch size: 190, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:23:34,924 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=228007.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:23:35,014 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6523, 2.7471, 2.6743, 4.8147, 3.5935, 4.1033, 1.6829, 2.9671], device='cuda:4'), covar=tensor([0.1388, 0.0821, 0.1204, 0.0143, 0.0297, 0.0367, 0.1619, 0.0838], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0194, 0.0192, 0.0205, 0.0214, 0.0202, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:23:45,412 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.279e+02 1.844e+02 2.189e+02 2.555e+02 4.222e+02, threshold=4.378e+02, percent-clipped=1.0 2023-05-01 15:24:34,811 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=228049.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:24:34,920 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2989, 5.3399, 5.6826, 5.6641, 5.7255, 5.3942, 5.2704, 5.0156], device='cuda:4'), covar=tensor([0.0298, 0.0441, 0.0320, 0.0363, 0.0600, 0.0298, 0.1186, 0.0425], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0455, 0.0441, 0.0410, 0.0488, 0.0462, 0.0548, 0.0371], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 15:24:41,826 INFO [train.py:904] (4/8) Epoch 23, batch 4750, loss[loss=0.161, simple_loss=0.2516, pruned_loss=0.0352, over 16169.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2755, pruned_loss=0.04993, over 3214519.70 frames. ], batch size: 35, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:25:02,129 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=228067.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:25:03,386 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=228068.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:25:19,478 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=228080.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:25:54,062 INFO [train.py:904] (4/8) Epoch 23, batch 4800, loss[loss=0.168, simple_loss=0.2555, pruned_loss=0.0403, over 17022.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2714, pruned_loss=0.04744, over 3208929.36 frames. ], batch size: 55, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:26:07,417 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1867, 3.4593, 3.6034, 2.0289, 2.9115, 2.3552, 3.5888, 3.6262], device='cuda:4'), covar=tensor([0.0262, 0.0831, 0.0615, 0.2155, 0.0965, 0.0996, 0.0650, 0.0981], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0168, 0.0169, 0.0154, 0.0146, 0.0130, 0.0144, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:26:10,773 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.852e+02 2.092e+02 2.435e+02 6.400e+02, threshold=4.184e+02, percent-clipped=1.0 2023-05-01 15:26:11,947 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=228115.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:26:50,597 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=228141.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:27:07,826 INFO [train.py:904] (4/8) Epoch 23, batch 4850, loss[loss=0.1818, simple_loss=0.2822, pruned_loss=0.04067, over 16363.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2725, pruned_loss=0.04668, over 3204427.90 frames. ], batch size: 146, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:27:54,516 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=228183.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:28:23,610 INFO [train.py:904] (4/8) Epoch 23, batch 4900, loss[loss=0.1658, simple_loss=0.2553, pruned_loss=0.03814, over 16653.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2719, pruned_loss=0.04557, over 3200092.08 frames. ], batch size: 62, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:28:42,122 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.282e+02 1.946e+02 2.242e+02 2.632e+02 4.949e+02, threshold=4.484e+02, percent-clipped=2.0 2023-05-01 15:29:36,710 INFO [train.py:904] (4/8) Epoch 23, batch 4950, loss[loss=0.184, simple_loss=0.2773, pruned_loss=0.04533, over 16386.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2713, pruned_loss=0.04459, over 3200538.48 frames. ], batch size: 146, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:29:43,699 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-05-01 15:30:51,760 INFO [train.py:904] (4/8) Epoch 23, batch 5000, loss[loss=0.1807, simple_loss=0.2708, pruned_loss=0.04531, over 17034.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2727, pruned_loss=0.04458, over 3202726.03 frames. ], batch size: 55, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:31:10,000 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.477e+02 1.949e+02 2.305e+02 2.743e+02 3.734e+02, threshold=4.611e+02, percent-clipped=0.0 2023-05-01 15:31:18,702 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8549, 3.9926, 2.5225, 4.8992, 3.0721, 4.6836, 2.9252, 3.2653], device='cuda:4'), covar=tensor([0.0285, 0.0325, 0.1591, 0.0088, 0.0806, 0.0443, 0.1162, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0175, 0.0192, 0.0164, 0.0175, 0.0216, 0.0200, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:31:30,345 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=228329.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:31:36,833 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=228333.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:31:58,855 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=228349.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:32:04,592 INFO [train.py:904] (4/8) Epoch 23, batch 5050, loss[loss=0.1867, simple_loss=0.2786, pruned_loss=0.04735, over 15589.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2732, pruned_loss=0.04468, over 3194269.19 frames. ], batch size: 191, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:32:19,748 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=228363.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:32:39,549 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4586, 3.4496, 2.0730, 3.9641, 2.6574, 3.9083, 2.3064, 2.8529], device='cuda:4'), covar=tensor([0.0251, 0.0373, 0.1635, 0.0142, 0.0811, 0.0456, 0.1409, 0.0727], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0175, 0.0192, 0.0164, 0.0175, 0.0216, 0.0200, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:32:58,917 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=228390.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:33:00,925 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9812, 2.1422, 2.2526, 3.5767, 2.0702, 2.4615, 2.2412, 2.3262], device='cuda:4'), covar=tensor([0.1509, 0.3632, 0.2977, 0.0621, 0.4100, 0.2605, 0.3674, 0.3212], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0455, 0.0371, 0.0330, 0.0436, 0.0523, 0.0424, 0.0530], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:33:04,496 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=228394.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:33:09,566 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=228397.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:33:17,238 INFO [train.py:904] (4/8) Epoch 23, batch 5100, loss[loss=0.1595, simple_loss=0.2565, pruned_loss=0.03129, over 16853.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2713, pruned_loss=0.04437, over 3198257.05 frames. ], batch size: 102, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:33:22,470 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3662, 4.2635, 4.4223, 4.5751, 4.7761, 4.3571, 4.7491, 4.7884], device='cuda:4'), covar=tensor([0.1824, 0.1315, 0.1591, 0.0737, 0.0481, 0.1005, 0.0613, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0642, 0.0796, 0.0918, 0.0802, 0.0611, 0.0637, 0.0658, 0.0766], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:33:34,756 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.346e+02 1.954e+02 2.359e+02 2.853e+02 6.747e+02, threshold=4.718e+02, percent-clipped=3.0 2023-05-01 15:33:55,467 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=228429.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 15:34:06,444 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=228436.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:34:30,926 INFO [train.py:904] (4/8) Epoch 23, batch 5150, loss[loss=0.1912, simple_loss=0.2766, pruned_loss=0.05292, over 12527.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2719, pruned_loss=0.04432, over 3171536.45 frames. ], batch size: 247, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:35:13,089 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4930, 4.1024, 4.0912, 2.7360, 3.6079, 4.1220, 3.6099, 2.3731], device='cuda:4'), covar=tensor([0.0540, 0.0047, 0.0045, 0.0395, 0.0103, 0.0101, 0.0115, 0.0460], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0085, 0.0087, 0.0135, 0.0100, 0.0112, 0.0096, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 15:35:15,345 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=228483.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:35:25,952 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=228490.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 15:35:44,365 INFO [train.py:904] (4/8) Epoch 23, batch 5200, loss[loss=0.1642, simple_loss=0.2562, pruned_loss=0.03614, over 16676.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2704, pruned_loss=0.04372, over 3180518.58 frames. ], batch size: 89, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:35:49,167 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9460, 2.2572, 2.2958, 2.7614, 1.9258, 3.2372, 1.7365, 2.7849], device='cuda:4'), covar=tensor([0.1216, 0.0732, 0.1130, 0.0152, 0.0110, 0.0350, 0.1513, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0176, 0.0196, 0.0192, 0.0205, 0.0215, 0.0204, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:36:01,315 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.301e+02 2.022e+02 2.336e+02 2.717e+02 6.475e+02, threshold=4.673e+02, percent-clipped=1.0 2023-05-01 15:36:25,656 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=228531.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:36:46,335 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6248, 2.6115, 1.7638, 2.7604, 2.1244, 2.8040, 2.0529, 2.3553], device='cuda:4'), covar=tensor([0.0277, 0.0337, 0.1307, 0.0218, 0.0728, 0.0488, 0.1175, 0.0639], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0176, 0.0193, 0.0164, 0.0175, 0.0216, 0.0200, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:36:57,628 INFO [train.py:904] (4/8) Epoch 23, batch 5250, loss[loss=0.1764, simple_loss=0.2698, pruned_loss=0.0415, over 15543.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2688, pruned_loss=0.04389, over 3181679.13 frames. ], batch size: 191, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:37:04,339 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9188, 2.9462, 2.9501, 5.1365, 4.0579, 4.3420, 1.9634, 3.1301], device='cuda:4'), covar=tensor([0.1242, 0.0764, 0.1081, 0.0112, 0.0302, 0.0363, 0.1495, 0.0818], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0175, 0.0195, 0.0191, 0.0205, 0.0214, 0.0203, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:37:49,377 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-05-01 15:38:10,576 INFO [train.py:904] (4/8) Epoch 23, batch 5300, loss[loss=0.1511, simple_loss=0.2345, pruned_loss=0.03389, over 16622.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2648, pruned_loss=0.04245, over 3193456.78 frames. ], batch size: 62, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:38:28,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.456e+02 1.965e+02 2.226e+02 2.604e+02 5.338e+02, threshold=4.452e+02, percent-clipped=3.0 2023-05-01 15:39:23,396 INFO [train.py:904] (4/8) Epoch 23, batch 5350, loss[loss=0.1782, simple_loss=0.2702, pruned_loss=0.04314, over 17246.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2632, pruned_loss=0.04187, over 3198819.74 frames. ], batch size: 52, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:39:38,567 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=228663.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:40:10,714 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=228685.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:40:16,732 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=228689.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:40:36,813 INFO [train.py:904] (4/8) Epoch 23, batch 5400, loss[loss=0.1903, simple_loss=0.2876, pruned_loss=0.04647, over 16210.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2658, pruned_loss=0.04235, over 3206675.69 frames. ], batch size: 165, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:40:38,240 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-01 15:40:48,792 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=228711.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:40:54,348 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.419e+02 1.938e+02 2.190e+02 2.538e+02 4.586e+02, threshold=4.379e+02, percent-clipped=1.0 2023-05-01 15:40:55,520 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0030, 3.5426, 3.5213, 2.2417, 3.1999, 3.5443, 3.2725, 1.9415], device='cuda:4'), covar=tensor([0.0587, 0.0054, 0.0057, 0.0429, 0.0110, 0.0090, 0.0108, 0.0510], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0084, 0.0086, 0.0135, 0.0099, 0.0110, 0.0096, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 15:41:27,281 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=228736.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:41:54,048 INFO [train.py:904] (4/8) Epoch 23, batch 5450, loss[loss=0.1822, simple_loss=0.2798, pruned_loss=0.04232, over 16812.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2693, pruned_loss=0.04378, over 3203621.68 frames. ], batch size: 102, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:42:42,364 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=228784.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:42:44,357 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=228785.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 15:43:12,590 INFO [train.py:904] (4/8) Epoch 23, batch 5500, loss[loss=0.2141, simple_loss=0.3, pruned_loss=0.06412, over 16727.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2764, pruned_loss=0.0479, over 3191141.97 frames. ], batch size: 89, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:43:32,425 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.710e+02 2.602e+02 3.171e+02 3.709e+02 7.104e+02, threshold=6.343e+02, percent-clipped=12.0 2023-05-01 15:44:31,569 INFO [train.py:904] (4/8) Epoch 23, batch 5550, loss[loss=0.2017, simple_loss=0.2896, pruned_loss=0.05686, over 16475.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2838, pruned_loss=0.05322, over 3149516.54 frames. ], batch size: 68, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:45:52,971 INFO [train.py:904] (4/8) Epoch 23, batch 5600, loss[loss=0.1859, simple_loss=0.2729, pruned_loss=0.04948, over 16577.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.2881, pruned_loss=0.05709, over 3108497.30 frames. ], batch size: 57, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:46:10,853 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.072e+02 3.210e+02 3.734e+02 4.915e+02 1.270e+03, threshold=7.469e+02, percent-clipped=7.0 2023-05-01 15:46:32,423 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7890, 1.3807, 1.7335, 1.7336, 1.8462, 1.9178, 1.6521, 1.8290], device='cuda:4'), covar=tensor([0.0235, 0.0372, 0.0192, 0.0256, 0.0255, 0.0151, 0.0392, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0194, 0.0181, 0.0186, 0.0199, 0.0156, 0.0198, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:46:45,615 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=228935.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:47:13,516 INFO [train.py:904] (4/8) Epoch 23, batch 5650, loss[loss=0.2007, simple_loss=0.2975, pruned_loss=0.05198, over 16648.00 frames. ], tot_loss[loss=0.2071, simple_loss=0.2928, pruned_loss=0.06071, over 3077800.50 frames. ], batch size: 89, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:48:04,622 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=228985.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:48:10,747 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=228989.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:48:22,409 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=228996.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:48:32,721 INFO [train.py:904] (4/8) Epoch 23, batch 5700, loss[loss=0.2777, simple_loss=0.3322, pruned_loss=0.1116, over 11458.00 frames. ], tot_loss[loss=0.2091, simple_loss=0.2938, pruned_loss=0.06226, over 3068047.61 frames. ], batch size: 248, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:48:50,704 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.137e+02 3.035e+02 3.772e+02 4.443e+02 6.175e+02, threshold=7.544e+02, percent-clipped=0.0 2023-05-01 15:49:18,829 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=229033.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:49:22,012 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6158, 2.5486, 2.3934, 3.7607, 2.8367, 3.7732, 1.4745, 2.9419], device='cuda:4'), covar=tensor([0.1424, 0.0824, 0.1343, 0.0209, 0.0226, 0.0402, 0.1782, 0.0765], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0196, 0.0192, 0.0206, 0.0215, 0.0204, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:49:25,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=229037.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:49:30,399 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-05-01 15:49:31,504 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-05-01 15:49:50,896 INFO [train.py:904] (4/8) Epoch 23, batch 5750, loss[loss=0.2308, simple_loss=0.3093, pruned_loss=0.07612, over 16899.00 frames. ], tot_loss[loss=0.2117, simple_loss=0.2962, pruned_loss=0.06359, over 3053912.22 frames. ], batch size: 109, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:50:44,842 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=229085.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 15:51:13,969 INFO [train.py:904] (4/8) Epoch 23, batch 5800, loss[loss=0.1981, simple_loss=0.2977, pruned_loss=0.04928, over 16313.00 frames. ], tot_loss[loss=0.2117, simple_loss=0.2967, pruned_loss=0.06332, over 3033867.05 frames. ], batch size: 146, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:51:32,213 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.995e+02 2.918e+02 3.323e+02 4.052e+02 8.500e+02, threshold=6.645e+02, percent-clipped=1.0 2023-05-01 15:52:01,459 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=229133.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 15:52:31,388 INFO [train.py:904] (4/8) Epoch 23, batch 5850, loss[loss=0.2065, simple_loss=0.2788, pruned_loss=0.06715, over 11464.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2939, pruned_loss=0.0613, over 3032136.54 frames. ], batch size: 247, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:53:22,327 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-05-01 15:53:53,140 INFO [train.py:904] (4/8) Epoch 23, batch 5900, loss[loss=0.2191, simple_loss=0.2885, pruned_loss=0.07487, over 11432.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2932, pruned_loss=0.06075, over 3054456.65 frames. ], batch size: 247, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:54:16,064 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.674e+02 2.520e+02 3.106e+02 3.673e+02 6.247e+02, threshold=6.213e+02, percent-clipped=0.0 2023-05-01 15:55:17,145 INFO [train.py:904] (4/8) Epoch 23, batch 5950, loss[loss=0.1803, simple_loss=0.2793, pruned_loss=0.04062, over 16725.00 frames. ], tot_loss[loss=0.2068, simple_loss=0.2943, pruned_loss=0.05962, over 3078050.26 frames. ], batch size: 89, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:56:18,199 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=229291.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:56:33,851 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=229300.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:56:37,756 INFO [train.py:904] (4/8) Epoch 23, batch 6000, loss[loss=0.1767, simple_loss=0.262, pruned_loss=0.04567, over 17212.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2929, pruned_loss=0.05863, over 3095626.62 frames. ], batch size: 44, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:56:37,756 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 15:56:49,494 INFO [train.py:938] (4/8) Epoch 23, validation: loss=0.1497, simple_loss=0.2623, pruned_loss=0.01859, over 944034.00 frames. 2023-05-01 15:56:49,495 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 15:57:07,745 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.923e+02 2.769e+02 3.211e+02 3.808e+02 9.716e+02, threshold=6.422e+02, percent-clipped=3.0 2023-05-01 15:57:08,374 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8541, 3.5930, 4.0962, 2.0647, 4.2632, 4.2691, 3.1339, 3.1327], device='cuda:4'), covar=tensor([0.0720, 0.0280, 0.0196, 0.1246, 0.0068, 0.0146, 0.0433, 0.0482], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0099, 0.0140, 0.0082, 0.0127, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 15:57:13,067 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0235, 5.0551, 4.8955, 4.5258, 4.5480, 4.9424, 4.7986, 4.6743], device='cuda:4'), covar=tensor([0.0692, 0.0666, 0.0316, 0.0364, 0.0962, 0.0516, 0.0456, 0.0748], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0439, 0.0346, 0.0343, 0.0351, 0.0402, 0.0236, 0.0412], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:57:27,126 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3182, 1.6769, 2.0681, 2.3325, 2.4067, 2.6178, 1.9074, 2.5977], device='cuda:4'), covar=tensor([0.0241, 0.0521, 0.0326, 0.0398, 0.0331, 0.0213, 0.0488, 0.0145], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0193, 0.0180, 0.0184, 0.0198, 0.0155, 0.0197, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:58:06,098 INFO [train.py:904] (4/8) Epoch 23, batch 6050, loss[loss=0.1947, simple_loss=0.2834, pruned_loss=0.05295, over 15438.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2916, pruned_loss=0.05825, over 3104634.69 frames. ], batch size: 191, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:58:20,397 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=229361.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 15:58:54,755 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6090, 5.9368, 5.6108, 5.7367, 5.3567, 5.3005, 5.2714, 6.0708], device='cuda:4'), covar=tensor([0.1398, 0.0857, 0.1186, 0.0890, 0.0876, 0.0727, 0.1366, 0.0855], device='cuda:4'), in_proj_covar=tensor([0.0674, 0.0820, 0.0678, 0.0624, 0.0519, 0.0528, 0.0690, 0.0644], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 15:59:09,392 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-05-01 15:59:21,989 INFO [train.py:904] (4/8) Epoch 23, batch 6100, loss[loss=0.2018, simple_loss=0.288, pruned_loss=0.05782, over 17036.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2911, pruned_loss=0.05727, over 3128152.74 frames. ], batch size: 53, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 15:59:40,506 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.924e+02 2.664e+02 3.486e+02 4.048e+02 8.551e+02, threshold=6.973e+02, percent-clipped=3.0 2023-05-01 16:00:22,339 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=229443.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:00:36,341 INFO [train.py:904] (4/8) Epoch 23, batch 6150, loss[loss=0.1883, simple_loss=0.2755, pruned_loss=0.05061, over 16832.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2889, pruned_loss=0.05667, over 3128478.72 frames. ], batch size: 116, lr: 2.93e-03, grad_scale: 8.0 2023-05-01 16:01:15,180 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=229477.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:01:53,538 INFO [train.py:904] (4/8) Epoch 23, batch 6200, loss[loss=0.1781, simple_loss=0.2667, pruned_loss=0.04475, over 16475.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2869, pruned_loss=0.056, over 3131911.75 frames. ], batch size: 68, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:01:55,795 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=229504.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:02:14,504 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.866e+02 2.788e+02 3.260e+02 3.956e+02 6.763e+02, threshold=6.520e+02, percent-clipped=0.0 2023-05-01 16:02:38,191 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0608, 4.0231, 3.9778, 3.1914, 3.9835, 1.8139, 3.7817, 3.4871], device='cuda:4'), covar=tensor([0.0136, 0.0123, 0.0193, 0.0321, 0.0110, 0.2850, 0.0163, 0.0285], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0160, 0.0201, 0.0179, 0.0178, 0.0208, 0.0190, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:02:50,025 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=229538.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:03:11,715 INFO [train.py:904] (4/8) Epoch 23, batch 6250, loss[loss=0.2073, simple_loss=0.2952, pruned_loss=0.05966, over 16631.00 frames. ], tot_loss[loss=0.199, simple_loss=0.2867, pruned_loss=0.05562, over 3129705.11 frames. ], batch size: 62, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:03:48,910 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=229577.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:04:09,692 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=229591.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:04:28,569 INFO [train.py:904] (4/8) Epoch 23, batch 6300, loss[loss=0.1973, simple_loss=0.2865, pruned_loss=0.05407, over 16946.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2863, pruned_loss=0.05508, over 3134836.60 frames. ], batch size: 109, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:04:50,640 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.872e+02 2.619e+02 3.224e+02 4.167e+02 1.378e+03, threshold=6.449e+02, percent-clipped=8.0 2023-05-01 16:05:27,268 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=229638.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:05:28,546 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=229639.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:05:48,006 INFO [train.py:904] (4/8) Epoch 23, batch 6350, loss[loss=0.2121, simple_loss=0.2976, pruned_loss=0.06332, over 16261.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2872, pruned_loss=0.05653, over 3113612.52 frames. ], batch size: 165, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:05:49,827 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5413, 4.5318, 4.9138, 4.8814, 4.8781, 4.5929, 4.5590, 4.4497], device='cuda:4'), covar=tensor([0.0354, 0.0641, 0.0381, 0.0429, 0.0488, 0.0414, 0.0904, 0.0529], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0461, 0.0448, 0.0417, 0.0493, 0.0469, 0.0555, 0.0375], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 16:05:53,574 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=229656.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:05:57,433 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=229659.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 16:07:04,888 INFO [train.py:904] (4/8) Epoch 23, batch 6400, loss[loss=0.1751, simple_loss=0.2663, pruned_loss=0.04198, over 16782.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2878, pruned_loss=0.0579, over 3101203.94 frames. ], batch size: 89, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:07:24,870 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.907e+02 2.905e+02 3.417e+02 4.252e+02 8.377e+02, threshold=6.835e+02, percent-clipped=2.0 2023-05-01 16:07:32,412 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=229720.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 16:08:06,150 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-05-01 16:08:09,043 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4663, 3.3069, 2.7334, 2.1935, 2.2662, 2.3630, 3.4331, 3.1179], device='cuda:4'), covar=tensor([0.2839, 0.0722, 0.1761, 0.2974, 0.3081, 0.2153, 0.0506, 0.1387], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0268, 0.0304, 0.0314, 0.0296, 0.0260, 0.0297, 0.0338], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 16:08:21,137 INFO [train.py:904] (4/8) Epoch 23, batch 6450, loss[loss=0.1806, simple_loss=0.2773, pruned_loss=0.04199, over 16523.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2882, pruned_loss=0.0576, over 3075445.24 frames. ], batch size: 75, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:09:34,335 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=229799.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:09:40,106 INFO [train.py:904] (4/8) Epoch 23, batch 6500, loss[loss=0.2141, simple_loss=0.2926, pruned_loss=0.06781, over 15444.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2867, pruned_loss=0.05746, over 3064271.17 frames. ], batch size: 191, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:09:59,295 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.807e+02 2.714e+02 3.280e+02 3.954e+02 7.541e+02, threshold=6.561e+02, percent-clipped=1.0 2023-05-01 16:10:25,441 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=229833.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:10:57,908 INFO [train.py:904] (4/8) Epoch 23, batch 6550, loss[loss=0.1923, simple_loss=0.2903, pruned_loss=0.04715, over 16370.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2897, pruned_loss=0.05855, over 3063955.13 frames. ], batch size: 35, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:12:02,427 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=229894.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:12:15,578 INFO [train.py:904] (4/8) Epoch 23, batch 6600, loss[loss=0.1825, simple_loss=0.2759, pruned_loss=0.04459, over 16659.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2911, pruned_loss=0.05847, over 3069927.87 frames. ], batch size: 89, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:12:35,466 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.769e+02 2.706e+02 3.377e+02 4.261e+02 8.660e+02, threshold=6.754e+02, percent-clipped=4.0 2023-05-01 16:13:02,444 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=229933.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:13:33,860 INFO [train.py:904] (4/8) Epoch 23, batch 6650, loss[loss=0.2185, simple_loss=0.3011, pruned_loss=0.06798, over 15316.00 frames. ], tot_loss[loss=0.2057, simple_loss=0.292, pruned_loss=0.05974, over 3049148.38 frames. ], batch size: 190, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:13:37,612 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=229955.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:13:38,753 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=229956.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:14:21,275 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9970, 4.0917, 3.8953, 3.6498, 3.6153, 4.0205, 3.6559, 3.7888], device='cuda:4'), covar=tensor([0.0605, 0.0560, 0.0329, 0.0333, 0.0816, 0.0494, 0.1247, 0.0619], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0437, 0.0342, 0.0340, 0.0348, 0.0398, 0.0233, 0.0410], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:14:26,030 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=229987.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:14:45,034 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-05-01 16:14:52,371 INFO [train.py:904] (4/8) Epoch 23, batch 6700, loss[loss=0.179, simple_loss=0.2738, pruned_loss=0.04212, over 15210.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2906, pruned_loss=0.05976, over 3042276.90 frames. ], batch size: 190, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:14:54,788 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230004.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:15:11,258 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=230015.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 16:15:12,083 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.848e+02 2.536e+02 3.063e+02 3.747e+02 8.042e+02, threshold=6.126e+02, percent-clipped=2.0 2023-05-01 16:16:02,844 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=230048.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:16:09,690 INFO [train.py:904] (4/8) Epoch 23, batch 6750, loss[loss=0.183, simple_loss=0.272, pruned_loss=0.04694, over 16721.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2894, pruned_loss=0.05939, over 3076150.31 frames. ], batch size: 62, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:17:19,741 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=230099.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:17:21,147 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9934, 2.1257, 2.1632, 3.5584, 2.0799, 2.4440, 2.2348, 2.2717], device='cuda:4'), covar=tensor([0.1472, 0.3544, 0.3064, 0.0615, 0.4376, 0.2465, 0.3534, 0.3504], device='cuda:4'), in_proj_covar=tensor([0.0404, 0.0452, 0.0369, 0.0327, 0.0436, 0.0519, 0.0423, 0.0528], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:17:24,163 INFO [train.py:904] (4/8) Epoch 23, batch 6800, loss[loss=0.2024, simple_loss=0.2896, pruned_loss=0.05756, over 17103.00 frames. ], tot_loss[loss=0.2042, simple_loss=0.2894, pruned_loss=0.05949, over 3074261.52 frames. ], batch size: 49, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:17:43,640 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.754e+02 2.682e+02 3.194e+02 4.039e+02 7.212e+02, threshold=6.387e+02, percent-clipped=3.0 2023-05-01 16:18:11,075 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=230133.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:18:31,928 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230147.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:18:36,432 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=230150.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:18:40,518 INFO [train.py:904] (4/8) Epoch 23, batch 6850, loss[loss=0.2054, simple_loss=0.3034, pruned_loss=0.0537, over 16401.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2905, pruned_loss=0.05912, over 3093908.75 frames. ], batch size: 35, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:19:23,350 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230181.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:19:56,852 INFO [train.py:904] (4/8) Epoch 23, batch 6900, loss[loss=0.2582, simple_loss=0.3205, pruned_loss=0.09794, over 11808.00 frames. ], tot_loss[loss=0.2049, simple_loss=0.2928, pruned_loss=0.05846, over 3103414.22 frames. ], batch size: 247, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:20:11,446 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=230211.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:20:20,076 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.486e+02 2.571e+02 3.126e+02 3.927e+02 7.395e+02, threshold=6.253e+02, percent-clipped=1.0 2023-05-01 16:20:45,495 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=230233.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:20:58,925 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=230241.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:21:13,082 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=230250.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:21:17,992 INFO [train.py:904] (4/8) Epoch 23, batch 6950, loss[loss=0.2065, simple_loss=0.2919, pruned_loss=0.0605, over 16738.00 frames. ], tot_loss[loss=0.2072, simple_loss=0.2941, pruned_loss=0.06011, over 3079803.95 frames. ], batch size: 134, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:21:56,019 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-05-01 16:22:01,068 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230281.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:22:20,888 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0349, 3.2849, 3.5007, 1.9800, 2.9887, 2.3811, 3.5242, 3.5406], device='cuda:4'), covar=tensor([0.0281, 0.0879, 0.0629, 0.2213, 0.0839, 0.1002, 0.0577, 0.0868], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0166, 0.0168, 0.0154, 0.0146, 0.0131, 0.0143, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:22:34,183 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=230302.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:22:34,894 INFO [train.py:904] (4/8) Epoch 23, batch 7000, loss[loss=0.1889, simple_loss=0.2783, pruned_loss=0.04973, over 16799.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2936, pruned_loss=0.05865, over 3098452.39 frames. ], batch size: 39, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:22:53,483 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=230315.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 16:22:56,093 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.940e+02 2.703e+02 3.306e+02 4.193e+02 6.685e+02, threshold=6.612e+02, percent-clipped=2.0 2023-05-01 16:23:35,805 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=230343.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:23:50,811 INFO [train.py:904] (4/8) Epoch 23, batch 7050, loss[loss=0.2291, simple_loss=0.2961, pruned_loss=0.08105, over 11329.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2947, pruned_loss=0.05907, over 3097569.62 frames. ], batch size: 248, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:23:56,769 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3710, 4.0479, 4.0005, 2.4960, 3.5703, 4.0437, 3.6301, 2.3341], device='cuda:4'), covar=tensor([0.0553, 0.0047, 0.0056, 0.0483, 0.0115, 0.0118, 0.0102, 0.0454], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0086, 0.0087, 0.0136, 0.0100, 0.0113, 0.0097, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 16:24:05,842 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230363.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 16:24:37,546 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5377, 3.5415, 2.7790, 2.1422, 2.3694, 2.3028, 3.7555, 3.2790], device='cuda:4'), covar=tensor([0.3057, 0.0668, 0.1811, 0.2779, 0.2738, 0.2193, 0.0447, 0.1247], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0269, 0.0305, 0.0315, 0.0297, 0.0260, 0.0297, 0.0338], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 16:24:50,158 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7539, 4.7222, 5.1230, 5.0693, 5.1280, 4.8211, 4.7871, 4.6153], device='cuda:4'), covar=tensor([0.0447, 0.1243, 0.0666, 0.0686, 0.0653, 0.0712, 0.1023, 0.0735], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0464, 0.0450, 0.0418, 0.0494, 0.0472, 0.0559, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 16:25:07,472 INFO [train.py:904] (4/8) Epoch 23, batch 7100, loss[loss=0.1851, simple_loss=0.279, pruned_loss=0.04564, over 16808.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2937, pruned_loss=0.05962, over 3062165.96 frames. ], batch size: 102, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:25:30,796 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.944e+02 2.925e+02 3.270e+02 3.973e+02 7.550e+02, threshold=6.541e+02, percent-clipped=2.0 2023-05-01 16:26:24,957 INFO [train.py:904] (4/8) Epoch 23, batch 7150, loss[loss=0.22, simple_loss=0.3026, pruned_loss=0.06874, over 16624.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2915, pruned_loss=0.0589, over 3085376.03 frames. ], batch size: 57, lr: 2.92e-03, grad_scale: 4.0 2023-05-01 16:26:49,352 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=230468.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:27:42,210 INFO [train.py:904] (4/8) Epoch 23, batch 7200, loss[loss=0.177, simple_loss=0.2743, pruned_loss=0.03986, over 16684.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2898, pruned_loss=0.05754, over 3075871.96 frames. ], batch size: 76, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:27:47,860 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=230506.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:28:03,476 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.863e+02 2.554e+02 3.131e+02 3.804e+02 6.997e+02, threshold=6.261e+02, percent-clipped=1.0 2023-05-01 16:28:16,799 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=230525.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:28:23,105 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=230529.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 16:28:30,040 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=230533.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:28:42,319 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2212, 2.9773, 3.2829, 1.6841, 3.3648, 3.4836, 2.7126, 2.6068], device='cuda:4'), covar=tensor([0.0874, 0.0298, 0.0206, 0.1343, 0.0100, 0.0193, 0.0487, 0.0535], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0108, 0.0099, 0.0138, 0.0082, 0.0125, 0.0128, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:28:42,338 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0903, 3.1725, 1.8977, 3.4589, 2.4148, 3.4919, 2.0598, 2.5400], device='cuda:4'), covar=tensor([0.0324, 0.0430, 0.1807, 0.0220, 0.0908, 0.0617, 0.1632, 0.0886], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0176, 0.0194, 0.0163, 0.0176, 0.0217, 0.0203, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:28:45,775 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6907, 2.8212, 2.3714, 2.5440, 3.1668, 2.7482, 3.2899, 3.3056], device='cuda:4'), covar=tensor([0.0106, 0.0375, 0.0467, 0.0398, 0.0230, 0.0359, 0.0220, 0.0251], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0232, 0.0223, 0.0226, 0.0234, 0.0232, 0.0232, 0.0230], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:28:52,809 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8002, 4.8167, 4.6616, 3.9242, 4.7238, 1.6774, 4.5209, 4.2664], device='cuda:4'), covar=tensor([0.0094, 0.0086, 0.0176, 0.0351, 0.0087, 0.2949, 0.0113, 0.0273], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0159, 0.0199, 0.0177, 0.0175, 0.0206, 0.0188, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:28:58,818 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=230550.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:29:00,806 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-05-01 16:29:02,529 INFO [train.py:904] (4/8) Epoch 23, batch 7250, loss[loss=0.1818, simple_loss=0.2602, pruned_loss=0.05169, over 16561.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2871, pruned_loss=0.05639, over 3076528.79 frames. ], batch size: 62, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:29:51,423 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=230586.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:30:01,059 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0844, 2.3347, 2.2928, 2.6467, 1.9517, 3.1410, 1.8518, 2.6706], device='cuda:4'), covar=tensor([0.1139, 0.0612, 0.1079, 0.0191, 0.0129, 0.0369, 0.1440, 0.0749], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0177, 0.0199, 0.0194, 0.0208, 0.0218, 0.0206, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:30:03,496 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=230594.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:30:08,112 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=230597.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:30:10,099 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230598.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:30:18,070 INFO [train.py:904] (4/8) Epoch 23, batch 7300, loss[loss=0.2458, simple_loss=0.309, pruned_loss=0.09136, over 11454.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2871, pruned_loss=0.05695, over 3060799.69 frames. ], batch size: 248, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:30:39,629 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.739e+02 2.620e+02 3.228e+02 4.011e+02 8.387e+02, threshold=6.456e+02, percent-clipped=2.0 2023-05-01 16:31:19,746 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=230643.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:31:33,672 INFO [train.py:904] (4/8) Epoch 23, batch 7350, loss[loss=0.2475, simple_loss=0.3134, pruned_loss=0.09084, over 11099.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.288, pruned_loss=0.05759, over 3073835.62 frames. ], batch size: 246, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:31:54,140 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6928, 2.7168, 2.4049, 4.0635, 2.9712, 3.8610, 1.5467, 2.7706], device='cuda:4'), covar=tensor([0.1442, 0.0784, 0.1350, 0.0218, 0.0260, 0.0406, 0.1779, 0.0922], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0177, 0.0199, 0.0193, 0.0208, 0.0218, 0.0206, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:32:33,454 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230691.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:32:33,835 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-01 16:32:53,493 INFO [train.py:904] (4/8) Epoch 23, batch 7400, loss[loss=0.1892, simple_loss=0.2859, pruned_loss=0.04621, over 16843.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.289, pruned_loss=0.05787, over 3091837.25 frames. ], batch size: 102, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:33:16,062 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.016e+02 2.706e+02 3.245e+02 4.052e+02 5.999e+02, threshold=6.490e+02, percent-clipped=0.0 2023-05-01 16:34:13,425 INFO [train.py:904] (4/8) Epoch 23, batch 7450, loss[loss=0.2216, simple_loss=0.3102, pruned_loss=0.06647, over 15251.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2902, pruned_loss=0.05916, over 3094046.31 frames. ], batch size: 190, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:34:43,046 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1158, 1.4727, 1.9417, 2.0913, 2.2084, 2.3752, 1.7093, 2.2540], device='cuda:4'), covar=tensor([0.0257, 0.0548, 0.0304, 0.0344, 0.0340, 0.0216, 0.0548, 0.0180], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0193, 0.0179, 0.0185, 0.0199, 0.0156, 0.0197, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:34:50,025 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9065, 3.9963, 2.4940, 4.7156, 3.1428, 4.6103, 2.6767, 3.1552], device='cuda:4'), covar=tensor([0.0265, 0.0342, 0.1669, 0.0206, 0.0762, 0.0503, 0.1383, 0.0840], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0177, 0.0195, 0.0164, 0.0176, 0.0217, 0.0203, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:35:02,398 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3340, 4.2280, 4.4063, 4.5485, 4.6994, 4.2673, 4.6562, 4.7304], device='cuda:4'), covar=tensor([0.2078, 0.1423, 0.1614, 0.0785, 0.0647, 0.1167, 0.0830, 0.0704], device='cuda:4'), in_proj_covar=tensor([0.0621, 0.0773, 0.0891, 0.0779, 0.0593, 0.0620, 0.0647, 0.0749], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:35:19,267 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=230792.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:35:35,336 INFO [train.py:904] (4/8) Epoch 23, batch 7500, loss[loss=0.2021, simple_loss=0.2985, pruned_loss=0.05278, over 16774.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2903, pruned_loss=0.05838, over 3084126.87 frames. ], batch size: 102, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:35:40,978 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=230806.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:35:58,127 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.010e+02 2.839e+02 3.579e+02 4.414e+02 1.002e+03, threshold=7.158e+02, percent-clipped=2.0 2023-05-01 16:36:09,657 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=230824.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 16:36:52,030 INFO [train.py:904] (4/8) Epoch 23, batch 7550, loss[loss=0.2159, simple_loss=0.2997, pruned_loss=0.06604, over 16236.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2897, pruned_loss=0.05906, over 3065243.69 frames. ], batch size: 165, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:36:52,469 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=230853.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:36:53,595 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230854.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:37:33,344 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=230881.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:37:45,737 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=230889.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:37:47,087 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=230890.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:37:58,262 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=230897.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:38:06,491 INFO [train.py:904] (4/8) Epoch 23, batch 7600, loss[loss=0.2017, simple_loss=0.2865, pruned_loss=0.05846, over 16873.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.289, pruned_loss=0.05905, over 3052924.28 frames. ], batch size: 109, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:38:27,912 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.909e+02 2.696e+02 3.168e+02 3.892e+02 6.127e+02, threshold=6.336e+02, percent-clipped=0.0 2023-05-01 16:39:09,876 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-05-01 16:39:10,332 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=230945.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:39:19,838 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=230951.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:39:21,933 INFO [train.py:904] (4/8) Epoch 23, batch 7650, loss[loss=0.2301, simple_loss=0.3108, pruned_loss=0.0747, over 15347.00 frames. ], tot_loss[loss=0.2042, simple_loss=0.2896, pruned_loss=0.05946, over 3055748.24 frames. ], batch size: 190, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:39:30,813 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0966, 3.4920, 3.5310, 2.2746, 3.2269, 3.5331, 3.2310, 2.0299], device='cuda:4'), covar=tensor([0.0549, 0.0072, 0.0060, 0.0438, 0.0121, 0.0122, 0.0116, 0.0478], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0085, 0.0086, 0.0134, 0.0098, 0.0111, 0.0096, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 16:40:35,935 INFO [train.py:904] (4/8) Epoch 23, batch 7700, loss[loss=0.2104, simple_loss=0.2951, pruned_loss=0.06279, over 15315.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2892, pruned_loss=0.05937, over 3072797.10 frames. ], batch size: 191, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:40:57,684 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.025e+02 2.849e+02 3.493e+02 4.384e+02 8.642e+02, threshold=6.986e+02, percent-clipped=7.0 2023-05-01 16:41:53,915 INFO [train.py:904] (4/8) Epoch 23, batch 7750, loss[loss=0.1958, simple_loss=0.2872, pruned_loss=0.05217, over 16329.00 frames. ], tot_loss[loss=0.2033, simple_loss=0.2891, pruned_loss=0.05873, over 3088286.13 frames. ], batch size: 146, lr: 2.92e-03, grad_scale: 8.0 2023-05-01 16:43:09,655 INFO [train.py:904] (4/8) Epoch 23, batch 7800, loss[loss=0.184, simple_loss=0.2771, pruned_loss=0.0454, over 16921.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2894, pruned_loss=0.05908, over 3082598.33 frames. ], batch size: 96, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:43:30,321 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.091e+02 2.907e+02 3.422e+02 4.020e+02 9.124e+02, threshold=6.845e+02, percent-clipped=1.0 2023-05-01 16:43:41,684 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=231124.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 16:43:52,510 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7089, 2.6527, 2.3674, 3.9198, 2.3799, 3.8873, 1.5851, 2.7709], device='cuda:4'), covar=tensor([0.1472, 0.0855, 0.1337, 0.0239, 0.0245, 0.0450, 0.1823, 0.0899], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0176, 0.0197, 0.0192, 0.0207, 0.0216, 0.0205, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:44:17,267 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=231148.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:44:24,895 INFO [train.py:904] (4/8) Epoch 23, batch 7850, loss[loss=0.1816, simple_loss=0.2716, pruned_loss=0.04575, over 16865.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.29, pruned_loss=0.05894, over 3073740.26 frames. ], batch size: 39, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:44:40,796 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.88 vs. limit=5.0 2023-05-01 16:44:52,236 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=231172.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:45:05,945 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=231181.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:45:07,308 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4508, 2.2141, 1.8501, 2.0255, 2.5201, 2.1823, 2.2792, 2.6466], device='cuda:4'), covar=tensor([0.0222, 0.0449, 0.0551, 0.0479, 0.0278, 0.0396, 0.0203, 0.0256], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0234, 0.0225, 0.0227, 0.0235, 0.0232, 0.0233, 0.0231], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:45:16,798 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=231189.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:45:37,201 INFO [train.py:904] (4/8) Epoch 23, batch 7900, loss[loss=0.271, simple_loss=0.3253, pruned_loss=0.1083, over 11306.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.2896, pruned_loss=0.05889, over 3064844.33 frames. ], batch size: 250, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:45:57,446 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.744e+02 2.682e+02 3.251e+02 4.238e+02 6.670e+02, threshold=6.501e+02, percent-clipped=0.0 2023-05-01 16:46:16,872 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=231229.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:46:31,296 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=231237.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:46:44,778 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=231246.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:46:52,740 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=231251.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:46:54,844 INFO [train.py:904] (4/8) Epoch 23, batch 7950, loss[loss=0.1927, simple_loss=0.2774, pruned_loss=0.05395, over 16692.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.2895, pruned_loss=0.05895, over 3072978.20 frames. ], batch size: 76, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:47:06,083 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.10 vs. limit=5.0 2023-05-01 16:47:19,481 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3249, 2.5073, 2.0323, 2.3337, 2.8624, 2.5190, 2.8949, 3.0478], device='cuda:4'), covar=tensor([0.0163, 0.0420, 0.0640, 0.0519, 0.0338, 0.0408, 0.0283, 0.0274], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0234, 0.0225, 0.0227, 0.0235, 0.0232, 0.0233, 0.0231], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:48:11,518 INFO [train.py:904] (4/8) Epoch 23, batch 8000, loss[loss=0.2021, simple_loss=0.3041, pruned_loss=0.05003, over 16780.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2905, pruned_loss=0.05961, over 3082500.68 frames. ], batch size: 124, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:48:26,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=231312.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:48:33,449 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.913e+02 2.670e+02 3.387e+02 4.030e+02 6.156e+02, threshold=6.774e+02, percent-clipped=0.0 2023-05-01 16:49:20,128 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6152, 4.5039, 4.6822, 4.8267, 5.0009, 4.5087, 4.9794, 5.0137], device='cuda:4'), covar=tensor([0.1985, 0.1215, 0.1490, 0.0692, 0.0563, 0.0972, 0.0608, 0.0686], device='cuda:4'), in_proj_covar=tensor([0.0625, 0.0779, 0.0895, 0.0783, 0.0599, 0.0623, 0.0653, 0.0757], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:49:26,911 INFO [train.py:904] (4/8) Epoch 23, batch 8050, loss[loss=0.236, simple_loss=0.315, pruned_loss=0.07846, over 11297.00 frames. ], tot_loss[loss=0.2042, simple_loss=0.2903, pruned_loss=0.05905, over 3071655.43 frames. ], batch size: 247, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:49:38,698 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-01 16:50:04,143 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0598, 4.2399, 2.3575, 4.9277, 3.2600, 4.7724, 2.4656, 3.3363], device='cuda:4'), covar=tensor([0.0268, 0.0339, 0.2067, 0.0214, 0.0808, 0.0392, 0.2034, 0.0814], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0174, 0.0193, 0.0163, 0.0175, 0.0216, 0.0201, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:50:09,653 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7282, 3.0106, 3.2388, 1.9544, 2.9005, 2.1487, 3.2883, 3.3124], device='cuda:4'), covar=tensor([0.0269, 0.0886, 0.0602, 0.2277, 0.0872, 0.1036, 0.0642, 0.0978], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0165, 0.0168, 0.0154, 0.0146, 0.0130, 0.0143, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:50:11,501 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2324, 4.2067, 4.1302, 3.3374, 4.1667, 1.5947, 3.9610, 3.7480], device='cuda:4'), covar=tensor([0.0134, 0.0126, 0.0202, 0.0362, 0.0109, 0.3005, 0.0156, 0.0308], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0159, 0.0200, 0.0177, 0.0176, 0.0208, 0.0188, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:50:42,758 INFO [train.py:904] (4/8) Epoch 23, batch 8100, loss[loss=0.256, simple_loss=0.3153, pruned_loss=0.09835, over 11377.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2901, pruned_loss=0.05882, over 3066118.43 frames. ], batch size: 246, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:51:04,332 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.823e+02 2.502e+02 3.046e+02 3.676e+02 7.002e+02, threshold=6.092e+02, percent-clipped=1.0 2023-05-01 16:51:16,719 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9825, 5.5093, 5.7265, 5.3525, 5.5382, 5.9761, 5.4632, 5.2451], device='cuda:4'), covar=tensor([0.1060, 0.1733, 0.2194, 0.1855, 0.2099, 0.0934, 0.1579, 0.2382], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0602, 0.0665, 0.0497, 0.0662, 0.0691, 0.0518, 0.0665], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 16:51:51,628 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=231448.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:51:59,308 INFO [train.py:904] (4/8) Epoch 23, batch 8150, loss[loss=0.1915, simple_loss=0.2727, pruned_loss=0.05518, over 16918.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2881, pruned_loss=0.0582, over 3062169.26 frames. ], batch size: 116, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:53:05,171 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=231496.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:53:14,921 INFO [train.py:904] (4/8) Epoch 23, batch 8200, loss[loss=0.2251, simple_loss=0.2951, pruned_loss=0.07755, over 12074.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2856, pruned_loss=0.05779, over 3075865.38 frames. ], batch size: 246, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:53:38,106 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.702e+02 2.780e+02 3.369e+02 4.151e+02 6.479e+02, threshold=6.737e+02, percent-clipped=3.0 2023-05-01 16:54:18,754 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-01 16:54:26,264 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=231546.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:54:37,501 INFO [train.py:904] (4/8) Epoch 23, batch 8250, loss[loss=0.1714, simple_loss=0.2715, pruned_loss=0.03567, over 15358.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2841, pruned_loss=0.05462, over 3074356.65 frames. ], batch size: 190, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:55:42,941 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=231594.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:55:57,538 INFO [train.py:904] (4/8) Epoch 23, batch 8300, loss[loss=0.1738, simple_loss=0.2725, pruned_loss=0.03757, over 16417.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2811, pruned_loss=0.05168, over 3068259.12 frames. ], batch size: 146, lr: 2.91e-03, grad_scale: 4.0 2023-05-01 16:56:04,251 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=231607.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:56:22,159 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.405e+02 2.236e+02 2.512e+02 2.973e+02 5.364e+02, threshold=5.024e+02, percent-clipped=0.0 2023-05-01 16:56:29,058 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5239, 3.4567, 3.4858, 2.5919, 3.3064, 2.0769, 3.0709, 2.8044], device='cuda:4'), covar=tensor([0.0152, 0.0161, 0.0183, 0.0213, 0.0117, 0.2490, 0.0152, 0.0260], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0159, 0.0200, 0.0177, 0.0176, 0.0208, 0.0188, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 16:56:42,866 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4012, 3.0923, 3.3037, 1.8999, 3.4226, 3.4980, 2.9006, 2.8401], device='cuda:4'), covar=tensor([0.0693, 0.0248, 0.0212, 0.1132, 0.0100, 0.0203, 0.0400, 0.0406], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0106, 0.0096, 0.0136, 0.0080, 0.0123, 0.0124, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 16:57:11,218 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=231647.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 16:57:19,274 INFO [train.py:904] (4/8) Epoch 23, batch 8350, loss[loss=0.1938, simple_loss=0.2753, pruned_loss=0.05609, over 12057.00 frames. ], tot_loss[loss=0.1908, simple_loss=0.2809, pruned_loss=0.05034, over 3057612.04 frames. ], batch size: 246, lr: 2.91e-03, grad_scale: 4.0 2023-05-01 16:58:23,425 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0442, 3.3930, 3.6222, 2.1245, 3.0850, 2.3962, 3.5790, 3.6374], device='cuda:4'), covar=tensor([0.0303, 0.0821, 0.0514, 0.2075, 0.0808, 0.0992, 0.0622, 0.0945], device='cuda:4'), in_proj_covar=tensor([0.0154, 0.0163, 0.0165, 0.0152, 0.0144, 0.0128, 0.0141, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 16:58:41,949 INFO [train.py:904] (4/8) Epoch 23, batch 8400, loss[loss=0.166, simple_loss=0.2644, pruned_loss=0.03382, over 16526.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2781, pruned_loss=0.04855, over 3036649.91 frames. ], batch size: 62, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 16:58:50,812 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=231708.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 16:59:02,037 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1981, 4.1688, 4.5216, 4.4905, 4.5175, 4.2670, 4.1987, 4.2404], device='cuda:4'), covar=tensor([0.0359, 0.0741, 0.0464, 0.0490, 0.0471, 0.0468, 0.0935, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0456, 0.0443, 0.0410, 0.0488, 0.0463, 0.0547, 0.0371], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 16:59:06,587 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.434e+02 2.191e+02 2.741e+02 3.304e+02 6.516e+02, threshold=5.483e+02, percent-clipped=5.0 2023-05-01 16:59:08,282 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-05-01 16:59:37,740 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-05-01 17:00:04,864 INFO [train.py:904] (4/8) Epoch 23, batch 8450, loss[loss=0.1728, simple_loss=0.2802, pruned_loss=0.03264, over 15217.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.276, pruned_loss=0.04673, over 3040638.86 frames. ], batch size: 190, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:00:52,188 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9801, 4.2450, 4.0887, 4.1166, 3.7910, 3.8452, 3.8628, 4.2605], device='cuda:4'), covar=tensor([0.1120, 0.0994, 0.1021, 0.0813, 0.0870, 0.1692, 0.1072, 0.1019], device='cuda:4'), in_proj_covar=tensor([0.0670, 0.0812, 0.0672, 0.0622, 0.0515, 0.0528, 0.0683, 0.0640], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:01:01,669 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-05-01 17:01:24,904 INFO [train.py:904] (4/8) Epoch 23, batch 8500, loss[loss=0.1507, simple_loss=0.2492, pruned_loss=0.02606, over 16646.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2723, pruned_loss=0.04399, over 3050293.07 frames. ], batch size: 62, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:01:48,516 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.272e+02 2.174e+02 2.671e+02 3.276e+02 7.093e+02, threshold=5.341e+02, percent-clipped=2.0 2023-05-01 17:02:40,881 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5166, 3.4293, 3.4981, 2.7050, 3.3694, 2.1017, 3.0848, 2.8536], device='cuda:4'), covar=tensor([0.0154, 0.0164, 0.0180, 0.0216, 0.0114, 0.2260, 0.0139, 0.0249], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0158, 0.0199, 0.0176, 0.0175, 0.0207, 0.0187, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:02:48,223 INFO [train.py:904] (4/8) Epoch 23, batch 8550, loss[loss=0.18, simple_loss=0.2616, pruned_loss=0.04915, over 12300.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2698, pruned_loss=0.04323, over 3017304.46 frames. ], batch size: 247, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:04:28,278 INFO [train.py:904] (4/8) Epoch 23, batch 8600, loss[loss=0.1719, simple_loss=0.2715, pruned_loss=0.03613, over 16779.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2698, pruned_loss=0.04239, over 2993621.76 frames. ], batch size: 124, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:04:36,583 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=231907.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:05:00,379 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.562e+02 2.309e+02 2.625e+02 3.302e+02 5.565e+02, threshold=5.250e+02, percent-clipped=1.0 2023-05-01 17:06:03,590 INFO [train.py:904] (4/8) Epoch 23, batch 8650, loss[loss=0.1699, simple_loss=0.273, pruned_loss=0.03342, over 16247.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2678, pruned_loss=0.04052, over 3016086.47 frames. ], batch size: 165, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:06:10,284 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=231955.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:07:08,950 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6255, 3.6956, 3.5035, 3.1466, 3.3032, 3.5999, 3.3521, 3.4354], device='cuda:4'), covar=tensor([0.0608, 0.0843, 0.0286, 0.0270, 0.0489, 0.0578, 0.1269, 0.0521], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0426, 0.0331, 0.0331, 0.0336, 0.0387, 0.0228, 0.0399], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:07:53,209 INFO [train.py:904] (4/8) Epoch 23, batch 8700, loss[loss=0.1642, simple_loss=0.2607, pruned_loss=0.0339, over 16221.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.265, pruned_loss=0.03919, over 3025982.84 frames. ], batch size: 165, lr: 2.91e-03, grad_scale: 4.0 2023-05-01 17:07:54,763 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=232003.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 17:08:23,058 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.274e+02 2.142e+02 2.520e+02 3.226e+02 6.058e+02, threshold=5.039e+02, percent-clipped=2.0 2023-05-01 17:09:27,783 INFO [train.py:904] (4/8) Epoch 23, batch 8750, loss[loss=0.1698, simple_loss=0.2772, pruned_loss=0.03117, over 16856.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2656, pruned_loss=0.03895, over 3039317.77 frames. ], batch size: 102, lr: 2.91e-03, grad_scale: 4.0 2023-05-01 17:11:15,140 INFO [train.py:904] (4/8) Epoch 23, batch 8800, loss[loss=0.1706, simple_loss=0.2627, pruned_loss=0.03931, over 12653.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.264, pruned_loss=0.0378, over 3042057.94 frames. ], batch size: 246, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:11:24,868 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 17:11:46,653 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.423e+02 2.145e+02 2.530e+02 3.020e+02 7.216e+02, threshold=5.060e+02, percent-clipped=1.0 2023-05-01 17:12:04,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6192, 3.5466, 3.5741, 2.7852, 3.5109, 2.0035, 3.3065, 2.9386], device='cuda:4'), covar=tensor([0.0125, 0.0114, 0.0175, 0.0203, 0.0099, 0.2537, 0.0127, 0.0268], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0157, 0.0198, 0.0174, 0.0174, 0.0206, 0.0186, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:12:57,922 INFO [train.py:904] (4/8) Epoch 23, batch 8850, loss[loss=0.1776, simple_loss=0.2848, pruned_loss=0.03519, over 16669.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2663, pruned_loss=0.0373, over 3025056.81 frames. ], batch size: 134, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:14:42,072 INFO [train.py:904] (4/8) Epoch 23, batch 8900, loss[loss=0.1796, simple_loss=0.2672, pruned_loss=0.04593, over 16202.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2671, pruned_loss=0.03689, over 3044019.14 frames. ], batch size: 165, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:15:00,696 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5090, 1.8231, 2.1162, 2.4721, 2.5513, 2.7596, 1.9809, 2.6524], device='cuda:4'), covar=tensor([0.0197, 0.0547, 0.0332, 0.0309, 0.0306, 0.0203, 0.0468, 0.0160], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0186, 0.0173, 0.0177, 0.0191, 0.0150, 0.0191, 0.0148], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:15:12,041 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 2.232e+02 2.655e+02 3.319e+02 5.405e+02, threshold=5.309e+02, percent-clipped=1.0 2023-05-01 17:15:29,402 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 17:15:54,028 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7422, 3.9263, 4.2505, 1.9162, 4.3730, 4.5458, 3.4210, 3.2788], device='cuda:4'), covar=tensor([0.0930, 0.0157, 0.0125, 0.1261, 0.0050, 0.0089, 0.0298, 0.0468], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0103, 0.0093, 0.0133, 0.0077, 0.0118, 0.0121, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 17:16:33,776 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=232247.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:16:45,324 INFO [train.py:904] (4/8) Epoch 23, batch 8950, loss[loss=0.1632, simple_loss=0.2579, pruned_loss=0.03424, over 15374.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2665, pruned_loss=0.03683, over 3062812.61 frames. ], batch size: 192, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:17:45,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3027, 3.0716, 3.3148, 1.7195, 3.4691, 3.5320, 2.8342, 2.7566], device='cuda:4'), covar=tensor([0.0770, 0.0262, 0.0204, 0.1295, 0.0083, 0.0153, 0.0423, 0.0461], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0103, 0.0094, 0.0134, 0.0078, 0.0119, 0.0122, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 17:18:01,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8418, 3.8246, 3.9738, 3.7817, 3.9029, 4.3191, 4.0159, 3.7137], device='cuda:4'), covar=tensor([0.2225, 0.2326, 0.2150, 0.2547, 0.2793, 0.1484, 0.1500, 0.2654], device='cuda:4'), in_proj_covar=tensor([0.0395, 0.0576, 0.0636, 0.0476, 0.0633, 0.0662, 0.0497, 0.0637], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 17:18:31,850 INFO [train.py:904] (4/8) Epoch 23, batch 9000, loss[loss=0.1681, simple_loss=0.2573, pruned_loss=0.0395, over 16932.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2635, pruned_loss=0.03583, over 3072080.65 frames. ], batch size: 116, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:18:31,850 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 17:18:42,672 INFO [train.py:938] (4/8) Epoch 23, validation: loss=0.1452, simple_loss=0.249, pruned_loss=0.02066, over 944034.00 frames. 2023-05-01 17:18:42,672 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 17:18:43,592 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=232303.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 17:18:54,249 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=232308.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:19:18,067 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.322e+02 1.963e+02 2.332e+02 2.758e+02 5.487e+02, threshold=4.665e+02, percent-clipped=1.0 2023-05-01 17:20:24,509 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=232351.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:20:28,118 INFO [train.py:904] (4/8) Epoch 23, batch 9050, loss[loss=0.1752, simple_loss=0.2669, pruned_loss=0.04175, over 13076.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2646, pruned_loss=0.03614, over 3086690.92 frames. ], batch size: 250, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:21:23,484 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=232381.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:22:01,465 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-05-01 17:22:03,800 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6925, 4.0071, 2.9361, 2.2187, 2.4602, 2.5312, 4.2707, 3.2999], device='cuda:4'), covar=tensor([0.2965, 0.0504, 0.1799, 0.3050, 0.2886, 0.2124, 0.0346, 0.1423], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0262, 0.0301, 0.0310, 0.0289, 0.0258, 0.0291, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 17:22:12,827 INFO [train.py:904] (4/8) Epoch 23, batch 9100, loss[loss=0.1771, simple_loss=0.2839, pruned_loss=0.03514, over 16349.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2637, pruned_loss=0.03639, over 3077516.17 frames. ], batch size: 146, lr: 2.91e-03, grad_scale: 4.0 2023-05-01 17:22:46,300 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.425e+02 2.157e+02 2.616e+02 3.171e+02 8.876e+02, threshold=5.231e+02, percent-clipped=3.0 2023-05-01 17:23:46,429 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=232442.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 17:24:08,718 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9518, 3.7640, 4.2245, 1.8842, 4.3704, 4.4756, 3.3765, 3.4125], device='cuda:4'), covar=tensor([0.0758, 0.0319, 0.0260, 0.1454, 0.0094, 0.0138, 0.0390, 0.0450], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0103, 0.0093, 0.0133, 0.0077, 0.0119, 0.0121, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 17:24:09,371 INFO [train.py:904] (4/8) Epoch 23, batch 9150, loss[loss=0.1687, simple_loss=0.2565, pruned_loss=0.04043, over 16683.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2634, pruned_loss=0.03606, over 3053809.43 frames. ], batch size: 57, lr: 2.91e-03, grad_scale: 4.0 2023-05-01 17:24:41,181 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-05-01 17:24:58,763 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4516, 3.3875, 3.4840, 3.5495, 3.5910, 3.3412, 3.5673, 3.6586], device='cuda:4'), covar=tensor([0.1182, 0.0915, 0.0997, 0.0640, 0.0637, 0.1973, 0.0869, 0.0751], device='cuda:4'), in_proj_covar=tensor([0.0617, 0.0763, 0.0876, 0.0773, 0.0587, 0.0614, 0.0641, 0.0747], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:25:00,453 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=232476.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:25:18,294 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1451, 3.3949, 3.4248, 2.3859, 3.0956, 3.4541, 3.2205, 2.0929], device='cuda:4'), covar=tensor([0.0502, 0.0057, 0.0057, 0.0363, 0.0131, 0.0097, 0.0095, 0.0464], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0083, 0.0084, 0.0133, 0.0097, 0.0108, 0.0093, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 17:25:52,923 INFO [train.py:904] (4/8) Epoch 23, batch 9200, loss[loss=0.176, simple_loss=0.2678, pruned_loss=0.04211, over 16341.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2596, pruned_loss=0.03507, over 3073796.50 frames. ], batch size: 146, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:26:21,316 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=232518.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:26:24,111 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.371e+02 2.032e+02 2.532e+02 2.962e+02 8.451e+02, threshold=5.064e+02, percent-clipped=2.0 2023-05-01 17:26:57,023 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=232537.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:27:27,732 INFO [train.py:904] (4/8) Epoch 23, batch 9250, loss[loss=0.1648, simple_loss=0.2596, pruned_loss=0.03502, over 16746.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2594, pruned_loss=0.03531, over 3071080.76 frames. ], batch size: 124, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:28:23,195 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=232579.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:29:17,460 INFO [train.py:904] (4/8) Epoch 23, batch 9300, loss[loss=0.1611, simple_loss=0.262, pruned_loss=0.03012, over 16118.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2582, pruned_loss=0.03498, over 3069896.21 frames. ], batch size: 165, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:29:18,375 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=232603.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:29:58,634 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.317e+02 2.074e+02 2.482e+02 2.986e+02 5.209e+02, threshold=4.963e+02, percent-clipped=1.0 2023-05-01 17:30:49,297 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8901, 4.6857, 4.9553, 5.0642, 5.3004, 4.7182, 5.3048, 5.2992], device='cuda:4'), covar=tensor([0.2144, 0.1447, 0.1768, 0.0853, 0.0640, 0.0833, 0.0576, 0.0734], device='cuda:4'), in_proj_covar=tensor([0.0613, 0.0758, 0.0868, 0.0769, 0.0584, 0.0611, 0.0637, 0.0743], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:31:04,797 INFO [train.py:904] (4/8) Epoch 23, batch 9350, loss[loss=0.1608, simple_loss=0.2525, pruned_loss=0.03453, over 16826.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2577, pruned_loss=0.03487, over 3070403.56 frames. ], batch size: 124, lr: 2.91e-03, grad_scale: 8.0 2023-05-01 17:31:50,477 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-05-01 17:32:47,938 INFO [train.py:904] (4/8) Epoch 23, batch 9400, loss[loss=0.1429, simple_loss=0.2309, pruned_loss=0.02745, over 12557.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2575, pruned_loss=0.03435, over 3066145.13 frames. ], batch size: 248, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:33:21,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.567e+02 2.048e+02 2.308e+02 2.875e+02 5.038e+02, threshold=4.615e+02, percent-clipped=1.0 2023-05-01 17:33:50,116 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 17:33:57,944 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=232737.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 17:34:29,649 INFO [train.py:904] (4/8) Epoch 23, batch 9450, loss[loss=0.16, simple_loss=0.2498, pruned_loss=0.03508, over 12474.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2589, pruned_loss=0.03453, over 3053464.17 frames. ], batch size: 246, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:35:04,032 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 17:36:09,220 INFO [train.py:904] (4/8) Epoch 23, batch 9500, loss[loss=0.1551, simple_loss=0.2551, pruned_loss=0.02756, over 15351.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2589, pruned_loss=0.03449, over 3077139.25 frames. ], batch size: 191, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:36:18,542 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4760, 3.4042, 3.5327, 3.6005, 3.6312, 3.3702, 3.6136, 3.6932], device='cuda:4'), covar=tensor([0.1277, 0.0944, 0.0966, 0.0597, 0.0606, 0.2192, 0.0827, 0.0752], device='cuda:4'), in_proj_covar=tensor([0.0610, 0.0756, 0.0866, 0.0765, 0.0581, 0.0607, 0.0635, 0.0740], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:36:44,543 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.549e+02 2.194e+02 2.525e+02 3.286e+02 6.197e+02, threshold=5.051e+02, percent-clipped=6.0 2023-05-01 17:37:08,360 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=232832.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:37:19,977 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-05-01 17:37:54,041 INFO [train.py:904] (4/8) Epoch 23, batch 9550, loss[loss=0.1615, simple_loss=0.2547, pruned_loss=0.03413, over 12490.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2588, pruned_loss=0.03465, over 3089541.91 frames. ], batch size: 250, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:38:38,052 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=232874.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:39:33,897 INFO [train.py:904] (4/8) Epoch 23, batch 9600, loss[loss=0.1759, simple_loss=0.2621, pruned_loss=0.04484, over 12627.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2604, pruned_loss=0.03552, over 3066812.09 frames. ], batch size: 249, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:39:34,627 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=232903.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:40:05,774 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.534e+02 2.131e+02 2.488e+02 2.932e+02 6.217e+02, threshold=4.975e+02, percent-clipped=3.0 2023-05-01 17:41:17,233 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=232951.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:41:20,697 INFO [train.py:904] (4/8) Epoch 23, batch 9650, loss[loss=0.1686, simple_loss=0.2666, pruned_loss=0.0353, over 16195.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.262, pruned_loss=0.0358, over 3069967.62 frames. ], batch size: 165, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:42:42,015 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=232990.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:43:09,181 INFO [train.py:904] (4/8) Epoch 23, batch 9700, loss[loss=0.1811, simple_loss=0.2845, pruned_loss=0.03883, over 16653.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2611, pruned_loss=0.03561, over 3072179.15 frames. ], batch size: 134, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:43:40,154 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=233018.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:43:43,784 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.588e+02 2.165e+02 2.462e+02 3.116e+02 5.593e+02, threshold=4.924e+02, percent-clipped=3.0 2023-05-01 17:44:22,523 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=233037.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:44:50,785 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=233051.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 17:44:53,544 INFO [train.py:904] (4/8) Epoch 23, batch 9750, loss[loss=0.1661, simple_loss=0.2653, pruned_loss=0.03349, over 16193.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2599, pruned_loss=0.03568, over 3052983.13 frames. ], batch size: 165, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:45:43,979 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=233079.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:45:59,982 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=233085.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:46:29,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6436, 4.7001, 4.5267, 4.1669, 4.2172, 4.5924, 4.3824, 4.2563], device='cuda:4'), covar=tensor([0.0576, 0.0696, 0.0313, 0.0325, 0.0910, 0.0548, 0.0461, 0.0685], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0415, 0.0328, 0.0326, 0.0329, 0.0380, 0.0224, 0.0391], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-05-01 17:46:31,580 INFO [train.py:904] (4/8) Epoch 23, batch 9800, loss[loss=0.1655, simple_loss=0.2665, pruned_loss=0.0323, over 16418.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.261, pruned_loss=0.03537, over 3071044.34 frames. ], batch size: 75, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:47:03,692 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.589e+02 2.119e+02 2.583e+02 3.356e+02 7.260e+02, threshold=5.167e+02, percent-clipped=1.0 2023-05-01 17:47:26,799 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=233132.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:47:35,941 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4269, 3.4640, 3.6792, 3.6603, 3.6836, 3.4972, 3.5461, 3.5700], device='cuda:4'), covar=tensor([0.0383, 0.0772, 0.0456, 0.0453, 0.0434, 0.0573, 0.0718, 0.0464], device='cuda:4'), in_proj_covar=tensor([0.0399, 0.0443, 0.0434, 0.0397, 0.0473, 0.0450, 0.0529, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 17:47:47,792 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2268, 4.3260, 4.1373, 3.8627, 3.9203, 4.2438, 3.9278, 3.9776], device='cuda:4'), covar=tensor([0.0534, 0.0448, 0.0293, 0.0287, 0.0622, 0.0420, 0.0750, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0412, 0.0327, 0.0324, 0.0327, 0.0378, 0.0222, 0.0389], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-05-01 17:48:10,564 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4372, 2.8956, 3.0856, 1.8981, 2.8091, 2.1996, 2.9565, 3.1280], device='cuda:4'), covar=tensor([0.0388, 0.0863, 0.0624, 0.2260, 0.0888, 0.1025, 0.0792, 0.0946], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0157, 0.0163, 0.0150, 0.0142, 0.0126, 0.0139, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 17:48:16,807 INFO [train.py:904] (4/8) Epoch 23, batch 9850, loss[loss=0.1632, simple_loss=0.2587, pruned_loss=0.03382, over 15319.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2625, pruned_loss=0.03529, over 3090493.34 frames. ], batch size: 191, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:49:01,100 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=233174.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:49:13,854 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=233180.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:50:08,889 INFO [train.py:904] (4/8) Epoch 23, batch 9900, loss[loss=0.1611, simple_loss=0.2465, pruned_loss=0.03786, over 12496.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2626, pruned_loss=0.0352, over 3085033.35 frames. ], batch size: 248, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:50:21,537 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7324, 2.6644, 1.9024, 2.8732, 2.1546, 2.8852, 2.1557, 2.4315], device='cuda:4'), covar=tensor([0.0330, 0.0420, 0.1420, 0.0269, 0.0714, 0.0515, 0.1324, 0.0669], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0168, 0.0186, 0.0156, 0.0169, 0.0205, 0.0197, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 17:50:46,347 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.375e+02 2.034e+02 2.487e+02 3.058e+02 5.815e+02, threshold=4.974e+02, percent-clipped=3.0 2023-05-01 17:50:52,784 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=233222.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:51:09,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2879, 3.2855, 1.9787, 3.7337, 2.4152, 3.6618, 2.1517, 2.6693], device='cuda:4'), covar=tensor([0.0344, 0.0426, 0.1751, 0.0224, 0.0903, 0.0608, 0.1591, 0.0827], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0168, 0.0186, 0.0155, 0.0169, 0.0205, 0.0197, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-01 17:51:45,251 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1859, 1.6064, 1.8616, 2.1553, 2.2288, 2.3307, 1.7580, 2.3120], device='cuda:4'), covar=tensor([0.0221, 0.0519, 0.0319, 0.0325, 0.0344, 0.0186, 0.0509, 0.0147], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0187, 0.0173, 0.0177, 0.0192, 0.0149, 0.0192, 0.0147], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:52:06,033 INFO [train.py:904] (4/8) Epoch 23, batch 9950, loss[loss=0.1672, simple_loss=0.2628, pruned_loss=0.03583, over 17100.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2652, pruned_loss=0.03589, over 3086624.30 frames. ], batch size: 53, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:52:10,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0035, 2.6629, 2.8564, 2.0011, 2.7046, 2.1008, 2.6850, 2.8926], device='cuda:4'), covar=tensor([0.0297, 0.0957, 0.0548, 0.1986, 0.0817, 0.0995, 0.0647, 0.0870], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0156, 0.0162, 0.0149, 0.0141, 0.0125, 0.0139, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 17:52:18,768 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6067, 2.0517, 1.7227, 1.7863, 2.3840, 2.0883, 2.0366, 2.5312], device='cuda:4'), covar=tensor([0.0254, 0.0528, 0.0691, 0.0588, 0.0289, 0.0430, 0.0316, 0.0280], device='cuda:4'), in_proj_covar=tensor([0.0205, 0.0229, 0.0221, 0.0223, 0.0229, 0.0230, 0.0223, 0.0223], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 17:52:20,180 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=233259.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:53:11,766 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.06 vs. limit=2.0 2023-05-01 17:54:01,243 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 17:54:07,452 INFO [train.py:904] (4/8) Epoch 23, batch 10000, loss[loss=0.1638, simple_loss=0.2646, pruned_loss=0.03149, over 16690.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2631, pruned_loss=0.03494, over 3108820.74 frames. ], batch size: 134, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:54:40,344 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.414e+02 2.128e+02 2.777e+02 3.344e+02 7.240e+02, threshold=5.555e+02, percent-clipped=2.0 2023-05-01 17:54:41,073 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=233320.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:55:26,143 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6338, 3.7032, 3.4936, 3.1612, 3.3171, 3.6066, 3.3533, 3.4250], device='cuda:4'), covar=tensor([0.0546, 0.0583, 0.0302, 0.0279, 0.0491, 0.0455, 0.1436, 0.0484], device='cuda:4'), in_proj_covar=tensor([0.0278, 0.0411, 0.0325, 0.0323, 0.0326, 0.0377, 0.0222, 0.0388], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-05-01 17:55:36,801 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=233346.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 17:55:47,972 INFO [train.py:904] (4/8) Epoch 23, batch 10050, loss[loss=0.1632, simple_loss=0.2544, pruned_loss=0.03598, over 12286.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2625, pruned_loss=0.03454, over 3085918.95 frames. ], batch size: 250, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:56:31,051 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=233374.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 17:57:20,830 INFO [train.py:904] (4/8) Epoch 23, batch 10100, loss[loss=0.1551, simple_loss=0.2461, pruned_loss=0.03205, over 16838.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2625, pruned_loss=0.03471, over 3073639.93 frames. ], batch size: 116, lr: 2.90e-03, grad_scale: 8.0 2023-05-01 17:57:53,980 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.457e+02 2.206e+02 2.577e+02 3.013e+02 4.747e+02, threshold=5.154e+02, percent-clipped=0.0 2023-05-01 17:59:06,334 INFO [train.py:904] (4/8) Epoch 24, batch 0, loss[loss=0.1606, simple_loss=0.2513, pruned_loss=0.03499, over 16860.00 frames. ], tot_loss[loss=0.1606, simple_loss=0.2513, pruned_loss=0.03499, over 16860.00 frames. ], batch size: 42, lr: 2.84e-03, grad_scale: 8.0 2023-05-01 17:59:06,335 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 17:59:14,243 INFO [train.py:938] (4/8) Epoch 24, validation: loss=0.145, simple_loss=0.2483, pruned_loss=0.02085, over 944034.00 frames. 2023-05-01 17:59:14,244 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 18:00:16,158 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-05-01 18:00:19,776 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0595, 2.2920, 2.5923, 2.9856, 2.8288, 3.5324, 2.2854, 3.5187], device='cuda:4'), covar=tensor([0.0267, 0.0525, 0.0361, 0.0382, 0.0371, 0.0213, 0.0570, 0.0205], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0190, 0.0176, 0.0179, 0.0195, 0.0151, 0.0194, 0.0149], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:00:23,694 INFO [train.py:904] (4/8) Epoch 24, batch 50, loss[loss=0.1741, simple_loss=0.2699, pruned_loss=0.03922, over 16762.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2691, pruned_loss=0.04846, over 750055.19 frames. ], batch size: 62, lr: 2.84e-03, grad_scale: 1.0 2023-05-01 18:00:44,679 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0297, 2.5448, 2.1581, 2.4326, 2.9700, 2.7137, 2.9977, 3.0381], device='cuda:4'), covar=tensor([0.0260, 0.0487, 0.0551, 0.0476, 0.0263, 0.0352, 0.0254, 0.0309], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0232, 0.0223, 0.0224, 0.0231, 0.0232, 0.0226, 0.0225], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:00:52,607 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.624e+02 2.410e+02 2.907e+02 3.501e+02 5.437e+02, threshold=5.814e+02, percent-clipped=4.0 2023-05-01 18:01:32,996 INFO [train.py:904] (4/8) Epoch 24, batch 100, loss[loss=0.1527, simple_loss=0.2419, pruned_loss=0.03173, over 15862.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2629, pruned_loss=0.04413, over 1316395.84 frames. ], batch size: 35, lr: 2.84e-03, grad_scale: 1.0 2023-05-01 18:01:53,761 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7473, 3.8909, 2.5117, 4.5125, 3.1581, 4.4115, 2.6457, 3.2882], device='cuda:4'), covar=tensor([0.0335, 0.0457, 0.1718, 0.0331, 0.0829, 0.0682, 0.1501, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0171, 0.0189, 0.0159, 0.0172, 0.0210, 0.0199, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 18:02:40,658 INFO [train.py:904] (4/8) Epoch 24, batch 150, loss[loss=0.1636, simple_loss=0.2499, pruned_loss=0.03867, over 16982.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.26, pruned_loss=0.04334, over 1766122.57 frames. ], batch size: 41, lr: 2.84e-03, grad_scale: 1.0 2023-05-01 18:02:56,017 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=233615.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:03:08,348 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.737e+02 2.160e+02 2.597e+02 3.268e+02 6.927e+02, threshold=5.194e+02, percent-clipped=3.0 2023-05-01 18:03:30,571 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=233639.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:03:39,355 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=233646.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 18:03:48,629 INFO [train.py:904] (4/8) Epoch 24, batch 200, loss[loss=0.1599, simple_loss=0.2601, pruned_loss=0.0299, over 17128.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.262, pruned_loss=0.04414, over 2106577.87 frames. ], batch size: 47, lr: 2.84e-03, grad_scale: 1.0 2023-05-01 18:04:17,983 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=233674.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:04:34,596 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6007, 2.4909, 1.9120, 2.1293, 2.8441, 2.5666, 3.1990, 3.1577], device='cuda:4'), covar=tensor([0.0205, 0.0647, 0.0792, 0.0732, 0.0410, 0.0564, 0.0334, 0.0370], device='cuda:4'), in_proj_covar=tensor([0.0215, 0.0238, 0.0228, 0.0230, 0.0237, 0.0237, 0.0234, 0.0232], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:04:45,848 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=233694.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:04:54,272 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=233700.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:04:58,753 INFO [train.py:904] (4/8) Epoch 24, batch 250, loss[loss=0.1739, simple_loss=0.2726, pruned_loss=0.0376, over 16755.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2609, pruned_loss=0.04457, over 2370707.76 frames. ], batch size: 57, lr: 2.84e-03, grad_scale: 1.0 2023-05-01 18:05:07,829 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=233710.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:05:25,969 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=233722.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:05:26,867 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.709e+02 2.284e+02 2.752e+02 3.292e+02 5.130e+02, threshold=5.503e+02, percent-clipped=0.0 2023-05-01 18:05:48,915 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 18:06:08,092 INFO [train.py:904] (4/8) Epoch 24, batch 300, loss[loss=0.1492, simple_loss=0.24, pruned_loss=0.02922, over 16831.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2582, pruned_loss=0.04344, over 2583839.03 frames. ], batch size: 90, lr: 2.84e-03, grad_scale: 1.0 2023-05-01 18:06:33,457 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=233771.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:07:16,410 INFO [train.py:904] (4/8) Epoch 24, batch 350, loss[loss=0.1648, simple_loss=0.2464, pruned_loss=0.04164, over 12275.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2555, pruned_loss=0.04175, over 2741887.03 frames. ], batch size: 246, lr: 2.84e-03, grad_scale: 1.0 2023-05-01 18:07:43,747 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.364e+02 2.126e+02 2.409e+02 2.975e+02 6.590e+02, threshold=4.818e+02, percent-clipped=2.0 2023-05-01 18:08:25,442 INFO [train.py:904] (4/8) Epoch 24, batch 400, loss[loss=0.1427, simple_loss=0.2339, pruned_loss=0.02574, over 17215.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2538, pruned_loss=0.04142, over 2875718.21 frames. ], batch size: 44, lr: 2.84e-03, grad_scale: 2.0 2023-05-01 18:09:08,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7528, 2.7236, 2.4678, 2.6765, 3.0046, 2.7763, 3.3149, 3.2155], device='cuda:4'), covar=tensor([0.0163, 0.0458, 0.0490, 0.0417, 0.0319, 0.0436, 0.0262, 0.0304], device='cuda:4'), in_proj_covar=tensor([0.0216, 0.0237, 0.0228, 0.0229, 0.0237, 0.0237, 0.0234, 0.0232], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:09:32,974 INFO [train.py:904] (4/8) Epoch 24, batch 450, loss[loss=0.1326, simple_loss=0.2185, pruned_loss=0.02334, over 17011.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2531, pruned_loss=0.04083, over 2978806.23 frames. ], batch size: 41, lr: 2.84e-03, grad_scale: 2.0 2023-05-01 18:09:44,087 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0645, 2.8439, 2.5999, 4.5199, 3.4561, 4.1568, 1.8708, 3.1448], device='cuda:4'), covar=tensor([0.1312, 0.0828, 0.1267, 0.0205, 0.0236, 0.0436, 0.1594, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0173, 0.0194, 0.0189, 0.0199, 0.0212, 0.0203, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 18:09:47,259 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3538, 1.7453, 2.0695, 2.1620, 2.3131, 2.2953, 1.8465, 2.3494], device='cuda:4'), covar=tensor([0.0242, 0.0467, 0.0292, 0.0356, 0.0309, 0.0364, 0.0491, 0.0198], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0194, 0.0180, 0.0184, 0.0200, 0.0156, 0.0198, 0.0153], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:09:50,097 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=233915.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:09:59,639 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 2.087e+02 2.389e+02 2.800e+02 4.652e+02, threshold=4.779e+02, percent-clipped=0.0 2023-05-01 18:10:38,994 INFO [train.py:904] (4/8) Epoch 24, batch 500, loss[loss=0.1541, simple_loss=0.2332, pruned_loss=0.03756, over 15680.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2512, pruned_loss=0.03961, over 3053353.15 frames. ], batch size: 190, lr: 2.83e-03, grad_scale: 2.0 2023-05-01 18:10:53,521 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=233963.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:11:37,695 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.7248, 6.1095, 5.8618, 5.8524, 5.5316, 5.4544, 5.5253, 6.2318], device='cuda:4'), covar=tensor([0.1427, 0.0977, 0.1005, 0.0820, 0.0819, 0.0634, 0.1205, 0.0916], device='cuda:4'), in_proj_covar=tensor([0.0686, 0.0831, 0.0685, 0.0638, 0.0528, 0.0536, 0.0703, 0.0655], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:11:37,699 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=233995.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:11:50,840 INFO [train.py:904] (4/8) Epoch 24, batch 550, loss[loss=0.1462, simple_loss=0.2346, pruned_loss=0.02889, over 17222.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2506, pruned_loss=0.03935, over 3109622.02 frames. ], batch size: 44, lr: 2.83e-03, grad_scale: 2.0 2023-05-01 18:12:17,098 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.523e+02 2.103e+02 2.436e+02 3.122e+02 5.008e+02, threshold=4.871e+02, percent-clipped=2.0 2023-05-01 18:12:20,416 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234025.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:12:57,925 INFO [train.py:904] (4/8) Epoch 24, batch 600, loss[loss=0.1749, simple_loss=0.2452, pruned_loss=0.0523, over 16791.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2493, pruned_loss=0.03953, over 3164473.18 frames. ], batch size: 124, lr: 2.83e-03, grad_scale: 2.0 2023-05-01 18:13:17,435 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=234066.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:13:45,322 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=234086.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:14:05,477 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-01 18:14:08,031 INFO [train.py:904] (4/8) Epoch 24, batch 650, loss[loss=0.1771, simple_loss=0.2654, pruned_loss=0.0444, over 17077.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.248, pruned_loss=0.03925, over 3192864.95 frames. ], batch size: 53, lr: 2.83e-03, grad_scale: 2.0 2023-05-01 18:14:11,006 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-05-01 18:14:36,190 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.365e+02 2.015e+02 2.288e+02 2.913e+02 5.815e+02, threshold=4.575e+02, percent-clipped=2.0 2023-05-01 18:15:16,420 INFO [train.py:904] (4/8) Epoch 24, batch 700, loss[loss=0.1741, simple_loss=0.2564, pruned_loss=0.04593, over 16716.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.248, pruned_loss=0.03887, over 3222250.21 frames. ], batch size: 62, lr: 2.83e-03, grad_scale: 2.0 2023-05-01 18:15:24,275 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6645, 5.0076, 4.7770, 4.7540, 4.4761, 4.4880, 4.4791, 5.0892], device='cuda:4'), covar=tensor([0.1382, 0.0992, 0.1080, 0.0908, 0.1011, 0.1268, 0.1259, 0.0988], device='cuda:4'), in_proj_covar=tensor([0.0685, 0.0833, 0.0686, 0.0638, 0.0529, 0.0536, 0.0705, 0.0655], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:16:23,766 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234202.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:16:24,493 INFO [train.py:904] (4/8) Epoch 24, batch 750, loss[loss=0.1503, simple_loss=0.24, pruned_loss=0.03025, over 17212.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2482, pruned_loss=0.03884, over 3245832.42 frames. ], batch size: 45, lr: 2.83e-03, grad_scale: 2.0 2023-05-01 18:16:44,349 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-05-01 18:16:45,080 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8481, 4.3881, 4.3998, 3.0176, 3.6006, 4.3582, 3.9177, 2.5782], device='cuda:4'), covar=tensor([0.0500, 0.0074, 0.0048, 0.0421, 0.0161, 0.0105, 0.0096, 0.0471], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0086, 0.0087, 0.0135, 0.0099, 0.0111, 0.0096, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 18:16:52,424 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.523e+02 2.040e+02 2.434e+02 2.884e+02 6.768e+02, threshold=4.867e+02, percent-clipped=4.0 2023-05-01 18:17:04,933 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234231.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:17:33,977 INFO [train.py:904] (4/8) Epoch 24, batch 800, loss[loss=0.1584, simple_loss=0.2397, pruned_loss=0.03853, over 16977.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2478, pruned_loss=0.03869, over 3268204.62 frames. ], batch size: 41, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:17:48,305 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=234263.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:17:59,546 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-05-01 18:18:29,131 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=234292.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:18:33,020 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=234295.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:18:43,643 INFO [train.py:904] (4/8) Epoch 24, batch 850, loss[loss=0.1824, simple_loss=0.2828, pruned_loss=0.04104, over 16678.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2478, pruned_loss=0.03841, over 3279359.36 frames. ], batch size: 62, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:19:11,878 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.464e+02 2.010e+02 2.479e+02 2.918e+02 4.237e+02, threshold=4.958e+02, percent-clipped=0.0 2023-05-01 18:19:40,588 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=234343.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:19:49,569 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2051, 5.1842, 5.0640, 4.5945, 4.7005, 5.0774, 5.0437, 4.7166], device='cuda:4'), covar=tensor([0.0594, 0.0529, 0.0327, 0.0325, 0.1093, 0.0492, 0.0360, 0.0774], device='cuda:4'), in_proj_covar=tensor([0.0300, 0.0444, 0.0351, 0.0350, 0.0353, 0.0407, 0.0238, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:19:52,516 INFO [train.py:904] (4/8) Epoch 24, batch 900, loss[loss=0.1582, simple_loss=0.2505, pruned_loss=0.03295, over 17132.00 frames. ], tot_loss[loss=0.1611, simple_loss=0.247, pruned_loss=0.03758, over 3297212.59 frames. ], batch size: 46, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:20:11,474 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=234366.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:20:33,657 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=234381.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:21:02,818 INFO [train.py:904] (4/8) Epoch 24, batch 950, loss[loss=0.1687, simple_loss=0.2475, pruned_loss=0.04492, over 16918.00 frames. ], tot_loss[loss=0.1611, simple_loss=0.2471, pruned_loss=0.03753, over 3302099.21 frames. ], batch size: 116, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:21:17,243 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=234414.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:21:30,309 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.144e+02 2.493e+02 3.109e+02 1.071e+03, threshold=4.986e+02, percent-clipped=5.0 2023-05-01 18:21:37,382 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7428, 2.0675, 2.3402, 2.5479, 2.6650, 2.5904, 1.9925, 2.7872], device='cuda:4'), covar=tensor([0.0193, 0.0450, 0.0311, 0.0316, 0.0305, 0.0343, 0.0473, 0.0189], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0195, 0.0181, 0.0186, 0.0200, 0.0158, 0.0198, 0.0155], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:21:38,433 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2317, 5.2897, 5.7064, 5.6665, 5.7128, 5.3319, 5.2844, 5.0866], device='cuda:4'), covar=tensor([0.0389, 0.0730, 0.0381, 0.0440, 0.0468, 0.0430, 0.0945, 0.0488], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0473, 0.0462, 0.0424, 0.0503, 0.0483, 0.0565, 0.0386], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 18:22:10,607 INFO [train.py:904] (4/8) Epoch 24, batch 1000, loss[loss=0.1612, simple_loss=0.235, pruned_loss=0.04371, over 16922.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2459, pruned_loss=0.03783, over 3309512.52 frames. ], batch size: 109, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:22:50,718 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6939, 3.8642, 2.3883, 4.4480, 2.9380, 4.3713, 2.5660, 3.0916], device='cuda:4'), covar=tensor([0.0395, 0.0417, 0.1763, 0.0371, 0.0940, 0.0568, 0.1595, 0.0899], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0179, 0.0196, 0.0169, 0.0178, 0.0220, 0.0205, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 18:23:22,528 INFO [train.py:904] (4/8) Epoch 24, batch 1050, loss[loss=0.1837, simple_loss=0.271, pruned_loss=0.04818, over 16646.00 frames. ], tot_loss[loss=0.1606, simple_loss=0.2452, pruned_loss=0.03806, over 3305766.62 frames. ], batch size: 57, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:23:28,487 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-05-01 18:23:50,904 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.512e+02 2.145e+02 2.609e+02 3.101e+02 1.491e+03, threshold=5.219e+02, percent-clipped=5.0 2023-05-01 18:24:30,874 INFO [train.py:904] (4/8) Epoch 24, batch 1100, loss[loss=0.1471, simple_loss=0.2418, pruned_loss=0.02617, over 17201.00 frames. ], tot_loss[loss=0.1602, simple_loss=0.245, pruned_loss=0.03771, over 3300243.53 frames. ], batch size: 44, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:24:37,558 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=234558.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:25:18,102 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=234587.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:25:38,659 INFO [train.py:904] (4/8) Epoch 24, batch 1150, loss[loss=0.164, simple_loss=0.2592, pruned_loss=0.03445, over 16689.00 frames. ], tot_loss[loss=0.1605, simple_loss=0.2453, pruned_loss=0.03782, over 3302525.73 frames. ], batch size: 62, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:25:53,357 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1433, 3.3657, 3.4965, 2.2570, 3.1321, 2.4808, 3.6865, 3.6658], device='cuda:4'), covar=tensor([0.0253, 0.0910, 0.0696, 0.2044, 0.0859, 0.1065, 0.0490, 0.0832], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0165, 0.0168, 0.0155, 0.0146, 0.0130, 0.0144, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 18:26:06,112 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.437e+02 2.089e+02 2.435e+02 2.920e+02 5.927e+02, threshold=4.869e+02, percent-clipped=1.0 2023-05-01 18:26:33,557 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 18:26:46,784 INFO [train.py:904] (4/8) Epoch 24, batch 1200, loss[loss=0.1681, simple_loss=0.2457, pruned_loss=0.04527, over 12474.00 frames. ], tot_loss[loss=0.16, simple_loss=0.2448, pruned_loss=0.03757, over 3307518.27 frames. ], batch size: 246, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:27:09,093 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234669.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:27:25,043 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=234681.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:27:53,729 INFO [train.py:904] (4/8) Epoch 24, batch 1250, loss[loss=0.1456, simple_loss=0.2304, pruned_loss=0.03038, over 15871.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2449, pruned_loss=0.03748, over 3315244.63 frames. ], batch size: 35, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:28:21,121 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.382e+02 2.044e+02 2.334e+02 2.766e+02 4.867e+02, threshold=4.669e+02, percent-clipped=0.0 2023-05-01 18:28:25,275 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234726.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:28:30,423 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=234729.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:28:31,740 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=234730.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:28:35,912 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7578, 4.0266, 2.8971, 2.3219, 2.6813, 2.4450, 4.1408, 3.4322], device='cuda:4'), covar=tensor([0.3068, 0.0588, 0.2018, 0.3034, 0.2726, 0.2257, 0.0571, 0.1558], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0271, 0.0309, 0.0318, 0.0298, 0.0266, 0.0298, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 18:28:37,182 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.09 vs. limit=2.0 2023-05-01 18:28:39,596 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.22 vs. limit=5.0 2023-05-01 18:29:02,399 INFO [train.py:904] (4/8) Epoch 24, batch 1300, loss[loss=0.1476, simple_loss=0.2338, pruned_loss=0.03065, over 17165.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2454, pruned_loss=0.03723, over 3320377.55 frames. ], batch size: 46, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:29:27,411 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1794, 5.7541, 5.8695, 5.5938, 5.6803, 6.2246, 5.8038, 5.4553], device='cuda:4'), covar=tensor([0.0955, 0.2033, 0.2584, 0.2125, 0.2607, 0.0869, 0.1533, 0.2458], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0611, 0.0675, 0.0508, 0.0673, 0.0705, 0.0526, 0.0674], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 18:29:41,663 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234782.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:29:47,319 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=234787.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:30:09,999 INFO [train.py:904] (4/8) Epoch 24, batch 1350, loss[loss=0.1428, simple_loss=0.2393, pruned_loss=0.0231, over 17210.00 frames. ], tot_loss[loss=0.1601, simple_loss=0.2456, pruned_loss=0.03728, over 3302432.82 frames. ], batch size: 44, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:30:38,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.453e+02 2.157e+02 2.428e+02 2.986e+02 5.268e+02, threshold=4.855e+02, percent-clipped=4.0 2023-05-01 18:30:41,843 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6012, 3.2113, 3.5096, 2.0349, 3.5775, 3.5790, 3.0002, 2.6674], device='cuda:4'), covar=tensor([0.0725, 0.0274, 0.0210, 0.1151, 0.0122, 0.0210, 0.0413, 0.0487], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0100, 0.0141, 0.0083, 0.0129, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 18:30:45,813 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-01 18:31:07,920 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=234843.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:31:19,624 INFO [train.py:904] (4/8) Epoch 24, batch 1400, loss[loss=0.1881, simple_loss=0.2628, pruned_loss=0.05673, over 16451.00 frames. ], tot_loss[loss=0.1597, simple_loss=0.2455, pruned_loss=0.037, over 3316563.47 frames. ], batch size: 75, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:31:27,956 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=234858.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:31:59,490 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-05-01 18:32:00,157 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234882.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:32:07,335 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=234887.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:32:29,519 INFO [train.py:904] (4/8) Epoch 24, batch 1450, loss[loss=0.1909, simple_loss=0.2622, pruned_loss=0.05983, over 16902.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2454, pruned_loss=0.03722, over 3318025.40 frames. ], batch size: 116, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:32:34,692 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=234906.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:32:42,101 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-05-01 18:32:51,053 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-05-01 18:32:58,960 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.369e+02 1.989e+02 2.305e+02 2.603e+02 6.556e+02, threshold=4.610e+02, percent-clipped=2.0 2023-05-01 18:33:14,529 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=234935.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:33:26,797 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=234943.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:33:38,963 INFO [train.py:904] (4/8) Epoch 24, batch 1500, loss[loss=0.1883, simple_loss=0.2575, pruned_loss=0.05954, over 16873.00 frames. ], tot_loss[loss=0.1604, simple_loss=0.2454, pruned_loss=0.03767, over 3323814.83 frames. ], batch size: 109, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:33:44,898 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-05-01 18:33:45,571 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234958.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:33:49,606 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=234960.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:34:06,070 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-05-01 18:34:13,044 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2610, 1.6040, 1.9869, 2.1014, 2.2465, 2.2885, 1.7786, 2.2822], device='cuda:4'), covar=tensor([0.0260, 0.0516, 0.0287, 0.0346, 0.0337, 0.0305, 0.0536, 0.0195], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0198, 0.0183, 0.0189, 0.0203, 0.0161, 0.0201, 0.0158], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:34:28,920 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8423, 2.1562, 2.4101, 3.1373, 2.1809, 2.3279, 2.3051, 2.2895], device='cuda:4'), covar=tensor([0.1538, 0.3483, 0.2795, 0.0830, 0.4056, 0.2519, 0.3347, 0.3468], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0460, 0.0378, 0.0333, 0.0443, 0.0526, 0.0431, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:34:49,720 INFO [train.py:904] (4/8) Epoch 24, batch 1550, loss[loss=0.1552, simple_loss=0.2449, pruned_loss=0.03276, over 17194.00 frames. ], tot_loss[loss=0.1614, simple_loss=0.2462, pruned_loss=0.03832, over 3321201.91 frames. ], batch size: 46, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:35:12,917 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=235019.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:35:15,777 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=235021.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:35:18,854 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.608e+02 2.203e+02 2.509e+02 3.045e+02 6.004e+02, threshold=5.019e+02, percent-clipped=5.0 2023-05-01 18:35:20,351 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=235025.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:35:21,523 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9948, 5.4107, 5.5360, 5.3357, 5.3040, 5.9616, 5.4790, 5.1466], device='cuda:4'), covar=tensor([0.1124, 0.2182, 0.2807, 0.2147, 0.3103, 0.1005, 0.1630, 0.2661], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0616, 0.0679, 0.0510, 0.0679, 0.0710, 0.0529, 0.0682], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 18:35:58,105 INFO [train.py:904] (4/8) Epoch 24, batch 1600, loss[loss=0.1493, simple_loss=0.2374, pruned_loss=0.03062, over 16840.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2478, pruned_loss=0.03917, over 3334288.62 frames. ], batch size: 42, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:36:37,992 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=235082.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:37:06,937 INFO [train.py:904] (4/8) Epoch 24, batch 1650, loss[loss=0.1683, simple_loss=0.2568, pruned_loss=0.03994, over 15919.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.249, pruned_loss=0.03942, over 3330151.95 frames. ], batch size: 35, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:37:35,281 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.609e+02 2.234e+02 2.602e+02 3.440e+02 9.714e+02, threshold=5.204e+02, percent-clipped=4.0 2023-05-01 18:37:36,746 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-01 18:37:54,827 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=235138.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:38:16,050 INFO [train.py:904] (4/8) Epoch 24, batch 1700, loss[loss=0.1483, simple_loss=0.243, pruned_loss=0.0268, over 17254.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2514, pruned_loss=0.04048, over 3317358.87 frames. ], batch size: 44, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:39:24,669 INFO [train.py:904] (4/8) Epoch 24, batch 1750, loss[loss=0.179, simple_loss=0.2686, pruned_loss=0.04472, over 16441.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2523, pruned_loss=0.04044, over 3319519.09 frames. ], batch size: 68, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:39:52,422 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.615e+02 2.256e+02 2.911e+02 3.628e+02 6.601e+02, threshold=5.823e+02, percent-clipped=7.0 2023-05-01 18:40:12,105 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6281, 2.7182, 2.6848, 4.4396, 2.5574, 3.0582, 2.7451, 2.8359], device='cuda:4'), covar=tensor([0.1375, 0.3382, 0.2955, 0.0577, 0.4132, 0.2520, 0.3322, 0.3450], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0462, 0.0379, 0.0334, 0.0443, 0.0528, 0.0433, 0.0539], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:40:12,950 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=235238.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:40:33,173 INFO [train.py:904] (4/8) Epoch 24, batch 1800, loss[loss=0.1872, simple_loss=0.2717, pruned_loss=0.05134, over 16473.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2524, pruned_loss=0.04019, over 3314476.76 frames. ], batch size: 146, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:40:36,417 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=235255.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:41:40,802 INFO [train.py:904] (4/8) Epoch 24, batch 1850, loss[loss=0.149, simple_loss=0.2302, pruned_loss=0.03384, over 16973.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2541, pruned_loss=0.04093, over 3312560.01 frames. ], batch size: 41, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:41:56,544 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=235314.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:41:59,455 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=235316.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:41:59,602 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=235316.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:42:11,306 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.696e+02 2.161e+02 2.587e+02 3.187e+02 4.962e+02, threshold=5.175e+02, percent-clipped=0.0 2023-05-01 18:42:12,352 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=235325.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:42:49,555 INFO [train.py:904] (4/8) Epoch 24, batch 1900, loss[loss=0.1703, simple_loss=0.2545, pruned_loss=0.043, over 16818.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2544, pruned_loss=0.04029, over 3315863.56 frames. ], batch size: 102, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:43:01,069 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=235360.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 18:43:17,587 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=235373.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:43:26,675 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-05-01 18:43:31,656 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=235382.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:44:00,955 INFO [train.py:904] (4/8) Epoch 24, batch 1950, loss[loss=0.146, simple_loss=0.2446, pruned_loss=0.02369, over 17133.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2542, pruned_loss=0.03986, over 3321816.29 frames. ], batch size: 47, lr: 2.83e-03, grad_scale: 4.0 2023-05-01 18:44:26,804 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=235421.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:44:30,592 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-01 18:44:31,114 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 2.100e+02 2.476e+02 2.993e+02 5.203e+02, threshold=4.952e+02, percent-clipped=1.0 2023-05-01 18:44:37,916 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=235430.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:44:48,977 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=235438.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:45:08,803 INFO [train.py:904] (4/8) Epoch 24, batch 2000, loss[loss=0.1496, simple_loss=0.2325, pruned_loss=0.03332, over 16563.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2538, pruned_loss=0.0401, over 3327304.19 frames. ], batch size: 68, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:45:53,863 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=235486.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:46:15,728 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1062, 4.5380, 4.4386, 3.3249, 3.6615, 4.4409, 3.9677, 2.6550], device='cuda:4'), covar=tensor([0.0462, 0.0055, 0.0055, 0.0349, 0.0156, 0.0097, 0.0104, 0.0489], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0087, 0.0087, 0.0135, 0.0100, 0.0112, 0.0097, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 18:46:16,401 INFO [train.py:904] (4/8) Epoch 24, batch 2050, loss[loss=0.1442, simple_loss=0.2363, pruned_loss=0.02608, over 17196.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2536, pruned_loss=0.03982, over 3320399.21 frames. ], batch size: 44, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:46:46,280 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.341e+02 2.111e+02 2.349e+02 3.134e+02 8.501e+02, threshold=4.699e+02, percent-clipped=2.0 2023-05-01 18:47:03,576 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=235538.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:47:14,212 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-05-01 18:47:24,545 INFO [train.py:904] (4/8) Epoch 24, batch 2100, loss[loss=0.1779, simple_loss=0.261, pruned_loss=0.04736, over 16870.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2543, pruned_loss=0.03953, over 3327075.40 frames. ], batch size: 96, lr: 2.83e-03, grad_scale: 8.0 2023-05-01 18:48:10,073 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=235586.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:48:31,088 INFO [train.py:904] (4/8) Epoch 24, batch 2150, loss[loss=0.2419, simple_loss=0.3141, pruned_loss=0.08489, over 11930.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2558, pruned_loss=0.0409, over 3319854.42 frames. ], batch size: 247, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:48:43,103 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=235611.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:48:46,868 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=235614.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:48:49,949 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=235616.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:49:02,117 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.381e+02 2.182e+02 2.677e+02 3.005e+02 4.991e+02, threshold=5.354e+02, percent-clipped=2.0 2023-05-01 18:49:41,310 INFO [train.py:904] (4/8) Epoch 24, batch 2200, loss[loss=0.1374, simple_loss=0.2262, pruned_loss=0.02429, over 16816.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2561, pruned_loss=0.04074, over 3317879.63 frames. ], batch size: 39, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:49:42,931 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-01 18:49:54,417 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=235662.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:49:57,658 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=235664.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:50:18,247 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=235679.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:50:50,798 INFO [train.py:904] (4/8) Epoch 24, batch 2250, loss[loss=0.1931, simple_loss=0.2658, pruned_loss=0.06015, over 16878.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2559, pruned_loss=0.04058, over 3320547.07 frames. ], batch size: 116, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:51:08,548 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=235716.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 18:51:20,155 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.445e+02 2.086e+02 2.517e+02 3.047e+02 4.896e+02, threshold=5.035e+02, percent-clipped=0.0 2023-05-01 18:51:41,300 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=235740.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:51:43,837 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-05-01 18:51:57,980 INFO [train.py:904] (4/8) Epoch 24, batch 2300, loss[loss=0.1706, simple_loss=0.2657, pruned_loss=0.03778, over 17058.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2566, pruned_loss=0.04094, over 3317813.59 frames. ], batch size: 55, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:52:50,346 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=235791.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:53:08,155 INFO [train.py:904] (4/8) Epoch 24, batch 2350, loss[loss=0.1387, simple_loss=0.2305, pruned_loss=0.02345, over 17223.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2573, pruned_loss=0.04126, over 3312279.83 frames. ], batch size: 44, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:53:22,232 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8387, 3.6471, 4.0036, 2.1451, 4.1097, 4.0655, 3.2795, 3.0218], device='cuda:4'), covar=tensor([0.0685, 0.0239, 0.0165, 0.1185, 0.0089, 0.0208, 0.0379, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0108, 0.0099, 0.0138, 0.0081, 0.0127, 0.0128, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 18:53:37,789 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.657e+02 2.284e+02 2.753e+02 3.332e+02 6.201e+02, threshold=5.507e+02, percent-clipped=2.0 2023-05-01 18:53:43,949 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3966, 2.5934, 2.1633, 2.3755, 2.9208, 2.5766, 3.0254, 3.0632], device='cuda:4'), covar=tensor([0.0197, 0.0476, 0.0595, 0.0484, 0.0288, 0.0418, 0.0290, 0.0316], device='cuda:4'), in_proj_covar=tensor([0.0227, 0.0245, 0.0234, 0.0235, 0.0245, 0.0245, 0.0245, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:54:14,692 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=235852.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:54:15,307 INFO [train.py:904] (4/8) Epoch 24, batch 2400, loss[loss=0.1564, simple_loss=0.252, pruned_loss=0.03043, over 17182.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2579, pruned_loss=0.04125, over 3324814.20 frames. ], batch size: 46, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:55:22,142 INFO [train.py:904] (4/8) Epoch 24, batch 2450, loss[loss=0.1797, simple_loss=0.2521, pruned_loss=0.05368, over 16773.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2579, pruned_loss=0.04095, over 3325744.21 frames. ], batch size: 124, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:55:33,241 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=235911.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:55:51,880 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.418e+02 2.168e+02 2.510e+02 2.858e+02 5.891e+02, threshold=5.019e+02, percent-clipped=1.0 2023-05-01 18:56:02,428 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3907, 2.3473, 2.3404, 4.1649, 2.2677, 2.7013, 2.4099, 2.5351], device='cuda:4'), covar=tensor([0.1413, 0.3977, 0.3225, 0.0589, 0.4209, 0.2937, 0.3675, 0.3687], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0462, 0.0379, 0.0334, 0.0442, 0.0530, 0.0433, 0.0539], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:56:28,796 INFO [train.py:904] (4/8) Epoch 24, batch 2500, loss[loss=0.1476, simple_loss=0.2461, pruned_loss=0.02454, over 17123.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2578, pruned_loss=0.0408, over 3328814.16 frames. ], batch size: 48, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:56:37,230 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=235959.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:57:00,751 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5776, 2.4225, 2.4701, 4.4189, 2.4337, 2.8129, 2.5437, 2.6602], device='cuda:4'), covar=tensor([0.1295, 0.3931, 0.3163, 0.0555, 0.4200, 0.2867, 0.3781, 0.3738], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0461, 0.0378, 0.0334, 0.0442, 0.0529, 0.0432, 0.0539], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:57:12,350 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.63 vs. limit=5.0 2023-05-01 18:57:41,344 INFO [train.py:904] (4/8) Epoch 24, batch 2550, loss[loss=0.1543, simple_loss=0.2534, pruned_loss=0.02763, over 17075.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2571, pruned_loss=0.04082, over 3326961.76 frames. ], batch size: 50, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:57:59,308 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=236016.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 18:58:01,830 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=236018.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:58:10,533 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.492e+02 2.266e+02 2.526e+02 3.100e+02 4.908e+02, threshold=5.053e+02, percent-clipped=0.0 2023-05-01 18:58:24,268 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=236035.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:58:30,724 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-01 18:58:36,687 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.91 vs. limit=5.0 2023-05-01 18:58:49,127 INFO [train.py:904] (4/8) Epoch 24, batch 2600, loss[loss=0.1496, simple_loss=0.2447, pruned_loss=0.02722, over 17131.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2574, pruned_loss=0.04039, over 3332036.27 frames. ], batch size: 47, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 18:58:52,642 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3592, 2.2721, 2.2486, 4.1756, 2.1855, 2.6273, 2.3523, 2.5140], device='cuda:4'), covar=tensor([0.1405, 0.3767, 0.3340, 0.0586, 0.4349, 0.2766, 0.3718, 0.3632], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0461, 0.0379, 0.0334, 0.0442, 0.0530, 0.0432, 0.0539], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 18:59:03,926 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=236064.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 18:59:25,544 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=236079.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 18:59:58,180 INFO [train.py:904] (4/8) Epoch 24, batch 2650, loss[loss=0.1438, simple_loss=0.2391, pruned_loss=0.02422, over 17192.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.258, pruned_loss=0.04027, over 3328677.98 frames. ], batch size: 46, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:00:05,024 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1084, 5.6277, 5.7863, 5.4846, 5.6447, 6.1597, 5.6953, 5.4256], device='cuda:4'), covar=tensor([0.0957, 0.1952, 0.2579, 0.2180, 0.2349, 0.0871, 0.1404, 0.2158], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0619, 0.0684, 0.0512, 0.0681, 0.0714, 0.0534, 0.0680], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 19:00:19,455 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.21 vs. limit=5.0 2023-05-01 19:00:27,606 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.526e+02 2.167e+02 2.518e+02 3.194e+02 5.192e+02, threshold=5.036e+02, percent-clipped=2.0 2023-05-01 19:00:42,117 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.95 vs. limit=5.0 2023-05-01 19:00:58,545 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=236147.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:01:05,908 INFO [train.py:904] (4/8) Epoch 24, batch 2700, loss[loss=0.1843, simple_loss=0.2668, pruned_loss=0.05086, over 16293.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2584, pruned_loss=0.0399, over 3328256.88 frames. ], batch size: 165, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:02:04,696 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-01 19:02:15,469 INFO [train.py:904] (4/8) Epoch 24, batch 2750, loss[loss=0.165, simple_loss=0.2583, pruned_loss=0.03587, over 17237.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2581, pruned_loss=0.03966, over 3328052.20 frames. ], batch size: 45, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:02:30,380 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0889, 4.9996, 4.9452, 4.4722, 4.6326, 4.9911, 4.9305, 4.6248], device='cuda:4'), covar=tensor([0.0535, 0.0584, 0.0296, 0.0356, 0.0978, 0.0470, 0.0340, 0.0807], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0465, 0.0365, 0.0365, 0.0370, 0.0423, 0.0249, 0.0441], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 19:02:38,256 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1095, 2.9882, 3.0084, 4.9879, 3.9239, 4.1388, 2.1854, 3.1995], device='cuda:4'), covar=tensor([0.1334, 0.0878, 0.1214, 0.0209, 0.0331, 0.0574, 0.1530, 0.0938], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0177, 0.0196, 0.0195, 0.0205, 0.0217, 0.0205, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 19:02:44,667 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.298e+02 1.972e+02 2.253e+02 2.676e+02 4.118e+02, threshold=4.507e+02, percent-clipped=0.0 2023-05-01 19:03:22,962 INFO [train.py:904] (4/8) Epoch 24, batch 2800, loss[loss=0.1549, simple_loss=0.2454, pruned_loss=0.03224, over 16770.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2572, pruned_loss=0.03936, over 3326951.67 frames. ], batch size: 83, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:04:32,914 INFO [train.py:904] (4/8) Epoch 24, batch 2850, loss[loss=0.151, simple_loss=0.2356, pruned_loss=0.03314, over 16766.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2566, pruned_loss=0.03937, over 3328457.56 frames. ], batch size: 39, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:05:04,031 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.602e+02 2.044e+02 2.504e+02 3.029e+02 5.589e+02, threshold=5.008e+02, percent-clipped=4.0 2023-05-01 19:05:17,588 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=236335.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:05:43,130 INFO [train.py:904] (4/8) Epoch 24, batch 2900, loss[loss=0.1601, simple_loss=0.2418, pruned_loss=0.03925, over 16010.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2558, pruned_loss=0.04005, over 3321636.41 frames. ], batch size: 35, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:05:52,227 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-01 19:06:09,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9484, 4.4875, 3.1025, 2.4620, 2.6801, 2.6312, 4.8414, 3.5483], device='cuda:4'), covar=tensor([0.2834, 0.0535, 0.1884, 0.2778, 0.2960, 0.2182, 0.0349, 0.1544], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0272, 0.0309, 0.0318, 0.0301, 0.0267, 0.0298, 0.0343], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 19:06:13,275 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=236374.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:06:25,486 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=236383.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:06:53,578 INFO [train.py:904] (4/8) Epoch 24, batch 2950, loss[loss=0.1675, simple_loss=0.2456, pruned_loss=0.04468, over 16755.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2555, pruned_loss=0.04053, over 3327762.42 frames. ], batch size: 83, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:07:24,081 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.806e+02 2.404e+02 2.771e+02 3.339e+02 1.037e+03, threshold=5.543e+02, percent-clipped=6.0 2023-05-01 19:07:54,769 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=236447.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:08:02,709 INFO [train.py:904] (4/8) Epoch 24, batch 3000, loss[loss=0.1633, simple_loss=0.2475, pruned_loss=0.03958, over 16829.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2552, pruned_loss=0.0405, over 3334183.45 frames. ], batch size: 102, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:08:02,709 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 19:08:12,051 INFO [train.py:938] (4/8) Epoch 24, validation: loss=0.1342, simple_loss=0.2393, pruned_loss=0.0145, over 944034.00 frames. 2023-05-01 19:08:12,052 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 19:09:12,137 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=236495.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:09:22,989 INFO [train.py:904] (4/8) Epoch 24, batch 3050, loss[loss=0.1741, simple_loss=0.2512, pruned_loss=0.04849, over 16856.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2556, pruned_loss=0.04144, over 3332403.11 frames. ], batch size: 116, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:09:30,239 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.99 vs. limit=5.0 2023-05-01 19:09:43,782 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1031, 3.9947, 4.4155, 2.5906, 4.6912, 4.7023, 3.4595, 3.6841], device='cuda:4'), covar=tensor([0.0724, 0.0257, 0.0214, 0.1020, 0.0069, 0.0215, 0.0396, 0.0390], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0109, 0.0099, 0.0139, 0.0082, 0.0128, 0.0128, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 19:09:53,422 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.422e+02 2.094e+02 2.423e+02 2.785e+02 4.383e+02, threshold=4.846e+02, percent-clipped=1.0 2023-05-01 19:10:32,486 INFO [train.py:904] (4/8) Epoch 24, batch 3100, loss[loss=0.1599, simple_loss=0.2338, pruned_loss=0.043, over 16661.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.255, pruned_loss=0.04104, over 3333776.46 frames. ], batch size: 134, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:10:36,388 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-01 19:10:51,127 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0929, 4.4109, 4.3724, 3.2286, 3.5732, 4.3777, 3.8816, 2.5271], device='cuda:4'), covar=tensor([0.0442, 0.0079, 0.0049, 0.0345, 0.0159, 0.0102, 0.0106, 0.0496], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0088, 0.0136, 0.0101, 0.0112, 0.0097, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 19:11:43,886 INFO [train.py:904] (4/8) Epoch 24, batch 3150, loss[loss=0.1638, simple_loss=0.2404, pruned_loss=0.0436, over 16741.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.254, pruned_loss=0.04048, over 3332984.50 frames. ], batch size: 83, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:12:13,745 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.596e+02 2.103e+02 2.556e+02 2.857e+02 5.533e+02, threshold=5.112e+02, percent-clipped=2.0 2023-05-01 19:12:52,317 INFO [train.py:904] (4/8) Epoch 24, batch 3200, loss[loss=0.1728, simple_loss=0.2514, pruned_loss=0.0471, over 15544.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2536, pruned_loss=0.04052, over 3326332.43 frames. ], batch size: 190, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:13:22,044 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=236674.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:14:01,577 INFO [train.py:904] (4/8) Epoch 24, batch 3250, loss[loss=0.1504, simple_loss=0.2412, pruned_loss=0.02973, over 17174.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2534, pruned_loss=0.04022, over 3329085.24 frames. ], batch size: 46, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:14:27,660 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=236722.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:14:31,667 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.578e+02 2.232e+02 2.615e+02 3.014e+02 5.902e+02, threshold=5.231e+02, percent-clipped=1.0 2023-05-01 19:15:11,542 INFO [train.py:904] (4/8) Epoch 24, batch 3300, loss[loss=0.1753, simple_loss=0.2622, pruned_loss=0.04413, over 16204.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2548, pruned_loss=0.04059, over 3327237.34 frames. ], batch size: 165, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:16:21,167 INFO [train.py:904] (4/8) Epoch 24, batch 3350, loss[loss=0.1684, simple_loss=0.25, pruned_loss=0.04336, over 16643.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2543, pruned_loss=0.04009, over 3332122.66 frames. ], batch size: 83, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:16:38,314 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8748, 2.8964, 2.5962, 4.5498, 3.5667, 4.0731, 1.8432, 2.8905], device='cuda:4'), covar=tensor([0.1281, 0.0720, 0.1210, 0.0218, 0.0261, 0.0547, 0.1463, 0.0940], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0178, 0.0197, 0.0197, 0.0207, 0.0219, 0.0205, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 19:16:51,747 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.239e+02 2.070e+02 2.526e+02 2.944e+02 6.359e+02, threshold=5.052e+02, percent-clipped=3.0 2023-05-01 19:17:17,972 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-01 19:17:25,333 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0163, 5.0223, 4.8054, 4.2229, 4.9196, 1.9763, 4.6840, 4.6498], device='cuda:4'), covar=tensor([0.0121, 0.0102, 0.0227, 0.0452, 0.0118, 0.2854, 0.0145, 0.0244], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0168, 0.0210, 0.0186, 0.0186, 0.0215, 0.0199, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:17:33,207 INFO [train.py:904] (4/8) Epoch 24, batch 3400, loss[loss=0.1891, simple_loss=0.2677, pruned_loss=0.05526, over 16898.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2541, pruned_loss=0.04017, over 3335136.95 frames. ], batch size: 116, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:17:55,370 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0972, 4.1315, 4.4107, 4.3885, 4.4268, 4.1574, 4.1844, 4.1281], device='cuda:4'), covar=tensor([0.0393, 0.0631, 0.0398, 0.0430, 0.0556, 0.0452, 0.0781, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0478, 0.0465, 0.0428, 0.0510, 0.0485, 0.0570, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 19:18:04,613 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.17 vs. limit=5.0 2023-05-01 19:18:23,651 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3172, 5.3265, 5.0797, 4.5051, 5.1732, 2.1126, 4.9440, 5.0335], device='cuda:4'), covar=tensor([0.0105, 0.0078, 0.0214, 0.0464, 0.0113, 0.2569, 0.0152, 0.0216], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0168, 0.0210, 0.0186, 0.0186, 0.0215, 0.0199, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:18:31,128 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=236892.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 19:18:45,596 INFO [train.py:904] (4/8) Epoch 24, batch 3450, loss[loss=0.1576, simple_loss=0.2324, pruned_loss=0.04141, over 16706.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.253, pruned_loss=0.03993, over 3326113.34 frames. ], batch size: 83, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:18:56,059 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-01 19:19:15,850 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.559e+02 2.094e+02 2.355e+02 2.776e+02 4.395e+02, threshold=4.710e+02, percent-clipped=0.0 2023-05-01 19:19:17,875 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-01 19:19:55,267 INFO [train.py:904] (4/8) Epoch 24, batch 3500, loss[loss=0.16, simple_loss=0.2622, pruned_loss=0.02891, over 17116.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2524, pruned_loss=0.03972, over 3330998.87 frames. ], batch size: 47, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:19:56,895 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=236953.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 19:20:41,620 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6892, 3.9778, 4.0581, 2.7457, 3.5698, 4.1343, 3.7375, 2.2605], device='cuda:4'), covar=tensor([0.0520, 0.0245, 0.0068, 0.0427, 0.0144, 0.0100, 0.0110, 0.0526], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0087, 0.0087, 0.0136, 0.0101, 0.0112, 0.0097, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 19:20:51,650 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-01 19:21:06,713 INFO [train.py:904] (4/8) Epoch 24, batch 3550, loss[loss=0.1681, simple_loss=0.2475, pruned_loss=0.04431, over 16434.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2514, pruned_loss=0.03945, over 3335827.89 frames. ], batch size: 146, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:21:09,733 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-05-01 19:21:18,569 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-05-01 19:21:24,112 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-05-01 19:21:35,731 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.420e+02 1.933e+02 2.243e+02 2.593e+02 4.523e+02, threshold=4.485e+02, percent-clipped=0.0 2023-05-01 19:22:04,825 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8800, 4.0734, 3.0939, 2.4178, 2.5863, 2.5691, 4.1725, 3.4494], device='cuda:4'), covar=tensor([0.2770, 0.0497, 0.1712, 0.3301, 0.3080, 0.2183, 0.0464, 0.1627], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0272, 0.0309, 0.0320, 0.0302, 0.0267, 0.0298, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 19:22:15,045 INFO [train.py:904] (4/8) Epoch 24, batch 3600, loss[loss=0.1675, simple_loss=0.2485, pruned_loss=0.04328, over 16311.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2506, pruned_loss=0.03926, over 3336524.81 frames. ], batch size: 165, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:22:20,519 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 19:22:37,075 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-01 19:23:26,120 INFO [train.py:904] (4/8) Epoch 24, batch 3650, loss[loss=0.1712, simple_loss=0.2448, pruned_loss=0.04876, over 16788.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2502, pruned_loss=0.03976, over 3322880.68 frames. ], batch size: 124, lr: 2.82e-03, grad_scale: 8.0 2023-05-01 19:23:58,725 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.153e+02 2.502e+02 3.114e+02 9.969e+02, threshold=5.004e+02, percent-clipped=5.0 2023-05-01 19:24:03,724 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8262, 4.8487, 5.0078, 4.8753, 4.9061, 5.4714, 4.9699, 4.6106], device='cuda:4'), covar=tensor([0.1399, 0.1871, 0.2117, 0.2115, 0.2494, 0.1012, 0.1709, 0.2664], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0626, 0.0688, 0.0517, 0.0687, 0.0718, 0.0539, 0.0681], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 19:24:39,858 INFO [train.py:904] (4/8) Epoch 24, batch 3700, loss[loss=0.1552, simple_loss=0.2362, pruned_loss=0.03706, over 16291.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2491, pruned_loss=0.04144, over 3300164.49 frames. ], batch size: 165, lr: 2.82e-03, grad_scale: 4.0 2023-05-01 19:25:53,323 INFO [train.py:904] (4/8) Epoch 24, batch 3750, loss[loss=0.1659, simple_loss=0.2474, pruned_loss=0.04217, over 16463.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2499, pruned_loss=0.04266, over 3276257.43 frames. ], batch size: 75, lr: 2.82e-03, grad_scale: 4.0 2023-05-01 19:26:10,249 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 19:26:25,688 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.587e+02 2.223e+02 2.609e+02 3.048e+02 5.921e+02, threshold=5.219e+02, percent-clipped=3.0 2023-05-01 19:26:57,488 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=237248.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 19:27:03,606 INFO [train.py:904] (4/8) Epoch 24, batch 3800, loss[loss=0.1514, simple_loss=0.2317, pruned_loss=0.03559, over 16659.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2508, pruned_loss=0.04374, over 3275960.78 frames. ], batch size: 89, lr: 2.82e-03, grad_scale: 4.0 2023-05-01 19:28:20,059 INFO [train.py:904] (4/8) Epoch 24, batch 3850, loss[loss=0.1777, simple_loss=0.2548, pruned_loss=0.05033, over 16907.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2516, pruned_loss=0.04487, over 3276935.04 frames. ], batch size: 116, lr: 2.81e-03, grad_scale: 4.0 2023-05-01 19:28:24,513 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4211, 1.7485, 2.1134, 2.3381, 2.5045, 2.3922, 1.8832, 2.5590], device='cuda:4'), covar=tensor([0.0200, 0.0548, 0.0341, 0.0311, 0.0340, 0.0398, 0.0544, 0.0183], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0195, 0.0184, 0.0188, 0.0203, 0.0163, 0.0200, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:28:52,884 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.689e+02 2.317e+02 2.601e+02 3.052e+02 4.766e+02, threshold=5.202e+02, percent-clipped=0.0 2023-05-01 19:29:30,846 INFO [train.py:904] (4/8) Epoch 24, batch 3900, loss[loss=0.1801, simple_loss=0.2476, pruned_loss=0.05637, over 16874.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2512, pruned_loss=0.04512, over 3285085.77 frames. ], batch size: 109, lr: 2.81e-03, grad_scale: 4.0 2023-05-01 19:30:43,384 INFO [train.py:904] (4/8) Epoch 24, batch 3950, loss[loss=0.1569, simple_loss=0.2282, pruned_loss=0.04278, over 16713.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2509, pruned_loss=0.04561, over 3267314.58 frames. ], batch size: 89, lr: 2.81e-03, grad_scale: 4.0 2023-05-01 19:31:17,988 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.570e+02 2.223e+02 2.645e+02 3.283e+02 6.262e+02, threshold=5.289e+02, percent-clipped=1.0 2023-05-01 19:31:57,060 INFO [train.py:904] (4/8) Epoch 24, batch 4000, loss[loss=0.197, simple_loss=0.2768, pruned_loss=0.05862, over 16813.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2509, pruned_loss=0.04563, over 3271735.59 frames. ], batch size: 124, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:32:53,207 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=237491.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:33:10,917 INFO [train.py:904] (4/8) Epoch 24, batch 4050, loss[loss=0.1654, simple_loss=0.2551, pruned_loss=0.03789, over 16844.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2511, pruned_loss=0.04451, over 3282392.10 frames. ], batch size: 116, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:33:43,734 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.357e+02 1.828e+02 2.122e+02 2.508e+02 6.067e+02, threshold=4.243e+02, percent-clipped=1.0 2023-05-01 19:34:17,714 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=237548.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 19:34:23,858 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=237552.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:34:24,477 INFO [train.py:904] (4/8) Epoch 24, batch 4100, loss[loss=0.1651, simple_loss=0.2622, pruned_loss=0.03407, over 16800.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2526, pruned_loss=0.04389, over 3283546.87 frames. ], batch size: 102, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:34:38,846 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=237563.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 19:35:30,061 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=237596.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 19:35:40,165 INFO [train.py:904] (4/8) Epoch 24, batch 4150, loss[loss=0.2046, simple_loss=0.3009, pruned_loss=0.05414, over 16661.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2603, pruned_loss=0.04675, over 3252289.51 frames. ], batch size: 134, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:36:14,092 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=237624.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 19:36:16,367 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.423e+02 2.139e+02 2.639e+02 3.206e+02 5.465e+02, threshold=5.279e+02, percent-clipped=4.0 2023-05-01 19:36:56,393 INFO [train.py:904] (4/8) Epoch 24, batch 4200, loss[loss=0.201, simple_loss=0.296, pruned_loss=0.05298, over 16574.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2668, pruned_loss=0.04811, over 3214932.20 frames. ], batch size: 62, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:36:59,483 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0273, 2.1306, 2.2448, 3.5187, 2.0867, 2.4621, 2.2408, 2.2869], device='cuda:4'), covar=tensor([0.1445, 0.3602, 0.2906, 0.0687, 0.4228, 0.2474, 0.3517, 0.3367], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0462, 0.0377, 0.0333, 0.0440, 0.0531, 0.0432, 0.0540], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:38:10,757 INFO [train.py:904] (4/8) Epoch 24, batch 4250, loss[loss=0.1682, simple_loss=0.2629, pruned_loss=0.03671, over 17241.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2707, pruned_loss=0.04801, over 3204793.58 frames. ], batch size: 45, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:38:16,753 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.42 vs. limit=5.0 2023-05-01 19:38:45,324 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.496e+02 2.159e+02 2.541e+02 2.876e+02 4.427e+02, threshold=5.081e+02, percent-clipped=0.0 2023-05-01 19:38:48,876 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5488, 3.6156, 2.6790, 2.2690, 2.3701, 2.4619, 3.7529, 3.1776], device='cuda:4'), covar=tensor([0.3090, 0.0676, 0.1945, 0.2882, 0.2809, 0.2108, 0.0590, 0.1357], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0272, 0.0309, 0.0320, 0.0304, 0.0267, 0.0299, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 19:39:16,332 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-05-01 19:39:26,152 INFO [train.py:904] (4/8) Epoch 24, batch 4300, loss[loss=0.1984, simple_loss=0.2977, pruned_loss=0.04955, over 16536.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2718, pruned_loss=0.04715, over 3174711.13 frames. ], batch size: 75, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:40:01,137 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-05-01 19:40:23,444 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=237790.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:40:41,734 INFO [train.py:904] (4/8) Epoch 24, batch 4350, loss[loss=0.1849, simple_loss=0.2726, pruned_loss=0.04853, over 16475.00 frames. ], tot_loss[loss=0.1857, simple_loss=0.2752, pruned_loss=0.04809, over 3177355.29 frames. ], batch size: 35, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:41:14,400 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.621e+02 2.331e+02 2.533e+02 3.016e+02 1.010e+03, threshold=5.065e+02, percent-clipped=1.0 2023-05-01 19:41:33,062 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3799, 4.4669, 4.7402, 4.6957, 4.7129, 4.4272, 4.4176, 4.2529], device='cuda:4'), covar=tensor([0.0302, 0.0475, 0.0297, 0.0358, 0.0430, 0.0360, 0.0806, 0.0546], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0468, 0.0453, 0.0417, 0.0498, 0.0474, 0.0555, 0.0381], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 19:41:45,862 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=237847.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:41:52,205 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=237851.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:41:54,769 INFO [train.py:904] (4/8) Epoch 24, batch 4400, loss[loss=0.2, simple_loss=0.2921, pruned_loss=0.05394, over 16761.00 frames. ], tot_loss[loss=0.188, simple_loss=0.2772, pruned_loss=0.04937, over 3178710.66 frames. ], batch size: 124, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:43:05,638 INFO [train.py:904] (4/8) Epoch 24, batch 4450, loss[loss=0.1808, simple_loss=0.2786, pruned_loss=0.04147, over 16774.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2809, pruned_loss=0.05086, over 3190089.56 frames. ], batch size: 83, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:43:18,864 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=237912.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:43:28,553 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=237919.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 19:43:38,899 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.342e+02 1.917e+02 2.137e+02 2.680e+02 5.470e+02, threshold=4.273e+02, percent-clipped=1.0 2023-05-01 19:44:16,769 INFO [train.py:904] (4/8) Epoch 24, batch 4500, loss[loss=0.1875, simple_loss=0.284, pruned_loss=0.04553, over 16896.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2818, pruned_loss=0.05172, over 3190300.83 frames. ], batch size: 96, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:44:37,790 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2872, 2.4721, 2.4285, 4.0138, 2.2905, 2.7804, 2.5132, 2.5530], device='cuda:4'), covar=tensor([0.1389, 0.3119, 0.2834, 0.0571, 0.4029, 0.2349, 0.3048, 0.3286], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0461, 0.0376, 0.0332, 0.0442, 0.0530, 0.0431, 0.0540], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:44:46,695 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=237973.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:45:02,421 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6881, 1.7915, 1.6115, 1.5201, 1.8858, 1.5893, 1.6572, 1.9259], device='cuda:4'), covar=tensor([0.0190, 0.0284, 0.0428, 0.0355, 0.0210, 0.0314, 0.0167, 0.0222], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0239, 0.0229, 0.0231, 0.0240, 0.0240, 0.0241, 0.0238], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:45:03,612 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5334, 5.4953, 5.3954, 5.0742, 5.1274, 5.4277, 5.2977, 5.0868], device='cuda:4'), covar=tensor([0.0461, 0.0353, 0.0186, 0.0201, 0.0707, 0.0313, 0.0238, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0446, 0.0351, 0.0350, 0.0353, 0.0405, 0.0239, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:45:32,218 INFO [train.py:904] (4/8) Epoch 24, batch 4550, loss[loss=0.1955, simple_loss=0.2846, pruned_loss=0.05326, over 17031.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2824, pruned_loss=0.05283, over 3187400.08 frames. ], batch size: 55, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:46:04,568 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.356e+02 1.751e+02 2.009e+02 2.366e+02 4.725e+02, threshold=4.018e+02, percent-clipped=1.0 2023-05-01 19:46:35,287 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7247, 3.7874, 4.0684, 2.2966, 3.4645, 2.6311, 3.9352, 4.0725], device='cuda:4'), covar=tensor([0.0183, 0.0744, 0.0474, 0.2074, 0.0753, 0.0923, 0.0512, 0.0831], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0167, 0.0168, 0.0154, 0.0146, 0.0131, 0.0143, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 19:46:44,383 INFO [train.py:904] (4/8) Epoch 24, batch 4600, loss[loss=0.1853, simple_loss=0.2751, pruned_loss=0.04771, over 16663.00 frames. ], tot_loss[loss=0.1948, simple_loss=0.2831, pruned_loss=0.05327, over 3184096.09 frames. ], batch size: 134, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:47:00,804 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=238064.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:47:15,198 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=238074.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:47:28,143 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6158, 2.5034, 1.8373, 2.7105, 2.0782, 2.7532, 2.1716, 2.3324], device='cuda:4'), covar=tensor([0.0331, 0.0367, 0.1408, 0.0218, 0.0713, 0.0381, 0.1172, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0179, 0.0195, 0.0169, 0.0177, 0.0220, 0.0203, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 19:47:35,621 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6591, 4.8045, 5.0153, 4.7273, 4.8466, 5.3995, 4.8821, 4.5973], device='cuda:4'), covar=tensor([0.1170, 0.1879, 0.2048, 0.2060, 0.2647, 0.0942, 0.1486, 0.2494], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0608, 0.0665, 0.0503, 0.0665, 0.0697, 0.0523, 0.0663], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 19:47:56,651 INFO [train.py:904] (4/8) Epoch 24, batch 4650, loss[loss=0.1823, simple_loss=0.2708, pruned_loss=0.04689, over 16785.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2823, pruned_loss=0.05331, over 3193497.37 frames. ], batch size: 89, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:48:25,477 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.89 vs. limit=5.0 2023-05-01 19:48:29,662 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=238125.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:48:30,388 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 1.898e+02 2.163e+02 2.416e+02 3.900e+02, threshold=4.325e+02, percent-clipped=0.0 2023-05-01 19:48:30,793 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=238126.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:48:43,748 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=238135.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:48:59,281 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=238146.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:49:00,458 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=238147.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:49:09,298 INFO [train.py:904] (4/8) Epoch 24, batch 4700, loss[loss=0.1857, simple_loss=0.2786, pruned_loss=0.0464, over 15304.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2795, pruned_loss=0.0521, over 3205386.93 frames. ], batch size: 190, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:49:58,183 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=238187.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:50:06,564 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-05-01 19:50:09,159 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=238195.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:50:20,662 INFO [train.py:904] (4/8) Epoch 24, batch 4750, loss[loss=0.1461, simple_loss=0.2366, pruned_loss=0.02784, over 16504.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2751, pruned_loss=0.0499, over 3210778.65 frames. ], batch size: 75, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:50:41,769 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-05-01 19:50:43,586 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=238219.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 19:50:53,740 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.372e+02 1.823e+02 2.174e+02 2.542e+02 5.265e+02, threshold=4.348e+02, percent-clipped=2.0 2023-05-01 19:51:31,473 INFO [train.py:904] (4/8) Epoch 24, batch 4800, loss[loss=0.1762, simple_loss=0.2754, pruned_loss=0.03848, over 16885.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2712, pruned_loss=0.04782, over 3216213.94 frames. ], batch size: 96, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:51:49,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3048, 3.3366, 2.0019, 3.6869, 2.5021, 3.6604, 2.2446, 2.6703], device='cuda:4'), covar=tensor([0.0313, 0.0400, 0.1739, 0.0183, 0.0952, 0.0580, 0.1531, 0.0846], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0179, 0.0195, 0.0169, 0.0177, 0.0220, 0.0203, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 19:51:52,869 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=238267.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 19:51:54,667 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=238268.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:52:25,765 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0693, 2.1854, 2.2056, 3.7833, 2.1506, 2.5296, 2.3055, 2.3544], device='cuda:4'), covar=tensor([0.1521, 0.3769, 0.3097, 0.0592, 0.4205, 0.2644, 0.3819, 0.3305], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0460, 0.0375, 0.0331, 0.0440, 0.0528, 0.0429, 0.0537], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:52:46,482 INFO [train.py:904] (4/8) Epoch 24, batch 4850, loss[loss=0.2, simple_loss=0.2824, pruned_loss=0.05877, over 16682.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2716, pruned_loss=0.04699, over 3215092.66 frames. ], batch size: 62, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:52:53,326 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6094, 4.7704, 5.0016, 4.9343, 4.9979, 4.7529, 4.5879, 4.5066], device='cuda:4'), covar=tensor([0.0397, 0.0522, 0.0451, 0.0509, 0.0574, 0.0392, 0.1240, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0460, 0.0449, 0.0413, 0.0495, 0.0468, 0.0552, 0.0376], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 19:53:22,733 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.326e+02 1.927e+02 2.251e+02 2.681e+02 4.052e+02, threshold=4.502e+02, percent-clipped=0.0 2023-05-01 19:54:04,040 INFO [train.py:904] (4/8) Epoch 24, batch 4900, loss[loss=0.174, simple_loss=0.2737, pruned_loss=0.0372, over 15430.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2708, pruned_loss=0.04535, over 3215025.32 frames. ], batch size: 191, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:55:17,228 INFO [train.py:904] (4/8) Epoch 24, batch 4950, loss[loss=0.1977, simple_loss=0.2905, pruned_loss=0.05244, over 16772.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2708, pruned_loss=0.04503, over 3210993.78 frames. ], batch size: 124, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:55:41,592 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=238420.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:55:41,738 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=238420.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 19:55:49,724 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.467e+02 2.002e+02 2.407e+02 2.827e+02 4.284e+02, threshold=4.815e+02, percent-clipped=0.0 2023-05-01 19:55:57,438 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=238430.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:55:59,976 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7704, 4.8360, 4.6365, 4.2804, 4.2775, 4.7174, 4.5204, 4.4398], device='cuda:4'), covar=tensor([0.0548, 0.0435, 0.0281, 0.0319, 0.0996, 0.0468, 0.0508, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0444, 0.0348, 0.0349, 0.0351, 0.0403, 0.0238, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:56:20,235 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=238446.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:56:29,585 INFO [train.py:904] (4/8) Epoch 24, batch 5000, loss[loss=0.1727, simple_loss=0.2662, pruned_loss=0.03961, over 16793.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2726, pruned_loss=0.04494, over 3209394.13 frames. ], batch size: 89, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:56:35,200 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9829, 4.8283, 5.0254, 5.2028, 5.4098, 4.8727, 5.4160, 5.4347], device='cuda:4'), covar=tensor([0.1817, 0.1318, 0.1769, 0.0775, 0.0499, 0.0781, 0.0512, 0.0592], device='cuda:4'), in_proj_covar=tensor([0.0644, 0.0797, 0.0918, 0.0806, 0.0615, 0.0636, 0.0665, 0.0775], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 19:57:11,125 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=238481.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 19:57:11,982 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=238482.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:57:31,131 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=238494.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 19:57:42,730 INFO [train.py:904] (4/8) Epoch 24, batch 5050, loss[loss=0.1675, simple_loss=0.262, pruned_loss=0.03652, over 16917.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2733, pruned_loss=0.04501, over 3206364.41 frames. ], batch size: 109, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:58:15,454 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.348e+02 2.006e+02 2.316e+02 2.692e+02 7.292e+02, threshold=4.632e+02, percent-clipped=1.0 2023-05-01 19:58:53,423 INFO [train.py:904] (4/8) Epoch 24, batch 5100, loss[loss=0.1745, simple_loss=0.2626, pruned_loss=0.04317, over 16757.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2714, pruned_loss=0.04432, over 3212014.32 frames. ], batch size: 83, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 19:59:16,751 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=238568.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:00:07,249 INFO [train.py:904] (4/8) Epoch 24, batch 5150, loss[loss=0.1924, simple_loss=0.2883, pruned_loss=0.04824, over 16785.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2715, pruned_loss=0.044, over 3215892.85 frames. ], batch size: 124, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 20:00:26,118 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=238616.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:00:38,784 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.286e+02 1.999e+02 2.272e+02 2.615e+02 4.924e+02, threshold=4.544e+02, percent-clipped=1.0 2023-05-01 20:00:53,618 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4253, 3.5649, 3.7759, 2.0585, 3.1730, 2.5106, 3.8530, 3.8123], device='cuda:4'), covar=tensor([0.0232, 0.0804, 0.0577, 0.2064, 0.0815, 0.0916, 0.0507, 0.0830], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0166, 0.0168, 0.0154, 0.0145, 0.0130, 0.0143, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:01:17,679 INFO [train.py:904] (4/8) Epoch 24, batch 5200, loss[loss=0.1791, simple_loss=0.2571, pruned_loss=0.05057, over 16648.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2692, pruned_loss=0.0435, over 3213986.29 frames. ], batch size: 57, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 20:01:57,601 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=238681.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:02:28,267 INFO [train.py:904] (4/8) Epoch 24, batch 5250, loss[loss=0.1669, simple_loss=0.2571, pruned_loss=0.03837, over 16477.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2667, pruned_loss=0.04292, over 3216534.19 frames. ], batch size: 68, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 20:02:52,175 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=238720.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:02:52,800 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-01 20:03:01,284 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.388e+02 1.983e+02 2.265e+02 2.655e+02 9.176e+02, threshold=4.530e+02, percent-clipped=2.0 2023-05-01 20:03:07,626 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=238730.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:03:25,284 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=238742.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:03:39,613 INFO [train.py:904] (4/8) Epoch 24, batch 5300, loss[loss=0.1388, simple_loss=0.2252, pruned_loss=0.02626, over 16479.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2635, pruned_loss=0.04166, over 3212462.90 frames. ], batch size: 75, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 20:04:01,109 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=238768.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:04:12,300 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=238776.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 20:04:15,073 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=238778.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:04:20,295 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=238782.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:04:50,565 INFO [train.py:904] (4/8) Epoch 24, batch 5350, loss[loss=0.1706, simple_loss=0.2612, pruned_loss=0.04004, over 16625.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2622, pruned_loss=0.04117, over 3216314.93 frames. ], batch size: 68, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 20:04:56,382 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8802, 4.7784, 4.9689, 5.1288, 5.3151, 4.7606, 5.3330, 5.3243], device='cuda:4'), covar=tensor([0.2032, 0.1288, 0.1754, 0.0754, 0.0529, 0.0856, 0.0550, 0.0673], device='cuda:4'), in_proj_covar=tensor([0.0647, 0.0797, 0.0918, 0.0808, 0.0616, 0.0637, 0.0664, 0.0776], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:05:17,665 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-01 20:05:21,927 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.474e+02 1.926e+02 2.251e+02 2.610e+02 1.024e+03, threshold=4.501e+02, percent-clipped=4.0 2023-05-01 20:05:27,686 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=238830.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:06:01,797 INFO [train.py:904] (4/8) Epoch 24, batch 5400, loss[loss=0.1851, simple_loss=0.2743, pruned_loss=0.04799, over 16887.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2648, pruned_loss=0.04176, over 3217200.00 frames. ], batch size: 109, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 20:07:18,012 INFO [train.py:904] (4/8) Epoch 24, batch 5450, loss[loss=0.2377, simple_loss=0.3249, pruned_loss=0.0753, over 16285.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2684, pruned_loss=0.04346, over 3206522.95 frames. ], batch size: 165, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 20:07:54,284 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 2.137e+02 2.656e+02 3.537e+02 7.468e+02, threshold=5.312e+02, percent-clipped=13.0 2023-05-01 20:08:10,986 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=238936.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 20:08:18,448 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6986, 2.5667, 2.2937, 3.6842, 2.5286, 3.8353, 1.4829, 2.7467], device='cuda:4'), covar=tensor([0.1328, 0.0791, 0.1317, 0.0228, 0.0233, 0.0371, 0.1744, 0.0836], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0177, 0.0196, 0.0194, 0.0205, 0.0216, 0.0205, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:08:35,278 INFO [train.py:904] (4/8) Epoch 24, batch 5500, loss[loss=0.191, simple_loss=0.2871, pruned_loss=0.04739, over 17120.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2746, pruned_loss=0.04674, over 3200944.97 frames. ], batch size: 47, lr: 2.81e-03, grad_scale: 8.0 2023-05-01 20:08:44,593 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6465, 3.5555, 4.0467, 2.0611, 4.2353, 4.2120, 3.0920, 3.1717], device='cuda:4'), covar=tensor([0.0843, 0.0275, 0.0199, 0.1238, 0.0068, 0.0151, 0.0444, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0100, 0.0139, 0.0082, 0.0127, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:09:32,513 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8691, 2.8051, 2.8247, 2.1364, 2.7066, 2.2050, 2.7620, 2.9877], device='cuda:4'), covar=tensor([0.0271, 0.0675, 0.0485, 0.1628, 0.0742, 0.0871, 0.0528, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0165, 0.0167, 0.0154, 0.0145, 0.0130, 0.0143, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:09:43,429 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=238997.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 20:09:53,111 INFO [train.py:904] (4/8) Epoch 24, batch 5550, loss[loss=0.2605, simple_loss=0.324, pruned_loss=0.0985, over 10871.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2814, pruned_loss=0.05157, over 3174013.91 frames. ], batch size: 246, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:10:30,505 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.205e+02 3.032e+02 3.628e+02 4.510e+02 7.943e+02, threshold=7.255e+02, percent-clipped=11.0 2023-05-01 20:10:48,340 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=239037.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:11:12,507 INFO [train.py:904] (4/8) Epoch 24, batch 5600, loss[loss=0.2209, simple_loss=0.3017, pruned_loss=0.06998, over 16697.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2862, pruned_loss=0.05576, over 3132863.45 frames. ], batch size: 134, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:11:21,219 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3955, 2.9659, 2.7202, 2.3470, 2.3638, 2.3234, 2.9998, 2.9575], device='cuda:4'), covar=tensor([0.2084, 0.0572, 0.1443, 0.2486, 0.2137, 0.2102, 0.0470, 0.1105], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0272, 0.0307, 0.0319, 0.0301, 0.0265, 0.0299, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 20:11:52,321 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=239076.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 20:12:02,257 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-05-01 20:12:14,908 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6126, 2.5062, 1.9226, 2.6722, 2.1361, 2.7291, 2.1512, 2.3827], device='cuda:4'), covar=tensor([0.0314, 0.0398, 0.1149, 0.0254, 0.0642, 0.0437, 0.1236, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0178, 0.0194, 0.0167, 0.0178, 0.0218, 0.0203, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:12:19,677 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3707, 3.0139, 2.8958, 1.9508, 2.6647, 2.0921, 3.0190, 3.2652], device='cuda:4'), covar=tensor([0.0354, 0.0720, 0.0736, 0.2109, 0.0983, 0.1089, 0.0687, 0.0747], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0165, 0.0167, 0.0153, 0.0145, 0.0130, 0.0143, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:12:36,912 INFO [train.py:904] (4/8) Epoch 24, batch 5650, loss[loss=0.2029, simple_loss=0.2973, pruned_loss=0.05429, over 16896.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2911, pruned_loss=0.05954, over 3112794.28 frames. ], batch size: 96, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:12:56,811 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2759, 3.4341, 3.5734, 3.5374, 3.5623, 3.3939, 3.4242, 3.4781], device='cuda:4'), covar=tensor([0.0434, 0.0783, 0.0493, 0.0479, 0.0568, 0.0600, 0.0833, 0.0583], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0467, 0.0453, 0.0415, 0.0499, 0.0473, 0.0557, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 20:12:59,844 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2536, 4.3038, 4.6055, 4.5535, 4.5852, 4.3129, 4.3008, 4.2703], device='cuda:4'), covar=tensor([0.0355, 0.0583, 0.0375, 0.0433, 0.0476, 0.0398, 0.0891, 0.0476], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0467, 0.0453, 0.0415, 0.0499, 0.0473, 0.0556, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 20:13:11,109 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=239124.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 20:13:14,765 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.394e+02 3.339e+02 4.323e+02 5.614e+02 1.229e+03, threshold=8.646e+02, percent-clipped=9.0 2023-05-01 20:13:55,697 INFO [train.py:904] (4/8) Epoch 24, batch 5700, loss[loss=0.2766, simple_loss=0.3267, pruned_loss=0.1133, over 11231.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2933, pruned_loss=0.06199, over 3081157.55 frames. ], batch size: 248, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:13:59,334 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0641, 2.4515, 2.2889, 2.9014, 1.9454, 3.1478, 1.7961, 2.7312], device='cuda:4'), covar=tensor([0.1164, 0.0574, 0.1122, 0.0203, 0.0128, 0.0391, 0.1561, 0.0743], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0177, 0.0196, 0.0194, 0.0205, 0.0216, 0.0204, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:14:02,398 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.5976, 2.6561, 2.3214, 3.8788, 2.7014, 3.7527, 1.3874, 2.6924], device='cuda:4'), covar=tensor([0.1483, 0.0789, 0.1394, 0.0199, 0.0283, 0.0477, 0.1910, 0.0979], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0177, 0.0196, 0.0194, 0.0205, 0.0216, 0.0204, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:15:14,246 INFO [train.py:904] (4/8) Epoch 24, batch 5750, loss[loss=0.1912, simple_loss=0.2862, pruned_loss=0.04809, over 16900.00 frames. ], tot_loss[loss=0.2115, simple_loss=0.2962, pruned_loss=0.0634, over 3066194.38 frames. ], batch size: 96, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:15:46,060 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3397, 3.3323, 3.7940, 1.7455, 3.9367, 3.9934, 2.9697, 2.8807], device='cuda:4'), covar=tensor([0.0974, 0.0310, 0.0239, 0.1392, 0.0099, 0.0172, 0.0464, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0100, 0.0139, 0.0082, 0.0128, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:15:53,491 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.857e+02 3.000e+02 3.580e+02 4.341e+02 9.766e+02, threshold=7.159e+02, percent-clipped=1.0 2023-05-01 20:15:58,019 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9769, 5.2640, 5.0493, 5.0600, 4.8222, 4.6946, 4.6626, 5.3558], device='cuda:4'), covar=tensor([0.1228, 0.0822, 0.0947, 0.0879, 0.0825, 0.0955, 0.1233, 0.0794], device='cuda:4'), in_proj_covar=tensor([0.0688, 0.0826, 0.0689, 0.0641, 0.0529, 0.0531, 0.0696, 0.0648], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:16:24,182 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-05-01 20:16:37,341 INFO [train.py:904] (4/8) Epoch 24, batch 5800, loss[loss=0.1784, simple_loss=0.2745, pruned_loss=0.04115, over 16337.00 frames. ], tot_loss[loss=0.211, simple_loss=0.2962, pruned_loss=0.0629, over 3044136.19 frames. ], batch size: 146, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:17:40,222 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=239292.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 20:17:56,792 INFO [train.py:904] (4/8) Epoch 24, batch 5850, loss[loss=0.2032, simple_loss=0.2966, pruned_loss=0.05488, over 16669.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2935, pruned_loss=0.06104, over 3043480.86 frames. ], batch size: 89, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:18:33,898 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.882e+02 2.891e+02 3.438e+02 4.490e+02 7.448e+02, threshold=6.875e+02, percent-clipped=1.0 2023-05-01 20:18:53,008 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=239337.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:19:19,097 INFO [train.py:904] (4/8) Epoch 24, batch 5900, loss[loss=0.2152, simple_loss=0.2975, pruned_loss=0.06644, over 15283.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2926, pruned_loss=0.06008, over 3077401.79 frames. ], batch size: 190, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:20:00,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3911, 3.2367, 3.5375, 1.7195, 3.7033, 3.7906, 2.9163, 2.7723], device='cuda:4'), covar=tensor([0.0896, 0.0314, 0.0258, 0.1444, 0.0127, 0.0233, 0.0488, 0.0511], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0109, 0.0100, 0.0139, 0.0083, 0.0128, 0.0129, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:20:14,707 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=239385.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:20:42,777 INFO [train.py:904] (4/8) Epoch 24, batch 5950, loss[loss=0.2058, simple_loss=0.2911, pruned_loss=0.06028, over 15286.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2939, pruned_loss=0.05904, over 3091135.55 frames. ], batch size: 190, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:21:21,253 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.785e+02 2.477e+02 3.014e+02 3.724e+02 7.833e+02, threshold=6.029e+02, percent-clipped=2.0 2023-05-01 20:22:03,333 INFO [train.py:904] (4/8) Epoch 24, batch 6000, loss[loss=0.1918, simple_loss=0.2781, pruned_loss=0.05269, over 16741.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2921, pruned_loss=0.05796, over 3097179.76 frames. ], batch size: 124, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:22:03,333 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 20:22:14,269 INFO [train.py:938] (4/8) Epoch 24, validation: loss=0.1493, simple_loss=0.2618, pruned_loss=0.01837, over 944034.00 frames. 2023-05-01 20:22:14,269 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 20:22:14,566 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3354, 4.3974, 4.7383, 4.6737, 4.7041, 4.4087, 4.4072, 4.2977], device='cuda:4'), covar=tensor([0.0372, 0.0573, 0.0396, 0.0429, 0.0498, 0.0455, 0.0952, 0.0547], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0468, 0.0455, 0.0416, 0.0501, 0.0474, 0.0557, 0.0379], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 20:22:50,571 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9909, 2.1077, 2.2586, 3.4450, 2.0191, 2.4354, 2.2102, 2.2384], device='cuda:4'), covar=tensor([0.1416, 0.3586, 0.2895, 0.0709, 0.4208, 0.2377, 0.3558, 0.3360], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0455, 0.0372, 0.0328, 0.0436, 0.0521, 0.0426, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:23:32,267 INFO [train.py:904] (4/8) Epoch 24, batch 6050, loss[loss=0.1898, simple_loss=0.2843, pruned_loss=0.04769, over 16907.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2907, pruned_loss=0.0577, over 3106569.09 frames. ], batch size: 116, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:23:59,221 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4927, 3.3701, 2.6844, 2.2027, 2.3011, 2.2997, 3.5256, 3.1134], device='cuda:4'), covar=tensor([0.2968, 0.0735, 0.1796, 0.2881, 0.2669, 0.2231, 0.0520, 0.1382], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0272, 0.0308, 0.0320, 0.0301, 0.0266, 0.0300, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 20:24:09,878 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.866e+02 2.713e+02 3.363e+02 4.065e+02 7.660e+02, threshold=6.725e+02, percent-clipped=3.0 2023-05-01 20:24:51,433 INFO [train.py:904] (4/8) Epoch 24, batch 6100, loss[loss=0.1676, simple_loss=0.2594, pruned_loss=0.03787, over 17017.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2896, pruned_loss=0.05632, over 3139031.85 frames. ], batch size: 41, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:25:28,698 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1351, 1.6478, 1.9874, 2.0754, 2.2510, 2.4011, 1.7433, 2.3246], device='cuda:4'), covar=tensor([0.0262, 0.0478, 0.0311, 0.0407, 0.0317, 0.0217, 0.0518, 0.0156], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0194, 0.0183, 0.0185, 0.0202, 0.0160, 0.0200, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:25:30,217 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3769, 3.3363, 3.3905, 3.4778, 3.5035, 3.2807, 3.5008, 3.5527], device='cuda:4'), covar=tensor([0.1230, 0.0980, 0.1088, 0.0618, 0.0657, 0.2270, 0.1097, 0.0884], device='cuda:4'), in_proj_covar=tensor([0.0643, 0.0791, 0.0913, 0.0801, 0.0611, 0.0632, 0.0664, 0.0771], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:25:38,111 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4004, 2.9142, 3.0535, 1.9950, 2.7697, 2.0722, 3.0664, 3.1572], device='cuda:4'), covar=tensor([0.0285, 0.0759, 0.0657, 0.2042, 0.0901, 0.1061, 0.0672, 0.0829], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0166, 0.0169, 0.0155, 0.0147, 0.0131, 0.0144, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:25:49,806 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8917, 2.1654, 2.4539, 3.1054, 2.2241, 2.3618, 2.3527, 2.2908], device='cuda:4'), covar=tensor([0.1346, 0.3213, 0.2484, 0.0742, 0.3985, 0.2431, 0.3025, 0.3095], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0455, 0.0372, 0.0328, 0.0436, 0.0522, 0.0426, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:25:56,112 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=239592.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 20:26:14,091 INFO [train.py:904] (4/8) Epoch 24, batch 6150, loss[loss=0.2001, simple_loss=0.275, pruned_loss=0.06261, over 11606.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2883, pruned_loss=0.05622, over 3113002.55 frames. ], batch size: 248, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:26:40,701 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-05-01 20:26:45,894 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=239622.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:26:53,236 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.890e+02 2.573e+02 3.097e+02 3.625e+02 7.083e+02, threshold=6.193e+02, percent-clipped=1.0 2023-05-01 20:27:14,517 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=239640.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 20:27:24,230 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4458, 2.5700, 2.5264, 4.2638, 2.3491, 2.8018, 2.5498, 2.6794], device='cuda:4'), covar=tensor([0.1281, 0.3114, 0.2812, 0.0497, 0.3961, 0.2437, 0.3165, 0.3004], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0456, 0.0373, 0.0329, 0.0437, 0.0522, 0.0427, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:27:34,956 INFO [train.py:904] (4/8) Epoch 24, batch 6200, loss[loss=0.1773, simple_loss=0.2622, pruned_loss=0.04616, over 16635.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2864, pruned_loss=0.05589, over 3110459.68 frames. ], batch size: 62, lr: 2.80e-03, grad_scale: 4.0 2023-05-01 20:27:49,707 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9192, 2.4260, 2.0227, 2.1373, 2.7144, 2.3957, 2.6353, 2.9162], device='cuda:4'), covar=tensor([0.0209, 0.0435, 0.0595, 0.0531, 0.0263, 0.0413, 0.0235, 0.0267], device='cuda:4'), in_proj_covar=tensor([0.0219, 0.0238, 0.0230, 0.0231, 0.0240, 0.0237, 0.0238, 0.0236], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:28:09,708 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-05-01 20:28:11,741 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6996, 3.7468, 2.9229, 2.2671, 2.4828, 2.4687, 4.0775, 3.3637], device='cuda:4'), covar=tensor([0.2894, 0.0662, 0.1799, 0.2765, 0.2632, 0.2115, 0.0446, 0.1249], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0273, 0.0308, 0.0320, 0.0301, 0.0266, 0.0300, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 20:28:22,558 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=239683.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:28:52,025 INFO [train.py:904] (4/8) Epoch 24, batch 6250, loss[loss=0.1957, simple_loss=0.2881, pruned_loss=0.05164, over 16597.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2858, pruned_loss=0.05552, over 3122968.74 frames. ], batch size: 62, lr: 2.80e-03, grad_scale: 4.0 2023-05-01 20:29:29,984 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.754e+02 2.614e+02 3.057e+02 3.659e+02 7.500e+02, threshold=6.114e+02, percent-clipped=2.0 2023-05-01 20:29:52,270 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-01 20:30:05,922 INFO [train.py:904] (4/8) Epoch 24, batch 6300, loss[loss=0.1955, simple_loss=0.291, pruned_loss=0.04999, over 16712.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2857, pruned_loss=0.05476, over 3147727.61 frames. ], batch size: 76, lr: 2.80e-03, grad_scale: 4.0 2023-05-01 20:31:24,178 INFO [train.py:904] (4/8) Epoch 24, batch 6350, loss[loss=0.2015, simple_loss=0.286, pruned_loss=0.0585, over 16940.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2862, pruned_loss=0.05571, over 3139299.92 frames. ], batch size: 109, lr: 2.80e-03, grad_scale: 4.0 2023-05-01 20:32:03,927 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.770e+02 2.790e+02 3.396e+02 4.270e+02 1.026e+03, threshold=6.791e+02, percent-clipped=1.0 2023-05-01 20:32:42,135 INFO [train.py:904] (4/8) Epoch 24, batch 6400, loss[loss=0.2205, simple_loss=0.3014, pruned_loss=0.06981, over 16869.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2871, pruned_loss=0.05748, over 3118936.59 frames. ], batch size: 116, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:33:58,190 INFO [train.py:904] (4/8) Epoch 24, batch 6450, loss[loss=0.1849, simple_loss=0.2725, pruned_loss=0.0486, over 16628.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2869, pruned_loss=0.05633, over 3123787.83 frames. ], batch size: 134, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:34:21,785 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4642, 2.2487, 3.0257, 3.2725, 3.1841, 3.8640, 2.5486, 3.8696], device='cuda:4'), covar=tensor([0.0192, 0.0522, 0.0318, 0.0316, 0.0297, 0.0143, 0.0513, 0.0115], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0193, 0.0181, 0.0184, 0.0201, 0.0159, 0.0199, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:34:37,432 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.694e+02 2.672e+02 3.087e+02 3.802e+02 7.341e+02, threshold=6.174e+02, percent-clipped=2.0 2023-05-01 20:35:16,098 INFO [train.py:904] (4/8) Epoch 24, batch 6500, loss[loss=0.1989, simple_loss=0.2866, pruned_loss=0.0556, over 16724.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2851, pruned_loss=0.05569, over 3139860.32 frames. ], batch size: 83, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:35:55,172 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=239978.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:36:27,587 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-05-01 20:36:39,172 INFO [train.py:904] (4/8) Epoch 24, batch 6550, loss[loss=0.1883, simple_loss=0.2857, pruned_loss=0.04543, over 16616.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2879, pruned_loss=0.05647, over 3129449.18 frames. ], batch size: 134, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:37:16,974 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.470e+02 2.521e+02 3.197e+02 3.720e+02 9.360e+02, threshold=6.395e+02, percent-clipped=2.0 2023-05-01 20:37:36,952 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6480, 2.5413, 1.9312, 2.6789, 2.1490, 2.7596, 2.1593, 2.3964], device='cuda:4'), covar=tensor([0.0324, 0.0344, 0.1176, 0.0279, 0.0615, 0.0467, 0.1121, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0178, 0.0195, 0.0167, 0.0177, 0.0217, 0.0203, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:37:54,309 INFO [train.py:904] (4/8) Epoch 24, batch 6600, loss[loss=0.1859, simple_loss=0.2797, pruned_loss=0.04605, over 16904.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2906, pruned_loss=0.05718, over 3135987.24 frames. ], batch size: 96, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:38:18,431 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5739, 4.5497, 4.4364, 3.6839, 4.4971, 1.6153, 4.2321, 4.0535], device='cuda:4'), covar=tensor([0.0121, 0.0108, 0.0216, 0.0359, 0.0103, 0.2918, 0.0143, 0.0251], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0165, 0.0207, 0.0182, 0.0181, 0.0212, 0.0193, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:39:11,701 INFO [train.py:904] (4/8) Epoch 24, batch 6650, loss[loss=0.2393, simple_loss=0.3087, pruned_loss=0.08493, over 11223.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.291, pruned_loss=0.05832, over 3112428.23 frames. ], batch size: 247, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:39:50,364 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.847e+02 2.925e+02 3.532e+02 4.365e+02 9.170e+02, threshold=7.065e+02, percent-clipped=1.0 2023-05-01 20:40:28,845 INFO [train.py:904] (4/8) Epoch 24, batch 6700, loss[loss=0.2173, simple_loss=0.3019, pruned_loss=0.06639, over 16762.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2886, pruned_loss=0.05784, over 3113214.99 frames. ], batch size: 124, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:41:45,713 INFO [train.py:904] (4/8) Epoch 24, batch 6750, loss[loss=0.1896, simple_loss=0.2785, pruned_loss=0.05033, over 15373.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.2893, pruned_loss=0.0591, over 3077353.42 frames. ], batch size: 190, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:42:23,537 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.933e+02 2.874e+02 3.326e+02 4.186e+02 6.587e+02, threshold=6.652e+02, percent-clipped=0.0 2023-05-01 20:43:01,407 INFO [train.py:904] (4/8) Epoch 24, batch 6800, loss[loss=0.1932, simple_loss=0.2838, pruned_loss=0.05132, over 16597.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2895, pruned_loss=0.05925, over 3074734.81 frames. ], batch size: 68, lr: 2.80e-03, grad_scale: 8.0 2023-05-01 20:43:33,064 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.97 vs. limit=5.0 2023-05-01 20:43:42,194 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=240278.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:44:06,644 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5802, 4.5316, 4.4310, 3.6038, 4.4942, 1.6644, 4.1879, 4.0477], device='cuda:4'), covar=tensor([0.0149, 0.0143, 0.0216, 0.0440, 0.0139, 0.2978, 0.0221, 0.0281], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0165, 0.0207, 0.0183, 0.0181, 0.0212, 0.0194, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:44:07,868 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0314, 2.3276, 2.3619, 2.7900, 2.0296, 3.1701, 1.8527, 2.7035], device='cuda:4'), covar=tensor([0.1182, 0.0636, 0.1081, 0.0196, 0.0129, 0.0372, 0.1448, 0.0713], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0176, 0.0195, 0.0193, 0.0205, 0.0216, 0.0203, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:44:21,118 INFO [train.py:904] (4/8) Epoch 24, batch 6850, loss[loss=0.2286, simple_loss=0.3016, pruned_loss=0.07776, over 11964.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2898, pruned_loss=0.05941, over 3068475.86 frames. ], batch size: 250, lr: 2.80e-03, grad_scale: 4.0 2023-05-01 20:44:56,066 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=240326.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:45:00,027 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.493e+02 2.623e+02 3.086e+02 3.760e+02 7.465e+02, threshold=6.171e+02, percent-clipped=2.0 2023-05-01 20:45:17,477 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=240340.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 20:45:35,447 INFO [train.py:904] (4/8) Epoch 24, batch 6900, loss[loss=0.2533, simple_loss=0.3258, pruned_loss=0.09041, over 11392.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.292, pruned_loss=0.05853, over 3083813.13 frames. ], batch size: 249, lr: 2.80e-03, grad_scale: 4.0 2023-05-01 20:46:26,529 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-05-01 20:46:33,684 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.62 vs. limit=5.0 2023-05-01 20:46:50,841 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=240401.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 20:46:52,761 INFO [train.py:904] (4/8) Epoch 24, batch 6950, loss[loss=0.2149, simple_loss=0.2979, pruned_loss=0.06594, over 16613.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2928, pruned_loss=0.05905, over 3099560.66 frames. ], batch size: 62, lr: 2.80e-03, grad_scale: 2.0 2023-05-01 20:47:33,339 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.139e+02 2.974e+02 3.564e+02 4.384e+02 8.202e+02, threshold=7.128e+02, percent-clipped=2.0 2023-05-01 20:48:07,678 INFO [train.py:904] (4/8) Epoch 24, batch 7000, loss[loss=0.2232, simple_loss=0.318, pruned_loss=0.06419, over 16221.00 frames. ], tot_loss[loss=0.205, simple_loss=0.2934, pruned_loss=0.0583, over 3116629.53 frames. ], batch size: 165, lr: 2.80e-03, grad_scale: 2.0 2023-05-01 20:48:31,505 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9540, 4.0207, 4.2956, 4.2625, 4.2833, 4.0417, 4.0237, 3.9865], device='cuda:4'), covar=tensor([0.0385, 0.0676, 0.0461, 0.0473, 0.0523, 0.0507, 0.0976, 0.0593], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0464, 0.0451, 0.0416, 0.0497, 0.0472, 0.0555, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 20:48:49,499 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1502, 2.8636, 3.1239, 1.8379, 3.2572, 3.2958, 2.6542, 2.5615], device='cuda:4'), covar=tensor([0.0865, 0.0301, 0.0239, 0.1154, 0.0110, 0.0214, 0.0533, 0.0474], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0109, 0.0100, 0.0138, 0.0082, 0.0128, 0.0129, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 20:49:13,686 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8774, 2.1264, 2.3190, 3.3876, 2.0354, 2.3500, 2.2369, 2.2406], device='cuda:4'), covar=tensor([0.1533, 0.3471, 0.3000, 0.0692, 0.4432, 0.2545, 0.3679, 0.3604], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0457, 0.0375, 0.0330, 0.0440, 0.0523, 0.0429, 0.0535], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:49:21,394 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2234, 5.2017, 4.9491, 4.2554, 5.1222, 1.7854, 4.8313, 4.6509], device='cuda:4'), covar=tensor([0.0110, 0.0084, 0.0221, 0.0449, 0.0095, 0.2939, 0.0135, 0.0279], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0165, 0.0207, 0.0183, 0.0181, 0.0213, 0.0194, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:49:23,902 INFO [train.py:904] (4/8) Epoch 24, batch 7050, loss[loss=0.213, simple_loss=0.288, pruned_loss=0.06903, over 11505.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2935, pruned_loss=0.05801, over 3128423.85 frames. ], batch size: 247, lr: 2.80e-03, grad_scale: 2.0 2023-05-01 20:49:24,937 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3842, 3.4821, 3.6461, 3.6106, 3.6329, 3.4586, 3.4928, 3.5249], device='cuda:4'), covar=tensor([0.0425, 0.0696, 0.0425, 0.0438, 0.0550, 0.0540, 0.0824, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0465, 0.0451, 0.0416, 0.0498, 0.0473, 0.0555, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 20:50:06,656 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.746e+02 2.536e+02 3.115e+02 4.008e+02 6.614e+02, threshold=6.229e+02, percent-clipped=0.0 2023-05-01 20:50:42,263 INFO [train.py:904] (4/8) Epoch 24, batch 7100, loss[loss=0.2436, simple_loss=0.3021, pruned_loss=0.0925, over 11280.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2924, pruned_loss=0.05808, over 3096158.22 frames. ], batch size: 246, lr: 2.80e-03, grad_scale: 2.0 2023-05-01 20:51:59,028 INFO [train.py:904] (4/8) Epoch 24, batch 7150, loss[loss=0.2045, simple_loss=0.284, pruned_loss=0.06251, over 15343.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2909, pruned_loss=0.05867, over 3070788.48 frames. ], batch size: 190, lr: 2.80e-03, grad_scale: 2.0 2023-05-01 20:52:39,163 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.962e+02 2.941e+02 3.518e+02 4.081e+02 6.999e+02, threshold=7.036e+02, percent-clipped=1.0 2023-05-01 20:52:49,484 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-01 20:53:12,641 INFO [train.py:904] (4/8) Epoch 24, batch 7200, loss[loss=0.1788, simple_loss=0.2795, pruned_loss=0.03907, over 16532.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2893, pruned_loss=0.05756, over 3055569.07 frames. ], batch size: 62, lr: 2.80e-03, grad_scale: 4.0 2023-05-01 20:54:20,528 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=240696.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 20:54:32,007 INFO [train.py:904] (4/8) Epoch 24, batch 7250, loss[loss=0.2052, simple_loss=0.2791, pruned_loss=0.06558, over 11396.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2868, pruned_loss=0.05637, over 3065316.52 frames. ], batch size: 246, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 20:54:44,339 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-05-01 20:55:12,201 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.572e+02 2.642e+02 3.264e+02 4.366e+02 7.559e+02, threshold=6.528e+02, percent-clipped=1.0 2023-05-01 20:55:45,057 INFO [train.py:904] (4/8) Epoch 24, batch 7300, loss[loss=0.2318, simple_loss=0.2955, pruned_loss=0.084, over 11506.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2867, pruned_loss=0.05661, over 3069575.35 frames. ], batch size: 248, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 20:56:03,608 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2960, 3.2131, 2.5857, 2.1119, 2.1937, 2.2217, 3.3590, 2.9265], device='cuda:4'), covar=tensor([0.3328, 0.0759, 0.2045, 0.2901, 0.3050, 0.2387, 0.0572, 0.1498], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0272, 0.0308, 0.0321, 0.0301, 0.0266, 0.0300, 0.0343], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 20:56:58,479 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4633, 3.4288, 3.4105, 2.6236, 3.3323, 2.1014, 3.0895, 2.7410], device='cuda:4'), covar=tensor([0.0135, 0.0119, 0.0191, 0.0215, 0.0100, 0.2271, 0.0132, 0.0259], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0163, 0.0204, 0.0180, 0.0178, 0.0209, 0.0191, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:57:02,841 INFO [train.py:904] (4/8) Epoch 24, batch 7350, loss[loss=0.2076, simple_loss=0.2901, pruned_loss=0.06256, over 16295.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2877, pruned_loss=0.0576, over 3053124.85 frames. ], batch size: 165, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 20:57:44,658 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.796e+02 2.735e+02 3.201e+02 4.098e+02 9.519e+02, threshold=6.402e+02, percent-clipped=6.0 2023-05-01 20:58:18,665 INFO [train.py:904] (4/8) Epoch 24, batch 7400, loss[loss=0.2298, simple_loss=0.3013, pruned_loss=0.07914, over 10972.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2884, pruned_loss=0.05818, over 3044631.74 frames. ], batch size: 248, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 20:58:29,794 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5259, 3.5828, 3.3592, 2.9613, 3.1728, 3.4892, 3.3309, 3.3372], device='cuda:4'), covar=tensor([0.0595, 0.0711, 0.0284, 0.0267, 0.0521, 0.0475, 0.1447, 0.0509], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0433, 0.0338, 0.0338, 0.0342, 0.0391, 0.0233, 0.0407], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:58:54,982 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3552, 2.6149, 2.1255, 2.3513, 2.9364, 2.5808, 2.9602, 3.0986], device='cuda:4'), covar=tensor([0.0156, 0.0424, 0.0581, 0.0497, 0.0287, 0.0410, 0.0248, 0.0291], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0232, 0.0225, 0.0226, 0.0234, 0.0232, 0.0232, 0.0230], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 20:59:34,863 INFO [train.py:904] (4/8) Epoch 24, batch 7450, loss[loss=0.2299, simple_loss=0.3, pruned_loss=0.07987, over 11473.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2894, pruned_loss=0.05889, over 3049661.11 frames. ], batch size: 248, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:00:19,560 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.977e+02 2.843e+02 3.471e+02 4.367e+02 9.441e+02, threshold=6.943e+02, percent-clipped=5.0 2023-05-01 21:00:22,549 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=240932.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 21:00:55,587 INFO [train.py:904] (4/8) Epoch 24, batch 7500, loss[loss=0.1638, simple_loss=0.2498, pruned_loss=0.03886, over 17116.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2892, pruned_loss=0.05802, over 3045160.15 frames. ], batch size: 47, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:01:35,391 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1044, 2.1406, 2.2986, 3.7291, 2.0595, 2.4379, 2.2429, 2.3118], device='cuda:4'), covar=tensor([0.1397, 0.3698, 0.2858, 0.0597, 0.4486, 0.2630, 0.3619, 0.3470], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0458, 0.0373, 0.0329, 0.0439, 0.0523, 0.0429, 0.0533], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:01:45,853 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=240986.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:01:57,390 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=240993.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 21:02:00,924 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=240996.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 21:02:11,647 INFO [train.py:904] (4/8) Epoch 24, batch 7550, loss[loss=0.1905, simple_loss=0.2781, pruned_loss=0.05147, over 16545.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2891, pruned_loss=0.05927, over 3005909.24 frames. ], batch size: 75, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:02:53,683 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.791e+02 2.697e+02 3.200e+02 3.848e+02 7.597e+02, threshold=6.400e+02, percent-clipped=1.0 2023-05-01 21:03:15,801 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=241044.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:03:20,286 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=241047.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:03:29,244 INFO [train.py:904] (4/8) Epoch 24, batch 7600, loss[loss=0.1983, simple_loss=0.2873, pruned_loss=0.05462, over 16436.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.2881, pruned_loss=0.0588, over 3028483.82 frames. ], batch size: 146, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:03:48,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7887, 2.5162, 2.2484, 3.2779, 2.2022, 3.5514, 1.4290, 2.7572], device='cuda:4'), covar=tensor([0.1392, 0.0754, 0.1350, 0.0212, 0.0193, 0.0405, 0.1852, 0.0807], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0177, 0.0197, 0.0194, 0.0206, 0.0216, 0.0205, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 21:04:22,309 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3394, 3.2649, 2.6242, 2.1309, 2.2395, 2.2737, 3.3707, 3.0057], device='cuda:4'), covar=tensor([0.3109, 0.0741, 0.1834, 0.2829, 0.2633, 0.2245, 0.0545, 0.1375], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0270, 0.0306, 0.0318, 0.0299, 0.0265, 0.0298, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 21:04:47,281 INFO [train.py:904] (4/8) Epoch 24, batch 7650, loss[loss=0.2639, simple_loss=0.3243, pruned_loss=0.1018, over 11190.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.288, pruned_loss=0.05856, over 3055045.53 frames. ], batch size: 247, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:05:30,616 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.802e+02 2.751e+02 3.488e+02 4.821e+02 1.001e+03, threshold=6.976e+02, percent-clipped=5.0 2023-05-01 21:06:04,717 INFO [train.py:904] (4/8) Epoch 24, batch 7700, loss[loss=0.2191, simple_loss=0.3041, pruned_loss=0.06708, over 16746.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2875, pruned_loss=0.05846, over 3049130.46 frames. ], batch size: 124, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:07:23,957 INFO [train.py:904] (4/8) Epoch 24, batch 7750, loss[loss=0.2181, simple_loss=0.3059, pruned_loss=0.06512, over 15313.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2881, pruned_loss=0.05871, over 3042915.72 frames. ], batch size: 190, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:08:06,386 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.032e+02 2.844e+02 3.408e+02 3.883e+02 6.664e+02, threshold=6.815e+02, percent-clipped=0.0 2023-05-01 21:08:38,516 INFO [train.py:904] (4/8) Epoch 24, batch 7800, loss[loss=0.1833, simple_loss=0.2787, pruned_loss=0.04393, over 16398.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.2888, pruned_loss=0.05934, over 3041028.69 frames. ], batch size: 35, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:08:53,129 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8831, 2.8734, 2.4813, 2.7891, 3.2370, 2.9645, 3.4260, 3.4994], device='cuda:4'), covar=tensor([0.0137, 0.0428, 0.0546, 0.0421, 0.0303, 0.0374, 0.0256, 0.0257], device='cuda:4'), in_proj_covar=tensor([0.0213, 0.0234, 0.0226, 0.0227, 0.0236, 0.0233, 0.0234, 0.0231], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:09:31,605 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.20 vs. limit=5.0 2023-05-01 21:09:34,096 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=241288.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 21:09:56,310 INFO [train.py:904] (4/8) Epoch 24, batch 7850, loss[loss=0.2295, simple_loss=0.3078, pruned_loss=0.07558, over 11737.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.2894, pruned_loss=0.05819, over 3072689.71 frames. ], batch size: 248, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:10:38,474 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.020e+02 2.608e+02 3.095e+02 3.566e+02 5.691e+02, threshold=6.191e+02, percent-clipped=0.0 2023-05-01 21:10:55,477 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=241342.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:10:59,471 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5238, 3.4692, 3.4288, 2.6284, 3.3523, 2.0108, 3.1656, 2.7541], device='cuda:4'), covar=tensor([0.0208, 0.0170, 0.0251, 0.0335, 0.0140, 0.2839, 0.0180, 0.0345], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0163, 0.0205, 0.0181, 0.0179, 0.0210, 0.0191, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:11:11,149 INFO [train.py:904] (4/8) Epoch 24, batch 7900, loss[loss=0.2064, simple_loss=0.2953, pruned_loss=0.0588, over 15270.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2884, pruned_loss=0.05746, over 3092020.38 frames. ], batch size: 190, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:12:24,156 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6750, 2.3971, 2.0540, 2.2466, 2.6478, 2.3786, 2.5226, 2.8405], device='cuda:4'), covar=tensor([0.0215, 0.0363, 0.0491, 0.0400, 0.0265, 0.0341, 0.0234, 0.0228], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0231, 0.0223, 0.0224, 0.0233, 0.0230, 0.0231, 0.0228], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:12:29,005 INFO [train.py:904] (4/8) Epoch 24, batch 7950, loss[loss=0.201, simple_loss=0.2852, pruned_loss=0.05842, over 16122.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2881, pruned_loss=0.05767, over 3082273.70 frames. ], batch size: 35, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:12:35,100 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=241406.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:13:12,618 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.937e+02 2.663e+02 3.306e+02 3.929e+02 7.099e+02, threshold=6.611e+02, percent-clipped=2.0 2023-05-01 21:13:16,041 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.73 vs. limit=5.0 2023-05-01 21:13:46,678 INFO [train.py:904] (4/8) Epoch 24, batch 8000, loss[loss=0.1896, simple_loss=0.2769, pruned_loss=0.05119, over 17034.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2888, pruned_loss=0.05838, over 3077283.42 frames. ], batch size: 53, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:14:09,552 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=241467.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 21:14:41,749 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.39 vs. limit=5.0 2023-05-01 21:14:44,995 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1007, 3.3521, 3.3614, 2.2247, 3.1414, 3.3998, 3.1686, 1.9982], device='cuda:4'), covar=tensor([0.0561, 0.0077, 0.0075, 0.0443, 0.0123, 0.0122, 0.0117, 0.0499], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0086, 0.0087, 0.0135, 0.0099, 0.0111, 0.0096, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 21:15:04,297 INFO [train.py:904] (4/8) Epoch 24, batch 8050, loss[loss=0.2062, simple_loss=0.2911, pruned_loss=0.06066, over 16388.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.2891, pruned_loss=0.05852, over 3065993.22 frames. ], batch size: 68, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:15:47,626 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.925e+02 2.671e+02 3.058e+02 3.888e+02 6.821e+02, threshold=6.115e+02, percent-clipped=1.0 2023-05-01 21:16:22,137 INFO [train.py:904] (4/8) Epoch 24, batch 8100, loss[loss=0.2316, simple_loss=0.3, pruned_loss=0.08165, over 11528.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.289, pruned_loss=0.05812, over 3069453.92 frames. ], batch size: 246, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:17:14,347 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=241588.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 21:17:20,022 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-05-01 21:17:38,324 INFO [train.py:904] (4/8) Epoch 24, batch 8150, loss[loss=0.2038, simple_loss=0.2747, pruned_loss=0.06645, over 11502.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2867, pruned_loss=0.05757, over 3071851.53 frames. ], batch size: 247, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:18:06,849 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=241621.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:18:22,058 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.026e+02 2.691e+02 3.345e+02 4.318e+02 1.089e+03, threshold=6.691e+02, percent-clipped=5.0 2023-05-01 21:18:30,849 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=241636.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 21:18:37,191 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1848, 2.2719, 2.2295, 3.8790, 2.2192, 2.5880, 2.3344, 2.4266], device='cuda:4'), covar=tensor([0.1424, 0.3556, 0.3166, 0.0585, 0.4065, 0.2614, 0.3652, 0.3328], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0459, 0.0374, 0.0331, 0.0441, 0.0524, 0.0429, 0.0535], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:18:39,436 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=241642.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:18:55,841 INFO [train.py:904] (4/8) Epoch 24, batch 8200, loss[loss=0.2041, simple_loss=0.2894, pruned_loss=0.05942, over 16698.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2842, pruned_loss=0.05657, over 3098237.82 frames. ], batch size: 134, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:19:38,359 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9268, 3.7968, 4.0240, 4.1297, 4.2536, 3.8758, 4.1816, 4.2786], device='cuda:4'), covar=tensor([0.1881, 0.1301, 0.1515, 0.0753, 0.0656, 0.1526, 0.0896, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0638, 0.0789, 0.0904, 0.0793, 0.0609, 0.0627, 0.0661, 0.0767], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:19:42,913 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=241682.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:19:53,125 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 21:19:56,098 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=241690.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:20:16,378 INFO [train.py:904] (4/8) Epoch 24, batch 8250, loss[loss=0.2044, simple_loss=0.2925, pruned_loss=0.0581, over 16776.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2832, pruned_loss=0.05451, over 3064817.16 frames. ], batch size: 124, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:21:03,230 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.735e+02 2.441e+02 2.851e+02 3.404e+02 7.117e+02, threshold=5.702e+02, percent-clipped=1.0 2023-05-01 21:21:05,690 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-05-01 21:21:20,204 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4381, 3.6041, 3.6488, 2.6163, 3.3507, 3.6500, 3.4031, 2.2489], device='cuda:4'), covar=tensor([0.0473, 0.0069, 0.0060, 0.0351, 0.0103, 0.0098, 0.0090, 0.0467], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0085, 0.0086, 0.0133, 0.0098, 0.0110, 0.0095, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 21:21:39,006 INFO [train.py:904] (4/8) Epoch 24, batch 8300, loss[loss=0.1808, simple_loss=0.285, pruned_loss=0.03831, over 16675.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2808, pruned_loss=0.05158, over 3060320.09 frames. ], batch size: 89, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:21:41,895 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6597, 4.6358, 4.4506, 3.7346, 4.5569, 1.6894, 4.2692, 4.1483], device='cuda:4'), covar=tensor([0.0089, 0.0103, 0.0219, 0.0372, 0.0094, 0.2953, 0.0157, 0.0266], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0162, 0.0203, 0.0179, 0.0178, 0.0208, 0.0191, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:21:54,681 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=241762.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 21:23:02,690 INFO [train.py:904] (4/8) Epoch 24, batch 8350, loss[loss=0.1829, simple_loss=0.2826, pruned_loss=0.04155, over 16193.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2802, pruned_loss=0.04977, over 3049994.48 frames. ], batch size: 165, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:23:03,880 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=241803.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:23:37,701 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6893, 3.3823, 3.5998, 1.9182, 3.8170, 3.8004, 3.0679, 3.0711], device='cuda:4'), covar=tensor([0.0654, 0.0229, 0.0203, 0.1229, 0.0069, 0.0182, 0.0380, 0.0411], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0107, 0.0097, 0.0135, 0.0080, 0.0125, 0.0126, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 21:23:48,197 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.476e+02 2.114e+02 2.559e+02 3.113e+02 5.727e+02, threshold=5.118e+02, percent-clipped=1.0 2023-05-01 21:23:48,769 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=241831.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 21:24:23,521 INFO [train.py:904] (4/8) Epoch 24, batch 8400, loss[loss=0.1872, simple_loss=0.2845, pruned_loss=0.04497, over 16451.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.278, pruned_loss=0.04772, over 3060746.44 frames. ], batch size: 146, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:24:42,781 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=241864.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:25:14,987 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9176, 2.1785, 2.4178, 3.2047, 2.1949, 2.3410, 2.3274, 2.2790], device='cuda:4'), covar=tensor([0.1326, 0.3633, 0.2771, 0.0729, 0.4482, 0.2684, 0.3618, 0.3566], device='cuda:4'), in_proj_covar=tensor([0.0401, 0.0450, 0.0368, 0.0324, 0.0433, 0.0513, 0.0421, 0.0525], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:25:28,730 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=241892.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 21:25:45,943 INFO [train.py:904] (4/8) Epoch 24, batch 8450, loss[loss=0.1597, simple_loss=0.2592, pruned_loss=0.0301, over 16724.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.2755, pruned_loss=0.04553, over 3071837.24 frames. ], batch size: 83, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:26:31,269 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.437e+02 2.077e+02 2.419e+02 2.958e+02 5.415e+02, threshold=4.839e+02, percent-clipped=2.0 2023-05-01 21:27:07,283 INFO [train.py:904] (4/8) Epoch 24, batch 8500, loss[loss=0.1496, simple_loss=0.2342, pruned_loss=0.03249, over 11483.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.272, pruned_loss=0.0436, over 3053217.90 frames. ], batch size: 248, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:27:47,682 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=241977.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:27:55,694 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7925, 2.7240, 2.4804, 4.1141, 2.6248, 4.0129, 1.4804, 2.9585], device='cuda:4'), covar=tensor([0.1345, 0.0724, 0.1158, 0.0186, 0.0136, 0.0369, 0.1754, 0.0704], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0175, 0.0195, 0.0190, 0.0202, 0.0213, 0.0203, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 21:28:33,821 INFO [train.py:904] (4/8) Epoch 24, batch 8550, loss[loss=0.1699, simple_loss=0.2554, pruned_loss=0.04221, over 12417.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.269, pruned_loss=0.04239, over 3030583.00 frames. ], batch size: 246, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:29:26,913 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.634e+02 2.227e+02 2.590e+02 3.113e+02 4.507e+02, threshold=5.180e+02, percent-clipped=0.0 2023-05-01 21:30:12,357 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9359, 1.9866, 2.3940, 2.9358, 2.6865, 3.3164, 2.2155, 3.3113], device='cuda:4'), covar=tensor([0.0232, 0.0566, 0.0380, 0.0311, 0.0356, 0.0212, 0.0524, 0.0169], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0192, 0.0178, 0.0181, 0.0197, 0.0156, 0.0195, 0.0154], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:30:13,213 INFO [train.py:904] (4/8) Epoch 24, batch 8600, loss[loss=0.1766, simple_loss=0.2679, pruned_loss=0.0426, over 16582.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2694, pruned_loss=0.04196, over 3026786.79 frames. ], batch size: 68, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:30:31,472 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=242062.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:31:15,376 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-05-01 21:31:51,616 INFO [train.py:904] (4/8) Epoch 24, batch 8650, loss[loss=0.1585, simple_loss=0.2502, pruned_loss=0.03334, over 12245.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2679, pruned_loss=0.04063, over 3022115.61 frames. ], batch size: 250, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:32:10,579 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=242110.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:32:56,837 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.586e+02 2.159e+02 2.524e+02 3.190e+02 5.272e+02, threshold=5.049e+02, percent-clipped=1.0 2023-05-01 21:33:17,288 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2629, 2.9998, 3.1207, 1.9460, 3.2887, 3.2961, 2.7408, 2.7074], device='cuda:4'), covar=tensor([0.0769, 0.0276, 0.0214, 0.1113, 0.0098, 0.0238, 0.0465, 0.0451], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0106, 0.0095, 0.0134, 0.0079, 0.0123, 0.0124, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 21:33:37,175 INFO [train.py:904] (4/8) Epoch 24, batch 8700, loss[loss=0.1622, simple_loss=0.2606, pruned_loss=0.03192, over 15306.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2656, pruned_loss=0.03949, over 3032046.91 frames. ], batch size: 191, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:33:38,226 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=242153.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:33:43,887 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.90 vs. limit=5.0 2023-05-01 21:33:50,497 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=242159.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:34:20,810 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2940, 2.9833, 3.0959, 1.9365, 3.3008, 3.3327, 2.7574, 2.7567], device='cuda:4'), covar=tensor([0.0731, 0.0281, 0.0218, 0.1056, 0.0086, 0.0199, 0.0467, 0.0409], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0105, 0.0095, 0.0133, 0.0079, 0.0123, 0.0124, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 21:34:42,494 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=242187.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 21:35:13,442 INFO [train.py:904] (4/8) Epoch 24, batch 8750, loss[loss=0.172, simple_loss=0.2751, pruned_loss=0.03445, over 16823.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2653, pruned_loss=0.03847, over 3053984.70 frames. ], batch size: 83, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:35:42,262 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=242214.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:36:10,033 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=242226.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:36:21,483 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.517e+02 2.077e+02 2.590e+02 3.155e+02 6.830e+02, threshold=5.181e+02, percent-clipped=2.0 2023-05-01 21:36:36,150 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.02 vs. limit=5.0 2023-05-01 21:36:50,901 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2709, 3.0194, 3.1342, 1.8974, 3.3057, 3.3408, 2.7606, 2.6987], device='cuda:4'), covar=tensor([0.0730, 0.0237, 0.0210, 0.1105, 0.0085, 0.0182, 0.0479, 0.0433], device='cuda:4'), in_proj_covar=tensor([0.0142, 0.0105, 0.0095, 0.0134, 0.0079, 0.0123, 0.0124, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-01 21:37:05,081 INFO [train.py:904] (4/8) Epoch 24, batch 8800, loss[loss=0.1672, simple_loss=0.2645, pruned_loss=0.03496, over 16380.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2632, pruned_loss=0.03757, over 3042265.84 frames. ], batch size: 146, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:37:56,145 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=242277.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:38:09,314 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0438, 4.3082, 4.1818, 4.1723, 3.8725, 3.8510, 3.9181, 4.2998], device='cuda:4'), covar=tensor([0.1019, 0.0931, 0.0825, 0.0719, 0.0705, 0.1892, 0.0973, 0.0893], device='cuda:4'), in_proj_covar=tensor([0.0677, 0.0811, 0.0674, 0.0630, 0.0516, 0.0525, 0.0684, 0.0637], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:38:17,961 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=242287.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:38:50,061 INFO [train.py:904] (4/8) Epoch 24, batch 8850, loss[loss=0.1452, simple_loss=0.241, pruned_loss=0.0247, over 12444.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2649, pruned_loss=0.03691, over 3022487.25 frames. ], batch size: 246, lr: 2.79e-03, grad_scale: 8.0 2023-05-01 21:39:03,873 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=242309.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:39:30,558 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-05-01 21:39:38,585 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=242325.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:39:52,554 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.544e+02 2.137e+02 2.486e+02 3.124e+02 5.230e+02, threshold=4.972e+02, percent-clipped=1.0 2023-05-01 21:40:38,015 INFO [train.py:904] (4/8) Epoch 24, batch 8900, loss[loss=0.1634, simple_loss=0.2553, pruned_loss=0.03574, over 12268.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.265, pruned_loss=0.0364, over 3004598.87 frames. ], batch size: 248, lr: 2.79e-03, grad_scale: 4.0 2023-05-01 21:41:12,532 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=242370.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:41:41,347 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3179, 1.7184, 1.9488, 2.2966, 2.3361, 2.5637, 1.8029, 2.4899], device='cuda:4'), covar=tensor([0.0221, 0.0533, 0.0378, 0.0321, 0.0371, 0.0220, 0.0573, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0190, 0.0177, 0.0180, 0.0195, 0.0155, 0.0194, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 21:42:44,441 INFO [train.py:904] (4/8) Epoch 24, batch 8950, loss[loss=0.1442, simple_loss=0.2469, pruned_loss=0.02077, over 16723.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2649, pruned_loss=0.03648, over 3039694.81 frames. ], batch size: 83, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 21:43:49,726 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.585e+02 2.299e+02 2.560e+02 2.923e+02 8.049e+02, threshold=5.120e+02, percent-clipped=2.0 2023-05-01 21:44:35,821 INFO [train.py:904] (4/8) Epoch 24, batch 9000, loss[loss=0.1491, simple_loss=0.2441, pruned_loss=0.0271, over 16858.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2621, pruned_loss=0.03521, over 3052925.40 frames. ], batch size: 124, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 21:44:35,821 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 21:44:45,532 INFO [train.py:938] (4/8) Epoch 24, validation: loss=0.1445, simple_loss=0.2484, pruned_loss=0.02026, over 944034.00 frames. 2023-05-01 21:44:45,533 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 21:44:58,942 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=242459.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:45:49,155 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=242483.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:45:58,862 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=242487.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 21:46:30,341 INFO [train.py:904] (4/8) Epoch 24, batch 9050, loss[loss=0.1603, simple_loss=0.2483, pruned_loss=0.03609, over 15392.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2626, pruned_loss=0.03557, over 3062699.50 frames. ], batch size: 191, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 21:46:40,247 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=242507.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:46:45,327 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=242509.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:47:20,224 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.98 vs. limit=5.0 2023-05-01 21:47:30,072 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.651e+02 2.227e+02 2.575e+02 3.137e+02 5.273e+02, threshold=5.150e+02, percent-clipped=1.0 2023-05-01 21:47:35,727 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=242535.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 21:48:00,791 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=242544.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:48:17,897 INFO [train.py:904] (4/8) Epoch 24, batch 9100, loss[loss=0.169, simple_loss=0.2524, pruned_loss=0.04283, over 12262.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2623, pruned_loss=0.0359, over 3084749.51 frames. ], batch size: 248, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 21:48:23,302 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=242555.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:49:27,659 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=242582.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:50:16,332 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=242602.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:50:16,952 INFO [train.py:904] (4/8) Epoch 24, batch 9150, loss[loss=0.1648, simple_loss=0.2599, pruned_loss=0.03484, over 16892.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2627, pruned_loss=0.03584, over 3056927.73 frames. ], batch size: 96, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 21:50:47,180 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=242616.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:51:21,272 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.667e+02 2.227e+02 2.578e+02 3.148e+02 6.071e+02, threshold=5.155e+02, percent-clipped=3.0 2023-05-01 21:52:01,740 INFO [train.py:904] (4/8) Epoch 24, batch 9200, loss[loss=0.1479, simple_loss=0.249, pruned_loss=0.02339, over 16907.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2587, pruned_loss=0.0351, over 3043031.64 frames. ], batch size: 102, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 21:52:22,359 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=242663.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:52:24,801 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=242665.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:53:38,114 INFO [train.py:904] (4/8) Epoch 24, batch 9250, loss[loss=0.147, simple_loss=0.2325, pruned_loss=0.03071, over 12169.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2585, pruned_loss=0.03513, over 3036326.52 frames. ], batch size: 247, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 21:54:40,869 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 2.148e+02 2.469e+02 2.996e+02 5.681e+02, threshold=4.938e+02, percent-clipped=3.0 2023-05-01 21:55:27,519 INFO [train.py:904] (4/8) Epoch 24, batch 9300, loss[loss=0.1564, simple_loss=0.2538, pruned_loss=0.02953, over 16239.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2577, pruned_loss=0.03476, over 3059470.57 frames. ], batch size: 165, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 21:56:13,801 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=242772.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:57:14,573 INFO [train.py:904] (4/8) Epoch 24, batch 9350, loss[loss=0.1614, simple_loss=0.2493, pruned_loss=0.03676, over 12482.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2575, pruned_loss=0.03462, over 3056406.58 frames. ], batch size: 247, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 21:57:27,692 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=242809.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:58:13,362 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.321e+02 1.952e+02 2.311e+02 3.086e+02 6.417e+02, threshold=4.623e+02, percent-clipped=1.0 2023-05-01 21:58:14,124 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=242833.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:58:25,325 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=242839.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:58:54,546 INFO [train.py:904] (4/8) Epoch 24, batch 9400, loss[loss=0.1763, simple_loss=0.2768, pruned_loss=0.03789, over 16400.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2573, pruned_loss=0.03452, over 3043821.42 frames. ], batch size: 146, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 21:59:03,672 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=242857.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 21:59:54,991 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=242882.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:00:35,986 INFO [train.py:904] (4/8) Epoch 24, batch 9450, loss[loss=0.1686, simple_loss=0.2623, pruned_loss=0.03746, over 16326.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2596, pruned_loss=0.0349, over 3060534.78 frames. ], batch size: 146, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:00:51,460 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=242911.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:01:08,921 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0462, 2.0355, 2.4739, 2.9726, 2.7118, 3.3157, 2.0508, 3.3217], device='cuda:4'), covar=tensor([0.0198, 0.0596, 0.0409, 0.0301, 0.0365, 0.0201, 0.0647, 0.0217], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0190, 0.0177, 0.0179, 0.0195, 0.0154, 0.0193, 0.0152], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 22:01:14,135 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=242921.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:01:32,438 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=242930.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:01:37,830 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.228e+02 2.255e+02 2.567e+02 3.168e+02 7.490e+02, threshold=5.133e+02, percent-clipped=6.0 2023-05-01 22:01:45,310 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8561, 5.1031, 5.2243, 5.0159, 5.0929, 5.5853, 5.0482, 4.7635], device='cuda:4'), covar=tensor([0.0862, 0.1781, 0.2145, 0.1674, 0.2210, 0.0800, 0.1505, 0.2192], device='cuda:4'), in_proj_covar=tensor([0.0394, 0.0586, 0.0646, 0.0480, 0.0633, 0.0673, 0.0501, 0.0637], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:01:53,970 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 22:02:16,733 INFO [train.py:904] (4/8) Epoch 24, batch 9500, loss[loss=0.1692, simple_loss=0.2674, pruned_loss=0.03551, over 16156.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2588, pruned_loss=0.03446, over 3076482.26 frames. ], batch size: 165, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:02:18,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5545, 4.6759, 4.8197, 4.5441, 4.6415, 5.1746, 4.7115, 4.3559], device='cuda:4'), covar=tensor([0.1265, 0.1825, 0.2187, 0.2164, 0.2464, 0.1009, 0.1549, 0.2586], device='cuda:4'), in_proj_covar=tensor([0.0393, 0.0585, 0.0645, 0.0480, 0.0633, 0.0672, 0.0501, 0.0637], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:02:30,099 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=242958.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:02:43,800 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=242965.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:03:05,394 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7209, 2.3593, 2.3595, 3.5296, 1.8513, 3.6424, 1.4728, 2.7007], device='cuda:4'), covar=tensor([0.1451, 0.0830, 0.1169, 0.0174, 0.0098, 0.0342, 0.1845, 0.0815], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0173, 0.0193, 0.0187, 0.0197, 0.0211, 0.0202, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:03:16,249 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=242982.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 22:04:01,168 INFO [train.py:904] (4/8) Epoch 24, batch 9550, loss[loss=0.1757, simple_loss=0.277, pruned_loss=0.03721, over 16121.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2586, pruned_loss=0.0347, over 3076563.25 frames. ], batch size: 165, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:04:24,732 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=243013.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:05:06,080 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 2.055e+02 2.389e+02 2.872e+02 5.971e+02, threshold=4.777e+02, percent-clipped=1.0 2023-05-01 22:05:43,247 INFO [train.py:904] (4/8) Epoch 24, batch 9600, loss[loss=0.1979, simple_loss=0.3011, pruned_loss=0.04741, over 16697.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2595, pruned_loss=0.03525, over 3065675.13 frames. ], batch size: 134, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:06:12,628 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6426, 2.5994, 2.3741, 4.1459, 2.3492, 3.8914, 1.4505, 2.8247], device='cuda:4'), covar=tensor([0.1438, 0.0796, 0.1246, 0.0154, 0.0130, 0.0368, 0.1744, 0.0791], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0173, 0.0193, 0.0187, 0.0197, 0.0211, 0.0202, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:06:16,758 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7924, 3.1138, 3.4011, 1.8917, 2.9304, 2.1703, 3.2511, 3.2555], device='cuda:4'), covar=tensor([0.0272, 0.0831, 0.0568, 0.2160, 0.0793, 0.1013, 0.0638, 0.0946], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0158, 0.0163, 0.0150, 0.0141, 0.0127, 0.0139, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:07:31,344 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4824, 3.3838, 3.5453, 3.6195, 3.6609, 3.3620, 3.6328, 3.6994], device='cuda:4'), covar=tensor([0.1265, 0.1037, 0.1040, 0.0644, 0.0640, 0.2262, 0.0874, 0.0775], device='cuda:4'), in_proj_covar=tensor([0.0617, 0.0762, 0.0874, 0.0769, 0.0590, 0.0607, 0.0640, 0.0741], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 22:07:32,094 INFO [train.py:904] (4/8) Epoch 24, batch 9650, loss[loss=0.1653, simple_loss=0.2579, pruned_loss=0.03629, over 12859.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2621, pruned_loss=0.03552, over 3078610.72 frames. ], batch size: 246, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:08:29,798 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=243128.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:08:37,734 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.514e+02 2.193e+02 2.679e+02 3.134e+02 6.217e+02, threshold=5.359e+02, percent-clipped=4.0 2023-05-01 22:08:50,203 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=243139.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:09:19,548 INFO [train.py:904] (4/8) Epoch 24, batch 9700, loss[loss=0.1833, simple_loss=0.2747, pruned_loss=0.04594, over 16154.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2611, pruned_loss=0.03513, over 3087795.14 frames. ], batch size: 165, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:09:24,356 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5596, 1.8208, 2.1256, 2.4709, 2.4366, 2.7989, 2.0781, 2.7673], device='cuda:4'), covar=tensor([0.0250, 0.0561, 0.0392, 0.0344, 0.0379, 0.0205, 0.0527, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0188, 0.0176, 0.0177, 0.0193, 0.0153, 0.0192, 0.0151], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 22:09:40,342 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.64 vs. limit=5.0 2023-05-01 22:10:08,777 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7282, 2.8835, 2.5790, 4.4259, 2.8317, 4.1210, 1.5767, 3.0859], device='cuda:4'), covar=tensor([0.1385, 0.0723, 0.1121, 0.0168, 0.0143, 0.0331, 0.1699, 0.0689], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0173, 0.0192, 0.0186, 0.0196, 0.0210, 0.0202, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:10:30,926 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=243187.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:11:01,230 INFO [train.py:904] (4/8) Epoch 24, batch 9750, loss[loss=0.1691, simple_loss=0.2509, pruned_loss=0.0436, over 12450.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.26, pruned_loss=0.0351, over 3074239.49 frames. ], batch size: 248, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 22:11:17,016 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=243211.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:12:00,635 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3802, 3.4621, 2.2497, 3.8241, 2.6086, 3.7172, 2.3423, 2.8745], device='cuda:4'), covar=tensor([0.0332, 0.0406, 0.1559, 0.0234, 0.0866, 0.0757, 0.1544, 0.0821], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0171, 0.0188, 0.0159, 0.0173, 0.0208, 0.0198, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:12:03,356 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.343e+02 1.977e+02 2.392e+02 3.038e+02 5.323e+02, threshold=4.783e+02, percent-clipped=0.0 2023-05-01 22:12:37,392 INFO [train.py:904] (4/8) Epoch 24, batch 9800, loss[loss=0.1829, simple_loss=0.2834, pruned_loss=0.04119, over 15253.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2609, pruned_loss=0.03485, over 3088526.97 frames. ], batch size: 190, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 22:12:49,025 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=243258.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:12:50,824 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=243259.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:13:22,146 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=243277.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 22:13:47,744 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5256, 3.7998, 3.8648, 2.7472, 3.4724, 3.8988, 3.6061, 2.3611], device='cuda:4'), covar=tensor([0.0490, 0.0051, 0.0048, 0.0350, 0.0115, 0.0081, 0.0081, 0.0439], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0084, 0.0084, 0.0132, 0.0097, 0.0107, 0.0093, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:13:58,404 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=243293.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:14:21,694 INFO [train.py:904] (4/8) Epoch 24, batch 9850, loss[loss=0.1542, simple_loss=0.2496, pruned_loss=0.02942, over 16310.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2611, pruned_loss=0.03412, over 3102728.34 frames. ], batch size: 146, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 22:14:28,945 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=243306.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:15:22,935 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.302e+02 1.939e+02 2.274e+02 2.876e+02 6.181e+02, threshold=4.547e+02, percent-clipped=2.0 2023-05-01 22:16:11,623 INFO [train.py:904] (4/8) Epoch 24, batch 9900, loss[loss=0.1672, simple_loss=0.2734, pruned_loss=0.03053, over 16268.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2622, pruned_loss=0.0343, over 3105368.47 frames. ], batch size: 165, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 22:16:15,194 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=243354.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:16:34,765 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9412, 3.1497, 3.6441, 1.9319, 3.0746, 2.2927, 3.4761, 3.3205], device='cuda:4'), covar=tensor([0.0231, 0.0889, 0.0427, 0.2201, 0.0745, 0.0986, 0.0597, 0.1034], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0157, 0.0163, 0.0150, 0.0141, 0.0126, 0.0139, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:18:09,540 INFO [train.py:904] (4/8) Epoch 24, batch 9950, loss[loss=0.1582, simple_loss=0.2546, pruned_loss=0.03087, over 16468.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2643, pruned_loss=0.03504, over 3088223.09 frames. ], batch size: 68, lr: 2.78e-03, grad_scale: 4.0 2023-05-01 22:19:11,942 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=243428.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:19:24,674 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.448e+02 2.110e+02 2.470e+02 3.179e+02 5.008e+02, threshold=4.941e+02, percent-clipped=4.0 2023-05-01 22:20:10,934 INFO [train.py:904] (4/8) Epoch 24, batch 10000, loss[loss=0.1694, simple_loss=0.2689, pruned_loss=0.03498, over 16613.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2626, pruned_loss=0.0346, over 3101401.92 frames. ], batch size: 134, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:20:55,587 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=243476.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:21:51,122 INFO [train.py:904] (4/8) Epoch 24, batch 10050, loss[loss=0.1722, simple_loss=0.2736, pruned_loss=0.03543, over 16316.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2626, pruned_loss=0.03451, over 3073672.39 frames. ], batch size: 165, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:22:52,100 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.427e+02 2.109e+02 2.585e+02 3.087e+02 5.366e+02, threshold=5.170e+02, percent-clipped=3.0 2023-05-01 22:23:17,284 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4228, 4.5481, 4.6966, 4.4742, 4.6034, 5.0541, 4.5772, 4.2810], device='cuda:4'), covar=tensor([0.1382, 0.1716, 0.1772, 0.1976, 0.2186, 0.0892, 0.1529, 0.2513], device='cuda:4'), in_proj_covar=tensor([0.0388, 0.0577, 0.0638, 0.0474, 0.0624, 0.0660, 0.0494, 0.0627], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:23:25,157 INFO [train.py:904] (4/8) Epoch 24, batch 10100, loss[loss=0.1464, simple_loss=0.24, pruned_loss=0.02639, over 15327.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2626, pruned_loss=0.03445, over 3081176.17 frames. ], batch size: 191, lr: 2.78e-03, grad_scale: 8.0 2023-05-01 22:24:02,830 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.77 vs. limit=5.0 2023-05-01 22:24:16,126 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=243577.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 22:25:09,202 INFO [train.py:904] (4/8) Epoch 25, batch 0, loss[loss=0.2218, simple_loss=0.2893, pruned_loss=0.07719, over 16809.00 frames. ], tot_loss[loss=0.2218, simple_loss=0.2893, pruned_loss=0.07719, over 16809.00 frames. ], batch size: 124, lr: 2.72e-03, grad_scale: 8.0 2023-05-01 22:25:09,202 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 22:25:16,828 INFO [train.py:938] (4/8) Epoch 25, validation: loss=0.1443, simple_loss=0.2477, pruned_loss=0.02048, over 944034.00 frames. 2023-05-01 22:25:16,828 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 22:25:48,588 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=243625.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:25:52,519 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8183, 3.8200, 4.1341, 4.0868, 4.1222, 3.8897, 3.9111, 3.8838], device='cuda:4'), covar=tensor([0.0432, 0.0750, 0.0432, 0.0504, 0.0550, 0.0545, 0.0899, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0397, 0.0443, 0.0435, 0.0400, 0.0476, 0.0454, 0.0528, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 22:26:03,176 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.600e+02 2.517e+02 2.877e+02 3.521e+02 6.755e+02, threshold=5.755e+02, percent-clipped=6.0 2023-05-01 22:26:21,433 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=243649.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:26:26,908 INFO [train.py:904] (4/8) Epoch 25, batch 50, loss[loss=0.1827, simple_loss=0.2601, pruned_loss=0.05266, over 16717.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2671, pruned_loss=0.04611, over 752021.17 frames. ], batch size: 83, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:26:30,737 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-01 22:26:42,184 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3767, 5.3520, 5.1008, 4.8263, 5.1890, 2.0044, 5.0116, 5.0864], device='cuda:4'), covar=tensor([0.0073, 0.0063, 0.0242, 0.0280, 0.0094, 0.2740, 0.0123, 0.0196], device='cuda:4'), in_proj_covar=tensor([0.0166, 0.0158, 0.0196, 0.0171, 0.0174, 0.0205, 0.0186, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 22:27:04,790 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-05-01 22:27:35,800 INFO [train.py:904] (4/8) Epoch 25, batch 100, loss[loss=0.1729, simple_loss=0.2657, pruned_loss=0.04003, over 16837.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.265, pruned_loss=0.04486, over 1317657.40 frames. ], batch size: 62, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:27:56,670 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-05-01 22:28:22,079 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.579e+02 2.266e+02 2.864e+02 3.484e+02 1.313e+03, threshold=5.728e+02, percent-clipped=5.0 2023-05-01 22:28:45,144 INFO [train.py:904] (4/8) Epoch 25, batch 150, loss[loss=0.1935, simple_loss=0.2692, pruned_loss=0.05894, over 16874.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2627, pruned_loss=0.04365, over 1767414.14 frames. ], batch size: 116, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:29:55,569 INFO [train.py:904] (4/8) Epoch 25, batch 200, loss[loss=0.1688, simple_loss=0.2718, pruned_loss=0.03288, over 17104.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2625, pruned_loss=0.04334, over 2111316.57 frames. ], batch size: 49, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:30:40,583 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.429e+02 2.064e+02 2.455e+02 3.172e+02 7.590e+02, threshold=4.909e+02, percent-clipped=1.0 2023-05-01 22:31:04,161 INFO [train.py:904] (4/8) Epoch 25, batch 250, loss[loss=0.1705, simple_loss=0.2492, pruned_loss=0.04584, over 12161.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.26, pruned_loss=0.04313, over 2384113.19 frames. ], batch size: 246, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:31:22,902 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0443, 3.1909, 3.4149, 2.1888, 2.9360, 2.3464, 3.5462, 3.5086], device='cuda:4'), covar=tensor([0.0269, 0.0952, 0.0658, 0.2013, 0.0899, 0.1064, 0.0581, 0.0947], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0161, 0.0166, 0.0152, 0.0143, 0.0129, 0.0142, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:32:08,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7438, 2.7002, 2.7148, 4.8015, 3.7746, 4.2802, 1.6561, 3.0578], device='cuda:4'), covar=tensor([0.1551, 0.0924, 0.1313, 0.0229, 0.0260, 0.0439, 0.1836, 0.0873], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0176, 0.0196, 0.0191, 0.0200, 0.0214, 0.0205, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:32:14,684 INFO [train.py:904] (4/8) Epoch 25, batch 300, loss[loss=0.1655, simple_loss=0.2716, pruned_loss=0.02967, over 17051.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2579, pruned_loss=0.04171, over 2591877.81 frames. ], batch size: 50, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:33:00,865 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.556e+02 2.231e+02 2.519e+02 2.986e+02 7.202e+02, threshold=5.038e+02, percent-clipped=2.0 2023-05-01 22:33:19,534 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=243949.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:33:24,546 INFO [train.py:904] (4/8) Epoch 25, batch 350, loss[loss=0.1523, simple_loss=0.2391, pruned_loss=0.03274, over 17204.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.255, pruned_loss=0.04109, over 2760495.42 frames. ], batch size: 44, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:33:54,421 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8077, 2.6617, 2.5669, 4.2744, 3.5175, 4.1447, 1.6790, 2.8959], device='cuda:4'), covar=tensor([0.1458, 0.0737, 0.1196, 0.0196, 0.0114, 0.0351, 0.1627, 0.0851], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0176, 0.0195, 0.0191, 0.0200, 0.0213, 0.0204, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:34:00,972 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9234, 4.0496, 2.6542, 4.6680, 3.1729, 4.5836, 2.7035, 3.3209], device='cuda:4'), covar=tensor([0.0342, 0.0398, 0.1570, 0.0350, 0.0828, 0.0569, 0.1505, 0.0833], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0177, 0.0194, 0.0167, 0.0178, 0.0216, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:34:04,693 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4425, 3.4814, 4.1766, 2.2186, 3.2374, 2.5646, 3.8215, 3.7435], device='cuda:4'), covar=tensor([0.0284, 0.1042, 0.0452, 0.2083, 0.0843, 0.0983, 0.0663, 0.1139], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0162, 0.0167, 0.0153, 0.0144, 0.0129, 0.0142, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:34:25,843 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=243997.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:34:38,353 INFO [train.py:904] (4/8) Epoch 25, batch 400, loss[loss=0.1473, simple_loss=0.234, pruned_loss=0.03028, over 16919.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2536, pruned_loss=0.04081, over 2892139.94 frames. ], batch size: 96, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:34:50,888 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9697, 2.6735, 2.6873, 4.3217, 3.3759, 4.0837, 1.6424, 2.9680], device='cuda:4'), covar=tensor([0.1364, 0.0827, 0.1212, 0.0240, 0.0218, 0.0468, 0.1755, 0.0845], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0176, 0.0195, 0.0191, 0.0200, 0.0213, 0.0204, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:34:54,342 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1644, 4.6057, 3.3128, 2.7300, 2.9047, 2.8955, 4.9805, 3.6701], device='cuda:4'), covar=tensor([0.2613, 0.0555, 0.1885, 0.2717, 0.2844, 0.2067, 0.0317, 0.1501], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0268, 0.0305, 0.0315, 0.0294, 0.0265, 0.0295, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:35:02,276 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-05-01 22:35:22,991 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.463e+02 2.193e+02 2.643e+02 3.297e+02 6.329e+02, threshold=5.285e+02, percent-clipped=2.0 2023-05-01 22:35:42,259 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.98 vs. limit=5.0 2023-05-01 22:35:47,032 INFO [train.py:904] (4/8) Epoch 25, batch 450, loss[loss=0.1705, simple_loss=0.2451, pruned_loss=0.04794, over 16432.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2515, pruned_loss=0.03983, over 2992690.71 frames. ], batch size: 146, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:36:50,464 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9627, 4.9236, 5.3660, 5.3470, 5.3745, 5.0859, 4.9846, 4.8143], device='cuda:4'), covar=tensor([0.0327, 0.0667, 0.0384, 0.0369, 0.0441, 0.0368, 0.0907, 0.0470], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0459, 0.0447, 0.0412, 0.0491, 0.0470, 0.0546, 0.0375], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 22:36:55,203 INFO [train.py:904] (4/8) Epoch 25, batch 500, loss[loss=0.1614, simple_loss=0.2371, pruned_loss=0.04285, over 16844.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.25, pruned_loss=0.03895, over 3065739.61 frames. ], batch size: 116, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:37:18,532 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0974, 4.4674, 4.4302, 3.2436, 3.8236, 4.4722, 4.0280, 2.7602], device='cuda:4'), covar=tensor([0.0481, 0.0072, 0.0051, 0.0363, 0.0129, 0.0092, 0.0085, 0.0474], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0088, 0.0087, 0.0136, 0.0100, 0.0111, 0.0096, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 22:37:37,504 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7605, 4.8829, 5.0479, 4.8083, 4.8269, 5.5032, 4.9790, 4.6680], device='cuda:4'), covar=tensor([0.1358, 0.2156, 0.2590, 0.2419, 0.2698, 0.1099, 0.1935, 0.2682], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0610, 0.0672, 0.0499, 0.0659, 0.0693, 0.0520, 0.0659], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:37:42,041 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.358e+02 2.128e+02 2.486e+02 2.973e+02 4.659e+02, threshold=4.972e+02, percent-clipped=0.0 2023-05-01 22:38:05,719 INFO [train.py:904] (4/8) Epoch 25, batch 550, loss[loss=0.1612, simple_loss=0.2558, pruned_loss=0.0333, over 17034.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2488, pruned_loss=0.03824, over 3121366.66 frames. ], batch size: 50, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:39:15,776 INFO [train.py:904] (4/8) Epoch 25, batch 600, loss[loss=0.1578, simple_loss=0.2361, pruned_loss=0.03978, over 16834.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2484, pruned_loss=0.03813, over 3171294.18 frames. ], batch size: 102, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:39:47,840 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3852, 2.9671, 2.6828, 2.2780, 2.2362, 2.3360, 3.0027, 2.7389], device='cuda:4'), covar=tensor([0.2646, 0.0698, 0.1692, 0.2365, 0.2311, 0.2274, 0.0547, 0.1348], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0268, 0.0306, 0.0317, 0.0296, 0.0266, 0.0297, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:40:02,986 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 2.116e+02 2.478e+02 2.900e+02 1.597e+03, threshold=4.957e+02, percent-clipped=3.0 2023-05-01 22:40:24,985 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.92 vs. limit=5.0 2023-05-01 22:40:25,349 INFO [train.py:904] (4/8) Epoch 25, batch 650, loss[loss=0.1572, simple_loss=0.245, pruned_loss=0.03469, over 16785.00 frames. ], tot_loss[loss=0.1617, simple_loss=0.2474, pruned_loss=0.03796, over 3203465.27 frames. ], batch size: 57, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:41:33,626 INFO [train.py:904] (4/8) Epoch 25, batch 700, loss[loss=0.1527, simple_loss=0.2312, pruned_loss=0.03706, over 15937.00 frames. ], tot_loss[loss=0.1616, simple_loss=0.2477, pruned_loss=0.03778, over 3225551.04 frames. ], batch size: 35, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:41:55,440 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-01 22:42:06,434 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2862, 3.4992, 3.9251, 2.1470, 3.1559, 2.5277, 3.6915, 3.6394], device='cuda:4'), covar=tensor([0.0317, 0.0961, 0.0457, 0.2106, 0.0831, 0.0949, 0.0678, 0.1188], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0165, 0.0169, 0.0155, 0.0146, 0.0131, 0.0144, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:42:14,150 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4617, 3.5242, 3.7025, 3.6745, 3.7001, 3.5188, 3.5630, 3.5788], device='cuda:4'), covar=tensor([0.0486, 0.0921, 0.0541, 0.0485, 0.0549, 0.0583, 0.0838, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0468, 0.0456, 0.0421, 0.0501, 0.0480, 0.0557, 0.0382], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 22:42:20,953 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.323e+02 2.146e+02 2.447e+02 2.875e+02 6.652e+02, threshold=4.894e+02, percent-clipped=4.0 2023-05-01 22:42:24,716 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-01 22:42:42,412 INFO [train.py:904] (4/8) Epoch 25, batch 750, loss[loss=0.1631, simple_loss=0.2596, pruned_loss=0.03334, over 17073.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2488, pruned_loss=0.03809, over 3250224.03 frames. ], batch size: 53, lr: 2.72e-03, grad_scale: 2.0 2023-05-01 22:42:42,762 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=244353.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:43:44,008 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8160, 4.0058, 3.0026, 2.3545, 2.5613, 2.4969, 4.1457, 3.4402], device='cuda:4'), covar=tensor([0.2588, 0.0547, 0.1716, 0.2987, 0.2706, 0.2110, 0.0454, 0.1431], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0270, 0.0309, 0.0319, 0.0299, 0.0268, 0.0299, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:43:52,151 INFO [train.py:904] (4/8) Epoch 25, batch 800, loss[loss=0.1648, simple_loss=0.2563, pruned_loss=0.03667, over 16683.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2485, pruned_loss=0.03825, over 3267063.73 frames. ], batch size: 62, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:43:52,791 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8103, 3.0177, 3.0701, 5.0718, 4.3195, 4.5016, 1.5810, 3.4461], device='cuda:4'), covar=tensor([0.1403, 0.0742, 0.1042, 0.0178, 0.0163, 0.0356, 0.1687, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0176, 0.0194, 0.0192, 0.0199, 0.0213, 0.0204, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:44:08,187 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=244414.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:44:39,329 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.613e+02 2.216e+02 2.518e+02 3.114e+02 7.971e+02, threshold=5.036e+02, percent-clipped=2.0 2023-05-01 22:45:03,354 INFO [train.py:904] (4/8) Epoch 25, batch 850, loss[loss=0.146, simple_loss=0.2399, pruned_loss=0.02602, over 17200.00 frames. ], tot_loss[loss=0.1619, simple_loss=0.248, pruned_loss=0.03793, over 3275810.93 frames. ], batch size: 45, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:46:12,668 INFO [train.py:904] (4/8) Epoch 25, batch 900, loss[loss=0.1423, simple_loss=0.2346, pruned_loss=0.02501, over 17200.00 frames. ], tot_loss[loss=0.161, simple_loss=0.2473, pruned_loss=0.03729, over 3291134.42 frames. ], batch size: 44, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:47:00,724 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.397e+02 1.999e+02 2.403e+02 2.874e+02 4.869e+02, threshold=4.806e+02, percent-clipped=0.0 2023-05-01 22:47:23,393 INFO [train.py:904] (4/8) Epoch 25, batch 950, loss[loss=0.1632, simple_loss=0.2482, pruned_loss=0.03909, over 16836.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2473, pruned_loss=0.03782, over 3304425.48 frames. ], batch size: 102, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:47:32,708 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6503, 3.3547, 3.8682, 1.9602, 3.9158, 3.9889, 3.2416, 2.9251], device='cuda:4'), covar=tensor([0.0822, 0.0318, 0.0212, 0.1274, 0.0124, 0.0235, 0.0394, 0.0488], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0111, 0.0100, 0.0140, 0.0084, 0.0130, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 22:48:33,896 INFO [train.py:904] (4/8) Epoch 25, batch 1000, loss[loss=0.2081, simple_loss=0.2758, pruned_loss=0.07022, over 16720.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.247, pruned_loss=0.03797, over 3311252.46 frames. ], batch size: 124, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:49:21,001 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.583e+02 2.078e+02 2.468e+02 2.868e+02 8.684e+02, threshold=4.936e+02, percent-clipped=5.0 2023-05-01 22:49:42,534 INFO [train.py:904] (4/8) Epoch 25, batch 1050, loss[loss=0.1434, simple_loss=0.2381, pruned_loss=0.02433, over 17123.00 frames. ], tot_loss[loss=0.161, simple_loss=0.2469, pruned_loss=0.03757, over 3315263.51 frames. ], batch size: 48, lr: 2.72e-03, grad_scale: 4.0 2023-05-01 22:49:56,500 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7758, 2.5040, 2.0802, 2.2928, 2.8254, 2.6043, 2.7945, 2.9350], device='cuda:4'), covar=tensor([0.0252, 0.0424, 0.0565, 0.0487, 0.0260, 0.0340, 0.0224, 0.0292], device='cuda:4'), in_proj_covar=tensor([0.0226, 0.0246, 0.0235, 0.0236, 0.0248, 0.0245, 0.0246, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 22:50:18,874 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 22:50:53,097 INFO [train.py:904] (4/8) Epoch 25, batch 1100, loss[loss=0.1629, simple_loss=0.2619, pruned_loss=0.03198, over 17030.00 frames. ], tot_loss[loss=0.1612, simple_loss=0.2474, pruned_loss=0.03751, over 3322491.52 frames. ], batch size: 50, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 22:51:01,221 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=244709.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:51:02,991 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.79 vs. limit=5.0 2023-05-01 22:51:13,207 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=244717.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 22:51:40,295 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.414e+02 2.068e+02 2.379e+02 2.713e+02 6.650e+02, threshold=4.759e+02, percent-clipped=1.0 2023-05-01 22:52:00,316 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5215, 3.3744, 2.7616, 2.2621, 2.2434, 2.3584, 3.4474, 2.9845], device='cuda:4'), covar=tensor([0.2860, 0.0645, 0.1815, 0.2830, 0.2893, 0.2269, 0.0520, 0.1681], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0272, 0.0310, 0.0320, 0.0300, 0.0270, 0.0300, 0.0345], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 22:52:02,014 INFO [train.py:904] (4/8) Epoch 25, batch 1150, loss[loss=0.163, simple_loss=0.2585, pruned_loss=0.0338, over 17044.00 frames. ], tot_loss[loss=0.1602, simple_loss=0.2463, pruned_loss=0.03708, over 3311299.41 frames. ], batch size: 53, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 22:52:37,154 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=244778.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 22:52:41,374 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2355, 4.2469, 4.5697, 4.5468, 4.5983, 4.3116, 4.3323, 4.2513], device='cuda:4'), covar=tensor([0.0424, 0.0757, 0.0447, 0.0480, 0.0498, 0.0537, 0.0819, 0.0626], device='cuda:4'), in_proj_covar=tensor([0.0431, 0.0478, 0.0466, 0.0431, 0.0512, 0.0491, 0.0569, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 22:53:11,753 INFO [train.py:904] (4/8) Epoch 25, batch 1200, loss[loss=0.1815, simple_loss=0.2621, pruned_loss=0.05048, over 16711.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2457, pruned_loss=0.037, over 3321228.41 frames. ], batch size: 62, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 22:53:30,384 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6424, 4.4378, 4.6762, 4.8338, 4.9166, 4.4520, 4.8416, 4.9164], device='cuda:4'), covar=tensor([0.1803, 0.1329, 0.1394, 0.0816, 0.0655, 0.1113, 0.1944, 0.0903], device='cuda:4'), in_proj_covar=tensor([0.0668, 0.0824, 0.0949, 0.0834, 0.0636, 0.0654, 0.0686, 0.0799], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 22:53:57,425 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.058e+02 2.429e+02 2.907e+02 5.779e+02, threshold=4.859e+02, percent-clipped=1.0 2023-05-01 22:54:19,233 INFO [train.py:904] (4/8) Epoch 25, batch 1250, loss[loss=0.1831, simple_loss=0.256, pruned_loss=0.05509, over 16408.00 frames. ], tot_loss[loss=0.1603, simple_loss=0.2457, pruned_loss=0.03742, over 3310205.16 frames. ], batch size: 146, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 22:54:22,849 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-05-01 22:55:06,283 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=244887.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:55:21,524 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2885, 4.1176, 4.3539, 4.4755, 4.5519, 4.1218, 4.3723, 4.5606], device='cuda:4'), covar=tensor([0.1636, 0.1186, 0.1251, 0.0664, 0.0594, 0.1314, 0.2819, 0.0722], device='cuda:4'), in_proj_covar=tensor([0.0669, 0.0826, 0.0951, 0.0836, 0.0638, 0.0657, 0.0687, 0.0801], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 22:55:26,307 INFO [train.py:904] (4/8) Epoch 25, batch 1300, loss[loss=0.1537, simple_loss=0.2319, pruned_loss=0.03777, over 15335.00 frames. ], tot_loss[loss=0.1605, simple_loss=0.2453, pruned_loss=0.03779, over 3311259.10 frames. ], batch size: 190, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 22:56:12,138 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.526e+02 2.186e+02 2.539e+02 3.062e+02 9.189e+02, threshold=5.078e+02, percent-clipped=4.0 2023-05-01 22:56:28,349 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=244948.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:56:35,068 INFO [train.py:904] (4/8) Epoch 25, batch 1350, loss[loss=0.1614, simple_loss=0.2424, pruned_loss=0.04019, over 12597.00 frames. ], tot_loss[loss=0.1592, simple_loss=0.2442, pruned_loss=0.03706, over 3306170.55 frames. ], batch size: 248, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 22:56:36,704 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=244954.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:57:17,903 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4468, 4.4983, 4.8219, 4.8068, 4.8503, 4.5650, 4.5736, 4.4051], device='cuda:4'), covar=tensor([0.0510, 0.0993, 0.0492, 0.0622, 0.0535, 0.0597, 0.0899, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0478, 0.0465, 0.0429, 0.0510, 0.0491, 0.0568, 0.0389], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 22:57:43,144 INFO [train.py:904] (4/8) Epoch 25, batch 1400, loss[loss=0.1573, simple_loss=0.2472, pruned_loss=0.03366, over 16744.00 frames. ], tot_loss[loss=0.1596, simple_loss=0.2447, pruned_loss=0.03728, over 3310663.02 frames. ], batch size: 62, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 22:57:52,395 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=245009.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:58:00,045 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=245015.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:58:28,649 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.516e+02 2.075e+02 2.428e+02 3.013e+02 6.709e+02, threshold=4.856e+02, percent-clipped=2.0 2023-05-01 22:58:38,864 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.08 vs. limit=2.0 2023-05-01 22:58:51,281 INFO [train.py:904] (4/8) Epoch 25, batch 1450, loss[loss=0.1597, simple_loss=0.2344, pruned_loss=0.04257, over 16477.00 frames. ], tot_loss[loss=0.1587, simple_loss=0.2437, pruned_loss=0.03683, over 3315492.55 frames. ], batch size: 75, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 22:58:56,916 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=245057.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 22:59:18,777 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=245073.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 22:59:40,984 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2630, 2.5047, 2.8459, 3.2227, 3.0204, 3.8019, 2.6548, 3.6764], device='cuda:4'), covar=tensor([0.0232, 0.0499, 0.0351, 0.0321, 0.0355, 0.0161, 0.0490, 0.0172], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0197, 0.0184, 0.0187, 0.0203, 0.0161, 0.0199, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 22:59:49,343 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2591, 4.6797, 4.5887, 3.4524, 3.8616, 4.5665, 4.1431, 2.9625], device='cuda:4'), covar=tensor([0.0431, 0.0050, 0.0043, 0.0334, 0.0139, 0.0101, 0.0082, 0.0417], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0088, 0.0136, 0.0100, 0.0112, 0.0097, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 23:00:00,089 INFO [train.py:904] (4/8) Epoch 25, batch 1500, loss[loss=0.1873, simple_loss=0.2644, pruned_loss=0.05509, over 15524.00 frames. ], tot_loss[loss=0.1596, simple_loss=0.2443, pruned_loss=0.03746, over 3306361.37 frames. ], batch size: 191, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:00:12,935 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-05-01 23:00:46,573 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.416e+02 2.188e+02 2.473e+02 2.927e+02 4.606e+02, threshold=4.945e+02, percent-clipped=0.0 2023-05-01 23:01:08,932 INFO [train.py:904] (4/8) Epoch 25, batch 1550, loss[loss=0.1529, simple_loss=0.2421, pruned_loss=0.03186, over 17176.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2461, pruned_loss=0.03839, over 3311461.61 frames. ], batch size: 46, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:01:36,830 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.79 vs. limit=5.0 2023-05-01 23:02:19,885 INFO [train.py:904] (4/8) Epoch 25, batch 1600, loss[loss=0.1934, simple_loss=0.2762, pruned_loss=0.05532, over 16932.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2486, pruned_loss=0.0395, over 3311411.66 frames. ], batch size: 109, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:03:06,948 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.282e+02 2.634e+02 3.263e+02 7.681e+02, threshold=5.268e+02, percent-clipped=4.0 2023-05-01 23:03:16,799 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=245243.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:03:29,744 INFO [train.py:904] (4/8) Epoch 25, batch 1650, loss[loss=0.162, simple_loss=0.2609, pruned_loss=0.03158, over 17144.00 frames. ], tot_loss[loss=0.165, simple_loss=0.25, pruned_loss=0.03999, over 3310203.52 frames. ], batch size: 48, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:04:41,928 INFO [train.py:904] (4/8) Epoch 25, batch 1700, loss[loss=0.1467, simple_loss=0.2337, pruned_loss=0.02981, over 16801.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2519, pruned_loss=0.04055, over 3302409.69 frames. ], batch size: 39, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:04:51,854 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=245310.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:05:30,603 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.540e+02 2.070e+02 2.579e+02 3.059e+02 5.844e+02, threshold=5.158e+02, percent-clipped=3.0 2023-05-01 23:05:44,916 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-05-01 23:05:50,196 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9626, 4.0622, 2.7796, 4.7258, 3.2271, 4.6480, 2.6989, 3.3807], device='cuda:4'), covar=tensor([0.0319, 0.0379, 0.1518, 0.0249, 0.0800, 0.0507, 0.1572, 0.0782], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0181, 0.0197, 0.0173, 0.0180, 0.0222, 0.0206, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 23:05:52,657 INFO [train.py:904] (4/8) Epoch 25, batch 1750, loss[loss=0.1721, simple_loss=0.2602, pruned_loss=0.04195, over 16568.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2528, pruned_loss=0.04046, over 3308955.12 frames. ], batch size: 68, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:06:20,904 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=245373.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 23:07:03,959 INFO [train.py:904] (4/8) Epoch 25, batch 1800, loss[loss=0.1934, simple_loss=0.2711, pruned_loss=0.05781, over 16740.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2535, pruned_loss=0.04008, over 3310624.43 frames. ], batch size: 134, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:07:29,357 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=245421.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 23:07:47,370 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0718, 5.0080, 4.9405, 4.4306, 4.5862, 4.9778, 4.9240, 4.6058], device='cuda:4'), covar=tensor([0.0629, 0.0670, 0.0341, 0.0391, 0.1076, 0.0490, 0.0349, 0.0774], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0462, 0.0358, 0.0361, 0.0363, 0.0417, 0.0246, 0.0433], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:07:51,079 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.502e+02 2.188e+02 2.630e+02 2.978e+02 6.380e+02, threshold=5.260e+02, percent-clipped=2.0 2023-05-01 23:08:09,967 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-01 23:08:12,915 INFO [train.py:904] (4/8) Epoch 25, batch 1850, loss[loss=0.1616, simple_loss=0.2613, pruned_loss=0.03098, over 17130.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2545, pruned_loss=0.0404, over 3314000.92 frames. ], batch size: 48, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:08:17,021 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.04 vs. limit=5.0 2023-05-01 23:08:23,447 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=245460.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:08:39,325 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.87 vs. limit=5.0 2023-05-01 23:09:11,856 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6100, 3.8318, 3.8535, 1.9797, 3.0872, 2.2296, 3.9217, 4.1078], device='cuda:4'), covar=tensor([0.0245, 0.0884, 0.0655, 0.2524, 0.1065, 0.1333, 0.0614, 0.1065], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0167, 0.0170, 0.0156, 0.0147, 0.0132, 0.0145, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 23:09:23,472 INFO [train.py:904] (4/8) Epoch 25, batch 1900, loss[loss=0.1585, simple_loss=0.2427, pruned_loss=0.03715, over 16835.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.254, pruned_loss=0.03984, over 3301854.79 frames. ], batch size: 42, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:09:31,432 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-05-01 23:09:49,806 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=245521.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:10:11,813 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.421e+02 2.101e+02 2.496e+02 2.952e+02 1.304e+03, threshold=4.992e+02, percent-clipped=2.0 2023-05-01 23:10:20,943 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=245543.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:10:25,728 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=245546.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:10:33,939 INFO [train.py:904] (4/8) Epoch 25, batch 1950, loss[loss=0.1872, simple_loss=0.2657, pruned_loss=0.05435, over 16388.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2544, pruned_loss=0.03952, over 3296788.50 frames. ], batch size: 146, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:11:25,762 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7454, 4.7848, 5.1811, 5.1604, 5.2193, 4.8850, 4.8180, 4.7481], device='cuda:4'), covar=tensor([0.0383, 0.0631, 0.0419, 0.0407, 0.0575, 0.0465, 0.1045, 0.0495], device='cuda:4'), in_proj_covar=tensor([0.0431, 0.0480, 0.0469, 0.0429, 0.0514, 0.0494, 0.0570, 0.0391], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-01 23:11:26,900 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=245591.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:11:41,387 INFO [train.py:904] (4/8) Epoch 25, batch 2000, loss[loss=0.1503, simple_loss=0.2384, pruned_loss=0.03107, over 15946.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2549, pruned_loss=0.03971, over 3284524.02 frames. ], batch size: 35, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:11:49,120 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=245607.0, num_to_drop=1, layers_to_drop={3} 2023-05-01 23:11:53,225 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=245610.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:12:05,610 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0694, 2.1202, 2.2943, 3.6317, 2.1909, 2.4203, 2.2682, 2.2787], device='cuda:4'), covar=tensor([0.1624, 0.4068, 0.3304, 0.0799, 0.4180, 0.2791, 0.3887, 0.3342], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0465, 0.0383, 0.0337, 0.0445, 0.0532, 0.0437, 0.0544], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:12:31,548 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.417e+02 2.177e+02 2.547e+02 3.013e+02 5.081e+02, threshold=5.093e+02, percent-clipped=1.0 2023-05-01 23:12:50,396 INFO [train.py:904] (4/8) Epoch 25, batch 2050, loss[loss=0.1753, simple_loss=0.2733, pruned_loss=0.03867, over 17049.00 frames. ], tot_loss[loss=0.167, simple_loss=0.255, pruned_loss=0.03952, over 3291325.25 frames. ], batch size: 53, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:12:57,322 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=245658.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:13:34,836 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=245686.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 23:13:58,102 INFO [train.py:904] (4/8) Epoch 25, batch 2100, loss[loss=0.1519, simple_loss=0.2433, pruned_loss=0.03025, over 17215.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2559, pruned_loss=0.03986, over 3290710.62 frames. ], batch size: 44, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:14:12,452 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7317, 3.7687, 2.3660, 4.1385, 2.9786, 4.0554, 2.5250, 3.1008], device='cuda:4'), covar=tensor([0.0302, 0.0398, 0.1578, 0.0367, 0.0786, 0.0706, 0.1392, 0.0737], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0182, 0.0197, 0.0173, 0.0181, 0.0224, 0.0207, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 23:14:49,018 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.544e+02 2.201e+02 2.533e+02 2.997e+02 6.005e+02, threshold=5.066e+02, percent-clipped=2.0 2023-05-01 23:14:59,117 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=245747.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 23:15:08,425 INFO [train.py:904] (4/8) Epoch 25, batch 2150, loss[loss=0.169, simple_loss=0.2495, pruned_loss=0.04421, over 16788.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2563, pruned_loss=0.04073, over 3287004.97 frames. ], batch size: 124, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:16:00,543 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=245792.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:16:15,987 INFO [train.py:904] (4/8) Epoch 25, batch 2200, loss[loss=0.1813, simple_loss=0.2697, pruned_loss=0.04644, over 16302.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2565, pruned_loss=0.0411, over 3292752.96 frames. ], batch size: 165, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:16:34,700 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=245816.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:17:06,517 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-01 23:17:06,932 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.642e+02 2.208e+02 2.555e+02 2.978e+02 8.197e+02, threshold=5.110e+02, percent-clipped=1.0 2023-05-01 23:17:15,213 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-05-01 23:17:23,656 INFO [train.py:904] (4/8) Epoch 25, batch 2250, loss[loss=0.1448, simple_loss=0.2234, pruned_loss=0.03312, over 16759.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2564, pruned_loss=0.04118, over 3292493.26 frames. ], batch size: 89, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:17:24,730 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=245853.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:17:39,333 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0020, 4.4462, 3.0970, 2.4599, 2.7143, 2.6187, 4.7929, 3.6257], device='cuda:4'), covar=tensor([0.2923, 0.0573, 0.1938, 0.3094, 0.3084, 0.2217, 0.0362, 0.1461], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0273, 0.0311, 0.0320, 0.0302, 0.0271, 0.0301, 0.0348], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 23:17:40,532 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5625, 2.6435, 2.1365, 2.4532, 2.9602, 2.6653, 3.1039, 3.1448], device='cuda:4'), covar=tensor([0.0182, 0.0459, 0.0623, 0.0474, 0.0296, 0.0427, 0.0322, 0.0301], device='cuda:4'), in_proj_covar=tensor([0.0228, 0.0248, 0.0235, 0.0237, 0.0248, 0.0247, 0.0248, 0.0244], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:18:16,591 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=245891.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:18:32,174 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=245902.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 23:18:32,939 INFO [train.py:904] (4/8) Epoch 25, batch 2300, loss[loss=0.1728, simple_loss=0.2744, pruned_loss=0.03555, over 17274.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2572, pruned_loss=0.0412, over 3294698.44 frames. ], batch size: 52, lr: 2.71e-03, grad_scale: 2.0 2023-05-01 23:19:24,138 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.666e+02 2.182e+02 2.492e+02 2.872e+02 5.312e+02, threshold=4.984e+02, percent-clipped=2.0 2023-05-01 23:19:42,181 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=245952.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:19:42,915 INFO [train.py:904] (4/8) Epoch 25, batch 2350, loss[loss=0.1736, simple_loss=0.2517, pruned_loss=0.04774, over 16441.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2571, pruned_loss=0.04136, over 3301101.23 frames. ], batch size: 146, lr: 2.71e-03, grad_scale: 2.0 2023-05-01 23:20:54,771 INFO [train.py:904] (4/8) Epoch 25, batch 2400, loss[loss=0.1524, simple_loss=0.2524, pruned_loss=0.02622, over 17137.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2585, pruned_loss=0.04195, over 3295693.40 frames. ], batch size: 47, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:21:14,190 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4344, 2.3313, 2.4251, 4.1920, 2.2763, 2.7222, 2.4210, 2.4990], device='cuda:4'), covar=tensor([0.1348, 0.3823, 0.3228, 0.0612, 0.4315, 0.2762, 0.3610, 0.3753], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0466, 0.0384, 0.0338, 0.0445, 0.0533, 0.0438, 0.0545], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:21:20,017 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2733, 2.2980, 2.4468, 3.9830, 2.3434, 2.6789, 2.3756, 2.4864], device='cuda:4'), covar=tensor([0.1552, 0.3848, 0.3060, 0.0682, 0.3979, 0.2659, 0.4063, 0.3132], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0466, 0.0384, 0.0338, 0.0445, 0.0533, 0.0438, 0.0546], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:21:46,453 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.525e+02 2.068e+02 2.493e+02 3.230e+02 1.350e+03, threshold=4.987e+02, percent-clipped=5.0 2023-05-01 23:21:49,227 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=246042.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 23:22:04,193 INFO [train.py:904] (4/8) Epoch 25, batch 2450, loss[loss=0.1483, simple_loss=0.2376, pruned_loss=0.02948, over 16829.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2588, pruned_loss=0.04157, over 3299402.48 frames. ], batch size: 42, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:23:10,958 INFO [train.py:904] (4/8) Epoch 25, batch 2500, loss[loss=0.1505, simple_loss=0.2392, pruned_loss=0.03096, over 17170.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2583, pruned_loss=0.04096, over 3307730.39 frames. ], batch size: 46, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:23:21,509 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=246111.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:23:28,038 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=246116.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:24:01,581 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 2.119e+02 2.579e+02 3.197e+02 6.008e+02, threshold=5.158e+02, percent-clipped=3.0 2023-05-01 23:24:12,503 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=246148.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:24:18,464 INFO [train.py:904] (4/8) Epoch 25, batch 2550, loss[loss=0.1725, simple_loss=0.2653, pruned_loss=0.03987, over 16719.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2581, pruned_loss=0.04078, over 3311426.95 frames. ], batch size: 62, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:24:35,188 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=246164.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:24:45,453 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=246172.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:25:17,781 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=246196.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 23:25:19,026 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8856, 4.8395, 4.6492, 3.5121, 4.7866, 1.7119, 4.3818, 4.3439], device='cuda:4'), covar=tensor([0.0184, 0.0179, 0.0356, 0.0822, 0.0179, 0.3687, 0.0249, 0.0460], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0170, 0.0210, 0.0185, 0.0187, 0.0216, 0.0199, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:25:26,253 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=246202.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 23:25:26,988 INFO [train.py:904] (4/8) Epoch 25, batch 2600, loss[loss=0.1703, simple_loss=0.2672, pruned_loss=0.03669, over 16618.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2574, pruned_loss=0.0402, over 3310253.82 frames. ], batch size: 62, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:25:52,175 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=246221.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:26:18,783 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 2.097e+02 2.510e+02 3.006e+02 6.140e+02, threshold=5.021e+02, percent-clipped=2.0 2023-05-01 23:26:28,307 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=246247.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:26:33,162 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=246250.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:26:37,197 INFO [train.py:904] (4/8) Epoch 25, batch 2650, loss[loss=0.1614, simple_loss=0.2584, pruned_loss=0.03218, over 17039.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2582, pruned_loss=0.04042, over 3308310.27 frames. ], batch size: 50, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:26:42,343 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=246257.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 23:27:17,679 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=246282.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:27:44,599 INFO [train.py:904] (4/8) Epoch 25, batch 2700, loss[loss=0.1803, simple_loss=0.2848, pruned_loss=0.0379, over 16734.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2582, pruned_loss=0.03985, over 3301140.75 frames. ], batch size: 57, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:28:34,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.322e+02 1.968e+02 2.327e+02 2.625e+02 5.550e+02, threshold=4.654e+02, percent-clipped=1.0 2023-05-01 23:28:36,943 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=246342.0, num_to_drop=1, layers_to_drop={1} 2023-05-01 23:28:52,088 INFO [train.py:904] (4/8) Epoch 25, batch 2750, loss[loss=0.1832, simple_loss=0.2575, pruned_loss=0.05445, over 16796.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2584, pruned_loss=0.03994, over 3300014.30 frames. ], batch size: 124, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:29:44,292 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=246390.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 23:30:01,769 INFO [train.py:904] (4/8) Epoch 25, batch 2800, loss[loss=0.1418, simple_loss=0.229, pruned_loss=0.02732, over 16782.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2571, pruned_loss=0.03962, over 3307160.80 frames. ], batch size: 39, lr: 2.71e-03, grad_scale: 8.0 2023-05-01 23:30:54,340 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 1.989e+02 2.330e+02 2.808e+02 5.234e+02, threshold=4.661e+02, percent-clipped=4.0 2023-05-01 23:31:04,185 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=246448.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:31:10,740 INFO [train.py:904] (4/8) Epoch 25, batch 2850, loss[loss=0.1742, simple_loss=0.2606, pruned_loss=0.0439, over 16470.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2556, pruned_loss=0.03886, over 3316842.90 frames. ], batch size: 68, lr: 2.71e-03, grad_scale: 4.0 2023-05-01 23:31:31,880 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=246467.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:31:36,921 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-05-01 23:32:10,938 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=246496.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:32:20,906 INFO [train.py:904] (4/8) Epoch 25, batch 2900, loss[loss=0.1423, simple_loss=0.2244, pruned_loss=0.03014, over 16702.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2544, pruned_loss=0.0393, over 3324570.70 frames. ], batch size: 89, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:33:14,957 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.640e+02 2.258e+02 2.678e+02 3.242e+02 4.768e+02, threshold=5.357e+02, percent-clipped=1.0 2023-05-01 23:33:25,339 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=246547.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:33:32,607 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=246552.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 23:33:33,511 INFO [train.py:904] (4/8) Epoch 25, batch 2950, loss[loss=0.1635, simple_loss=0.2596, pruned_loss=0.03368, over 17092.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2546, pruned_loss=0.03998, over 3328058.40 frames. ], batch size: 47, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:34:06,239 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=246577.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:34:32,301 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=246595.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:34:36,917 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9423, 5.2284, 4.9914, 5.0208, 4.7975, 4.7563, 4.6490, 5.3152], device='cuda:4'), covar=tensor([0.1307, 0.0943, 0.1084, 0.0881, 0.0848, 0.1047, 0.1283, 0.0898], device='cuda:4'), in_proj_covar=tensor([0.0723, 0.0875, 0.0716, 0.0678, 0.0556, 0.0555, 0.0736, 0.0686], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:34:41,071 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-01 23:34:42,716 INFO [train.py:904] (4/8) Epoch 25, batch 3000, loss[loss=0.1751, simple_loss=0.2558, pruned_loss=0.04722, over 16753.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2545, pruned_loss=0.04032, over 3330613.85 frames. ], batch size: 83, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:34:42,716 INFO [train.py:929] (4/8) Computing validation loss 2023-05-01 23:34:52,580 INFO [train.py:938] (4/8) Epoch 25, validation: loss=0.1341, simple_loss=0.239, pruned_loss=0.01457, over 944034.00 frames. 2023-05-01 23:34:52,581 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-01 23:35:46,635 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.472e+02 2.065e+02 2.466e+02 2.875e+02 5.514e+02, threshold=4.932e+02, percent-clipped=1.0 2023-05-01 23:36:05,621 INFO [train.py:904] (4/8) Epoch 25, batch 3050, loss[loss=0.1845, simple_loss=0.2571, pruned_loss=0.05593, over 16873.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.255, pruned_loss=0.04114, over 3324403.73 frames. ], batch size: 116, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:36:38,436 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=246677.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:37:13,556 INFO [train.py:904] (4/8) Epoch 25, batch 3100, loss[loss=0.1671, simple_loss=0.2586, pruned_loss=0.0378, over 17061.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2543, pruned_loss=0.04071, over 3329884.47 frames. ], batch size: 55, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:37:32,714 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.97 vs. limit=5.0 2023-05-01 23:38:04,953 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=246738.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:38:08,780 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.354e+02 2.159e+02 2.563e+02 3.074e+02 4.728e+02, threshold=5.126e+02, percent-clipped=0.0 2023-05-01 23:38:19,712 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-05-01 23:38:25,015 INFO [train.py:904] (4/8) Epoch 25, batch 3150, loss[loss=0.2009, simple_loss=0.2878, pruned_loss=0.05701, over 12282.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2539, pruned_loss=0.03984, over 3332271.75 frames. ], batch size: 247, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:38:37,708 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8929, 2.1844, 2.4835, 2.8578, 2.8013, 3.3805, 2.2214, 3.4609], device='cuda:4'), covar=tensor([0.0354, 0.0546, 0.0437, 0.0410, 0.0422, 0.0266, 0.0620, 0.0202], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0197, 0.0185, 0.0189, 0.0204, 0.0163, 0.0201, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:38:39,273 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-01 23:38:44,293 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=246767.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:39:34,236 INFO [train.py:904] (4/8) Epoch 25, batch 3200, loss[loss=0.1572, simple_loss=0.2407, pruned_loss=0.03687, over 16828.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.253, pruned_loss=0.03974, over 3323708.05 frames. ], batch size: 102, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:39:51,752 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=246815.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:40:01,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7699, 2.7208, 2.4726, 2.6221, 3.0541, 2.8414, 3.3237, 3.2171], device='cuda:4'), covar=tensor([0.0150, 0.0463, 0.0521, 0.0495, 0.0305, 0.0414, 0.0328, 0.0317], device='cuda:4'), in_proj_covar=tensor([0.0229, 0.0247, 0.0234, 0.0237, 0.0248, 0.0246, 0.0249, 0.0245], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:40:27,532 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.603e+02 2.068e+02 2.379e+02 2.953e+02 6.702e+02, threshold=4.759e+02, percent-clipped=1.0 2023-05-01 23:40:42,117 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=246852.0, num_to_drop=1, layers_to_drop={2} 2023-05-01 23:40:42,934 INFO [train.py:904] (4/8) Epoch 25, batch 3250, loss[loss=0.1807, simple_loss=0.2571, pruned_loss=0.05218, over 16855.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2522, pruned_loss=0.03932, over 3322578.93 frames. ], batch size: 116, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:41:16,794 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=246877.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:41:49,849 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=246900.0, num_to_drop=1, layers_to_drop={0} 2023-05-01 23:41:53,758 INFO [train.py:904] (4/8) Epoch 25, batch 3300, loss[loss=0.1524, simple_loss=0.2389, pruned_loss=0.033, over 16877.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2547, pruned_loss=0.04052, over 3312399.23 frames. ], batch size: 96, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:42:24,635 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=246925.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:42:36,857 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=246934.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:42:46,634 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.501e+02 2.089e+02 2.401e+02 2.822e+02 3.775e+02, threshold=4.802e+02, percent-clipped=0.0 2023-05-01 23:43:02,686 INFO [train.py:904] (4/8) Epoch 25, batch 3350, loss[loss=0.1603, simple_loss=0.2491, pruned_loss=0.03573, over 16733.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2554, pruned_loss=0.04055, over 3307926.09 frames. ], batch size: 102, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:43:26,723 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.09 vs. limit=2.0 2023-05-01 23:44:01,663 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=246995.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:44:13,546 INFO [train.py:904] (4/8) Epoch 25, batch 3400, loss[loss=0.2022, simple_loss=0.2779, pruned_loss=0.06327, over 15573.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2556, pruned_loss=0.0409, over 3305463.79 frames. ], batch size: 190, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:44:42,236 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4310, 2.2734, 1.9308, 2.1308, 2.6123, 2.4021, 2.4254, 2.6935], device='cuda:4'), covar=tensor([0.0285, 0.0496, 0.0575, 0.0513, 0.0267, 0.0375, 0.0252, 0.0331], device='cuda:4'), in_proj_covar=tensor([0.0230, 0.0248, 0.0235, 0.0238, 0.0249, 0.0247, 0.0250, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:44:56,893 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=247033.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:45:08,919 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.556e+02 2.118e+02 2.456e+02 2.904e+02 5.980e+02, threshold=4.912e+02, percent-clipped=3.0 2023-05-01 23:45:13,117 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7920, 2.0934, 2.4910, 2.7852, 2.7266, 3.3065, 2.3484, 3.3012], device='cuda:4'), covar=tensor([0.0345, 0.0565, 0.0384, 0.0399, 0.0384, 0.0228, 0.0545, 0.0181], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0196, 0.0184, 0.0189, 0.0205, 0.0163, 0.0201, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:45:26,146 INFO [train.py:904] (4/8) Epoch 25, batch 3450, loss[loss=0.1467, simple_loss=0.2345, pruned_loss=0.02944, over 15907.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2544, pruned_loss=0.04018, over 3300186.22 frames. ], batch size: 35, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:45:34,106 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1636, 4.4937, 4.4315, 3.4208, 3.6840, 4.3755, 3.9821, 2.7744], device='cuda:4'), covar=tensor([0.0447, 0.0059, 0.0055, 0.0319, 0.0153, 0.0109, 0.0098, 0.0450], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0089, 0.0090, 0.0137, 0.0102, 0.0113, 0.0099, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 23:45:48,421 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7531, 3.7459, 2.9303, 2.3012, 2.4874, 2.4854, 3.8753, 3.3366], device='cuda:4'), covar=tensor([0.2794, 0.0689, 0.1857, 0.3057, 0.2825, 0.2211, 0.0544, 0.1708], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0275, 0.0313, 0.0322, 0.0305, 0.0272, 0.0303, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-01 23:45:56,726 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0880, 4.7733, 5.1287, 5.2974, 5.5150, 4.8073, 5.4720, 5.4899], device='cuda:4'), covar=tensor([0.1779, 0.1418, 0.1837, 0.0888, 0.0570, 0.0826, 0.0572, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0692, 0.0855, 0.0986, 0.0863, 0.0661, 0.0682, 0.0707, 0.0828], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:46:35,593 INFO [train.py:904] (4/8) Epoch 25, batch 3500, loss[loss=0.1636, simple_loss=0.2452, pruned_loss=0.04095, over 16858.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.253, pruned_loss=0.03921, over 3311053.25 frames. ], batch size: 96, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:46:37,854 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6899, 3.6559, 3.9270, 2.8539, 3.5852, 3.9933, 3.6914, 2.4451], device='cuda:4'), covar=tensor([0.0489, 0.0253, 0.0062, 0.0377, 0.0108, 0.0096, 0.0100, 0.0443], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0089, 0.0089, 0.0136, 0.0102, 0.0113, 0.0099, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 23:47:30,016 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.591e+02 2.124e+02 2.460e+02 2.748e+02 9.578e+02, threshold=4.920e+02, percent-clipped=1.0 2023-05-01 23:47:45,775 INFO [train.py:904] (4/8) Epoch 25, batch 3550, loss[loss=0.1783, simple_loss=0.2793, pruned_loss=0.03866, over 16664.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2519, pruned_loss=0.03893, over 3314592.82 frames. ], batch size: 62, lr: 2.70e-03, grad_scale: 4.0 2023-05-01 23:48:22,212 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-01 23:48:35,158 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6996, 6.1091, 5.7547, 5.8713, 5.4557, 5.5437, 5.4057, 6.2126], device='cuda:4'), covar=tensor([0.1488, 0.0946, 0.1134, 0.1015, 0.0940, 0.0677, 0.1299, 0.1027], device='cuda:4'), in_proj_covar=tensor([0.0732, 0.0882, 0.0725, 0.0684, 0.0560, 0.0562, 0.0744, 0.0693], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:48:55,167 INFO [train.py:904] (4/8) Epoch 25, batch 3600, loss[loss=0.1589, simple_loss=0.2538, pruned_loss=0.03201, over 17130.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2503, pruned_loss=0.0384, over 3312248.08 frames. ], batch size: 48, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:49:34,988 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-05-01 23:49:49,109 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.664e+02 2.145e+02 2.569e+02 3.333e+02 7.279e+02, threshold=5.139e+02, percent-clipped=3.0 2023-05-01 23:50:05,473 INFO [train.py:904] (4/8) Epoch 25, batch 3650, loss[loss=0.1457, simple_loss=0.2328, pruned_loss=0.02928, over 15439.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2493, pruned_loss=0.03926, over 3306487.32 frames. ], batch size: 191, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:50:53,397 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-05-01 23:50:59,456 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=247290.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:51:17,357 INFO [train.py:904] (4/8) Epoch 25, batch 3700, loss[loss=0.189, simple_loss=0.2602, pruned_loss=0.0589, over 16347.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2483, pruned_loss=0.04078, over 3270586.50 frames. ], batch size: 165, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:52:01,354 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=247333.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:52:12,255 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7188, 5.0401, 4.7975, 4.8165, 4.6245, 4.5287, 4.4598, 5.1062], device='cuda:4'), covar=tensor([0.1267, 0.0860, 0.0960, 0.0875, 0.0788, 0.1275, 0.1190, 0.0905], device='cuda:4'), in_proj_covar=tensor([0.0727, 0.0876, 0.0719, 0.0680, 0.0555, 0.0558, 0.0738, 0.0687], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:52:12,944 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.806e+02 2.292e+02 2.615e+02 3.085e+02 5.082e+02, threshold=5.229e+02, percent-clipped=0.0 2023-05-01 23:52:30,272 INFO [train.py:904] (4/8) Epoch 25, batch 3750, loss[loss=0.1788, simple_loss=0.2629, pruned_loss=0.04731, over 15599.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2485, pruned_loss=0.04219, over 3255929.94 frames. ], batch size: 191, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:52:46,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9601, 2.9542, 2.5998, 4.6367, 3.7394, 4.2069, 1.8299, 3.1411], device='cuda:4'), covar=tensor([0.1316, 0.0724, 0.1235, 0.0182, 0.0280, 0.0419, 0.1607, 0.0858], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0179, 0.0197, 0.0198, 0.0206, 0.0219, 0.0206, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 23:53:02,524 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-01 23:53:06,784 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=247379.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:53:10,239 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=247381.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:53:42,915 INFO [train.py:904] (4/8) Epoch 25, batch 3800, loss[loss=0.164, simple_loss=0.2446, pruned_loss=0.0417, over 16439.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.25, pruned_loss=0.04315, over 3260066.97 frames. ], batch size: 146, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:54:07,802 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8184, 4.9832, 5.1623, 4.9453, 5.0219, 5.5811, 5.1109, 4.7644], device='cuda:4'), covar=tensor([0.1363, 0.1926, 0.2083, 0.1953, 0.2572, 0.1029, 0.1515, 0.2574], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0637, 0.0694, 0.0516, 0.0688, 0.0723, 0.0540, 0.0689], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-01 23:54:16,003 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=247426.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:54:36,489 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=247440.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:54:38,774 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.614e+02 2.181e+02 2.707e+02 3.148e+02 5.686e+02, threshold=5.414e+02, percent-clipped=1.0 2023-05-01 23:54:54,710 INFO [train.py:904] (4/8) Epoch 25, batch 3850, loss[loss=0.1592, simple_loss=0.2348, pruned_loss=0.04181, over 16682.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2505, pruned_loss=0.04388, over 3261907.75 frames. ], batch size: 83, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:54:55,629 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.81 vs. limit=5.0 2023-05-01 23:55:25,063 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-01 23:55:43,360 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=247487.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:56:02,932 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=247500.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:56:06,077 INFO [train.py:904] (4/8) Epoch 25, batch 3900, loss[loss=0.1509, simple_loss=0.2331, pruned_loss=0.03433, over 16267.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2504, pruned_loss=0.0444, over 3260391.53 frames. ], batch size: 165, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:56:39,822 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=247525.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:56:54,260 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-01 23:56:55,338 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8657, 2.7520, 2.6678, 1.8867, 2.6653, 2.7860, 2.6398, 1.8980], device='cuda:4'), covar=tensor([0.0483, 0.0101, 0.0089, 0.0418, 0.0121, 0.0138, 0.0136, 0.0416], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0136, 0.0101, 0.0112, 0.0098, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-01 23:57:03,124 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.630e+02 2.139e+02 2.490e+02 2.933e+02 5.052e+02, threshold=4.980e+02, percent-clipped=0.0 2023-05-01 23:57:17,841 INFO [train.py:904] (4/8) Epoch 25, batch 3950, loss[loss=0.1795, simple_loss=0.2551, pruned_loss=0.05198, over 16766.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2504, pruned_loss=0.04522, over 3257285.21 frames. ], batch size: 83, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:57:19,951 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=247554.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:57:30,908 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=247561.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:57:34,141 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3697, 4.6790, 4.4588, 4.4740, 4.2177, 4.1768, 4.1802, 4.7342], device='cuda:4'), covar=tensor([0.1307, 0.0829, 0.1044, 0.0917, 0.0887, 0.1581, 0.1167, 0.0843], device='cuda:4'), in_proj_covar=tensor([0.0727, 0.0876, 0.0722, 0.0680, 0.0558, 0.0559, 0.0742, 0.0686], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-01 23:58:06,006 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=247586.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:58:12,096 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=247590.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:58:30,186 INFO [train.py:904] (4/8) Epoch 25, batch 4000, loss[loss=0.1953, simple_loss=0.2569, pruned_loss=0.06682, over 16749.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2505, pruned_loss=0.04564, over 3255013.57 frames. ], batch size: 124, lr: 2.70e-03, grad_scale: 8.0 2023-05-01 23:58:42,457 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6905, 3.4186, 3.8894, 1.9794, 3.9910, 4.0185, 3.1096, 3.0102], device='cuda:4'), covar=tensor([0.0805, 0.0287, 0.0169, 0.1258, 0.0094, 0.0189, 0.0426, 0.0480], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0111, 0.0100, 0.0139, 0.0084, 0.0130, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-01 23:58:49,601 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=247615.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:59:21,866 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=247638.0, num_to_drop=0, layers_to_drop=set() 2023-05-01 23:59:27,471 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.424e+02 1.992e+02 2.395e+02 2.922e+02 5.946e+02, threshold=4.790e+02, percent-clipped=1.0 2023-05-01 23:59:44,160 INFO [train.py:904] (4/8) Epoch 25, batch 4050, loss[loss=0.1598, simple_loss=0.2441, pruned_loss=0.03776, over 17015.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2515, pruned_loss=0.04497, over 3252823.32 frames. ], batch size: 55, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:00:50,902 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5646, 3.6276, 2.6719, 2.3619, 2.3949, 2.4444, 3.8789, 3.2981], device='cuda:4'), covar=tensor([0.3118, 0.0674, 0.2033, 0.2684, 0.2574, 0.2096, 0.0487, 0.1243], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0273, 0.0311, 0.0319, 0.0304, 0.0270, 0.0302, 0.0348], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 00:00:59,188 INFO [train.py:904] (4/8) Epoch 25, batch 4100, loss[loss=0.2134, simple_loss=0.3017, pruned_loss=0.06257, over 16555.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.253, pruned_loss=0.04434, over 3257938.02 frames. ], batch size: 35, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:01:48,426 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=247735.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:01:57,927 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.293e+02 2.110e+02 2.475e+02 2.947e+02 8.069e+02, threshold=4.949e+02, percent-clipped=1.0 2023-05-02 00:02:14,951 INFO [train.py:904] (4/8) Epoch 25, batch 4150, loss[loss=0.2284, simple_loss=0.3089, pruned_loss=0.07391, over 11388.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2602, pruned_loss=0.04662, over 3236707.75 frames. ], batch size: 250, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:02:35,809 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1471, 4.0407, 4.2414, 4.3610, 4.4590, 4.0533, 4.4100, 4.5065], device='cuda:4'), covar=tensor([0.1637, 0.1213, 0.1298, 0.0673, 0.0564, 0.1281, 0.0859, 0.0663], device='cuda:4'), in_proj_covar=tensor([0.0678, 0.0834, 0.0960, 0.0842, 0.0645, 0.0664, 0.0692, 0.0808], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:02:43,656 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-02 00:03:01,850 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=247782.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:03:04,903 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=247784.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:03:32,729 INFO [train.py:904] (4/8) Epoch 25, batch 4200, loss[loss=0.2188, simple_loss=0.296, pruned_loss=0.07075, over 11534.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2668, pruned_loss=0.04816, over 3216150.53 frames. ], batch size: 246, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:04:30,297 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.539e+02 2.155e+02 2.612e+02 3.147e+02 5.112e+02, threshold=5.224e+02, percent-clipped=1.0 2023-05-02 00:04:35,236 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=247845.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:04:46,252 INFO [train.py:904] (4/8) Epoch 25, batch 4250, loss[loss=0.1771, simple_loss=0.2733, pruned_loss=0.04039, over 16785.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2709, pruned_loss=0.0484, over 3175988.08 frames. ], batch size: 39, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:04:52,105 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=247856.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:05:28,692 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=247881.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:05:36,025 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4978, 5.7948, 5.5304, 5.6156, 5.2731, 5.1426, 5.1141, 5.9172], device='cuda:4'), covar=tensor([0.1183, 0.0778, 0.1000, 0.0811, 0.0793, 0.0649, 0.1180, 0.0802], device='cuda:4'), in_proj_covar=tensor([0.0712, 0.0861, 0.0708, 0.0667, 0.0546, 0.0548, 0.0725, 0.0675], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:06:02,532 INFO [train.py:904] (4/8) Epoch 25, batch 4300, loss[loss=0.2019, simple_loss=0.2853, pruned_loss=0.05927, over 11917.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.272, pruned_loss=0.04734, over 3180363.49 frames. ], batch size: 248, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:06:10,516 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3919, 2.6106, 2.5291, 4.2478, 2.4269, 2.8959, 2.6255, 2.7021], device='cuda:4'), covar=tensor([0.1326, 0.3087, 0.2703, 0.0535, 0.3706, 0.2269, 0.2966, 0.3059], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0464, 0.0380, 0.0335, 0.0442, 0.0532, 0.0436, 0.0543], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:06:13,452 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=247910.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:06:15,891 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-05-02 00:06:57,363 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2242, 3.8689, 3.7330, 2.3574, 3.4510, 3.8392, 3.4307, 2.1948], device='cuda:4'), covar=tensor([0.0577, 0.0040, 0.0063, 0.0464, 0.0103, 0.0089, 0.0107, 0.0464], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0087, 0.0088, 0.0134, 0.0100, 0.0111, 0.0097, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 00:07:01,390 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.557e+02 2.163e+02 2.477e+02 3.039e+02 5.024e+02, threshold=4.955e+02, percent-clipped=0.0 2023-05-02 00:07:17,409 INFO [train.py:904] (4/8) Epoch 25, batch 4350, loss[loss=0.1893, simple_loss=0.2817, pruned_loss=0.04844, over 16337.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2747, pruned_loss=0.04814, over 3182936.81 frames. ], batch size: 165, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:08:05,121 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=247984.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:08:23,840 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=247996.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:08:36,577 INFO [train.py:904] (4/8) Epoch 25, batch 4400, loss[loss=0.1986, simple_loss=0.2886, pruned_loss=0.05429, over 16260.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2772, pruned_loss=0.04961, over 3169974.61 frames. ], batch size: 165, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:09:24,794 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248035.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:09:34,457 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 1.997e+02 2.337e+02 2.699e+02 4.708e+02, threshold=4.674e+02, percent-clipped=0.0 2023-05-02 00:09:39,323 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=248045.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:09:50,303 INFO [train.py:904] (4/8) Epoch 25, batch 4450, loss[loss=0.1971, simple_loss=0.2839, pruned_loss=0.05516, over 16720.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2808, pruned_loss=0.05097, over 3187952.51 frames. ], batch size: 62, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:09:56,453 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=248057.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:10:33,113 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248082.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:10:33,180 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=248082.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:10:34,920 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248083.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:11:04,506 INFO [train.py:904] (4/8) Epoch 25, batch 4500, loss[loss=0.2037, simple_loss=0.2758, pruned_loss=0.06576, over 11574.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.2812, pruned_loss=0.05185, over 3196752.27 frames. ], batch size: 246, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:11:44,503 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248130.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:11:59,979 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=248140.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:12:01,968 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.331e+02 1.873e+02 2.071e+02 2.399e+02 5.095e+02, threshold=4.142e+02, percent-clipped=1.0 2023-05-02 00:12:04,306 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=248143.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:12:17,508 INFO [train.py:904] (4/8) Epoch 25, batch 4550, loss[loss=0.2021, simple_loss=0.289, pruned_loss=0.05757, over 16258.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.2823, pruned_loss=0.05277, over 3210613.10 frames. ], batch size: 165, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:12:22,806 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248156.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:13:00,067 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248181.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:13:32,545 INFO [train.py:904] (4/8) Epoch 25, batch 4600, loss[loss=0.1725, simple_loss=0.2662, pruned_loss=0.03936, over 17172.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2825, pruned_loss=0.05295, over 3202786.22 frames. ], batch size: 46, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:13:34,235 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248204.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:13:42,788 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248210.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:14:11,936 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248229.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:14:29,191 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 1.740e+02 1.903e+02 2.276e+02 6.899e+02, threshold=3.806e+02, percent-clipped=0.0 2023-05-02 00:14:46,950 INFO [train.py:904] (4/8) Epoch 25, batch 4650, loss[loss=0.1802, simple_loss=0.2621, pruned_loss=0.04911, over 17065.00 frames. ], tot_loss[loss=0.1945, simple_loss=0.2824, pruned_loss=0.0533, over 3199687.71 frames. ], batch size: 53, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:14:54,068 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248258.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:16:01,798 INFO [train.py:904] (4/8) Epoch 25, batch 4700, loss[loss=0.1821, simple_loss=0.2639, pruned_loss=0.05018, over 16320.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.2795, pruned_loss=0.05175, over 3208718.78 frames. ], batch size: 35, lr: 2.70e-03, grad_scale: 8.0 2023-05-02 00:16:35,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6827, 2.7996, 2.5424, 4.8675, 3.6353, 4.2160, 1.6155, 2.8713], device='cuda:4'), covar=tensor([0.1430, 0.0830, 0.1353, 0.0151, 0.0294, 0.0355, 0.1749, 0.0958], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0178, 0.0197, 0.0195, 0.0205, 0.0217, 0.0206, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 00:16:44,307 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5041, 3.5116, 3.4621, 2.5870, 3.2997, 2.1183, 3.1053, 2.6938], device='cuda:4'), covar=tensor([0.0205, 0.0223, 0.0189, 0.0313, 0.0134, 0.2617, 0.0154, 0.0346], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0168, 0.0209, 0.0185, 0.0186, 0.0214, 0.0198, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:16:44,341 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=248332.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:16:56,010 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=248340.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:16:57,976 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 1.897e+02 2.230e+02 2.532e+02 4.236e+02, threshold=4.460e+02, percent-clipped=2.0 2023-05-02 00:17:13,840 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=248352.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:17:14,735 INFO [train.py:904] (4/8) Epoch 25, batch 4750, loss[loss=0.1489, simple_loss=0.2378, pruned_loss=0.02995, over 17227.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2753, pruned_loss=0.04963, over 3209072.75 frames. ], batch size: 44, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:17:20,286 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6904, 3.5697, 4.0179, 1.8779, 4.2472, 4.2329, 3.1667, 3.2192], device='cuda:4'), covar=tensor([0.0842, 0.0314, 0.0195, 0.1356, 0.0067, 0.0126, 0.0432, 0.0469], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0111, 0.0100, 0.0140, 0.0085, 0.0130, 0.0130, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 00:17:59,847 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=248383.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:18:13,861 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=248393.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:18:29,675 INFO [train.py:904] (4/8) Epoch 25, batch 4800, loss[loss=0.1631, simple_loss=0.2471, pruned_loss=0.03958, over 17085.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2712, pruned_loss=0.04742, over 3212811.91 frames. ], batch size: 55, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:19:22,230 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=248438.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:19:24,762 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248440.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:19:27,480 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.387e+02 1.792e+02 2.090e+02 2.399e+02 3.533e+02, threshold=4.179e+02, percent-clipped=0.0 2023-05-02 00:19:31,516 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=248444.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:19:33,118 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.95 vs. limit=5.0 2023-05-02 00:19:43,833 INFO [train.py:904] (4/8) Epoch 25, batch 4850, loss[loss=0.1919, simple_loss=0.2966, pruned_loss=0.0436, over 16266.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2717, pruned_loss=0.04649, over 3198687.45 frames. ], batch size: 165, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:20:37,397 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248488.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:20:52,042 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.47 vs. limit=2.0 2023-05-02 00:20:53,036 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8913, 2.1818, 2.4249, 3.1117, 2.2204, 2.3668, 2.3302, 2.3364], device='cuda:4'), covar=tensor([0.1502, 0.3353, 0.2552, 0.0769, 0.3899, 0.2441, 0.3583, 0.3059], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0464, 0.0380, 0.0335, 0.0442, 0.0532, 0.0436, 0.0541], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:20:58,428 INFO [train.py:904] (4/8) Epoch 25, batch 4900, loss[loss=0.1741, simple_loss=0.2671, pruned_loss=0.04058, over 16629.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2711, pruned_loss=0.04542, over 3173481.70 frames. ], batch size: 76, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:22:03,445 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 1.924e+02 2.287e+02 2.776e+02 8.047e+02, threshold=4.573e+02, percent-clipped=3.0 2023-05-02 00:22:20,178 INFO [train.py:904] (4/8) Epoch 25, batch 4950, loss[loss=0.188, simple_loss=0.2799, pruned_loss=0.04803, over 16739.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2703, pruned_loss=0.04464, over 3173185.11 frames. ], batch size: 124, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:23:03,258 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.81 vs. limit=5.0 2023-05-02 00:23:31,272 INFO [train.py:904] (4/8) Epoch 25, batch 5000, loss[loss=0.1833, simple_loss=0.2782, pruned_loss=0.04423, over 16694.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2722, pruned_loss=0.04491, over 3188758.63 frames. ], batch size: 134, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:23:44,456 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6523, 3.7212, 2.7888, 2.2887, 2.5347, 2.5485, 4.0770, 3.3543], device='cuda:4'), covar=tensor([0.2936, 0.0697, 0.1965, 0.2692, 0.2545, 0.1939, 0.0475, 0.1277], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0273, 0.0310, 0.0320, 0.0303, 0.0269, 0.0301, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 00:24:25,296 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248640.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:24:27,922 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.446e+02 2.126e+02 2.416e+02 2.928e+02 6.487e+02, threshold=4.832e+02, percent-clipped=1.0 2023-05-02 00:24:42,962 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248652.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:24:43,830 INFO [train.py:904] (4/8) Epoch 25, batch 5050, loss[loss=0.1818, simple_loss=0.279, pruned_loss=0.04237, over 16267.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.273, pruned_loss=0.04465, over 3207509.31 frames. ], batch size: 165, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:24:48,593 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7452, 4.7353, 5.0142, 5.0265, 5.0316, 4.7428, 4.6936, 4.5669], device='cuda:4'), covar=tensor([0.0287, 0.0446, 0.0354, 0.0342, 0.0447, 0.0326, 0.0966, 0.0452], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0463, 0.0454, 0.0414, 0.0498, 0.0474, 0.0552, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 00:24:51,575 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5227, 4.6427, 4.7950, 4.5114, 4.6553, 5.1710, 4.6756, 4.3604], device='cuda:4'), covar=tensor([0.1249, 0.1690, 0.1888, 0.1961, 0.2357, 0.0924, 0.1376, 0.2352], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0614, 0.0669, 0.0500, 0.0668, 0.0702, 0.0524, 0.0670], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 00:25:35,138 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248688.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:25:35,152 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=248688.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:25:39,544 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=248691.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:25:53,427 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248700.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:25:57,288 INFO [train.py:904] (4/8) Epoch 25, batch 5100, loss[loss=0.1791, simple_loss=0.2683, pruned_loss=0.04501, over 16445.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2713, pruned_loss=0.0439, over 3221971.04 frames. ], batch size: 146, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:26:11,045 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-05-02 00:26:15,839 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5797, 4.4993, 4.3647, 2.9368, 3.7566, 4.4407, 3.7731, 2.4370], device='cuda:4'), covar=tensor([0.0579, 0.0037, 0.0041, 0.0397, 0.0098, 0.0084, 0.0115, 0.0470], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0087, 0.0088, 0.0135, 0.0100, 0.0111, 0.0097, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 00:26:15,860 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=248715.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:26:50,064 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248738.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:26:52,083 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=248739.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:26:56,592 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.343e+02 1.974e+02 2.293e+02 2.738e+02 8.128e+02, threshold=4.585e+02, percent-clipped=1.0 2023-05-02 00:27:12,687 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=248752.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:27:13,424 INFO [train.py:904] (4/8) Epoch 25, batch 5150, loss[loss=0.1775, simple_loss=0.2757, pruned_loss=0.03963, over 16215.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2706, pruned_loss=0.04324, over 3200110.89 frames. ], batch size: 165, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:27:13,795 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0349, 5.4567, 5.6482, 5.3382, 5.4760, 5.9867, 5.4616, 5.1744], device='cuda:4'), covar=tensor([0.0866, 0.1538, 0.1877, 0.1922, 0.2159, 0.0786, 0.1339, 0.2361], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0612, 0.0666, 0.0498, 0.0665, 0.0700, 0.0521, 0.0667], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 00:27:47,495 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=248776.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:28:03,429 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=248786.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:28:07,430 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-05-02 00:28:28,598 INFO [train.py:904] (4/8) Epoch 25, batch 5200, loss[loss=0.1614, simple_loss=0.2367, pruned_loss=0.04306, over 16293.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2693, pruned_loss=0.04293, over 3207592.20 frames. ], batch size: 35, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:29:25,199 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.065e+02 2.411e+02 2.935e+02 6.100e+02, threshold=4.822e+02, percent-clipped=2.0 2023-05-02 00:29:41,113 INFO [train.py:904] (4/8) Epoch 25, batch 5250, loss[loss=0.1755, simple_loss=0.2643, pruned_loss=0.04338, over 16609.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2672, pruned_loss=0.04261, over 3211669.57 frames. ], batch size: 62, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:30:00,373 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-05-02 00:30:14,870 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9723, 2.2039, 2.5625, 2.9784, 2.9077, 3.4915, 2.2677, 3.3050], device='cuda:4'), covar=tensor([0.0221, 0.0526, 0.0369, 0.0366, 0.0320, 0.0176, 0.0557, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0194, 0.0182, 0.0187, 0.0201, 0.0160, 0.0199, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:30:53,680 INFO [train.py:904] (4/8) Epoch 25, batch 5300, loss[loss=0.1727, simple_loss=0.254, pruned_loss=0.04568, over 16850.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2635, pruned_loss=0.04131, over 3212236.19 frames. ], batch size: 116, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:31:51,341 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.416e+02 1.862e+02 2.129e+02 2.555e+02 4.739e+02, threshold=4.259e+02, percent-clipped=0.0 2023-05-02 00:32:07,997 INFO [train.py:904] (4/8) Epoch 25, batch 5350, loss[loss=0.1793, simple_loss=0.2766, pruned_loss=0.04098, over 15438.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2621, pruned_loss=0.04092, over 3209930.91 frames. ], batch size: 191, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:32:09,022 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3684, 5.6283, 5.3574, 5.4728, 5.1364, 5.0926, 4.9556, 5.7611], device='cuda:4'), covar=tensor([0.1300, 0.0789, 0.0973, 0.0771, 0.0746, 0.0693, 0.1216, 0.0808], device='cuda:4'), in_proj_covar=tensor([0.0698, 0.0840, 0.0694, 0.0648, 0.0535, 0.0534, 0.0708, 0.0658], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:33:01,226 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=248988.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:33:22,807 INFO [train.py:904] (4/8) Epoch 25, batch 5400, loss[loss=0.1839, simple_loss=0.2874, pruned_loss=0.04019, over 16561.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2642, pruned_loss=0.041, over 3229407.75 frames. ], batch size: 75, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:33:35,808 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-05-02 00:34:11,322 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=249036.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:34:17,282 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=249039.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:34:20,472 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.375e+02 1.925e+02 2.196e+02 2.578e+02 3.732e+02, threshold=4.392e+02, percent-clipped=0.0 2023-05-02 00:34:29,122 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=249047.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:34:40,342 INFO [train.py:904] (4/8) Epoch 25, batch 5450, loss[loss=0.1738, simple_loss=0.2692, pruned_loss=0.03922, over 17185.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.268, pruned_loss=0.04277, over 3222389.00 frames. ], batch size: 44, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:35:08,680 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=249071.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:35:15,018 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-02 00:35:34,926 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=249087.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:35:58,380 INFO [train.py:904] (4/8) Epoch 25, batch 5500, loss[loss=0.2645, simple_loss=0.3302, pruned_loss=0.09937, over 11468.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2756, pruned_loss=0.0478, over 3151420.50 frames. ], batch size: 248, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:37:00,872 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.871e+02 3.023e+02 3.674e+02 4.693e+02 9.094e+02, threshold=7.348e+02, percent-clipped=33.0 2023-05-02 00:37:17,247 INFO [train.py:904] (4/8) Epoch 25, batch 5550, loss[loss=0.2785, simple_loss=0.3381, pruned_loss=0.1095, over 11044.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2823, pruned_loss=0.05173, over 3157414.41 frames. ], batch size: 248, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:38:41,394 INFO [train.py:904] (4/8) Epoch 25, batch 5600, loss[loss=0.2097, simple_loss=0.3021, pruned_loss=0.05868, over 16683.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2862, pruned_loss=0.05558, over 3117752.26 frames. ], batch size: 89, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:38:43,345 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2138, 3.1093, 3.5084, 1.8067, 3.6652, 3.6761, 2.7190, 2.6842], device='cuda:4'), covar=tensor([0.0907, 0.0318, 0.0233, 0.1250, 0.0092, 0.0201, 0.0547, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0112, 0.0101, 0.0141, 0.0086, 0.0131, 0.0131, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 00:39:02,145 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=249215.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:39:06,160 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-05-02 00:39:47,610 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.932e+02 3.336e+02 4.056e+02 4.975e+02 8.511e+02, threshold=8.112e+02, percent-clipped=2.0 2023-05-02 00:40:04,579 INFO [train.py:904] (4/8) Epoch 25, batch 5650, loss[loss=0.2176, simple_loss=0.3004, pruned_loss=0.06738, over 16928.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2914, pruned_loss=0.05985, over 3076590.61 frames. ], batch size: 109, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:40:12,874 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3868, 5.6893, 5.3854, 5.4693, 5.1415, 5.0383, 5.0973, 5.8249], device='cuda:4'), covar=tensor([0.1207, 0.0897, 0.1117, 0.0933, 0.0855, 0.0769, 0.1156, 0.0859], device='cuda:4'), in_proj_covar=tensor([0.0692, 0.0835, 0.0691, 0.0644, 0.0531, 0.0532, 0.0702, 0.0653], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:40:41,238 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=249276.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:41:22,856 INFO [train.py:904] (4/8) Epoch 25, batch 5700, loss[loss=0.1967, simple_loss=0.2883, pruned_loss=0.05258, over 15548.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2935, pruned_loss=0.06144, over 3063783.76 frames. ], batch size: 191, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:41:55,372 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8274, 3.8854, 4.1053, 4.0796, 4.1106, 3.8860, 3.8991, 3.9381], device='cuda:4'), covar=tensor([0.0373, 0.0595, 0.0465, 0.0472, 0.0491, 0.0485, 0.0964, 0.0526], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0470, 0.0459, 0.0418, 0.0502, 0.0480, 0.0559, 0.0382], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 00:42:25,335 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.887e+02 2.916e+02 3.392e+02 3.948e+02 9.464e+02, threshold=6.785e+02, percent-clipped=1.0 2023-05-02 00:42:34,463 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=249347.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:42:41,357 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5593, 2.2132, 1.8845, 1.9899, 2.5055, 2.1505, 2.3101, 2.6540], device='cuda:4'), covar=tensor([0.0223, 0.0437, 0.0553, 0.0507, 0.0270, 0.0394, 0.0240, 0.0295], device='cuda:4'), in_proj_covar=tensor([0.0219, 0.0239, 0.0227, 0.0230, 0.0239, 0.0238, 0.0239, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:42:43,861 INFO [train.py:904] (4/8) Epoch 25, batch 5750, loss[loss=0.2259, simple_loss=0.2964, pruned_loss=0.07767, over 11069.00 frames. ], tot_loss[loss=0.211, simple_loss=0.2962, pruned_loss=0.06293, over 3038461.30 frames. ], batch size: 247, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:43:13,851 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=249371.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:43:53,995 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=249395.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:44:07,472 INFO [train.py:904] (4/8) Epoch 25, batch 5800, loss[loss=0.1785, simple_loss=0.2763, pruned_loss=0.04032, over 16363.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.2953, pruned_loss=0.06166, over 3043572.26 frames. ], batch size: 146, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:44:32,817 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=249419.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:44:44,171 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9784, 3.8859, 4.0436, 4.1476, 4.2711, 3.8752, 4.2054, 4.2904], device='cuda:4'), covar=tensor([0.1818, 0.1221, 0.1451, 0.0762, 0.0600, 0.1600, 0.0918, 0.0755], device='cuda:4'), in_proj_covar=tensor([0.0660, 0.0812, 0.0934, 0.0819, 0.0628, 0.0646, 0.0673, 0.0786], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:44:54,472 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0485, 3.3944, 3.4908, 2.2306, 3.2388, 3.5334, 3.2760, 1.9651], device='cuda:4'), covar=tensor([0.0633, 0.0076, 0.0072, 0.0500, 0.0110, 0.0111, 0.0106, 0.0542], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0136, 0.0100, 0.0112, 0.0097, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 00:45:09,430 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.796e+02 2.976e+02 3.356e+02 4.382e+02 8.536e+02, threshold=6.712e+02, percent-clipped=3.0 2023-05-02 00:45:15,511 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.40 vs. limit=5.0 2023-05-02 00:45:26,283 INFO [train.py:904] (4/8) Epoch 25, batch 5850, loss[loss=0.2099, simple_loss=0.2873, pruned_loss=0.06622, over 11406.00 frames. ], tot_loss[loss=0.2068, simple_loss=0.2933, pruned_loss=0.06017, over 3041838.96 frames. ], batch size: 248, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:46:46,942 INFO [train.py:904] (4/8) Epoch 25, batch 5900, loss[loss=0.1899, simple_loss=0.2809, pruned_loss=0.04945, over 16917.00 frames. ], tot_loss[loss=0.2057, simple_loss=0.2925, pruned_loss=0.05942, over 3069812.39 frames. ], batch size: 109, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:47:10,889 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0667, 2.3139, 2.3682, 2.6923, 1.9586, 3.1717, 1.9031, 2.8001], device='cuda:4'), covar=tensor([0.1055, 0.0643, 0.1016, 0.0162, 0.0105, 0.0305, 0.1349, 0.0625], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0178, 0.0197, 0.0195, 0.0206, 0.0216, 0.0206, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 00:47:52,231 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.570e+02 2.624e+02 3.247e+02 4.027e+02 8.250e+02, threshold=6.495e+02, percent-clipped=2.0 2023-05-02 00:48:08,068 INFO [train.py:904] (4/8) Epoch 25, batch 5950, loss[loss=0.1908, simple_loss=0.2825, pruned_loss=0.04956, over 16489.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2928, pruned_loss=0.05821, over 3075537.70 frames. ], batch size: 146, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:48:08,573 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8506, 3.8011, 3.9176, 4.0098, 4.0979, 3.7155, 4.0712, 4.1279], device='cuda:4'), covar=tensor([0.1676, 0.1124, 0.1253, 0.0702, 0.0687, 0.1962, 0.0858, 0.0802], device='cuda:4'), in_proj_covar=tensor([0.0660, 0.0812, 0.0934, 0.0818, 0.0627, 0.0645, 0.0673, 0.0786], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:48:33,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0517, 3.0653, 1.8644, 3.2668, 2.2872, 3.3082, 2.1518, 2.5778], device='cuda:4'), covar=tensor([0.0326, 0.0439, 0.1628, 0.0277, 0.0797, 0.0655, 0.1410, 0.0759], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0180, 0.0194, 0.0168, 0.0177, 0.0219, 0.0203, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 00:48:36,947 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=249571.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:48:41,572 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5852, 4.8003, 4.9251, 4.7099, 4.7999, 5.2866, 4.7760, 4.5479], device='cuda:4'), covar=tensor([0.1288, 0.1776, 0.2050, 0.1945, 0.2253, 0.0928, 0.1620, 0.2382], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0617, 0.0676, 0.0503, 0.0668, 0.0706, 0.0524, 0.0673], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 00:48:50,692 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 00:49:28,143 INFO [train.py:904] (4/8) Epoch 25, batch 6000, loss[loss=0.1982, simple_loss=0.2878, pruned_loss=0.05431, over 17032.00 frames. ], tot_loss[loss=0.2035, simple_loss=0.2915, pruned_loss=0.05777, over 3085957.22 frames. ], batch size: 55, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:49:28,143 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 00:49:38,598 INFO [train.py:938] (4/8) Epoch 25, validation: loss=0.1487, simple_loss=0.2613, pruned_loss=0.01811, over 944034.00 frames. 2023-05-02 00:49:38,598 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 00:49:42,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5633, 3.4987, 1.8233, 4.1372, 2.5799, 3.9893, 1.8840, 2.7514], device='cuda:4'), covar=tensor([0.0313, 0.0402, 0.2121, 0.0295, 0.0923, 0.0594, 0.2149, 0.0894], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0181, 0.0195, 0.0169, 0.0178, 0.0220, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 00:50:36,538 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.971e+02 2.903e+02 3.391e+02 3.955e+02 5.249e+02, threshold=6.782e+02, percent-clipped=0.0 2023-05-02 00:50:38,181 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-05-02 00:50:54,725 INFO [train.py:904] (4/8) Epoch 25, batch 6050, loss[loss=0.1876, simple_loss=0.2859, pruned_loss=0.04462, over 16252.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2904, pruned_loss=0.05734, over 3094057.90 frames. ], batch size: 165, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:51:35,285 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1279, 5.1807, 4.9836, 4.5707, 4.5946, 5.0541, 4.9286, 4.7150], device='cuda:4'), covar=tensor([0.0812, 0.0892, 0.0400, 0.0449, 0.1166, 0.0720, 0.0506, 0.0893], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0449, 0.0349, 0.0354, 0.0354, 0.0408, 0.0239, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:51:43,349 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 00:52:12,397 INFO [train.py:904] (4/8) Epoch 25, batch 6100, loss[loss=0.2083, simple_loss=0.2942, pruned_loss=0.06119, over 16608.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2895, pruned_loss=0.05606, over 3110676.92 frames. ], batch size: 62, lr: 2.69e-03, grad_scale: 16.0 2023-05-02 00:53:15,558 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.704e+02 2.711e+02 3.338e+02 3.699e+02 9.166e+02, threshold=6.676e+02, percent-clipped=2.0 2023-05-02 00:53:29,950 INFO [train.py:904] (4/8) Epoch 25, batch 6150, loss[loss=0.1785, simple_loss=0.2697, pruned_loss=0.04366, over 16718.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2875, pruned_loss=0.05561, over 3099913.06 frames. ], batch size: 76, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:54:51,553 INFO [train.py:904] (4/8) Epoch 25, batch 6200, loss[loss=0.2388, simple_loss=0.3075, pruned_loss=0.08507, over 11493.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2853, pruned_loss=0.05499, over 3115240.20 frames. ], batch size: 247, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:54:57,700 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.55 vs. limit=5.0 2023-05-02 00:55:03,133 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0977, 5.0984, 4.9041, 4.2060, 5.0068, 1.8978, 4.7574, 4.5782], device='cuda:4'), covar=tensor([0.0085, 0.0079, 0.0185, 0.0396, 0.0082, 0.2758, 0.0117, 0.0250], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0167, 0.0207, 0.0184, 0.0184, 0.0213, 0.0196, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 00:55:10,549 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=249814.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:55:55,230 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.970e+02 2.750e+02 3.284e+02 4.072e+02 6.772e+02, threshold=6.569e+02, percent-clipped=1.0 2023-05-02 00:56:10,145 INFO [train.py:904] (4/8) Epoch 25, batch 6250, loss[loss=0.1933, simple_loss=0.2922, pruned_loss=0.04719, over 16533.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2846, pruned_loss=0.05441, over 3128280.87 frames. ], batch size: 68, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:56:38,948 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=249871.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:56:40,167 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2206, 2.8604, 3.1292, 1.8152, 3.2419, 3.3094, 2.7323, 2.5834], device='cuda:4'), covar=tensor([0.0853, 0.0327, 0.0201, 0.1215, 0.0127, 0.0224, 0.0468, 0.0487], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0101, 0.0140, 0.0085, 0.0130, 0.0130, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 00:56:44,480 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=249875.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:57:26,264 INFO [train.py:904] (4/8) Epoch 25, batch 6300, loss[loss=0.2073, simple_loss=0.2908, pruned_loss=0.06194, over 16709.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.285, pruned_loss=0.05413, over 3121939.21 frames. ], batch size: 134, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:57:52,579 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=249919.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 00:58:29,092 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.822e+02 2.746e+02 3.246e+02 3.923e+02 7.422e+02, threshold=6.492e+02, percent-clipped=1.0 2023-05-02 00:58:45,042 INFO [train.py:904] (4/8) Epoch 25, batch 6350, loss[loss=0.1843, simple_loss=0.2751, pruned_loss=0.04677, over 16387.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.285, pruned_loss=0.05473, over 3113678.62 frames. ], batch size: 146, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 00:58:48,568 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=249955.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 00:58:54,471 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-05-02 00:58:59,135 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3081, 3.4397, 3.5929, 3.5762, 3.5951, 3.3928, 3.4611, 3.4905], device='cuda:4'), covar=tensor([0.0547, 0.1217, 0.0715, 0.0705, 0.0753, 0.1004, 0.0930, 0.0831], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0471, 0.0458, 0.0418, 0.0503, 0.0480, 0.0558, 0.0383], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 00:59:01,756 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-05-02 00:59:20,619 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=249976.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:00:04,120 INFO [train.py:904] (4/8) Epoch 25, batch 6400, loss[loss=0.1679, simple_loss=0.2621, pruned_loss=0.03686, over 16902.00 frames. ], tot_loss[loss=0.198, simple_loss=0.2851, pruned_loss=0.05544, over 3110166.98 frames. ], batch size: 96, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 01:00:13,466 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-02 01:00:23,779 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=250016.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 01:00:46,619 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1687, 2.4312, 2.5831, 2.0450, 2.7115, 2.8212, 2.4630, 2.4283], device='cuda:4'), covar=tensor([0.0657, 0.0280, 0.0286, 0.0899, 0.0151, 0.0332, 0.0458, 0.0444], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0101, 0.0141, 0.0085, 0.0131, 0.0131, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 01:00:56,492 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=250037.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:01:01,986 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9482, 4.1848, 3.9937, 4.0789, 3.7248, 3.8032, 3.8882, 4.1760], device='cuda:4'), covar=tensor([0.1128, 0.0891, 0.1056, 0.0859, 0.0857, 0.1677, 0.0975, 0.1064], device='cuda:4'), in_proj_covar=tensor([0.0694, 0.0838, 0.0692, 0.0647, 0.0533, 0.0534, 0.0705, 0.0656], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:01:05,937 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.084e+02 2.850e+02 3.461e+02 4.247e+02 7.668e+02, threshold=6.921e+02, percent-clipped=1.0 2023-05-02 01:01:16,860 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-05-02 01:01:20,199 INFO [train.py:904] (4/8) Epoch 25, batch 6450, loss[loss=0.1826, simple_loss=0.2749, pruned_loss=0.04512, over 17014.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.285, pruned_loss=0.05465, over 3117106.84 frames. ], batch size: 53, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 01:01:33,122 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3828, 1.6783, 2.0528, 2.3155, 2.3683, 2.6206, 1.8239, 2.4844], device='cuda:4'), covar=tensor([0.0247, 0.0541, 0.0346, 0.0377, 0.0356, 0.0222, 0.0558, 0.0172], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0195, 0.0183, 0.0187, 0.0202, 0.0161, 0.0200, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:02:19,751 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4005, 2.1836, 1.8333, 1.9817, 2.4327, 2.0700, 2.1162, 2.5424], device='cuda:4'), covar=tensor([0.0253, 0.0416, 0.0557, 0.0501, 0.0274, 0.0419, 0.0254, 0.0290], device='cuda:4'), in_proj_covar=tensor([0.0220, 0.0239, 0.0228, 0.0230, 0.0239, 0.0239, 0.0240, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:02:32,298 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6092, 4.6603, 4.9487, 4.9356, 4.9678, 4.6746, 4.6494, 4.5036], device='cuda:4'), covar=tensor([0.0363, 0.0563, 0.0401, 0.0428, 0.0460, 0.0403, 0.0938, 0.0521], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0471, 0.0458, 0.0418, 0.0502, 0.0480, 0.0556, 0.0383], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 01:02:36,760 INFO [train.py:904] (4/8) Epoch 25, batch 6500, loss[loss=0.2337, simple_loss=0.2982, pruned_loss=0.08461, over 11544.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2833, pruned_loss=0.05442, over 3115857.86 frames. ], batch size: 247, lr: 2.69e-03, grad_scale: 8.0 2023-05-02 01:03:40,563 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.844e+02 2.643e+02 3.230e+02 3.830e+02 1.038e+03, threshold=6.461e+02, percent-clipped=2.0 2023-05-02 01:03:52,680 INFO [train.py:904] (4/8) Epoch 25, batch 6550, loss[loss=0.2016, simple_loss=0.3003, pruned_loss=0.05141, over 17030.00 frames. ], tot_loss[loss=0.1979, simple_loss=0.2857, pruned_loss=0.05508, over 3127451.66 frames. ], batch size: 50, lr: 2.69e-03, grad_scale: 4.0 2023-05-02 01:04:18,810 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=250170.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:04:30,982 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 01:05:05,603 INFO [train.py:904] (4/8) Epoch 25, batch 6600, loss[loss=0.201, simple_loss=0.2869, pruned_loss=0.05756, over 15429.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2881, pruned_loss=0.05613, over 3110601.50 frames. ], batch size: 191, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:05:13,160 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5137, 4.5567, 4.8964, 4.8712, 4.8966, 4.5956, 4.5654, 4.4613], device='cuda:4'), covar=tensor([0.0336, 0.0557, 0.0385, 0.0409, 0.0404, 0.0428, 0.0939, 0.0545], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0470, 0.0457, 0.0418, 0.0501, 0.0478, 0.0557, 0.0382], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 01:06:06,338 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5770, 1.7395, 2.2469, 2.6328, 2.5550, 2.9716, 1.9098, 2.8782], device='cuda:4'), covar=tensor([0.0258, 0.0606, 0.0367, 0.0364, 0.0372, 0.0213, 0.0615, 0.0162], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0195, 0.0182, 0.0186, 0.0202, 0.0161, 0.0199, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:06:08,494 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.914e+02 2.823e+02 3.325e+02 3.947e+02 6.947e+02, threshold=6.650e+02, percent-clipped=1.0 2023-05-02 01:06:10,512 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.21 vs. limit=5.0 2023-05-02 01:06:21,898 INFO [train.py:904] (4/8) Epoch 25, batch 6650, loss[loss=0.1831, simple_loss=0.2741, pruned_loss=0.04604, over 17191.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2881, pruned_loss=0.05655, over 3113971.50 frames. ], batch size: 46, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:06:49,324 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6905, 2.7723, 2.4333, 2.6107, 3.0641, 2.6523, 3.2008, 3.3123], device='cuda:4'), covar=tensor([0.0116, 0.0416, 0.0512, 0.0442, 0.0298, 0.0443, 0.0269, 0.0272], device='cuda:4'), in_proj_covar=tensor([0.0218, 0.0237, 0.0227, 0.0228, 0.0238, 0.0237, 0.0237, 0.0235], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:07:37,549 INFO [train.py:904] (4/8) Epoch 25, batch 6700, loss[loss=0.2584, simple_loss=0.3222, pruned_loss=0.09725, over 11565.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2869, pruned_loss=0.05664, over 3115852.46 frames. ], batch size: 247, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:07:50,661 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=250311.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 01:08:23,220 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=250332.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:08:41,137 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.782e+02 2.721e+02 3.128e+02 3.707e+02 8.776e+02, threshold=6.256e+02, percent-clipped=1.0 2023-05-02 01:08:53,997 INFO [train.py:904] (4/8) Epoch 25, batch 6750, loss[loss=0.1834, simple_loss=0.2767, pruned_loss=0.04503, over 16780.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2867, pruned_loss=0.05762, over 3093820.39 frames. ], batch size: 102, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:09:28,314 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2929, 5.9294, 6.1371, 5.7326, 5.8755, 6.3751, 5.9363, 5.6861], device='cuda:4'), covar=tensor([0.0794, 0.1582, 0.1704, 0.1816, 0.1897, 0.0776, 0.1329, 0.2133], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0614, 0.0674, 0.0502, 0.0666, 0.0702, 0.0520, 0.0671], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 01:10:10,574 INFO [train.py:904] (4/8) Epoch 25, batch 6800, loss[loss=0.1893, simple_loss=0.28, pruned_loss=0.04933, over 16581.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2871, pruned_loss=0.05735, over 3104707.63 frames. ], batch size: 75, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:11:16,510 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.894e+02 2.847e+02 3.369e+02 4.013e+02 7.021e+02, threshold=6.738e+02, percent-clipped=2.0 2023-05-02 01:11:27,491 INFO [train.py:904] (4/8) Epoch 25, batch 6850, loss[loss=0.2068, simple_loss=0.3022, pruned_loss=0.05574, over 17119.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.2878, pruned_loss=0.05732, over 3101536.12 frames. ], batch size: 49, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:11:32,407 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0595, 2.2331, 2.1412, 3.7794, 2.1079, 2.5886, 2.2282, 2.3825], device='cuda:4'), covar=tensor([0.1595, 0.4026, 0.3378, 0.0604, 0.4562, 0.2820, 0.4373, 0.3253], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0460, 0.0377, 0.0331, 0.0441, 0.0528, 0.0432, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:11:52,924 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=250470.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:12:34,846 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4859, 3.2916, 3.7983, 1.7475, 3.8934, 3.9548, 2.9416, 2.8180], device='cuda:4'), covar=tensor([0.0825, 0.0302, 0.0183, 0.1309, 0.0093, 0.0182, 0.0450, 0.0512], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0101, 0.0140, 0.0085, 0.0130, 0.0130, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 01:12:43,352 INFO [train.py:904] (4/8) Epoch 25, batch 6900, loss[loss=0.2012, simple_loss=0.2964, pruned_loss=0.05299, over 16022.00 frames. ], tot_loss[loss=0.2019, simple_loss=0.2904, pruned_loss=0.05674, over 3122385.10 frames. ], batch size: 35, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:13:06,142 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=250518.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:13:47,067 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.918e+02 2.786e+02 3.277e+02 4.153e+02 1.055e+03, threshold=6.554e+02, percent-clipped=5.0 2023-05-02 01:13:52,735 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4142, 3.7895, 3.8351, 2.5157, 3.4968, 3.8692, 3.4954, 2.1521], device='cuda:4'), covar=tensor([0.0528, 0.0062, 0.0057, 0.0429, 0.0101, 0.0109, 0.0104, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0087, 0.0087, 0.0135, 0.0100, 0.0111, 0.0097, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 01:14:00,192 INFO [train.py:904] (4/8) Epoch 25, batch 6950, loss[loss=0.2182, simple_loss=0.3018, pruned_loss=0.06728, over 16275.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.292, pruned_loss=0.05842, over 3113441.43 frames. ], batch size: 165, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:14:34,535 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3690, 3.4309, 1.7925, 3.7930, 2.5555, 3.7341, 2.1072, 2.7133], device='cuda:4'), covar=tensor([0.0300, 0.0412, 0.1990, 0.0252, 0.0863, 0.0679, 0.1693, 0.0808], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0180, 0.0195, 0.0168, 0.0177, 0.0218, 0.0202, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 01:15:15,025 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-02 01:15:18,378 INFO [train.py:904] (4/8) Epoch 25, batch 7000, loss[loss=0.1958, simple_loss=0.2911, pruned_loss=0.05031, over 17043.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2918, pruned_loss=0.05796, over 3098899.27 frames. ], batch size: 53, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:15:28,663 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6398, 3.4061, 3.8162, 1.9354, 3.9251, 4.0050, 3.0534, 2.9188], device='cuda:4'), covar=tensor([0.0756, 0.0278, 0.0172, 0.1206, 0.0087, 0.0158, 0.0398, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0110, 0.0100, 0.0139, 0.0085, 0.0129, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 01:15:31,313 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=250611.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 01:16:03,416 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=250632.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:16:03,865 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-02 01:16:22,009 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.860e+02 2.716e+02 3.192e+02 3.909e+02 7.863e+02, threshold=6.384e+02, percent-clipped=2.0 2023-05-02 01:16:35,090 INFO [train.py:904] (4/8) Epoch 25, batch 7050, loss[loss=0.1743, simple_loss=0.2705, pruned_loss=0.03907, over 16415.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2928, pruned_loss=0.05802, over 3095312.96 frames. ], batch size: 68, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:16:44,698 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=250659.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 01:16:49,609 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=250662.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:17:17,436 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=250680.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:17:32,896 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-02 01:17:51,094 INFO [train.py:904] (4/8) Epoch 25, batch 7100, loss[loss=0.2337, simple_loss=0.2893, pruned_loss=0.08904, over 11182.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2912, pruned_loss=0.05802, over 3084316.29 frames. ], batch size: 247, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:18:23,408 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=250723.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:18:56,841 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.881e+02 2.742e+02 3.283e+02 4.161e+02 6.949e+02, threshold=6.565e+02, percent-clipped=2.0 2023-05-02 01:19:09,288 INFO [train.py:904] (4/8) Epoch 25, batch 7150, loss[loss=0.1855, simple_loss=0.2737, pruned_loss=0.04866, over 16754.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2894, pruned_loss=0.05789, over 3080685.30 frames. ], batch size: 83, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:20:22,683 INFO [train.py:904] (4/8) Epoch 25, batch 7200, loss[loss=0.1704, simple_loss=0.2596, pruned_loss=0.04056, over 16655.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2879, pruned_loss=0.05698, over 3067221.41 frames. ], batch size: 62, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:21:08,915 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=250833.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:21:23,971 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8240, 4.8863, 4.7298, 4.3423, 4.3858, 4.7979, 4.5933, 4.5013], device='cuda:4'), covar=tensor([0.0598, 0.0499, 0.0284, 0.0333, 0.0994, 0.0450, 0.0451, 0.0669], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0446, 0.0346, 0.0352, 0.0352, 0.0404, 0.0238, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:21:28,120 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=250845.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:21:28,842 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.479e+02 2.397e+02 2.821e+02 3.422e+02 6.552e+02, threshold=5.642e+02, percent-clipped=0.0 2023-05-02 01:21:41,038 INFO [train.py:904] (4/8) Epoch 25, batch 7250, loss[loss=0.1639, simple_loss=0.2549, pruned_loss=0.03648, over 16738.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.2854, pruned_loss=0.05559, over 3067194.15 frames. ], batch size: 83, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:21:44,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6095, 4.5861, 4.4390, 3.7296, 4.5134, 1.7283, 4.2395, 4.0276], device='cuda:4'), covar=tensor([0.0089, 0.0090, 0.0188, 0.0326, 0.0090, 0.2839, 0.0125, 0.0277], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0165, 0.0205, 0.0182, 0.0182, 0.0211, 0.0193, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:22:00,229 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.48 vs. limit=5.0 2023-05-02 01:22:42,143 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=250894.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:22:55,036 INFO [train.py:904] (4/8) Epoch 25, batch 7300, loss[loss=0.2347, simple_loss=0.3213, pruned_loss=0.07402, over 16413.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.2845, pruned_loss=0.05491, over 3083705.97 frames. ], batch size: 146, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:23:00,843 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=250906.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 01:23:59,648 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.919e+02 2.893e+02 3.388e+02 3.918e+02 7.629e+02, threshold=6.776e+02, percent-clipped=7.0 2023-05-02 01:24:00,712 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-05-02 01:24:09,711 INFO [train.py:904] (4/8) Epoch 25, batch 7350, loss[loss=0.1882, simple_loss=0.2801, pruned_loss=0.04815, over 16442.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2858, pruned_loss=0.05626, over 3054570.92 frames. ], batch size: 146, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:25:15,592 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9037, 5.2271, 5.4011, 5.1300, 5.2327, 5.7400, 5.1670, 4.9734], device='cuda:4'), covar=tensor([0.0963, 0.1619, 0.2181, 0.1896, 0.2152, 0.0849, 0.1627, 0.2359], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0611, 0.0675, 0.0502, 0.0665, 0.0699, 0.0524, 0.0672], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 01:25:27,064 INFO [train.py:904] (4/8) Epoch 25, batch 7400, loss[loss=0.2599, simple_loss=0.3209, pruned_loss=0.09941, over 11030.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.2875, pruned_loss=0.05744, over 3041873.78 frames. ], batch size: 248, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:25:50,966 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=251018.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:26:35,984 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.828e+02 2.726e+02 3.311e+02 3.902e+02 7.761e+02, threshold=6.622e+02, percent-clipped=1.0 2023-05-02 01:26:46,721 INFO [train.py:904] (4/8) Epoch 25, batch 7450, loss[loss=0.2067, simple_loss=0.3018, pruned_loss=0.05586, over 15479.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.2881, pruned_loss=0.05811, over 3030908.16 frames. ], batch size: 191, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:27:37,626 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8473, 3.7469, 3.8794, 3.9839, 4.0875, 3.7040, 4.0332, 4.1175], device='cuda:4'), covar=tensor([0.1610, 0.1217, 0.1394, 0.0751, 0.0658, 0.1912, 0.0944, 0.0752], device='cuda:4'), in_proj_covar=tensor([0.0644, 0.0796, 0.0917, 0.0803, 0.0618, 0.0635, 0.0667, 0.0775], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:28:06,267 INFO [train.py:904] (4/8) Epoch 25, batch 7500, loss[loss=0.1948, simple_loss=0.2795, pruned_loss=0.05503, over 16870.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.2878, pruned_loss=0.05728, over 3044349.07 frames. ], batch size: 83, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:28:46,871 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3450, 5.6235, 5.3864, 5.4061, 5.0982, 5.0068, 5.0407, 5.7600], device='cuda:4'), covar=tensor([0.1204, 0.0840, 0.1067, 0.0995, 0.0823, 0.0805, 0.1226, 0.0798], device='cuda:4'), in_proj_covar=tensor([0.0693, 0.0840, 0.0692, 0.0647, 0.0532, 0.0538, 0.0703, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:29:12,775 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.915e+02 2.911e+02 3.474e+02 4.178e+02 7.715e+02, threshold=6.948e+02, percent-clipped=3.0 2023-05-02 01:29:23,951 INFO [train.py:904] (4/8) Epoch 25, batch 7550, loss[loss=0.1879, simple_loss=0.2727, pruned_loss=0.05158, over 16680.00 frames. ], tot_loss[loss=0.2003, simple_loss=0.2864, pruned_loss=0.05707, over 3061863.44 frames. ], batch size: 134, lr: 2.68e-03, grad_scale: 4.0 2023-05-02 01:29:24,679 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5382, 3.6231, 2.6547, 2.1768, 2.4069, 2.3580, 3.8413, 3.2267], device='cuda:4'), covar=tensor([0.3090, 0.0637, 0.1998, 0.2673, 0.2613, 0.2175, 0.0449, 0.1363], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0273, 0.0311, 0.0321, 0.0304, 0.0270, 0.0301, 0.0345], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 01:29:37,774 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-05-02 01:30:02,768 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2181, 3.4190, 3.5568, 3.5335, 3.5561, 3.3828, 3.3005, 3.4566], device='cuda:4'), covar=tensor([0.0647, 0.0957, 0.0672, 0.0709, 0.0771, 0.0768, 0.1271, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0472, 0.0459, 0.0422, 0.0505, 0.0483, 0.0560, 0.0386], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 01:30:19,164 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=251189.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:30:32,699 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=251198.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:30:37,923 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=251201.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 01:30:39,949 INFO [train.py:904] (4/8) Epoch 25, batch 7600, loss[loss=0.193, simple_loss=0.2821, pruned_loss=0.05193, over 16737.00 frames. ], tot_loss[loss=0.2003, simple_loss=0.286, pruned_loss=0.05727, over 3079497.40 frames. ], batch size: 39, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:31:43,658 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.097e+02 2.965e+02 3.524e+02 4.371e+02 1.071e+03, threshold=7.047e+02, percent-clipped=4.0 2023-05-02 01:31:53,103 INFO [train.py:904] (4/8) Epoch 25, batch 7650, loss[loss=0.2177, simple_loss=0.3011, pruned_loss=0.06709, over 15402.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.287, pruned_loss=0.05799, over 3086491.74 frames. ], batch size: 191, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:32:02,544 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=251259.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:32:58,471 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5307, 3.5849, 3.3221, 3.0017, 3.2018, 3.4701, 3.3239, 3.2974], device='cuda:4'), covar=tensor([0.0559, 0.0674, 0.0288, 0.0300, 0.0513, 0.0530, 0.1315, 0.0547], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0445, 0.0344, 0.0350, 0.0350, 0.0402, 0.0238, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:33:06,999 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=251301.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:33:08,935 INFO [train.py:904] (4/8) Epoch 25, batch 7700, loss[loss=0.2134, simple_loss=0.2923, pruned_loss=0.06728, over 15480.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2869, pruned_loss=0.05832, over 3073683.94 frames. ], batch size: 191, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:33:32,364 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=251318.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:34:14,480 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.002e+02 2.953e+02 3.490e+02 4.215e+02 8.631e+02, threshold=6.980e+02, percent-clipped=3.0 2023-05-02 01:34:25,530 INFO [train.py:904] (4/8) Epoch 25, batch 7750, loss[loss=0.1998, simple_loss=0.2894, pruned_loss=0.0551, over 16821.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2872, pruned_loss=0.05838, over 3065091.03 frames. ], batch size: 96, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:34:39,731 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=251362.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:34:45,579 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=251366.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:35:35,157 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4103, 2.6347, 2.2108, 2.4221, 2.9838, 2.5754, 2.9874, 3.1794], device='cuda:4'), covar=tensor([0.0158, 0.0413, 0.0570, 0.0440, 0.0288, 0.0401, 0.0263, 0.0268], device='cuda:4'), in_proj_covar=tensor([0.0216, 0.0235, 0.0225, 0.0227, 0.0236, 0.0234, 0.0234, 0.0233], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:35:40,750 INFO [train.py:904] (4/8) Epoch 25, batch 7800, loss[loss=0.2498, simple_loss=0.3181, pruned_loss=0.09072, over 11113.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2877, pruned_loss=0.05855, over 3079513.19 frames. ], batch size: 248, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:35:42,441 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=251404.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:36:45,154 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.753e+02 2.864e+02 3.394e+02 4.001e+02 5.962e+02, threshold=6.789e+02, percent-clipped=0.0 2023-05-02 01:36:55,300 INFO [train.py:904] (4/8) Epoch 25, batch 7850, loss[loss=0.1795, simple_loss=0.269, pruned_loss=0.04495, over 17117.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.289, pruned_loss=0.05842, over 3067062.39 frames. ], batch size: 48, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:36:58,725 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7867, 3.1126, 3.3925, 1.9852, 2.9422, 2.1995, 3.3170, 3.3635], device='cuda:4'), covar=tensor([0.0250, 0.0818, 0.0537, 0.2165, 0.0828, 0.1018, 0.0622, 0.0851], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0167, 0.0169, 0.0155, 0.0147, 0.0131, 0.0145, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 01:37:14,491 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=251465.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:37:28,290 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1492, 2.0143, 1.6884, 1.7990, 2.2093, 1.9450, 1.8682, 2.3497], device='cuda:4'), covar=tensor([0.0244, 0.0413, 0.0542, 0.0448, 0.0277, 0.0363, 0.0212, 0.0265], device='cuda:4'), in_proj_covar=tensor([0.0215, 0.0235, 0.0225, 0.0226, 0.0236, 0.0234, 0.0234, 0.0233], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:37:50,393 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=251489.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:38:07,398 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=251501.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 01:38:09,238 INFO [train.py:904] (4/8) Epoch 25, batch 7900, loss[loss=0.2213, simple_loss=0.3049, pruned_loss=0.06885, over 15383.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2882, pruned_loss=0.05789, over 3082607.69 frames. ], batch size: 190, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:38:40,896 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1786, 5.7535, 5.9594, 5.6268, 5.7078, 6.2300, 5.6331, 5.4707], device='cuda:4'), covar=tensor([0.0911, 0.1709, 0.2095, 0.1804, 0.2134, 0.0882, 0.1505, 0.2353], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0615, 0.0679, 0.0504, 0.0670, 0.0703, 0.0528, 0.0677], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 01:39:04,226 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=251537.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:39:17,277 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.008e+02 2.635e+02 3.162e+02 3.936e+02 7.872e+02, threshold=6.324e+02, percent-clipped=1.0 2023-05-02 01:39:22,129 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=251549.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:39:28,550 INFO [train.py:904] (4/8) Epoch 25, batch 7950, loss[loss=0.2048, simple_loss=0.2893, pruned_loss=0.06014, over 15263.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2881, pruned_loss=0.05774, over 3080617.50 frames. ], batch size: 190, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:39:30,109 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=251554.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:40:46,595 INFO [train.py:904] (4/8) Epoch 25, batch 8000, loss[loss=0.188, simple_loss=0.2763, pruned_loss=0.04983, over 16481.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2886, pruned_loss=0.05868, over 3071751.34 frames. ], batch size: 62, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:41:44,809 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=251641.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:41:51,620 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.813e+02 2.777e+02 3.202e+02 3.994e+02 6.409e+02, threshold=6.404e+02, percent-clipped=2.0 2023-05-02 01:42:01,886 INFO [train.py:904] (4/8) Epoch 25, batch 8050, loss[loss=0.1864, simple_loss=0.2785, pruned_loss=0.04718, over 16896.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2888, pruned_loss=0.0586, over 3082835.99 frames. ], batch size: 109, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:42:09,018 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=251657.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:42:54,710 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8976, 4.9634, 5.3062, 5.2699, 5.2880, 4.9603, 4.9403, 4.7113], device='cuda:4'), covar=tensor([0.0315, 0.0535, 0.0331, 0.0380, 0.0523, 0.0366, 0.0974, 0.0487], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0475, 0.0460, 0.0423, 0.0507, 0.0486, 0.0561, 0.0387], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 01:43:17,011 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=251702.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:43:17,637 INFO [train.py:904] (4/8) Epoch 25, batch 8100, loss[loss=0.1993, simple_loss=0.2865, pruned_loss=0.05611, over 16195.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2888, pruned_loss=0.05816, over 3093781.04 frames. ], batch size: 165, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:43:27,272 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=251709.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:44:04,240 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5470, 3.6268, 2.1331, 4.0788, 2.7053, 4.0014, 2.3522, 2.9153], device='cuda:4'), covar=tensor([0.0317, 0.0466, 0.1787, 0.0298, 0.0839, 0.0754, 0.1542, 0.0793], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0180, 0.0196, 0.0168, 0.0178, 0.0219, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 01:44:22,959 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.895e+02 2.808e+02 3.404e+02 4.395e+02 1.324e+03, threshold=6.808e+02, percent-clipped=5.0 2023-05-02 01:44:28,957 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-05-02 01:44:33,793 INFO [train.py:904] (4/8) Epoch 25, batch 8150, loss[loss=0.1754, simple_loss=0.2636, pruned_loss=0.04362, over 16854.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2861, pruned_loss=0.05666, over 3099897.71 frames. ], batch size: 116, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:44:45,086 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=251760.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:45:00,608 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=251770.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:45:01,009 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-05-02 01:45:50,905 INFO [train.py:904] (4/8) Epoch 25, batch 8200, loss[loss=0.2017, simple_loss=0.2893, pruned_loss=0.05703, over 16910.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2833, pruned_loss=0.0558, over 3096069.88 frames. ], batch size: 109, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:46:59,192 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.400e+02 2.618e+02 3.043e+02 3.668e+02 8.118e+02, threshold=6.086e+02, percent-clipped=2.0 2023-05-02 01:47:01,853 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0646, 3.1174, 2.0448, 3.2934, 2.3524, 3.3072, 2.0970, 2.6368], device='cuda:4'), covar=tensor([0.0320, 0.0365, 0.1441, 0.0353, 0.0824, 0.0608, 0.1539, 0.0685], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0179, 0.0195, 0.0167, 0.0177, 0.0218, 0.0204, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 01:47:10,655 INFO [train.py:904] (4/8) Epoch 25, batch 8250, loss[loss=0.165, simple_loss=0.2647, pruned_loss=0.03263, over 16786.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2822, pruned_loss=0.05344, over 3071950.15 frames. ], batch size: 102, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:47:12,512 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=251854.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:47:15,449 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6438, 4.6823, 4.5031, 4.0827, 4.1608, 4.5720, 4.3742, 4.2545], device='cuda:4'), covar=tensor([0.0612, 0.0635, 0.0364, 0.0414, 0.1033, 0.0583, 0.0513, 0.0811], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0445, 0.0344, 0.0349, 0.0349, 0.0400, 0.0238, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:48:29,881 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=251902.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:48:30,708 INFO [train.py:904] (4/8) Epoch 25, batch 8300, loss[loss=0.172, simple_loss=0.2711, pruned_loss=0.03643, over 16911.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.2791, pruned_loss=0.05051, over 3050170.54 frames. ], batch size: 116, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:49:41,162 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.412e+02 2.151e+02 2.560e+02 3.132e+02 5.384e+02, threshold=5.121e+02, percent-clipped=0.0 2023-05-02 01:49:52,634 INFO [train.py:904] (4/8) Epoch 25, batch 8350, loss[loss=0.2023, simple_loss=0.2822, pruned_loss=0.06115, over 12207.00 frames. ], tot_loss[loss=0.1878, simple_loss=0.2786, pruned_loss=0.04854, over 3050206.83 frames. ], batch size: 246, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:49:59,406 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=251957.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:50:41,537 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0895, 3.3102, 3.6856, 2.1502, 3.1562, 2.3433, 3.6083, 3.4847], device='cuda:4'), covar=tensor([0.0276, 0.0873, 0.0531, 0.2162, 0.0753, 0.1052, 0.0600, 0.0967], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0164, 0.0167, 0.0152, 0.0144, 0.0129, 0.0142, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 01:51:05,348 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=251997.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:51:17,313 INFO [train.py:904] (4/8) Epoch 25, batch 8400, loss[loss=0.1675, simple_loss=0.2671, pruned_loss=0.03395, over 16893.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2764, pruned_loss=0.0465, over 3054972.96 frames. ], batch size: 116, lr: 2.68e-03, grad_scale: 8.0 2023-05-02 01:51:22,061 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=252005.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:52:27,817 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.337e+02 2.253e+02 2.668e+02 3.204e+02 6.662e+02, threshold=5.337e+02, percent-clipped=3.0 2023-05-02 01:52:40,176 INFO [train.py:904] (4/8) Epoch 25, batch 8450, loss[loss=0.1919, simple_loss=0.2804, pruned_loss=0.05166, over 16880.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2745, pruned_loss=0.04513, over 3035227.95 frames. ], batch size: 116, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 01:52:51,805 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=252060.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:52:59,910 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=252065.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:53:14,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5702, 3.1463, 3.3383, 1.9391, 3.4996, 3.5055, 2.9672, 2.9926], device='cuda:4'), covar=tensor([0.0599, 0.0248, 0.0231, 0.1138, 0.0095, 0.0244, 0.0408, 0.0362], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0108, 0.0097, 0.0136, 0.0082, 0.0126, 0.0126, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 01:53:32,170 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9595, 4.2185, 4.0352, 4.0932, 3.7903, 3.8611, 3.8328, 4.2182], device='cuda:4'), covar=tensor([0.1200, 0.0913, 0.1011, 0.0850, 0.0763, 0.1543, 0.0980, 0.1011], device='cuda:4'), in_proj_covar=tensor([0.0683, 0.0824, 0.0681, 0.0637, 0.0523, 0.0531, 0.0691, 0.0645], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:54:01,927 INFO [train.py:904] (4/8) Epoch 25, batch 8500, loss[loss=0.1564, simple_loss=0.2397, pruned_loss=0.03659, over 11796.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2712, pruned_loss=0.04344, over 3026987.27 frames. ], batch size: 248, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 01:54:11,542 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=252108.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 01:55:04,242 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4371, 3.0782, 3.3243, 1.7823, 3.4684, 3.4965, 2.9278, 2.8353], device='cuda:4'), covar=tensor([0.0684, 0.0280, 0.0211, 0.1259, 0.0096, 0.0206, 0.0402, 0.0423], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0108, 0.0096, 0.0135, 0.0082, 0.0125, 0.0125, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 01:55:06,824 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6713, 3.9188, 2.7228, 2.2071, 2.3565, 2.3151, 4.1126, 3.2630], device='cuda:4'), covar=tensor([0.3041, 0.0542, 0.2126, 0.3262, 0.3330, 0.2463, 0.0418, 0.1441], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0267, 0.0305, 0.0315, 0.0297, 0.0266, 0.0294, 0.0338], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 01:55:15,324 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.653e+02 2.299e+02 2.680e+02 3.253e+02 6.153e+02, threshold=5.360e+02, percent-clipped=1.0 2023-05-02 01:55:28,836 INFO [train.py:904] (4/8) Epoch 25, batch 8550, loss[loss=0.173, simple_loss=0.2685, pruned_loss=0.03875, over 16406.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2689, pruned_loss=0.0423, over 3034976.25 frames. ], batch size: 68, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 01:56:10,253 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-05-02 01:57:01,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0467, 3.9158, 4.1238, 4.2357, 4.3884, 3.9249, 4.3120, 4.3993], device='cuda:4'), covar=tensor([0.1837, 0.1307, 0.1524, 0.0809, 0.0589, 0.1375, 0.0816, 0.0712], device='cuda:4'), in_proj_covar=tensor([0.0631, 0.0778, 0.0896, 0.0786, 0.0604, 0.0623, 0.0655, 0.0761], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 01:57:02,329 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-05-02 01:57:09,035 INFO [train.py:904] (4/8) Epoch 25, batch 8600, loss[loss=0.166, simple_loss=0.2573, pruned_loss=0.03731, over 16623.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2688, pruned_loss=0.04122, over 3027934.01 frames. ], batch size: 57, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 01:58:34,356 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.652e+02 2.200e+02 2.678e+02 3.278e+02 9.763e+02, threshold=5.356e+02, percent-clipped=4.0 2023-05-02 01:58:48,836 INFO [train.py:904] (4/8) Epoch 25, batch 8650, loss[loss=0.1516, simple_loss=0.2548, pruned_loss=0.02415, over 16534.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2667, pruned_loss=0.03952, over 3026346.60 frames. ], batch size: 62, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:00:23,904 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=252297.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:00:33,683 INFO [train.py:904] (4/8) Epoch 25, batch 8700, loss[loss=0.1637, simple_loss=0.2535, pruned_loss=0.03696, over 16900.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2646, pruned_loss=0.0384, over 3048895.88 frames. ], batch size: 109, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:01:53,637 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=252345.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:01:54,399 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.271e+02 2.533e+02 3.010e+02 6.763e+02, threshold=5.067e+02, percent-clipped=2.0 2023-05-02 02:02:09,358 INFO [train.py:904] (4/8) Epoch 25, batch 8750, loss[loss=0.1794, simple_loss=0.262, pruned_loss=0.04836, over 12206.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2637, pruned_loss=0.0377, over 3048432.91 frames. ], batch size: 250, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:02:41,192 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=252365.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:04:01,511 INFO [train.py:904] (4/8) Epoch 25, batch 8800, loss[loss=0.1924, simple_loss=0.2779, pruned_loss=0.05346, over 12669.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2626, pruned_loss=0.03687, over 3055515.81 frames. ], batch size: 248, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:04:12,163 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3531, 3.4301, 2.1302, 3.7237, 2.5651, 3.6814, 2.2148, 2.7761], device='cuda:4'), covar=tensor([0.0311, 0.0375, 0.1603, 0.0247, 0.0866, 0.0620, 0.1652, 0.0820], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0174, 0.0191, 0.0162, 0.0173, 0.0211, 0.0200, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 02:04:21,826 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=252413.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:04:30,822 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=252417.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:05:07,188 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=252434.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:05:31,487 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.474e+02 2.215e+02 2.676e+02 3.061e+02 6.379e+02, threshold=5.352e+02, percent-clipped=4.0 2023-05-02 02:05:46,035 INFO [train.py:904] (4/8) Epoch 25, batch 8850, loss[loss=0.1703, simple_loss=0.2767, pruned_loss=0.03192, over 16403.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2656, pruned_loss=0.03649, over 3061916.19 frames. ], batch size: 146, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:06:14,436 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-02 02:06:39,730 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=252478.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:07:16,636 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=252495.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:07:31,607 INFO [train.py:904] (4/8) Epoch 25, batch 8900, loss[loss=0.1814, simple_loss=0.27, pruned_loss=0.04638, over 12665.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2658, pruned_loss=0.03588, over 3045333.35 frames. ], batch size: 248, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:07:57,603 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-05-02 02:09:18,926 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.583e+02 2.220e+02 2.582e+02 3.143e+02 6.266e+02, threshold=5.163e+02, percent-clipped=2.0 2023-05-02 02:09:34,571 INFO [train.py:904] (4/8) Epoch 25, batch 8950, loss[loss=0.1674, simple_loss=0.2661, pruned_loss=0.03436, over 15271.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.265, pruned_loss=0.03573, over 3069197.34 frames. ], batch size: 191, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:10:59,698 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=252592.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:11:09,743 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4102, 4.1776, 4.2925, 4.5588, 4.7131, 4.3408, 4.7923, 4.7533], device='cuda:4'), covar=tensor([0.2021, 0.1543, 0.2242, 0.1029, 0.0898, 0.1190, 0.0823, 0.1034], device='cuda:4'), in_proj_covar=tensor([0.0621, 0.0764, 0.0880, 0.0776, 0.0593, 0.0613, 0.0644, 0.0745], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:11:21,898 INFO [train.py:904] (4/8) Epoch 25, batch 9000, loss[loss=0.1547, simple_loss=0.2509, pruned_loss=0.02923, over 15387.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2618, pruned_loss=0.03453, over 3081327.78 frames. ], batch size: 191, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:11:21,898 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 02:11:31,593 INFO [train.py:938] (4/8) Epoch 25, validation: loss=0.1442, simple_loss=0.248, pruned_loss=0.02014, over 944034.00 frames. 2023-05-02 02:11:31,593 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 02:11:53,527 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=252613.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:12:20,797 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0544, 4.0276, 3.9316, 3.0876, 3.9759, 1.7514, 3.7626, 3.5934], device='cuda:4'), covar=tensor([0.0117, 0.0109, 0.0200, 0.0326, 0.0115, 0.3005, 0.0151, 0.0311], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0163, 0.0201, 0.0177, 0.0178, 0.0209, 0.0190, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:13:01,734 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.469e+02 2.125e+02 2.594e+02 3.127e+02 6.449e+02, threshold=5.187e+02, percent-clipped=3.0 2023-05-02 02:13:14,920 INFO [train.py:904] (4/8) Epoch 25, batch 9050, loss[loss=0.1651, simple_loss=0.2573, pruned_loss=0.03644, over 12689.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2629, pruned_loss=0.03543, over 3077071.69 frames. ], batch size: 248, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:13:15,976 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=252653.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:13:45,256 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-02 02:13:47,191 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=252668.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:13:56,961 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=252674.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:14:58,382 INFO [train.py:904] (4/8) Epoch 25, batch 9100, loss[loss=0.1638, simple_loss=0.2705, pruned_loss=0.02852, over 16914.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2626, pruned_loss=0.03582, over 3075390.69 frames. ], batch size: 102, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:16:01,394 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=252729.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:16:41,919 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.565e+02 2.125e+02 2.555e+02 2.905e+02 5.383e+02, threshold=5.110e+02, percent-clipped=1.0 2023-05-02 02:16:57,795 INFO [train.py:904] (4/8) Epoch 25, batch 9150, loss[loss=0.1717, simple_loss=0.259, pruned_loss=0.04221, over 11867.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2631, pruned_loss=0.03556, over 3072001.99 frames. ], batch size: 248, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:17:42,773 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=252773.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:17:55,976 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3163, 4.4070, 4.2270, 3.9460, 3.9249, 4.3339, 4.0272, 4.0760], device='cuda:4'), covar=tensor([0.0651, 0.0629, 0.0359, 0.0324, 0.0868, 0.0511, 0.0766, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0438, 0.0341, 0.0344, 0.0342, 0.0395, 0.0235, 0.0409], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:18:20,260 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=252790.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:18:43,278 INFO [train.py:904] (4/8) Epoch 25, batch 9200, loss[loss=0.1776, simple_loss=0.2741, pruned_loss=0.04053, over 15303.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2592, pruned_loss=0.03466, over 3072220.98 frames. ], batch size: 191, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:18:53,735 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6106, 2.6853, 2.3915, 2.4333, 2.9915, 2.6555, 3.1931, 3.2458], device='cuda:4'), covar=tensor([0.0180, 0.0415, 0.0485, 0.0483, 0.0290, 0.0408, 0.0252, 0.0269], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0234, 0.0224, 0.0225, 0.0234, 0.0233, 0.0229, 0.0230], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:20:05,202 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.526e+02 2.337e+02 2.810e+02 3.410e+02 9.767e+02, threshold=5.620e+02, percent-clipped=4.0 2023-05-02 02:20:19,098 INFO [train.py:904] (4/8) Epoch 25, batch 9250, loss[loss=0.1509, simple_loss=0.2501, pruned_loss=0.02587, over 15383.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2589, pruned_loss=0.0345, over 3074548.06 frames. ], batch size: 191, lr: 2.67e-03, grad_scale: 16.0 2023-05-02 02:20:28,248 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=252857.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:22:11,609 INFO [train.py:904] (4/8) Epoch 25, batch 9300, loss[loss=0.1463, simple_loss=0.2445, pruned_loss=0.02406, over 15490.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2574, pruned_loss=0.03416, over 3066027.63 frames. ], batch size: 191, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:22:47,885 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=252918.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:23:45,075 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.307e+02 1.953e+02 2.173e+02 2.787e+02 5.118e+02, threshold=4.346e+02, percent-clipped=0.0 2023-05-02 02:23:47,874 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=252948.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:23:55,315 INFO [train.py:904] (4/8) Epoch 25, batch 9350, loss[loss=0.1724, simple_loss=0.2591, pruned_loss=0.04281, over 12272.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2571, pruned_loss=0.03427, over 3054795.82 frames. ], batch size: 248, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:24:01,499 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3128, 4.1065, 4.0660, 4.4568, 4.6009, 4.2153, 4.5388, 4.6117], device='cuda:4'), covar=tensor([0.1804, 0.1403, 0.2341, 0.1030, 0.0917, 0.1554, 0.1090, 0.1078], device='cuda:4'), in_proj_covar=tensor([0.0623, 0.0764, 0.0880, 0.0777, 0.0595, 0.0615, 0.0644, 0.0748], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:24:29,895 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=252969.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:24:54,822 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5467, 3.8609, 2.7167, 2.1501, 2.2974, 2.3943, 4.1154, 3.1653], device='cuda:4'), covar=tensor([0.3330, 0.0593, 0.2049, 0.3329, 0.3390, 0.2325, 0.0389, 0.1568], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0264, 0.0303, 0.0311, 0.0292, 0.0263, 0.0292, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 02:25:16,733 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=252993.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:25:36,656 INFO [train.py:904] (4/8) Epoch 25, batch 9400, loss[loss=0.1404, simple_loss=0.2283, pruned_loss=0.02622, over 12441.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2564, pruned_loss=0.03404, over 3039570.26 frames. ], batch size: 247, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:26:19,759 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=253024.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:27:02,201 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5634, 3.0395, 3.2831, 1.8804, 2.8182, 2.2090, 3.0979, 3.1780], device='cuda:4'), covar=tensor([0.0383, 0.0924, 0.0549, 0.2278, 0.0910, 0.1098, 0.0774, 0.1113], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0159, 0.0163, 0.0149, 0.0141, 0.0126, 0.0139, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 02:27:05,786 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.393e+02 2.088e+02 2.613e+02 3.190e+02 8.435e+02, threshold=5.225e+02, percent-clipped=5.0 2023-05-02 02:27:17,027 INFO [train.py:904] (4/8) Epoch 25, batch 9450, loss[loss=0.1628, simple_loss=0.254, pruned_loss=0.03574, over 16704.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2584, pruned_loss=0.03416, over 3046752.75 frames. ], batch size: 62, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:27:19,783 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=253054.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:27:57,645 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=253073.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:28:31,274 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=253090.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:28:54,659 INFO [train.py:904] (4/8) Epoch 25, batch 9500, loss[loss=0.165, simple_loss=0.2608, pruned_loss=0.0346, over 15429.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2581, pruned_loss=0.03405, over 3057679.69 frames. ], batch size: 191, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:29:33,158 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=253121.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:30:02,663 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9175, 2.8306, 2.6594, 1.9736, 2.5615, 2.8228, 2.6721, 1.9958], device='cuda:4'), covar=tensor([0.0447, 0.0071, 0.0073, 0.0377, 0.0137, 0.0100, 0.0101, 0.0438], device='cuda:4'), in_proj_covar=tensor([0.0132, 0.0084, 0.0085, 0.0131, 0.0098, 0.0107, 0.0094, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 02:30:06,532 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=253138.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:30:26,750 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.402e+02 2.108e+02 2.493e+02 2.917e+02 7.333e+02, threshold=4.986e+02, percent-clipped=1.0 2023-05-02 02:30:40,902 INFO [train.py:904] (4/8) Epoch 25, batch 9550, loss[loss=0.1876, simple_loss=0.2874, pruned_loss=0.0439, over 16169.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2581, pruned_loss=0.03462, over 3051191.87 frames. ], batch size: 165, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:31:08,041 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9392, 2.7059, 2.7617, 1.8912, 2.5939, 1.9555, 2.8098, 2.9085], device='cuda:4'), covar=tensor([0.0279, 0.1019, 0.0694, 0.2416, 0.1042, 0.1335, 0.0682, 0.0848], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0159, 0.0164, 0.0150, 0.0142, 0.0127, 0.0140, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 02:32:04,919 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4286, 2.8708, 3.0911, 1.9661, 2.7813, 2.0422, 3.0367, 3.0892], device='cuda:4'), covar=tensor([0.0298, 0.0940, 0.0569, 0.2166, 0.0870, 0.1148, 0.0657, 0.0957], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0159, 0.0163, 0.0150, 0.0141, 0.0127, 0.0139, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 02:32:22,177 INFO [train.py:904] (4/8) Epoch 25, batch 9600, loss[loss=0.1619, simple_loss=0.2617, pruned_loss=0.03107, over 16430.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2597, pruned_loss=0.03531, over 3045820.30 frames. ], batch size: 68, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:32:42,584 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=253213.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:32:58,002 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-02 02:33:09,497 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6375, 2.6843, 1.8526, 2.8240, 2.0537, 2.8291, 2.1249, 2.3636], device='cuda:4'), covar=tensor([0.0338, 0.0366, 0.1417, 0.0340, 0.0738, 0.0488, 0.1389, 0.0677], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0171, 0.0190, 0.0160, 0.0171, 0.0208, 0.0197, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 02:33:55,685 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.643e+02 2.388e+02 2.773e+02 3.482e+02 6.416e+02, threshold=5.546e+02, percent-clipped=8.0 2023-05-02 02:33:59,535 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=253248.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:34:10,799 INFO [train.py:904] (4/8) Epoch 25, batch 9650, loss[loss=0.1749, simple_loss=0.2721, pruned_loss=0.03888, over 16244.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2616, pruned_loss=0.03537, over 3051354.03 frames. ], batch size: 165, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:34:52,564 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=253269.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:35:07,646 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-05-02 02:35:46,089 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=253296.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:36:00,301 INFO [train.py:904] (4/8) Epoch 25, batch 9700, loss[loss=0.1564, simple_loss=0.2464, pruned_loss=0.03315, over 12229.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2602, pruned_loss=0.03525, over 3044520.16 frames. ], batch size: 250, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:36:01,248 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-05-02 02:36:27,595 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=253317.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:36:42,663 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=253324.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:36:56,093 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=253329.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:37:34,496 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.439e+02 2.159e+02 2.493e+02 2.931e+02 5.551e+02, threshold=4.985e+02, percent-clipped=1.0 2023-05-02 02:37:37,140 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=253349.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:37:44,442 INFO [train.py:904] (4/8) Epoch 25, batch 9750, loss[loss=0.1669, simple_loss=0.2643, pruned_loss=0.03478, over 16341.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2592, pruned_loss=0.03524, over 3059005.08 frames. ], batch size: 146, lr: 2.67e-03, grad_scale: 4.0 2023-05-02 02:38:07,240 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1393, 4.0647, 4.4331, 4.4136, 4.4300, 4.1647, 4.1873, 4.1528], device='cuda:4'), covar=tensor([0.0377, 0.0741, 0.0492, 0.0476, 0.0493, 0.0448, 0.0832, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0457, 0.0447, 0.0408, 0.0491, 0.0469, 0.0540, 0.0375], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 02:38:21,281 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=253372.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:38:32,017 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4404, 3.3307, 3.4636, 3.5898, 3.6079, 3.2647, 3.5772, 3.6478], device='cuda:4'), covar=tensor([0.1398, 0.1139, 0.1243, 0.0758, 0.0769, 0.3394, 0.1025, 0.0930], device='cuda:4'), in_proj_covar=tensor([0.0625, 0.0765, 0.0881, 0.0778, 0.0598, 0.0618, 0.0646, 0.0749], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:38:38,914 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=253380.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:39:00,967 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=253390.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:39:01,199 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-05-02 02:39:23,677 INFO [train.py:904] (4/8) Epoch 25, batch 9800, loss[loss=0.1539, simple_loss=0.2555, pruned_loss=0.02612, over 16521.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2599, pruned_loss=0.03455, over 3073937.31 frames. ], batch size: 75, lr: 2.67e-03, grad_scale: 4.0 2023-05-02 02:39:53,082 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=253419.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:40:38,465 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=253441.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:40:55,714 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.418e+02 1.999e+02 2.298e+02 2.745e+02 1.319e+03, threshold=4.597e+02, percent-clipped=1.0 2023-05-02 02:41:06,296 INFO [train.py:904] (4/8) Epoch 25, batch 9850, loss[loss=0.1691, simple_loss=0.2671, pruned_loss=0.0355, over 16798.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2606, pruned_loss=0.03399, over 3066098.82 frames. ], batch size: 124, lr: 2.67e-03, grad_scale: 4.0 2023-05-02 02:41:20,596 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9047, 2.1496, 2.3545, 3.2016, 2.2226, 2.3100, 2.3239, 2.2285], device='cuda:4'), covar=tensor([0.1362, 0.3830, 0.2974, 0.0801, 0.4398, 0.2740, 0.3605, 0.4006], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0452, 0.0372, 0.0323, 0.0434, 0.0514, 0.0424, 0.0527], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:42:03,928 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=253480.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:42:08,456 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5037, 1.7687, 2.2591, 2.5565, 2.4860, 2.8382, 2.0987, 2.7866], device='cuda:4'), covar=tensor([0.0287, 0.0632, 0.0368, 0.0347, 0.0368, 0.0233, 0.0533, 0.0180], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0193, 0.0181, 0.0182, 0.0199, 0.0157, 0.0196, 0.0156], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:43:00,285 INFO [train.py:904] (4/8) Epoch 25, batch 9900, loss[loss=0.1711, simple_loss=0.2728, pruned_loss=0.03476, over 16266.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2606, pruned_loss=0.03385, over 3057674.27 frames. ], batch size: 165, lr: 2.67e-03, grad_scale: 4.0 2023-05-02 02:43:25,451 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=253513.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:43:41,967 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=253521.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:44:46,255 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.406e+02 2.031e+02 2.358e+02 2.986e+02 6.654e+02, threshold=4.715e+02, percent-clipped=4.0 2023-05-02 02:44:50,155 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4848, 3.7807, 2.7872, 2.1802, 2.3624, 2.4821, 3.9925, 3.3381], device='cuda:4'), covar=tensor([0.3162, 0.0572, 0.1954, 0.3079, 0.2660, 0.2055, 0.0403, 0.1189], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0263, 0.0302, 0.0310, 0.0290, 0.0261, 0.0291, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 02:44:56,510 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6444, 2.9690, 3.2682, 1.8462, 2.8972, 2.1371, 3.2684, 3.2513], device='cuda:4'), covar=tensor([0.0240, 0.0877, 0.0582, 0.2212, 0.0756, 0.1028, 0.0578, 0.0900], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0159, 0.0163, 0.0150, 0.0141, 0.0127, 0.0139, 0.0171], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 02:44:59,712 INFO [train.py:904] (4/8) Epoch 25, batch 9950, loss[loss=0.1495, simple_loss=0.2482, pruned_loss=0.02544, over 16514.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2631, pruned_loss=0.03427, over 3072424.06 frames. ], batch size: 68, lr: 2.67e-03, grad_scale: 4.0 2023-05-02 02:45:18,702 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=253561.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:45:34,736 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=253567.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:46:14,394 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=253582.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:47:02,510 INFO [train.py:904] (4/8) Epoch 25, batch 10000, loss[loss=0.1512, simple_loss=0.2392, pruned_loss=0.03164, over 12672.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2615, pruned_loss=0.03407, over 3066525.20 frames. ], batch size: 250, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:47:52,950 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=253628.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:47:56,027 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9966, 4.2353, 4.0908, 4.1287, 3.8164, 3.8660, 3.8524, 4.2506], device='cuda:4'), covar=tensor([0.1042, 0.0831, 0.0904, 0.0800, 0.0781, 0.1802, 0.0952, 0.0894], device='cuda:4'), in_proj_covar=tensor([0.0673, 0.0812, 0.0668, 0.0626, 0.0515, 0.0521, 0.0678, 0.0634], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:48:26,396 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7393, 2.9955, 3.4988, 1.9879, 2.9318, 2.0887, 3.3553, 3.2111], device='cuda:4'), covar=tensor([0.0258, 0.0991, 0.0459, 0.2155, 0.0800, 0.1101, 0.0583, 0.0969], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0158, 0.0163, 0.0150, 0.0141, 0.0126, 0.0139, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 02:48:36,795 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.253e+02 2.008e+02 2.354e+02 2.749e+02 5.133e+02, threshold=4.708e+02, percent-clipped=1.0 2023-05-02 02:48:39,841 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=253649.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:48:41,271 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=253650.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:48:43,142 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4583, 3.3787, 3.5071, 3.5727, 3.6185, 3.3178, 3.5819, 3.6596], device='cuda:4'), covar=tensor([0.1239, 0.0875, 0.1055, 0.0642, 0.0598, 0.2183, 0.0771, 0.0720], device='cuda:4'), in_proj_covar=tensor([0.0620, 0.0758, 0.0873, 0.0773, 0.0593, 0.0611, 0.0642, 0.0744], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:48:46,085 INFO [train.py:904] (4/8) Epoch 25, batch 10050, loss[loss=0.1612, simple_loss=0.2567, pruned_loss=0.03286, over 16858.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2613, pruned_loss=0.03381, over 3084399.01 frames. ], batch size: 76, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:49:22,630 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.39 vs. limit=5.0 2023-05-02 02:49:48,421 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=253685.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:50:11,405 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=253697.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:50:20,880 INFO [train.py:904] (4/8) Epoch 25, batch 10100, loss[loss=0.1502, simple_loss=0.2396, pruned_loss=0.03034, over 12371.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2615, pruned_loss=0.03401, over 3093247.54 frames. ], batch size: 248, lr: 2.67e-03, grad_scale: 8.0 2023-05-02 02:50:35,701 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=253711.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:51:21,521 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=253736.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:51:29,988 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3053, 3.4052, 1.9956, 3.7466, 2.5168, 3.6868, 2.2433, 2.7370], device='cuda:4'), covar=tensor([0.0363, 0.0414, 0.1901, 0.0265, 0.0985, 0.0669, 0.1624, 0.0887], device='cuda:4'), in_proj_covar=tensor([0.0167, 0.0171, 0.0191, 0.0160, 0.0173, 0.0207, 0.0198, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 02:51:33,767 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.529e+02 2.192e+02 2.668e+02 3.131e+02 7.050e+02, threshold=5.336e+02, percent-clipped=1.0 2023-05-02 02:52:07,291 INFO [train.py:904] (4/8) Epoch 26, batch 0, loss[loss=0.1508, simple_loss=0.2396, pruned_loss=0.03096, over 16822.00 frames. ], tot_loss[loss=0.1508, simple_loss=0.2396, pruned_loss=0.03096, over 16822.00 frames. ], batch size: 42, lr: 2.61e-03, grad_scale: 8.0 2023-05-02 02:52:07,292 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 02:52:14,676 INFO [train.py:938] (4/8) Epoch 26, validation: loss=0.1437, simple_loss=0.2472, pruned_loss=0.02009, over 944034.00 frames. 2023-05-02 02:52:14,677 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 02:52:44,911 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=253775.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:53:23,651 INFO [train.py:904] (4/8) Epoch 26, batch 50, loss[loss=0.186, simple_loss=0.2669, pruned_loss=0.0525, over 16418.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2655, pruned_loss=0.047, over 755302.51 frames. ], batch size: 146, lr: 2.61e-03, grad_scale: 1.0 2023-05-02 02:54:28,065 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.564e+02 2.375e+02 2.827e+02 3.414e+02 5.395e+02, threshold=5.654e+02, percent-clipped=1.0 2023-05-02 02:54:30,988 INFO [train.py:904] (4/8) Epoch 26, batch 100, loss[loss=0.1491, simple_loss=0.2299, pruned_loss=0.03416, over 16772.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2645, pruned_loss=0.04516, over 1322330.54 frames. ], batch size: 39, lr: 2.61e-03, grad_scale: 1.0 2023-05-02 02:54:51,323 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-05-02 02:55:03,716 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=253877.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:55:39,316 INFO [train.py:904] (4/8) Epoch 26, batch 150, loss[loss=0.1993, simple_loss=0.2766, pruned_loss=0.06096, over 16286.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2621, pruned_loss=0.04369, over 1764521.79 frames. ], batch size: 165, lr: 2.61e-03, grad_scale: 1.0 2023-05-02 02:55:51,776 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9216, 4.3863, 3.0899, 2.3153, 2.6032, 2.4264, 4.7597, 3.4977], device='cuda:4'), covar=tensor([0.2985, 0.0548, 0.1960, 0.3086, 0.3287, 0.2386, 0.0367, 0.1600], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0267, 0.0307, 0.0315, 0.0296, 0.0266, 0.0296, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 02:56:08,374 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=253923.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:56:12,904 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0196, 2.0627, 2.2500, 3.5564, 2.1246, 2.3281, 2.1469, 2.1742], device='cuda:4'), covar=tensor([0.1520, 0.4086, 0.3233, 0.0817, 0.4268, 0.2914, 0.4027, 0.3394], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0457, 0.0377, 0.0328, 0.0439, 0.0520, 0.0429, 0.0534], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:56:46,628 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.138e+02 2.556e+02 3.055e+02 5.694e+02, threshold=5.112e+02, percent-clipped=1.0 2023-05-02 02:56:49,068 INFO [train.py:904] (4/8) Epoch 26, batch 200, loss[loss=0.1285, simple_loss=0.2206, pruned_loss=0.01816, over 16816.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2616, pruned_loss=0.04354, over 2108098.29 frames. ], batch size: 39, lr: 2.61e-03, grad_scale: 1.0 2023-05-02 02:57:34,337 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=253985.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:58:03,909 INFO [train.py:904] (4/8) Epoch 26, batch 250, loss[loss=0.1719, simple_loss=0.2444, pruned_loss=0.0497, over 16899.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2606, pruned_loss=0.04326, over 2374115.50 frames. ], batch size: 109, lr: 2.61e-03, grad_scale: 1.0 2023-05-02 02:58:07,779 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=254006.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:58:47,173 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=254033.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:58:50,907 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=254036.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:59:10,748 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.700e+02 2.097e+02 2.455e+02 3.037e+02 5.274e+02, threshold=4.910e+02, percent-clipped=1.0 2023-05-02 02:59:14,066 INFO [train.py:904] (4/8) Epoch 26, batch 300, loss[loss=0.1546, simple_loss=0.2464, pruned_loss=0.03142, over 17078.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2578, pruned_loss=0.04225, over 2571321.71 frames. ], batch size: 55, lr: 2.61e-03, grad_scale: 1.0 2023-05-02 02:59:33,955 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8179, 2.0312, 2.3565, 3.0717, 2.1268, 2.1470, 2.2201, 2.1336], device='cuda:4'), covar=tensor([0.1773, 0.4046, 0.2914, 0.0976, 0.4819, 0.3240, 0.3604, 0.4169], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0460, 0.0379, 0.0330, 0.0441, 0.0524, 0.0432, 0.0537], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 02:59:45,411 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=254075.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 02:59:57,064 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=254084.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:00:23,586 INFO [train.py:904] (4/8) Epoch 26, batch 350, loss[loss=0.1562, simple_loss=0.2404, pruned_loss=0.03604, over 15484.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2551, pruned_loss=0.04107, over 2733257.09 frames. ], batch size: 191, lr: 2.61e-03, grad_scale: 1.0 2023-05-02 03:00:51,986 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=254123.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:01:30,800 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.391e+02 2.013e+02 2.457e+02 3.049e+02 5.935e+02, threshold=4.915e+02, percent-clipped=1.0 2023-05-02 03:01:33,889 INFO [train.py:904] (4/8) Epoch 26, batch 400, loss[loss=0.172, simple_loss=0.2596, pruned_loss=0.04218, over 16685.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.253, pruned_loss=0.04034, over 2868532.39 frames. ], batch size: 57, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:02:01,759 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7778, 4.1798, 4.1429, 3.0198, 3.6672, 4.2058, 3.7926, 2.4996], device='cuda:4'), covar=tensor([0.0531, 0.0097, 0.0063, 0.0379, 0.0142, 0.0113, 0.0100, 0.0489], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0087, 0.0088, 0.0135, 0.0101, 0.0111, 0.0096, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 03:02:07,865 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=254177.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:02:16,825 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.87 vs. limit=5.0 2023-05-02 03:02:44,705 INFO [train.py:904] (4/8) Epoch 26, batch 450, loss[loss=0.1464, simple_loss=0.233, pruned_loss=0.02987, over 17011.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2521, pruned_loss=0.03996, over 2968929.22 frames. ], batch size: 41, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:03:08,903 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 03:03:12,798 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=254223.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:03:15,095 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=254225.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:03:50,701 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.473e+02 2.133e+02 2.439e+02 2.881e+02 4.883e+02, threshold=4.879e+02, percent-clipped=0.0 2023-05-02 03:03:52,994 INFO [train.py:904] (4/8) Epoch 26, batch 500, loss[loss=0.178, simple_loss=0.2514, pruned_loss=0.05231, over 12196.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2504, pruned_loss=0.03902, over 3047681.30 frames. ], batch size: 247, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:04:18,987 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=254271.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:04:57,428 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6887, 3.7453, 2.4368, 3.9776, 3.0362, 3.9598, 2.5532, 3.1498], device='cuda:4'), covar=tensor([0.0300, 0.0397, 0.1505, 0.0475, 0.0707, 0.0713, 0.1394, 0.0684], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0179, 0.0197, 0.0169, 0.0178, 0.0217, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:05:01,740 INFO [train.py:904] (4/8) Epoch 26, batch 550, loss[loss=0.161, simple_loss=0.2399, pruned_loss=0.04107, over 16625.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2506, pruned_loss=0.03892, over 3111759.57 frames. ], batch size: 89, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:05:05,512 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=254306.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:05:47,967 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-05-02 03:06:08,395 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.696e+02 2.164e+02 2.429e+02 2.784e+02 6.302e+02, threshold=4.858e+02, percent-clipped=1.0 2023-05-02 03:06:11,644 INFO [train.py:904] (4/8) Epoch 26, batch 600, loss[loss=0.1642, simple_loss=0.2609, pruned_loss=0.03374, over 16673.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2497, pruned_loss=0.03909, over 3153530.35 frames. ], batch size: 57, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:06:13,035 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=254354.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:07:21,734 INFO [train.py:904] (4/8) Epoch 26, batch 650, loss[loss=0.1572, simple_loss=0.2533, pruned_loss=0.03051, over 16671.00 frames. ], tot_loss[loss=0.1617, simple_loss=0.2477, pruned_loss=0.03789, over 3180563.82 frames. ], batch size: 57, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:07:22,587 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-05-02 03:07:34,415 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.83 vs. limit=5.0 2023-05-02 03:08:28,770 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.596e+02 2.176e+02 2.565e+02 3.138e+02 6.256e+02, threshold=5.131e+02, percent-clipped=2.0 2023-05-02 03:08:30,991 INFO [train.py:904] (4/8) Epoch 26, batch 700, loss[loss=0.1634, simple_loss=0.2625, pruned_loss=0.03214, over 17027.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2481, pruned_loss=0.03812, over 3215793.37 frames. ], batch size: 50, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:08:34,500 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2218, 2.5760, 2.1276, 2.3873, 2.9386, 2.6229, 2.9252, 2.9865], device='cuda:4'), covar=tensor([0.0285, 0.0499, 0.0631, 0.0598, 0.0343, 0.0452, 0.0366, 0.0373], device='cuda:4'), in_proj_covar=tensor([0.0226, 0.0246, 0.0234, 0.0234, 0.0245, 0.0244, 0.0241, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:09:41,471 INFO [train.py:904] (4/8) Epoch 26, batch 750, loss[loss=0.1374, simple_loss=0.2251, pruned_loss=0.02486, over 16779.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2487, pruned_loss=0.03833, over 3240591.29 frames. ], batch size: 39, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:10:27,464 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=254536.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:10:46,994 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.386e+02 2.194e+02 2.558e+02 3.127e+02 5.314e+02, threshold=5.115e+02, percent-clipped=1.0 2023-05-02 03:10:50,620 INFO [train.py:904] (4/8) Epoch 26, batch 800, loss[loss=0.1591, simple_loss=0.2515, pruned_loss=0.03338, over 17197.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.249, pruned_loss=0.03832, over 3265195.34 frames. ], batch size: 46, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:10:50,953 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2713, 5.8131, 5.8785, 5.5827, 5.7244, 6.2648, 5.7878, 5.4738], device='cuda:4'), covar=tensor([0.0994, 0.1944, 0.2795, 0.2284, 0.2642, 0.1008, 0.1610, 0.2519], device='cuda:4'), in_proj_covar=tensor([0.0423, 0.0622, 0.0691, 0.0512, 0.0677, 0.0713, 0.0535, 0.0681], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 03:11:20,517 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=254574.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:11:51,498 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=254597.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:11:59,172 INFO [train.py:904] (4/8) Epoch 26, batch 850, loss[loss=0.1519, simple_loss=0.2489, pruned_loss=0.02744, over 17216.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2479, pruned_loss=0.03758, over 3290260.46 frames. ], batch size: 45, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:12:45,183 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=254635.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:13:00,379 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9912, 4.4301, 4.4292, 3.2323, 3.6819, 4.4155, 3.9354, 2.8410], device='cuda:4'), covar=tensor([0.0491, 0.0067, 0.0050, 0.0362, 0.0149, 0.0088, 0.0104, 0.0432], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0087, 0.0088, 0.0135, 0.0101, 0.0112, 0.0097, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 03:13:08,051 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.339e+02 2.023e+02 2.316e+02 2.771e+02 4.972e+02, threshold=4.632e+02, percent-clipped=1.0 2023-05-02 03:13:10,208 INFO [train.py:904] (4/8) Epoch 26, batch 900, loss[loss=0.1983, simple_loss=0.2802, pruned_loss=0.05813, over 16768.00 frames. ], tot_loss[loss=0.1602, simple_loss=0.2465, pruned_loss=0.0369, over 3304524.09 frames. ], batch size: 102, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:13:49,478 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2770, 5.8361, 5.9576, 5.6692, 5.8259, 6.3457, 5.8588, 5.5114], device='cuda:4'), covar=tensor([0.0869, 0.1930, 0.2712, 0.2320, 0.2525, 0.0881, 0.1647, 0.2552], device='cuda:4'), in_proj_covar=tensor([0.0424, 0.0623, 0.0694, 0.0514, 0.0680, 0.0714, 0.0538, 0.0683], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 03:14:19,297 INFO [train.py:904] (4/8) Epoch 26, batch 950, loss[loss=0.1438, simple_loss=0.2255, pruned_loss=0.03104, over 16850.00 frames. ], tot_loss[loss=0.1602, simple_loss=0.2465, pruned_loss=0.03698, over 3301108.45 frames. ], batch size: 42, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:14:24,833 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-02 03:14:46,160 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8918, 2.0033, 2.4179, 2.7956, 2.7576, 2.7996, 2.0388, 3.0323], device='cuda:4'), covar=tensor([0.0203, 0.0576, 0.0423, 0.0312, 0.0377, 0.0358, 0.0626, 0.0221], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0199, 0.0186, 0.0188, 0.0205, 0.0163, 0.0201, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:15:24,094 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.276e+02 2.119e+02 2.483e+02 2.968e+02 5.109e+02, threshold=4.965e+02, percent-clipped=3.0 2023-05-02 03:15:27,107 INFO [train.py:904] (4/8) Epoch 26, batch 1000, loss[loss=0.1436, simple_loss=0.2226, pruned_loss=0.0323, over 16782.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2457, pruned_loss=0.037, over 3304078.55 frames. ], batch size: 102, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:15:31,194 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=254755.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:15:48,841 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5102, 3.7294, 4.1962, 2.3536, 3.2995, 2.6856, 3.8832, 3.9338], device='cuda:4'), covar=tensor([0.0277, 0.0879, 0.0470, 0.1947, 0.0841, 0.0936, 0.0673, 0.0985], device='cuda:4'), in_proj_covar=tensor([0.0159, 0.0167, 0.0169, 0.0156, 0.0148, 0.0131, 0.0146, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:16:06,345 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-05-02 03:16:28,378 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=254797.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:16:35,898 INFO [train.py:904] (4/8) Epoch 26, batch 1050, loss[loss=0.1778, simple_loss=0.2509, pruned_loss=0.05239, over 16892.00 frames. ], tot_loss[loss=0.1601, simple_loss=0.2461, pruned_loss=0.03704, over 3302341.29 frames. ], batch size: 96, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:16:55,092 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=254816.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:17:21,675 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-02 03:17:42,898 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=254851.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:17:43,590 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.253e+02 2.171e+02 2.721e+02 3.139e+02 5.089e+02, threshold=5.442e+02, percent-clipped=1.0 2023-05-02 03:17:44,777 INFO [train.py:904] (4/8) Epoch 26, batch 1100, loss[loss=0.1493, simple_loss=0.2294, pruned_loss=0.03462, over 12241.00 frames. ], tot_loss[loss=0.1595, simple_loss=0.2457, pruned_loss=0.03672, over 3301365.99 frames. ], batch size: 248, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:17:53,539 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=254858.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:18:36,859 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=254892.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:18:53,028 INFO [train.py:904] (4/8) Epoch 26, batch 1150, loss[loss=0.1588, simple_loss=0.2534, pruned_loss=0.03207, over 17143.00 frames. ], tot_loss[loss=0.1592, simple_loss=0.2454, pruned_loss=0.03648, over 3302370.16 frames. ], batch size: 46, lr: 2.61e-03, grad_scale: 2.0 2023-05-02 03:19:04,330 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=254912.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:19:29,841 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=254930.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:19:51,495 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2382, 4.2985, 4.4333, 4.2476, 4.3112, 4.8290, 4.3457, 4.0284], device='cuda:4'), covar=tensor([0.1887, 0.2084, 0.2574, 0.2342, 0.2901, 0.1280, 0.1759, 0.2750], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0621, 0.0691, 0.0512, 0.0678, 0.0714, 0.0538, 0.0680], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 03:19:59,610 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.804e+02 2.093e+02 2.455e+02 3.027e+02 6.541e+02, threshold=4.911e+02, percent-clipped=2.0 2023-05-02 03:20:00,665 INFO [train.py:904] (4/8) Epoch 26, batch 1200, loss[loss=0.1574, simple_loss=0.2383, pruned_loss=0.03824, over 16866.00 frames. ], tot_loss[loss=0.1581, simple_loss=0.2442, pruned_loss=0.03603, over 3313707.75 frames. ], batch size: 102, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:21:10,675 INFO [train.py:904] (4/8) Epoch 26, batch 1250, loss[loss=0.1679, simple_loss=0.2413, pruned_loss=0.04731, over 16833.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2452, pruned_loss=0.03736, over 3313215.37 frames. ], batch size: 83, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:21:11,236 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8838, 2.5992, 2.0720, 2.4338, 2.9929, 2.7319, 2.9694, 3.0043], device='cuda:4'), covar=tensor([0.0242, 0.0425, 0.0601, 0.0427, 0.0239, 0.0339, 0.0229, 0.0314], device='cuda:4'), in_proj_covar=tensor([0.0228, 0.0247, 0.0234, 0.0235, 0.0246, 0.0245, 0.0244, 0.0243], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:21:45,811 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5277, 3.6157, 3.7404, 1.9034, 3.0419, 2.2009, 3.8629, 3.8599], device='cuda:4'), covar=tensor([0.0262, 0.0919, 0.0659, 0.2543, 0.1058, 0.1254, 0.0637, 0.1059], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0168, 0.0170, 0.0157, 0.0148, 0.0132, 0.0146, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 03:21:50,197 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-05-02 03:22:19,973 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.482e+02 2.021e+02 2.372e+02 2.873e+02 4.328e+02, threshold=4.744e+02, percent-clipped=0.0 2023-05-02 03:22:21,123 INFO [train.py:904] (4/8) Epoch 26, batch 1300, loss[loss=0.1501, simple_loss=0.2394, pruned_loss=0.03045, over 16596.00 frames. ], tot_loss[loss=0.1597, simple_loss=0.245, pruned_loss=0.03715, over 3324201.22 frames. ], batch size: 62, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:22:41,380 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-02 03:23:04,585 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0197, 3.9081, 4.0987, 4.2126, 4.2719, 3.8601, 4.1241, 4.2936], device='cuda:4'), covar=tensor([0.1573, 0.1084, 0.1165, 0.0621, 0.0614, 0.1584, 0.2094, 0.0695], device='cuda:4'), in_proj_covar=tensor([0.0674, 0.0824, 0.0954, 0.0842, 0.0638, 0.0662, 0.0695, 0.0807], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:23:30,835 INFO [train.py:904] (4/8) Epoch 26, batch 1350, loss[loss=0.1488, simple_loss=0.2354, pruned_loss=0.03109, over 15484.00 frames. ], tot_loss[loss=0.1597, simple_loss=0.2452, pruned_loss=0.03711, over 3305148.93 frames. ], batch size: 191, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:23:42,882 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=255111.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:24:38,901 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.606e+02 2.137e+02 2.466e+02 2.927e+02 7.701e+02, threshold=4.932e+02, percent-clipped=1.0 2023-05-02 03:24:40,859 INFO [train.py:904] (4/8) Epoch 26, batch 1400, loss[loss=0.1468, simple_loss=0.2417, pruned_loss=0.02595, over 17112.00 frames. ], tot_loss[loss=0.1593, simple_loss=0.2444, pruned_loss=0.03706, over 3308178.86 frames. ], batch size: 49, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:24:41,769 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=255153.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:25:23,419 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9027, 2.1206, 2.6540, 2.9564, 2.8018, 3.4848, 2.5279, 3.4519], device='cuda:4'), covar=tensor([0.0348, 0.0593, 0.0401, 0.0410, 0.0417, 0.0223, 0.0508, 0.0216], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0200, 0.0188, 0.0190, 0.0206, 0.0164, 0.0202, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:25:34,104 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=255192.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:25:48,426 INFO [train.py:904] (4/8) Epoch 26, batch 1450, loss[loss=0.1664, simple_loss=0.2335, pruned_loss=0.04963, over 16794.00 frames. ], tot_loss[loss=0.1592, simple_loss=0.2447, pruned_loss=0.03686, over 3315432.22 frames. ], batch size: 124, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:25:54,157 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=255207.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:26:25,717 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6231, 1.7942, 2.3008, 2.4985, 2.5622, 2.6421, 1.9748, 2.7578], device='cuda:4'), covar=tensor([0.0236, 0.0580, 0.0381, 0.0319, 0.0353, 0.0316, 0.0595, 0.0214], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0200, 0.0188, 0.0190, 0.0206, 0.0164, 0.0202, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:26:26,884 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=255230.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:26:40,748 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=255240.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:26:56,772 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.573e+02 2.211e+02 2.586e+02 2.983e+02 9.031e+02, threshold=5.172e+02, percent-clipped=4.0 2023-05-02 03:26:57,913 INFO [train.py:904] (4/8) Epoch 26, batch 1500, loss[loss=0.1581, simple_loss=0.2384, pruned_loss=0.03894, over 16880.00 frames. ], tot_loss[loss=0.1593, simple_loss=0.2444, pruned_loss=0.03704, over 3318411.41 frames. ], batch size: 96, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:27:31,539 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=255278.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:27:36,855 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4081, 3.3846, 3.4478, 3.5101, 3.5816, 3.3062, 3.4795, 3.6403], device='cuda:4'), covar=tensor([0.1318, 0.1006, 0.1178, 0.0713, 0.0656, 0.2180, 0.1266, 0.0843], device='cuda:4'), in_proj_covar=tensor([0.0683, 0.0835, 0.0964, 0.0851, 0.0645, 0.0670, 0.0704, 0.0816], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:28:04,972 INFO [train.py:904] (4/8) Epoch 26, batch 1550, loss[loss=0.1565, simple_loss=0.2554, pruned_loss=0.02873, over 17304.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2467, pruned_loss=0.03813, over 3315144.94 frames. ], batch size: 52, lr: 2.61e-03, grad_scale: 4.0 2023-05-02 03:29:12,928 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.706e+02 2.260e+02 2.709e+02 3.380e+02 1.198e+03, threshold=5.419e+02, percent-clipped=4.0 2023-05-02 03:29:14,103 INFO [train.py:904] (4/8) Epoch 26, batch 1600, loss[loss=0.162, simple_loss=0.2595, pruned_loss=0.03224, over 17088.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2486, pruned_loss=0.03866, over 3319985.37 frames. ], batch size: 48, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:30:03,502 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=255388.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:30:12,103 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.69 vs. limit=5.0 2023-05-02 03:30:23,416 INFO [train.py:904] (4/8) Epoch 26, batch 1650, loss[loss=0.206, simple_loss=0.2897, pruned_loss=0.06119, over 11986.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2501, pruned_loss=0.03964, over 3314005.35 frames. ], batch size: 246, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:30:35,816 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=255411.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:31:09,743 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9350, 5.2142, 5.4252, 5.1482, 5.2750, 5.8554, 5.3604, 5.0744], device='cuda:4'), covar=tensor([0.1121, 0.2107, 0.2534, 0.2280, 0.2609, 0.1073, 0.1594, 0.2363], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0633, 0.0701, 0.0519, 0.0686, 0.0726, 0.0545, 0.0690], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 03:31:28,968 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=255449.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:31:32,745 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.657e+02 2.212e+02 2.578e+02 3.028e+02 5.571e+02, threshold=5.156e+02, percent-clipped=1.0 2023-05-02 03:31:33,284 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9530, 4.1289, 2.7471, 4.7020, 3.2568, 4.6593, 2.8665, 3.3641], device='cuda:4'), covar=tensor([0.0315, 0.0381, 0.1513, 0.0331, 0.0805, 0.0549, 0.1448, 0.0789], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0183, 0.0201, 0.0175, 0.0181, 0.0223, 0.0208, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:31:34,037 INFO [train.py:904] (4/8) Epoch 26, batch 1700, loss[loss=0.1334, simple_loss=0.2172, pruned_loss=0.02476, over 16945.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2513, pruned_loss=0.03982, over 3316320.44 frames. ], batch size: 41, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:31:34,397 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=255453.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:31:43,178 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=255459.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:32:13,996 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6506, 2.6685, 2.1981, 2.5251, 2.9736, 2.7411, 3.2634, 3.1612], device='cuda:4'), covar=tensor([0.0190, 0.0506, 0.0646, 0.0510, 0.0330, 0.0426, 0.0308, 0.0338], device='cuda:4'), in_proj_covar=tensor([0.0231, 0.0249, 0.0236, 0.0237, 0.0248, 0.0247, 0.0246, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:32:40,140 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=255501.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:32:42,740 INFO [train.py:904] (4/8) Epoch 26, batch 1750, loss[loss=0.1716, simple_loss=0.2686, pruned_loss=0.03724, over 17047.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2524, pruned_loss=0.04017, over 3316596.47 frames. ], batch size: 53, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:32:47,932 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=255507.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:33:40,584 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8214, 3.3713, 3.8835, 2.1228, 3.9501, 3.9766, 3.1857, 3.0085], device='cuda:4'), covar=tensor([0.0694, 0.0279, 0.0177, 0.1118, 0.0120, 0.0195, 0.0413, 0.0430], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0113, 0.0101, 0.0141, 0.0086, 0.0132, 0.0131, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:33:49,124 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 2.153e+02 2.652e+02 3.222e+02 7.615e+02, threshold=5.303e+02, percent-clipped=5.0 2023-05-02 03:33:51,387 INFO [train.py:904] (4/8) Epoch 26, batch 1800, loss[loss=0.1771, simple_loss=0.2634, pruned_loss=0.04539, over 16689.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2529, pruned_loss=0.03975, over 3324612.89 frames. ], batch size: 62, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:33:54,879 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=255555.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:34:18,269 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3097, 5.2748, 5.0212, 4.5799, 5.1084, 1.9056, 4.8460, 4.8261], device='cuda:4'), covar=tensor([0.0085, 0.0081, 0.0235, 0.0384, 0.0105, 0.2933, 0.0155, 0.0261], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0172, 0.0209, 0.0185, 0.0187, 0.0218, 0.0199, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:34:59,037 INFO [train.py:904] (4/8) Epoch 26, batch 1850, loss[loss=0.1818, simple_loss=0.287, pruned_loss=0.0383, over 16690.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.254, pruned_loss=0.04004, over 3317333.81 frames. ], batch size: 57, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:35:26,673 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2887, 3.4139, 3.8279, 2.3244, 3.2439, 2.5499, 3.7053, 3.6903], device='cuda:4'), covar=tensor([0.0267, 0.0945, 0.0552, 0.1995, 0.0812, 0.1002, 0.0580, 0.0975], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0169, 0.0171, 0.0158, 0.0148, 0.0133, 0.0147, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 03:35:46,009 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0957, 3.0215, 1.9626, 3.2261, 2.3755, 3.2557, 2.1429, 2.5605], device='cuda:4'), covar=tensor([0.0374, 0.0539, 0.1721, 0.0359, 0.0912, 0.0783, 0.1509, 0.0836], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0184, 0.0201, 0.0176, 0.0182, 0.0223, 0.0208, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:36:05,023 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.481e+02 2.077e+02 2.353e+02 2.960e+02 6.152e+02, threshold=4.707e+02, percent-clipped=2.0 2023-05-02 03:36:06,157 INFO [train.py:904] (4/8) Epoch 26, batch 1900, loss[loss=0.1842, simple_loss=0.2655, pruned_loss=0.05139, over 15401.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2538, pruned_loss=0.03932, over 3301697.17 frames. ], batch size: 190, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:36:46,128 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.08 vs. limit=5.0 2023-05-02 03:37:01,477 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0621, 3.2744, 3.3653, 2.2860, 3.1775, 3.5289, 3.2384, 1.9505], device='cuda:4'), covar=tensor([0.0664, 0.0172, 0.0085, 0.0499, 0.0149, 0.0118, 0.0128, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0136, 0.0102, 0.0113, 0.0097, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 03:37:16,451 INFO [train.py:904] (4/8) Epoch 26, batch 1950, loss[loss=0.168, simple_loss=0.2615, pruned_loss=0.03725, over 17083.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2531, pruned_loss=0.03828, over 3312792.26 frames. ], batch size: 55, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:37:47,519 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 03:37:59,928 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1734, 3.2541, 3.2461, 5.2516, 4.3434, 4.4574, 2.1491, 3.5965], device='cuda:4'), covar=tensor([0.1236, 0.0733, 0.0995, 0.0185, 0.0224, 0.0374, 0.1434, 0.0682], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0180, 0.0199, 0.0199, 0.0206, 0.0220, 0.0210, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:38:12,955 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=255744.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:38:24,265 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 2.052e+02 2.408e+02 2.894e+02 5.578e+02, threshold=4.816e+02, percent-clipped=3.0 2023-05-02 03:38:25,464 INFO [train.py:904] (4/8) Epoch 26, batch 2000, loss[loss=0.1697, simple_loss=0.256, pruned_loss=0.04173, over 16747.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2518, pruned_loss=0.03782, over 3326077.54 frames. ], batch size: 83, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:38:28,257 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2298, 3.3702, 3.8100, 2.1920, 3.2096, 2.4700, 3.6133, 3.5832], device='cuda:4'), covar=tensor([0.0273, 0.0957, 0.0509, 0.2069, 0.0815, 0.1026, 0.0651, 0.0988], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0170, 0.0171, 0.0158, 0.0149, 0.0133, 0.0147, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 03:39:35,297 INFO [train.py:904] (4/8) Epoch 26, batch 2050, loss[loss=0.1574, simple_loss=0.2536, pruned_loss=0.0306, over 17148.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2527, pruned_loss=0.03838, over 3314261.28 frames. ], batch size: 47, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:40:03,846 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1926, 3.7652, 4.3507, 2.2468, 4.5650, 4.6549, 3.3991, 3.6126], device='cuda:4'), covar=tensor([0.0663, 0.0297, 0.0204, 0.1132, 0.0073, 0.0169, 0.0421, 0.0391], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0112, 0.0100, 0.0140, 0.0086, 0.0131, 0.0130, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:40:44,482 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.586e+02 2.146e+02 2.478e+02 2.962e+02 5.196e+02, threshold=4.956e+02, percent-clipped=1.0 2023-05-02 03:40:45,698 INFO [train.py:904] (4/8) Epoch 26, batch 2100, loss[loss=0.1802, simple_loss=0.2698, pruned_loss=0.04535, over 16731.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2534, pruned_loss=0.03861, over 3312633.83 frames. ], batch size: 57, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:41:53,127 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4373, 4.0322, 4.4714, 2.5759, 4.7672, 4.7952, 3.5176, 3.8437], device='cuda:4'), covar=tensor([0.0655, 0.0267, 0.0221, 0.1017, 0.0075, 0.0210, 0.0429, 0.0368], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0112, 0.0101, 0.0140, 0.0086, 0.0131, 0.0130, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:41:54,461 INFO [train.py:904] (4/8) Epoch 26, batch 2150, loss[loss=0.1483, simple_loss=0.2347, pruned_loss=0.03098, over 16059.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2543, pruned_loss=0.03912, over 3313950.07 frames. ], batch size: 35, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:42:43,997 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4311, 5.4131, 5.2060, 4.6756, 5.2840, 2.1709, 4.9951, 5.0841], device='cuda:4'), covar=tensor([0.0091, 0.0074, 0.0226, 0.0391, 0.0101, 0.2554, 0.0145, 0.0205], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0172, 0.0209, 0.0185, 0.0186, 0.0217, 0.0199, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:43:04,641 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.569e+02 2.196e+02 2.655e+02 3.354e+02 7.063e+02, threshold=5.310e+02, percent-clipped=7.0 2023-05-02 03:43:05,921 INFO [train.py:904] (4/8) Epoch 26, batch 2200, loss[loss=0.1511, simple_loss=0.246, pruned_loss=0.02809, over 17201.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2544, pruned_loss=0.03937, over 3316227.78 frames. ], batch size: 44, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:43:10,518 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=255956.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:43:49,377 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0008, 5.1081, 5.5140, 5.4666, 5.4718, 5.1718, 5.0645, 4.9295], device='cuda:4'), covar=tensor([0.0362, 0.0574, 0.0379, 0.0409, 0.0445, 0.0409, 0.0977, 0.0471], device='cuda:4'), in_proj_covar=tensor([0.0435, 0.0490, 0.0474, 0.0434, 0.0522, 0.0499, 0.0576, 0.0399], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 03:44:19,344 INFO [train.py:904] (4/8) Epoch 26, batch 2250, loss[loss=0.163, simple_loss=0.2517, pruned_loss=0.03711, over 16469.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2543, pruned_loss=0.03891, over 3322644.84 frames. ], batch size: 68, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:44:22,409 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.41 vs. limit=5.0 2023-05-02 03:44:40,040 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256017.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:44:48,574 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.49 vs. limit=2.0 2023-05-02 03:44:52,357 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-05-02 03:45:09,622 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256038.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:45:17,805 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=256044.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:45:27,436 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.334e+02 2.041e+02 2.398e+02 2.796e+02 4.429e+02, threshold=4.795e+02, percent-clipped=0.0 2023-05-02 03:45:29,157 INFO [train.py:904] (4/8) Epoch 26, batch 2300, loss[loss=0.1811, simple_loss=0.2606, pruned_loss=0.0508, over 16765.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2543, pruned_loss=0.03934, over 3316654.59 frames. ], batch size: 83, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:46:00,745 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3711, 2.3790, 2.4596, 4.1152, 2.2919, 2.7271, 2.4486, 2.5716], device='cuda:4'), covar=tensor([0.1452, 0.3839, 0.3202, 0.0651, 0.4334, 0.2718, 0.3862, 0.3782], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0470, 0.0385, 0.0338, 0.0446, 0.0536, 0.0441, 0.0549], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:46:09,918 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256081.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:46:14,183 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256084.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:46:24,423 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=256092.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:46:32,812 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9706, 2.3948, 2.4011, 2.7478, 2.3522, 3.0972, 1.7636, 2.8060], device='cuda:4'), covar=tensor([0.1168, 0.0653, 0.0993, 0.0196, 0.0153, 0.0360, 0.1421, 0.0700], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0181, 0.0200, 0.0200, 0.0207, 0.0221, 0.0210, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 03:46:34,517 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256099.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:46:39,719 INFO [train.py:904] (4/8) Epoch 26, batch 2350, loss[loss=0.1807, simple_loss=0.2739, pruned_loss=0.04376, over 16772.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2546, pruned_loss=0.04015, over 3306760.84 frames. ], batch size: 102, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:47:19,414 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256132.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:47:32,083 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256142.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:47:37,012 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256145.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:47:45,351 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 2.073e+02 2.391e+02 2.953e+02 7.650e+02, threshold=4.783e+02, percent-clipped=2.0 2023-05-02 03:47:46,517 INFO [train.py:904] (4/8) Epoch 26, batch 2400, loss[loss=0.1941, simple_loss=0.2838, pruned_loss=0.05225, over 17065.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2549, pruned_loss=0.03998, over 3321263.25 frames. ], batch size: 53, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:48:42,709 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256193.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 03:48:55,675 INFO [train.py:904] (4/8) Epoch 26, batch 2450, loss[loss=0.1765, simple_loss=0.2708, pruned_loss=0.04107, over 17031.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2561, pruned_loss=0.04055, over 3317429.87 frames. ], batch size: 55, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:49:33,768 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4117, 2.3702, 2.3877, 4.1091, 2.2699, 2.8321, 2.4642, 2.4945], device='cuda:4'), covar=tensor([0.1394, 0.3896, 0.3329, 0.0642, 0.4254, 0.2582, 0.3775, 0.3748], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0470, 0.0386, 0.0339, 0.0447, 0.0537, 0.0441, 0.0550], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:50:01,724 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.511e+02 2.193e+02 2.630e+02 3.165e+02 5.747e+02, threshold=5.261e+02, percent-clipped=2.0 2023-05-02 03:50:03,664 INFO [train.py:904] (4/8) Epoch 26, batch 2500, loss[loss=0.1607, simple_loss=0.2432, pruned_loss=0.03911, over 16816.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2559, pruned_loss=0.04016, over 3320469.79 frames. ], batch size: 102, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:50:20,861 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8608, 2.1066, 2.6158, 2.8805, 2.7139, 3.3905, 2.3365, 3.3960], device='cuda:4'), covar=tensor([0.0293, 0.0600, 0.0376, 0.0375, 0.0417, 0.0206, 0.0560, 0.0202], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0201, 0.0189, 0.0192, 0.0208, 0.0166, 0.0203, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:50:48,291 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256285.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:51:13,207 INFO [train.py:904] (4/8) Epoch 26, batch 2550, loss[loss=0.2063, simple_loss=0.2865, pruned_loss=0.06309, over 16280.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2563, pruned_loss=0.04023, over 3319338.43 frames. ], batch size: 165, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:51:26,214 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=256312.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:52:14,146 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256346.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:52:22,559 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.565e+02 2.118e+02 2.541e+02 2.890e+02 5.508e+02, threshold=5.081e+02, percent-clipped=1.0 2023-05-02 03:52:23,667 INFO [train.py:904] (4/8) Epoch 26, batch 2600, loss[loss=0.1746, simple_loss=0.2747, pruned_loss=0.03725, over 16613.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2566, pruned_loss=0.04013, over 3323054.03 frames. ], batch size: 57, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:52:51,382 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256372.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:53:21,458 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=256394.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:53:33,340 INFO [train.py:904] (4/8) Epoch 26, batch 2650, loss[loss=0.1671, simple_loss=0.253, pruned_loss=0.04056, over 16675.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2571, pruned_loss=0.04009, over 3329459.68 frames. ], batch size: 134, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:54:14,539 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256433.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:54:20,790 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=256437.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:54:24,140 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=256440.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:54:40,362 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.530e+02 2.157e+02 2.407e+02 2.834e+02 4.989e+02, threshold=4.815e+02, percent-clipped=0.0 2023-05-02 03:54:41,406 INFO [train.py:904] (4/8) Epoch 26, batch 2700, loss[loss=0.1917, simple_loss=0.2825, pruned_loss=0.05043, over 12506.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2569, pruned_loss=0.03943, over 3332919.97 frames. ], batch size: 246, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:55:06,483 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256471.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:55:28,868 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=256488.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 03:55:48,614 INFO [train.py:904] (4/8) Epoch 26, batch 2750, loss[loss=0.1866, simple_loss=0.2815, pruned_loss=0.04587, over 16989.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2567, pruned_loss=0.03887, over 3330666.92 frames. ], batch size: 55, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:56:05,790 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3654, 4.4225, 4.5600, 4.3217, 4.4070, 4.9950, 4.5195, 4.2152], device='cuda:4'), covar=tensor([0.1867, 0.2377, 0.2850, 0.2270, 0.3065, 0.1237, 0.1669, 0.2746], device='cuda:4'), in_proj_covar=tensor([0.0436, 0.0641, 0.0708, 0.0525, 0.0695, 0.0730, 0.0548, 0.0701], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 03:56:30,003 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256532.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:56:47,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1001, 4.5557, 3.3531, 2.4960, 2.8355, 2.8281, 4.9180, 3.7622], device='cuda:4'), covar=tensor([0.2742, 0.0636, 0.1728, 0.3144, 0.3174, 0.2027, 0.0364, 0.1579], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0275, 0.0312, 0.0323, 0.0304, 0.0273, 0.0302, 0.0350], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 03:56:56,714 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.369e+02 2.155e+02 2.522e+02 3.116e+02 8.730e+02, threshold=5.045e+02, percent-clipped=2.0 2023-05-02 03:56:58,711 INFO [train.py:904] (4/8) Epoch 26, batch 2800, loss[loss=0.1671, simple_loss=0.2628, pruned_loss=0.03566, over 17101.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2566, pruned_loss=0.03876, over 3336972.96 frames. ], batch size: 47, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:57:10,883 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2570, 3.3709, 3.5434, 2.4493, 3.2536, 3.6672, 3.3689, 2.1208], device='cuda:4'), covar=tensor([0.0528, 0.0140, 0.0066, 0.0399, 0.0126, 0.0097, 0.0110, 0.0483], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0136, 0.0101, 0.0113, 0.0097, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 03:58:07,641 INFO [train.py:904] (4/8) Epoch 26, batch 2850, loss[loss=0.1924, simple_loss=0.2903, pruned_loss=0.04721, over 16635.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2562, pruned_loss=0.03885, over 3327604.34 frames. ], batch size: 57, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:58:21,107 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=256612.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:58:30,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3473, 5.3516, 5.0610, 4.6129, 5.1931, 1.9980, 4.8921, 5.0313], device='cuda:4'), covar=tensor([0.0102, 0.0096, 0.0241, 0.0399, 0.0106, 0.2774, 0.0154, 0.0221], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0173, 0.0210, 0.0185, 0.0187, 0.0217, 0.0200, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 03:59:00,645 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=256641.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:59:15,039 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.454e+02 2.116e+02 2.591e+02 2.981e+02 5.904e+02, threshold=5.182e+02, percent-clipped=1.0 2023-05-02 03:59:16,864 INFO [train.py:904] (4/8) Epoch 26, batch 2900, loss[loss=0.1709, simple_loss=0.2479, pruned_loss=0.04693, over 16504.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2554, pruned_loss=0.03952, over 3319161.32 frames. ], batch size: 146, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 03:59:27,183 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=256660.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 03:59:33,254 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8581, 3.8036, 3.9442, 4.0424, 4.0833, 3.6985, 3.9480, 4.0943], device='cuda:4'), covar=tensor([0.1665, 0.1141, 0.1175, 0.0709, 0.0755, 0.1990, 0.2304, 0.0770], device='cuda:4'), in_proj_covar=tensor([0.0690, 0.0848, 0.0982, 0.0862, 0.0651, 0.0684, 0.0713, 0.0827], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:00:13,265 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=256694.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:00:25,265 INFO [train.py:904] (4/8) Epoch 26, batch 2950, loss[loss=0.1707, simple_loss=0.2534, pruned_loss=0.04401, over 16655.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2548, pruned_loss=0.03967, over 3317512.05 frames. ], batch size: 134, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:01:01,150 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=256728.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:01:13,434 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8923, 4.3041, 4.3607, 3.1648, 3.6540, 4.3177, 3.8891, 2.6520], device='cuda:4'), covar=tensor([0.0496, 0.0083, 0.0051, 0.0371, 0.0166, 0.0109, 0.0106, 0.0462], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0136, 0.0101, 0.0113, 0.0098, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 04:01:14,562 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=256737.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:01:18,102 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=256740.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:01:18,479 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.92 vs. limit=5.0 2023-05-02 04:01:20,854 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=256742.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:01:25,751 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9686, 2.8364, 2.6455, 4.4759, 3.4851, 4.1699, 1.6762, 3.0040], device='cuda:4'), covar=tensor([0.1265, 0.0695, 0.1117, 0.0167, 0.0195, 0.0398, 0.1569, 0.0777], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0179, 0.0198, 0.0200, 0.0205, 0.0218, 0.0208, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:01:35,074 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.668e+02 2.243e+02 2.622e+02 3.035e+02 4.998e+02, threshold=5.244e+02, percent-clipped=0.0 2023-05-02 04:01:35,089 INFO [train.py:904] (4/8) Epoch 26, batch 3000, loss[loss=0.1833, simple_loss=0.2632, pruned_loss=0.05169, over 16559.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2557, pruned_loss=0.0409, over 3309025.86 frames. ], batch size: 62, lr: 2.60e-03, grad_scale: 4.0 2023-05-02 04:01:35,090 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 04:01:44,000 INFO [train.py:938] (4/8) Epoch 26, validation: loss=0.1339, simple_loss=0.2392, pruned_loss=0.01435, over 944034.00 frames. 2023-05-02 04:01:44,001 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 04:01:57,418 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6642, 4.4931, 4.7039, 4.8547, 5.0040, 4.5443, 4.8921, 4.9647], device='cuda:4'), covar=tensor([0.1868, 0.1577, 0.1613, 0.0876, 0.0793, 0.1090, 0.1910, 0.1356], device='cuda:4'), in_proj_covar=tensor([0.0692, 0.0850, 0.0984, 0.0864, 0.0652, 0.0685, 0.0714, 0.0828], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:02:26,227 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1630, 4.1666, 4.4908, 4.4671, 4.5222, 4.2244, 4.2605, 4.1864], device='cuda:4'), covar=tensor([0.0365, 0.0658, 0.0446, 0.0465, 0.0493, 0.0487, 0.0851, 0.0646], device='cuda:4'), in_proj_covar=tensor([0.0444, 0.0500, 0.0485, 0.0444, 0.0532, 0.0511, 0.0592, 0.0409], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 04:02:27,415 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=256785.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:02:31,512 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=256788.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:02:31,592 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=256788.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:02:53,044 INFO [train.py:904] (4/8) Epoch 26, batch 3050, loss[loss=0.181, simple_loss=0.2698, pruned_loss=0.04612, over 16693.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2565, pruned_loss=0.04097, over 3314130.04 frames. ], batch size: 62, lr: 2.60e-03, grad_scale: 4.0 2023-05-02 04:03:27,135 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=256827.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:03:27,228 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256827.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 04:03:31,441 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5254, 3.7059, 3.8861, 2.6343, 3.5411, 3.9736, 3.6493, 2.2532], device='cuda:4'), covar=tensor([0.0564, 0.0194, 0.0071, 0.0462, 0.0125, 0.0116, 0.0118, 0.0542], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0136, 0.0102, 0.0113, 0.0098, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 04:03:38,906 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=256836.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:03:52,057 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-02 04:04:02,821 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.494e+02 2.111e+02 2.504e+02 3.018e+02 1.623e+03, threshold=5.009e+02, percent-clipped=3.0 2023-05-02 04:04:02,836 INFO [train.py:904] (4/8) Epoch 26, batch 3100, loss[loss=0.1672, simple_loss=0.2457, pruned_loss=0.04438, over 16276.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2554, pruned_loss=0.04065, over 3319228.51 frames. ], batch size: 165, lr: 2.60e-03, grad_scale: 4.0 2023-05-02 04:04:06,745 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3030, 4.3513, 4.6464, 4.6262, 4.6643, 4.3842, 4.3920, 4.3011], device='cuda:4'), covar=tensor([0.0370, 0.0733, 0.0386, 0.0409, 0.0528, 0.0453, 0.0817, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0445, 0.0501, 0.0485, 0.0444, 0.0533, 0.0512, 0.0593, 0.0410], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 04:04:19,871 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7796, 2.7267, 2.5201, 2.6119, 2.9948, 2.7997, 3.3220, 3.2272], device='cuda:4'), covar=tensor([0.0164, 0.0497, 0.0538, 0.0522, 0.0353, 0.0467, 0.0268, 0.0306], device='cuda:4'), in_proj_covar=tensor([0.0234, 0.0248, 0.0237, 0.0238, 0.0249, 0.0248, 0.0248, 0.0248], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:04:44,629 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-05-02 04:04:47,258 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=256883.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 04:04:53,573 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256888.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 04:05:13,354 INFO [train.py:904] (4/8) Epoch 26, batch 3150, loss[loss=0.1667, simple_loss=0.2571, pruned_loss=0.03818, over 17011.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2544, pruned_loss=0.04014, over 3323820.82 frames. ], batch size: 50, lr: 2.60e-03, grad_scale: 4.0 2023-05-02 04:06:05,091 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=256941.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:06:09,417 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=256944.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 04:06:20,702 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8787, 4.8146, 4.7950, 4.4388, 4.4900, 4.8508, 4.6621, 4.5740], device='cuda:4'), covar=tensor([0.0728, 0.0758, 0.0351, 0.0349, 0.0927, 0.0532, 0.0475, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0472, 0.0369, 0.0374, 0.0372, 0.0426, 0.0253, 0.0445], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 04:06:21,405 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.612e+02 2.162e+02 2.494e+02 3.116e+02 1.108e+03, threshold=4.988e+02, percent-clipped=4.0 2023-05-02 04:06:21,420 INFO [train.py:904] (4/8) Epoch 26, batch 3200, loss[loss=0.177, simple_loss=0.2597, pruned_loss=0.04715, over 16811.00 frames. ], tot_loss[loss=0.167, simple_loss=0.254, pruned_loss=0.03998, over 3316533.48 frames. ], batch size: 83, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:07:11,954 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=256989.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:07:29,905 INFO [train.py:904] (4/8) Epoch 26, batch 3250, loss[loss=0.143, simple_loss=0.2406, pruned_loss=0.02264, over 17113.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2537, pruned_loss=0.04013, over 3325530.13 frames. ], batch size: 47, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:07:35,817 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-05-02 04:07:43,699 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257013.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:08:03,348 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=257028.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:08:38,569 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.590e+02 2.275e+02 2.633e+02 3.156e+02 7.653e+02, threshold=5.267e+02, percent-clipped=3.0 2023-05-02 04:08:38,584 INFO [train.py:904] (4/8) Epoch 26, batch 3300, loss[loss=0.1792, simple_loss=0.2781, pruned_loss=0.04014, over 17117.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2548, pruned_loss=0.04026, over 3326021.28 frames. ], batch size: 48, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:09:07,541 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=257074.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:09:09,485 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=257076.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:09:48,082 INFO [train.py:904] (4/8) Epoch 26, batch 3350, loss[loss=0.1801, simple_loss=0.2629, pruned_loss=0.04871, over 16327.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2553, pruned_loss=0.0401, over 3322852.50 frames. ], batch size: 165, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:10:20,821 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=257127.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:10:24,189 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-05-02 04:10:28,549 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9205, 2.2879, 2.3054, 2.6632, 2.0390, 3.1397, 1.7955, 2.6036], device='cuda:4'), covar=tensor([0.1272, 0.0763, 0.1154, 0.0180, 0.0119, 0.0319, 0.1600, 0.0806], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0180, 0.0199, 0.0201, 0.0207, 0.0219, 0.0208, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:10:56,576 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.602e+02 2.179e+02 2.494e+02 3.258e+02 7.595e+02, threshold=4.988e+02, percent-clipped=3.0 2023-05-02 04:10:56,591 INFO [train.py:904] (4/8) Epoch 26, batch 3400, loss[loss=0.1564, simple_loss=0.2525, pruned_loss=0.03016, over 16620.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2545, pruned_loss=0.03956, over 3317112.83 frames. ], batch size: 62, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:11:02,202 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0623, 4.0848, 3.9689, 3.6782, 3.7542, 4.0818, 3.6682, 3.8335], device='cuda:4'), covar=tensor([0.0581, 0.0606, 0.0296, 0.0300, 0.0649, 0.0457, 0.1228, 0.0563], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0474, 0.0371, 0.0375, 0.0373, 0.0428, 0.0253, 0.0446], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 04:11:13,582 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257165.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:11:27,518 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=257175.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:11:39,234 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=257183.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 04:12:06,434 INFO [train.py:904] (4/8) Epoch 26, batch 3450, loss[loss=0.1731, simple_loss=0.2571, pruned_loss=0.04452, over 15482.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2528, pruned_loss=0.03911, over 3318373.42 frames. ], batch size: 190, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:12:34,024 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257222.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:12:39,832 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=257226.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:12:57,438 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=257239.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 04:12:57,849 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-05-02 04:12:59,798 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257241.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:13:16,457 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.387e+02 2.059e+02 2.361e+02 2.793e+02 7.239e+02, threshold=4.722e+02, percent-clipped=1.0 2023-05-02 04:13:16,472 INFO [train.py:904] (4/8) Epoch 26, batch 3500, loss[loss=0.1914, simple_loss=0.2747, pruned_loss=0.05404, over 16378.00 frames. ], tot_loss[loss=0.165, simple_loss=0.252, pruned_loss=0.03903, over 3317546.97 frames. ], batch size: 145, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:13:58,929 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=257283.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:14:25,661 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=257302.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:14:26,377 INFO [train.py:904] (4/8) Epoch 26, batch 3550, loss[loss=0.171, simple_loss=0.2473, pruned_loss=0.04735, over 16658.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2508, pruned_loss=0.03875, over 3318836.16 frames. ], batch size: 89, lr: 2.60e-03, grad_scale: 8.0 2023-05-02 04:14:30,694 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0973, 4.4876, 4.4680, 3.3772, 3.6849, 4.4803, 4.0229, 2.7500], device='cuda:4'), covar=tensor([0.0454, 0.0068, 0.0047, 0.0346, 0.0151, 0.0099, 0.0090, 0.0466], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0135, 0.0102, 0.0113, 0.0098, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 04:14:33,864 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 04:14:57,870 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2056, 2.8365, 2.6370, 4.5025, 3.6734, 4.1112, 1.7887, 3.0536], device='cuda:4'), covar=tensor([0.1112, 0.0743, 0.1139, 0.0264, 0.0239, 0.0587, 0.1467, 0.0830], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0180, 0.0198, 0.0200, 0.0207, 0.0219, 0.0208, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:15:02,083 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9071, 4.3973, 4.4038, 3.2100, 3.6768, 4.3516, 3.9622, 2.7840], device='cuda:4'), covar=tensor([0.0499, 0.0069, 0.0046, 0.0365, 0.0150, 0.0118, 0.0089, 0.0421], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0135, 0.0102, 0.0113, 0.0098, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 04:15:30,035 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-05-02 04:15:34,859 INFO [train.py:904] (4/8) Epoch 26, batch 3600, loss[loss=0.1536, simple_loss=0.2502, pruned_loss=0.02849, over 17128.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2497, pruned_loss=0.03853, over 3325052.45 frames. ], batch size: 47, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:15:35,977 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.460e+02 2.009e+02 2.326e+02 2.677e+02 5.900e+02, threshold=4.651e+02, percent-clipped=2.0 2023-05-02 04:15:58,478 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=257369.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:15:58,635 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3635, 4.0165, 4.4920, 2.2735, 4.7126, 4.7615, 3.5217, 3.6907], device='cuda:4'), covar=tensor([0.0621, 0.0288, 0.0189, 0.1205, 0.0072, 0.0193, 0.0395, 0.0395], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0112, 0.0100, 0.0140, 0.0086, 0.0132, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:16:07,804 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1825, 5.1338, 4.9000, 4.0839, 5.0314, 1.8342, 4.6814, 4.6880], device='cuda:4'), covar=tensor([0.0116, 0.0108, 0.0292, 0.0579, 0.0130, 0.3230, 0.0184, 0.0327], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0174, 0.0212, 0.0187, 0.0188, 0.0218, 0.0201, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:16:28,054 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9691, 2.6279, 2.5209, 1.9682, 2.5485, 2.7520, 2.5706, 1.9047], device='cuda:4'), covar=tensor([0.0421, 0.0112, 0.0100, 0.0397, 0.0162, 0.0156, 0.0138, 0.0436], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0135, 0.0102, 0.0113, 0.0098, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 04:16:47,814 INFO [train.py:904] (4/8) Epoch 26, batch 3650, loss[loss=0.2029, simple_loss=0.2779, pruned_loss=0.06393, over 11716.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2489, pruned_loss=0.03921, over 3294601.27 frames. ], batch size: 248, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:16:49,339 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-05-02 04:17:58,062 INFO [train.py:904] (4/8) Epoch 26, batch 3700, loss[loss=0.1823, simple_loss=0.2522, pruned_loss=0.05617, over 16898.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2476, pruned_loss=0.04029, over 3284176.18 frames. ], batch size: 109, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:17:59,896 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.445e+02 2.076e+02 2.563e+02 3.144e+02 6.860e+02, threshold=5.126e+02, percent-clipped=4.0 2023-05-02 04:18:14,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3451, 4.3145, 4.2673, 4.0256, 4.0518, 4.3636, 4.0344, 4.1058], device='cuda:4'), covar=tensor([0.0689, 0.0849, 0.0348, 0.0337, 0.0688, 0.0482, 0.0750, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0476, 0.0371, 0.0376, 0.0374, 0.0431, 0.0254, 0.0448], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 04:18:15,070 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8157, 3.1214, 2.9782, 2.0014, 2.6142, 1.9579, 3.3743, 3.4505], device='cuda:4'), covar=tensor([0.0261, 0.0878, 0.0734, 0.2349, 0.1137, 0.1286, 0.0613, 0.0872], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0171, 0.0171, 0.0157, 0.0148, 0.0133, 0.0147, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 04:18:18,534 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257467.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:18:32,558 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2915, 3.9139, 4.3890, 2.3326, 4.6657, 4.7090, 3.3666, 3.7001], device='cuda:4'), covar=tensor([0.0655, 0.0321, 0.0212, 0.1202, 0.0063, 0.0100, 0.0428, 0.0409], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0112, 0.0100, 0.0139, 0.0086, 0.0131, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:18:41,443 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=257483.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 04:19:09,128 INFO [train.py:904] (4/8) Epoch 26, batch 3750, loss[loss=0.1613, simple_loss=0.2437, pruned_loss=0.0395, over 16312.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2487, pruned_loss=0.04147, over 3271759.53 frames. ], batch size: 165, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:19:21,896 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-02 04:19:33,996 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=257521.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:19:44,839 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=257528.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 04:19:49,560 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=257531.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 04:19:57,808 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-02 04:20:01,919 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=257539.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 04:20:20,023 INFO [train.py:904] (4/8) Epoch 26, batch 3800, loss[loss=0.1878, simple_loss=0.2631, pruned_loss=0.05622, over 16749.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2506, pruned_loss=0.04297, over 3278627.30 frames. ], batch size: 124, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:20:22,164 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.842e+02 2.272e+02 2.465e+02 2.848e+02 4.858e+02, threshold=4.931e+02, percent-clipped=0.0 2023-05-02 04:20:47,419 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4561, 5.4700, 5.3235, 4.9276, 4.9730, 5.4144, 5.2051, 5.0521], device='cuda:4'), covar=tensor([0.0536, 0.0324, 0.0265, 0.0285, 0.0974, 0.0308, 0.0320, 0.0562], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0476, 0.0371, 0.0375, 0.0374, 0.0431, 0.0254, 0.0448], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 04:20:55,722 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=257578.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:20:55,849 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257578.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 04:21:09,654 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=257587.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 04:21:09,966 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-02 04:21:21,688 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-05-02 04:21:22,614 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=257597.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:21:23,283 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-02 04:21:25,432 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-02 04:21:30,542 INFO [train.py:904] (4/8) Epoch 26, batch 3850, loss[loss=0.1534, simple_loss=0.2323, pruned_loss=0.03719, over 16816.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2507, pruned_loss=0.04361, over 3276150.01 frames. ], batch size: 102, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:21:37,849 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-05-02 04:22:10,061 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9806, 4.2409, 4.0565, 4.1188, 3.7805, 3.8443, 3.8982, 4.2256], device='cuda:4'), covar=tensor([0.1218, 0.0939, 0.1058, 0.0885, 0.0879, 0.1759, 0.0979, 0.1026], device='cuda:4'), in_proj_covar=tensor([0.0724, 0.0880, 0.0719, 0.0680, 0.0559, 0.0557, 0.0739, 0.0687], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:22:20,363 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=257639.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 04:22:41,227 INFO [train.py:904] (4/8) Epoch 26, batch 3900, loss[loss=0.1586, simple_loss=0.2386, pruned_loss=0.03928, over 16797.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2507, pruned_loss=0.044, over 3270135.43 frames. ], batch size: 90, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:22:42,469 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.694e+02 2.236e+02 2.511e+02 3.000e+02 4.952e+02, threshold=5.021e+02, percent-clipped=1.0 2023-05-02 04:23:04,483 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=257669.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:23:51,421 INFO [train.py:904] (4/8) Epoch 26, batch 3950, loss[loss=0.1583, simple_loss=0.2343, pruned_loss=0.04112, over 16713.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2502, pruned_loss=0.0446, over 3277832.87 frames. ], batch size: 134, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:23:59,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5549, 3.5678, 2.2903, 3.7691, 2.9048, 3.7252, 2.4363, 2.9418], device='cuda:4'), covar=tensor([0.0245, 0.0359, 0.1395, 0.0314, 0.0674, 0.0785, 0.1357, 0.0658], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0182, 0.0198, 0.0175, 0.0181, 0.0223, 0.0206, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:24:12,553 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=257717.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:25:02,900 INFO [train.py:904] (4/8) Epoch 26, batch 4000, loss[loss=0.1758, simple_loss=0.2505, pruned_loss=0.05054, over 16719.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2497, pruned_loss=0.04486, over 3289849.32 frames. ], batch size: 124, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:25:03,988 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.551e+02 2.101e+02 2.446e+02 2.886e+02 5.630e+02, threshold=4.893e+02, percent-clipped=0.0 2023-05-02 04:25:41,168 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9131, 2.1458, 2.2415, 3.3707, 2.0811, 2.4262, 2.2658, 2.3047], device='cuda:4'), covar=tensor([0.1573, 0.3619, 0.2920, 0.0726, 0.4130, 0.2515, 0.3673, 0.3185], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0474, 0.0387, 0.0340, 0.0448, 0.0542, 0.0444, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:26:12,849 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-05-02 04:26:13,221 INFO [train.py:904] (4/8) Epoch 26, batch 4050, loss[loss=0.1725, simple_loss=0.256, pruned_loss=0.04454, over 16398.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2504, pruned_loss=0.04395, over 3289888.87 frames. ], batch size: 146, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:26:14,325 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4204, 4.2361, 4.1922, 2.7789, 3.6623, 4.2563, 3.7627, 2.6026], device='cuda:4'), covar=tensor([0.0530, 0.0039, 0.0038, 0.0393, 0.0100, 0.0078, 0.0090, 0.0384], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0088, 0.0089, 0.0135, 0.0101, 0.0112, 0.0097, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 04:26:40,023 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=257821.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:26:42,417 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=257823.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 04:27:25,363 INFO [train.py:904] (4/8) Epoch 26, batch 4100, loss[loss=0.1904, simple_loss=0.2709, pruned_loss=0.05492, over 16388.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2521, pruned_loss=0.0437, over 3286866.09 frames. ], batch size: 35, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:27:26,539 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.240e+02 1.750e+02 2.045e+02 2.393e+02 4.321e+02, threshold=4.090e+02, percent-clipped=1.0 2023-05-02 04:27:35,058 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257859.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:27:49,113 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=257869.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:28:02,814 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=257878.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:28:10,483 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257883.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:28:30,697 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=257897.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:28:39,711 INFO [train.py:904] (4/8) Epoch 26, batch 4150, loss[loss=0.1601, simple_loss=0.2518, pruned_loss=0.03421, over 16686.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2585, pruned_loss=0.04532, over 3266917.35 frames. ], batch size: 62, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:29:07,267 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=257920.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:29:15,910 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=257926.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:29:28,836 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=257934.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 04:29:44,220 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=257944.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 04:29:45,301 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=257945.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:29:53,465 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=257950.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:29:56,804 INFO [train.py:904] (4/8) Epoch 26, batch 4200, loss[loss=0.2031, simple_loss=0.2956, pruned_loss=0.05523, over 15334.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2651, pruned_loss=0.04698, over 3233055.80 frames. ], batch size: 190, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:29:58,474 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.624e+02 2.083e+02 2.573e+02 2.968e+02 5.903e+02, threshold=5.147e+02, percent-clipped=5.0 2023-05-02 04:30:23,959 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7490, 1.8385, 1.6684, 1.4957, 1.9371, 1.6038, 1.5991, 1.8918], device='cuda:4'), covar=tensor([0.0215, 0.0326, 0.0481, 0.0402, 0.0257, 0.0325, 0.0175, 0.0239], device='cuda:4'), in_proj_covar=tensor([0.0232, 0.0246, 0.0234, 0.0236, 0.0247, 0.0245, 0.0246, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:30:51,200 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.22 vs. limit=5.0 2023-05-02 04:31:15,317 INFO [train.py:904] (4/8) Epoch 26, batch 4250, loss[loss=0.1692, simple_loss=0.2667, pruned_loss=0.03583, over 16563.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2682, pruned_loss=0.04633, over 3222062.62 frames. ], batch size: 68, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:31:29,166 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=258011.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:32:29,606 INFO [train.py:904] (4/8) Epoch 26, batch 4300, loss[loss=0.1834, simple_loss=0.283, pruned_loss=0.04183, over 16813.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2698, pruned_loss=0.04579, over 3206567.02 frames. ], batch size: 102, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:32:31,420 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.484e+02 2.134e+02 2.518e+02 2.964e+02 4.474e+02, threshold=5.035e+02, percent-clipped=0.0 2023-05-02 04:32:46,290 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-05-02 04:33:02,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4789, 3.6493, 2.6442, 2.2836, 2.4926, 2.4716, 3.9184, 3.2455], device='cuda:4'), covar=tensor([0.3323, 0.0719, 0.2119, 0.2623, 0.2701, 0.2115, 0.0532, 0.1399], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0275, 0.0311, 0.0322, 0.0306, 0.0271, 0.0302, 0.0349], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 04:33:34,725 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-05-02 04:33:45,781 INFO [train.py:904] (4/8) Epoch 26, batch 4350, loss[loss=0.174, simple_loss=0.2754, pruned_loss=0.03632, over 16821.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2732, pruned_loss=0.04687, over 3198612.35 frames. ], batch size: 102, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:33:46,527 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0191, 2.1984, 2.1970, 3.5543, 2.1390, 2.5116, 2.3341, 2.3090], device='cuda:4'), covar=tensor([0.1459, 0.3299, 0.3015, 0.0647, 0.3988, 0.2405, 0.3075, 0.3356], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0468, 0.0382, 0.0336, 0.0444, 0.0537, 0.0439, 0.0548], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:34:10,087 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9977, 3.5228, 3.4766, 2.2407, 3.2137, 3.5338, 3.2306, 1.9787], device='cuda:4'), covar=tensor([0.0603, 0.0054, 0.0065, 0.0463, 0.0115, 0.0105, 0.0128, 0.0509], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0135, 0.0101, 0.0113, 0.0098, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 04:34:15,593 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=258123.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:34:58,989 INFO [train.py:904] (4/8) Epoch 26, batch 4400, loss[loss=0.1837, simple_loss=0.2725, pruned_loss=0.04744, over 16780.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2752, pruned_loss=0.04775, over 3202103.75 frames. ], batch size: 76, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:35:00,098 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 2.313e+02 2.752e+02 3.228e+02 5.813e+02, threshold=5.503e+02, percent-clipped=1.0 2023-05-02 04:35:18,207 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-05-02 04:35:25,198 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=258171.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:35:46,878 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5540, 3.5680, 3.5290, 2.8891, 3.3697, 2.0902, 3.2026, 2.8236], device='cuda:4'), covar=tensor([0.0138, 0.0105, 0.0172, 0.0249, 0.0098, 0.2578, 0.0124, 0.0302], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0172, 0.0210, 0.0185, 0.0187, 0.0216, 0.0200, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:36:06,179 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2805, 4.1925, 4.3559, 4.4644, 4.5554, 4.1968, 4.5407, 4.6124], device='cuda:4'), covar=tensor([0.1517, 0.1083, 0.1169, 0.0572, 0.0478, 0.1106, 0.0642, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0678, 0.0831, 0.0956, 0.0841, 0.0638, 0.0668, 0.0693, 0.0806], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:36:11,040 INFO [train.py:904] (4/8) Epoch 26, batch 4450, loss[loss=0.1823, simple_loss=0.2851, pruned_loss=0.03978, over 16831.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.2787, pruned_loss=0.04941, over 3198672.25 frames. ], batch size: 102, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:36:28,200 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=258215.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:36:55,448 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=258234.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 04:37:03,959 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=258239.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 04:37:22,664 INFO [train.py:904] (4/8) Epoch 26, batch 4500, loss[loss=0.1839, simple_loss=0.2726, pruned_loss=0.0476, over 16833.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2794, pruned_loss=0.05031, over 3203314.13 frames. ], batch size: 83, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:37:23,847 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.439e+02 1.900e+02 2.188e+02 2.602e+02 4.588e+02, threshold=4.376e+02, percent-clipped=0.0 2023-05-02 04:37:54,640 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1035, 2.3870, 2.6079, 1.9706, 2.7139, 2.7640, 2.4155, 2.3666], device='cuda:4'), covar=tensor([0.0671, 0.0270, 0.0214, 0.0866, 0.0124, 0.0238, 0.0462, 0.0424], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0111, 0.0099, 0.0138, 0.0085, 0.0129, 0.0129, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:38:05,642 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=258282.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 04:38:35,271 INFO [train.py:904] (4/8) Epoch 26, batch 4550, loss[loss=0.1923, simple_loss=0.2915, pruned_loss=0.0466, over 16852.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2805, pruned_loss=0.05138, over 3214648.86 frames. ], batch size: 102, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:38:39,768 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=258306.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:39:48,617 INFO [train.py:904] (4/8) Epoch 26, batch 4600, loss[loss=0.1868, simple_loss=0.2749, pruned_loss=0.04932, over 16421.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2813, pruned_loss=0.05167, over 3229691.43 frames. ], batch size: 35, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:39:50,255 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.394e+02 1.782e+02 2.088e+02 2.418e+02 3.840e+02, threshold=4.176e+02, percent-clipped=0.0 2023-05-02 04:41:03,103 INFO [train.py:904] (4/8) Epoch 26, batch 4650, loss[loss=0.1819, simple_loss=0.2672, pruned_loss=0.04832, over 16498.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2801, pruned_loss=0.05154, over 3240031.88 frames. ], batch size: 68, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:42:10,333 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9072, 4.9214, 4.7382, 4.3641, 4.3928, 4.8421, 4.6851, 4.5192], device='cuda:4'), covar=tensor([0.0576, 0.0498, 0.0319, 0.0359, 0.1060, 0.0463, 0.0421, 0.0639], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0456, 0.0357, 0.0361, 0.0359, 0.0413, 0.0244, 0.0430], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:42:14,211 INFO [train.py:904] (4/8) Epoch 26, batch 4700, loss[loss=0.1719, simple_loss=0.2564, pruned_loss=0.0437, over 16331.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2778, pruned_loss=0.05065, over 3244370.79 frames. ], batch size: 68, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:42:16,007 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.404e+02 1.737e+02 2.039e+02 2.431e+02 7.214e+02, threshold=4.078e+02, percent-clipped=1.0 2023-05-02 04:42:16,495 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=258454.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:43:26,795 INFO [train.py:904] (4/8) Epoch 26, batch 4750, loss[loss=0.1631, simple_loss=0.2501, pruned_loss=0.03801, over 16440.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2746, pruned_loss=0.04903, over 3224847.81 frames. ], batch size: 68, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:43:38,898 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=258511.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:43:44,992 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=258515.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:43:45,130 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=258515.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:44:02,221 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7582, 2.8417, 2.5698, 5.0570, 3.8893, 4.1611, 1.6691, 2.9357], device='cuda:4'), covar=tensor([0.1469, 0.0896, 0.1424, 0.0219, 0.0332, 0.0440, 0.1786, 0.0969], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0180, 0.0198, 0.0199, 0.0207, 0.0217, 0.0207, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:44:07,275 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1523, 4.3255, 3.1179, 2.7934, 3.0713, 2.9589, 4.7970, 3.7903], device='cuda:4'), covar=tensor([0.2609, 0.0631, 0.1853, 0.2442, 0.2498, 0.1777, 0.0441, 0.1239], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0273, 0.0310, 0.0321, 0.0304, 0.0270, 0.0301, 0.0348], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 04:44:21,174 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=258539.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:44:41,655 INFO [train.py:904] (4/8) Epoch 26, batch 4800, loss[loss=0.1639, simple_loss=0.2501, pruned_loss=0.03881, over 16345.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2703, pruned_loss=0.04669, over 3230819.21 frames. ], batch size: 35, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:44:43,327 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.278e+02 1.748e+02 2.158e+02 2.549e+02 7.061e+02, threshold=4.316e+02, percent-clipped=3.0 2023-05-02 04:44:56,481 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=258563.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:45:11,403 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=258572.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:45:25,648 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1700, 1.5137, 1.9065, 2.0725, 2.2134, 2.3076, 1.7364, 2.2957], device='cuda:4'), covar=tensor([0.0260, 0.0595, 0.0344, 0.0413, 0.0376, 0.0269, 0.0624, 0.0188], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0200, 0.0188, 0.0193, 0.0208, 0.0166, 0.0205, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:45:34,302 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=258587.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:45:57,754 INFO [train.py:904] (4/8) Epoch 26, batch 4850, loss[loss=0.1754, simple_loss=0.2751, pruned_loss=0.03785, over 15515.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2708, pruned_loss=0.04582, over 3197839.40 frames. ], batch size: 191, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:46:03,254 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=258606.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:46:15,529 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4197, 3.5277, 3.6466, 3.6230, 3.6425, 3.4937, 3.5306, 3.5305], device='cuda:4'), covar=tensor([0.0363, 0.0617, 0.0477, 0.0472, 0.0501, 0.0478, 0.0718, 0.0510], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0474, 0.0460, 0.0423, 0.0509, 0.0485, 0.0563, 0.0391], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 04:47:14,826 INFO [train.py:904] (4/8) Epoch 26, batch 4900, loss[loss=0.1728, simple_loss=0.2664, pruned_loss=0.03957, over 16740.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2699, pruned_loss=0.04458, over 3184883.15 frames. ], batch size: 124, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:47:16,710 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.370e+02 1.994e+02 2.195e+02 2.536e+02 4.973e+02, threshold=4.389e+02, percent-clipped=2.0 2023-05-02 04:47:17,074 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=258654.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:47:42,417 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0337, 2.3907, 2.6072, 1.8817, 2.7610, 2.8185, 2.4809, 2.3345], device='cuda:4'), covar=tensor([0.0737, 0.0252, 0.0205, 0.1015, 0.0120, 0.0221, 0.0437, 0.0492], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0111, 0.0100, 0.0139, 0.0086, 0.0129, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:48:29,883 INFO [train.py:904] (4/8) Epoch 26, batch 4950, loss[loss=0.1869, simple_loss=0.2834, pruned_loss=0.04521, over 16618.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2694, pruned_loss=0.04429, over 3182669.15 frames. ], batch size: 76, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:49:41,062 INFO [train.py:904] (4/8) Epoch 26, batch 5000, loss[loss=0.1695, simple_loss=0.2617, pruned_loss=0.03864, over 16665.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2713, pruned_loss=0.04412, over 3191719.17 frames. ], batch size: 57, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:49:42,187 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.524e+02 2.044e+02 2.396e+02 2.809e+02 5.347e+02, threshold=4.792e+02, percent-clipped=3.0 2023-05-02 04:50:39,021 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-02 04:50:54,489 INFO [train.py:904] (4/8) Epoch 26, batch 5050, loss[loss=0.1763, simple_loss=0.2677, pruned_loss=0.04249, over 16837.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.272, pruned_loss=0.04411, over 3187947.72 frames. ], batch size: 42, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:51:05,555 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=258810.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:51:14,765 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-05-02 04:52:07,189 INFO [train.py:904] (4/8) Epoch 26, batch 5100, loss[loss=0.2109, simple_loss=0.2859, pruned_loss=0.06797, over 11930.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2706, pruned_loss=0.04395, over 3182960.55 frames. ], batch size: 246, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:52:08,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.295e+02 1.999e+02 2.336e+02 2.714e+02 4.372e+02, threshold=4.672e+02, percent-clipped=0.0 2023-05-02 04:52:23,882 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-02 04:52:27,818 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=258867.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:52:44,083 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7773, 4.5825, 4.7355, 4.9629, 5.1169, 4.6774, 5.1668, 5.1703], device='cuda:4'), covar=tensor([0.1791, 0.1401, 0.1967, 0.0862, 0.0592, 0.0872, 0.0533, 0.0642], device='cuda:4'), in_proj_covar=tensor([0.0660, 0.0810, 0.0933, 0.0818, 0.0621, 0.0649, 0.0676, 0.0785], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:53:21,028 INFO [train.py:904] (4/8) Epoch 26, batch 5150, loss[loss=0.1796, simple_loss=0.2789, pruned_loss=0.04017, over 16393.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2705, pruned_loss=0.04317, over 3183835.89 frames. ], batch size: 146, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:53:38,339 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-02 04:54:08,960 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7519, 2.3406, 2.1639, 3.1591, 1.6699, 3.4459, 1.5244, 2.5867], device='cuda:4'), covar=tensor([0.1508, 0.0965, 0.1510, 0.0210, 0.0125, 0.0394, 0.1924, 0.0941], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0179, 0.0197, 0.0198, 0.0205, 0.0215, 0.0206, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:54:35,102 INFO [train.py:904] (4/8) Epoch 26, batch 5200, loss[loss=0.155, simple_loss=0.245, pruned_loss=0.03257, over 16574.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2696, pruned_loss=0.04272, over 3185408.27 frames. ], batch size: 68, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:54:36,751 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 1.846e+02 2.126e+02 2.536e+02 3.966e+02, threshold=4.251e+02, percent-clipped=0.0 2023-05-02 04:54:53,520 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=258965.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:55:12,589 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7726, 2.3064, 2.2640, 3.1720, 1.9161, 3.4456, 1.5006, 2.6725], device='cuda:4'), covar=tensor([0.1371, 0.0869, 0.1310, 0.0176, 0.0128, 0.0371, 0.1766, 0.0831], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0179, 0.0197, 0.0198, 0.0205, 0.0215, 0.0206, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 04:55:44,655 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0190, 4.2646, 4.0579, 4.1476, 3.8089, 3.8779, 3.8618, 4.2310], device='cuda:4'), covar=tensor([0.1199, 0.0955, 0.1025, 0.0767, 0.0806, 0.1609, 0.1036, 0.1083], device='cuda:4'), in_proj_covar=tensor([0.0699, 0.0848, 0.0695, 0.0653, 0.0539, 0.0536, 0.0715, 0.0663], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 04:55:48,688 INFO [train.py:904] (4/8) Epoch 26, batch 5250, loss[loss=0.1717, simple_loss=0.268, pruned_loss=0.03775, over 16856.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2668, pruned_loss=0.04199, over 3190460.90 frames. ], batch size: 96, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:56:23,018 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=259026.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:57:03,177 INFO [train.py:904] (4/8) Epoch 26, batch 5300, loss[loss=0.1479, simple_loss=0.2433, pruned_loss=0.02625, over 16842.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2637, pruned_loss=0.04119, over 3187769.08 frames. ], batch size: 102, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:57:04,410 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.339e+02 1.887e+02 2.180e+02 2.580e+02 4.921e+02, threshold=4.360e+02, percent-clipped=4.0 2023-05-02 04:58:06,108 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5081, 4.6469, 4.8140, 4.5898, 4.6642, 5.1719, 4.6335, 4.2747], device='cuda:4'), covar=tensor([0.1329, 0.1685, 0.1844, 0.1911, 0.2206, 0.0867, 0.1508, 0.2440], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0613, 0.0674, 0.0503, 0.0666, 0.0704, 0.0526, 0.0671], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 04:58:18,026 INFO [train.py:904] (4/8) Epoch 26, batch 5350, loss[loss=0.1875, simple_loss=0.2912, pruned_loss=0.04188, over 15313.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2623, pruned_loss=0.04069, over 3196857.73 frames. ], batch size: 190, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:58:27,832 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=259110.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:59:29,333 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.07 vs. limit=5.0 2023-05-02 04:59:31,425 INFO [train.py:904] (4/8) Epoch 26, batch 5400, loss[loss=0.1742, simple_loss=0.2669, pruned_loss=0.04077, over 17011.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2646, pruned_loss=0.04167, over 3192091.71 frames. ], batch size: 41, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 04:59:32,572 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 1.944e+02 2.298e+02 2.661e+02 4.475e+02, threshold=4.595e+02, percent-clipped=1.0 2023-05-02 04:59:38,343 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=259158.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 04:59:53,045 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=259167.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:00:48,641 INFO [train.py:904] (4/8) Epoch 26, batch 5450, loss[loss=0.2437, simple_loss=0.3284, pruned_loss=0.07954, over 15229.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2679, pruned_loss=0.04345, over 3165069.41 frames. ], batch size: 190, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 05:01:08,446 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=259215.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:01:25,007 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=259226.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:01:56,019 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5017, 3.3595, 3.8417, 1.9123, 3.9657, 3.9916, 2.9454, 2.9445], device='cuda:4'), covar=tensor([0.0846, 0.0305, 0.0199, 0.1246, 0.0078, 0.0149, 0.0456, 0.0496], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0113, 0.0101, 0.0140, 0.0087, 0.0130, 0.0130, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 05:02:05,202 INFO [train.py:904] (4/8) Epoch 26, batch 5500, loss[loss=0.2155, simple_loss=0.3055, pruned_loss=0.06281, over 16219.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.2743, pruned_loss=0.04722, over 3146519.33 frames. ], batch size: 165, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 05:02:07,104 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.528e+02 2.367e+02 2.972e+02 3.909e+02 6.600e+02, threshold=5.944e+02, percent-clipped=13.0 2023-05-02 05:02:23,228 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=259264.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:02:58,261 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=259287.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:03:22,300 INFO [train.py:904] (4/8) Epoch 26, batch 5550, loss[loss=0.2822, simple_loss=0.3377, pruned_loss=0.1134, over 11277.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2817, pruned_loss=0.05246, over 3108915.24 frames. ], batch size: 248, lr: 2.59e-03, grad_scale: 8.0 2023-05-02 05:03:50,539 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=259321.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:03:57,664 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=259325.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 05:04:40,425 INFO [train.py:904] (4/8) Epoch 26, batch 5600, loss[loss=0.2578, simple_loss=0.3242, pruned_loss=0.09572, over 11199.00 frames. ], tot_loss[loss=0.2003, simple_loss=0.2866, pruned_loss=0.05695, over 3055249.71 frames. ], batch size: 248, lr: 2.58e-03, grad_scale: 16.0 2023-05-02 05:04:41,807 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.957e+02 3.012e+02 3.705e+02 4.647e+02 1.087e+03, threshold=7.410e+02, percent-clipped=8.0 2023-05-02 05:05:21,017 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7902, 2.5890, 2.3599, 3.8449, 2.5869, 3.8090, 1.4527, 2.8104], device='cuda:4'), covar=tensor([0.1325, 0.0835, 0.1341, 0.0232, 0.0223, 0.0437, 0.1776, 0.0840], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0180, 0.0199, 0.0200, 0.0208, 0.0218, 0.0209, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 05:06:04,635 INFO [train.py:904] (4/8) Epoch 26, batch 5650, loss[loss=0.2267, simple_loss=0.3034, pruned_loss=0.07502, over 15229.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2917, pruned_loss=0.06056, over 3041619.66 frames. ], batch size: 190, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:06:15,073 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4899, 3.4576, 2.6552, 2.1893, 2.3479, 2.3667, 3.6765, 3.1963], device='cuda:4'), covar=tensor([0.3144, 0.0737, 0.2045, 0.2968, 0.2744, 0.2248, 0.0559, 0.1530], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0273, 0.0309, 0.0321, 0.0302, 0.0270, 0.0301, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 05:06:19,842 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1773, 4.1433, 4.5336, 4.4790, 4.5350, 4.2476, 4.2557, 4.2194], device='cuda:4'), covar=tensor([0.0368, 0.0714, 0.0391, 0.0491, 0.0444, 0.0477, 0.0931, 0.0549], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0473, 0.0459, 0.0422, 0.0508, 0.0485, 0.0561, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 05:06:22,315 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=259414.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:06:33,210 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9413, 3.8598, 3.9388, 4.1186, 4.1941, 3.8205, 4.1376, 4.2237], device='cuda:4'), covar=tensor([0.1868, 0.1327, 0.1814, 0.0869, 0.0827, 0.1760, 0.1154, 0.0866], device='cuda:4'), in_proj_covar=tensor([0.0665, 0.0816, 0.0941, 0.0825, 0.0630, 0.0656, 0.0680, 0.0793], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:07:22,187 INFO [train.py:904] (4/8) Epoch 26, batch 5700, loss[loss=0.2356, simple_loss=0.3224, pruned_loss=0.07435, over 16902.00 frames. ], tot_loss[loss=0.2086, simple_loss=0.2932, pruned_loss=0.06194, over 3043887.96 frames. ], batch size: 116, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:07:25,089 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.978e+02 3.137e+02 3.722e+02 4.714e+02 8.155e+02, threshold=7.444e+02, percent-clipped=2.0 2023-05-02 05:07:56,941 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=259475.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:08:11,957 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7812, 3.7348, 3.9266, 3.6474, 3.8484, 4.2333, 3.8696, 3.6172], device='cuda:4'), covar=tensor([0.2108, 0.2353, 0.2410, 0.2681, 0.2654, 0.1755, 0.1797, 0.2748], device='cuda:4'), in_proj_covar=tensor([0.0423, 0.0621, 0.0684, 0.0511, 0.0671, 0.0711, 0.0533, 0.0679], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 05:08:39,968 INFO [train.py:904] (4/8) Epoch 26, batch 5750, loss[loss=0.2284, simple_loss=0.3042, pruned_loss=0.07632, over 15235.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.2953, pruned_loss=0.06291, over 3033090.77 frames. ], batch size: 190, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:09:24,283 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-05-02 05:10:02,530 INFO [train.py:904] (4/8) Epoch 26, batch 5800, loss[loss=0.1693, simple_loss=0.2666, pruned_loss=0.03601, over 16769.00 frames. ], tot_loss[loss=0.2084, simple_loss=0.2942, pruned_loss=0.06131, over 3041610.93 frames. ], batch size: 124, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:10:05,673 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.039e+02 2.780e+02 3.397e+02 4.110e+02 5.922e+02, threshold=6.793e+02, percent-clipped=0.0 2023-05-02 05:10:11,596 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-05-02 05:10:24,500 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 05:10:47,019 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=259582.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:11:19,030 INFO [train.py:904] (4/8) Epoch 26, batch 5850, loss[loss=0.2444, simple_loss=0.317, pruned_loss=0.08596, over 11569.00 frames. ], tot_loss[loss=0.2067, simple_loss=0.2927, pruned_loss=0.06037, over 3022733.74 frames. ], batch size: 246, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:11:43,627 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-05-02 05:11:44,962 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=259620.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 05:11:47,021 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=259621.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:12:39,963 INFO [train.py:904] (4/8) Epoch 26, batch 5900, loss[loss=0.2023, simple_loss=0.2943, pruned_loss=0.05517, over 16451.00 frames. ], tot_loss[loss=0.2068, simple_loss=0.2929, pruned_loss=0.06036, over 3048562.96 frames. ], batch size: 75, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:12:43,693 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.729e+02 2.720e+02 3.451e+02 4.224e+02 7.867e+02, threshold=6.903e+02, percent-clipped=3.0 2023-05-02 05:13:08,191 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=259669.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:14:01,045 INFO [train.py:904] (4/8) Epoch 26, batch 5950, loss[loss=0.2416, simple_loss=0.3295, pruned_loss=0.07683, over 16836.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2943, pruned_loss=0.05907, over 3061118.58 frames. ], batch size: 116, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:14:46,899 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-02 05:15:12,613 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=259749.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 05:15:18,035 INFO [train.py:904] (4/8) Epoch 26, batch 6000, loss[loss=0.1787, simple_loss=0.2721, pruned_loss=0.04262, over 16911.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2929, pruned_loss=0.05816, over 3083595.74 frames. ], batch size: 96, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:15:18,036 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 05:15:28,182 INFO [train.py:938] (4/8) Epoch 26, validation: loss=0.1485, simple_loss=0.2607, pruned_loss=0.01818, over 944034.00 frames. 2023-05-02 05:15:28,183 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 05:15:30,551 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.937e+02 2.681e+02 3.578e+02 4.333e+02 8.656e+02, threshold=7.155e+02, percent-clipped=3.0 2023-05-02 05:15:55,086 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=259770.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:16:13,711 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=259782.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:16:46,426 INFO [train.py:904] (4/8) Epoch 26, batch 6050, loss[loss=0.1801, simple_loss=0.2798, pruned_loss=0.04016, over 16652.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2906, pruned_loss=0.0572, over 3088945.45 frames. ], batch size: 76, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:16:59,670 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=259810.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 05:17:04,426 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0457, 3.1750, 3.1949, 2.1044, 3.0225, 3.2182, 3.0575, 1.8515], device='cuda:4'), covar=tensor([0.0596, 0.0085, 0.0085, 0.0473, 0.0130, 0.0146, 0.0120, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0090, 0.0135, 0.0102, 0.0114, 0.0098, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 05:17:51,053 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=259843.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 05:18:05,774 INFO [train.py:904] (4/8) Epoch 26, batch 6100, loss[loss=0.1721, simple_loss=0.2688, pruned_loss=0.03766, over 16863.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2893, pruned_loss=0.05584, over 3110141.53 frames. ], batch size: 96, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:18:09,301 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.693e+02 2.757e+02 3.100e+02 3.601e+02 8.920e+02, threshold=6.201e+02, percent-clipped=2.0 2023-05-02 05:18:27,886 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=259866.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:18:44,802 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-05-02 05:18:53,012 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=259882.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:19:23,804 INFO [train.py:904] (4/8) Epoch 26, batch 6150, loss[loss=0.2039, simple_loss=0.2804, pruned_loss=0.06371, over 11609.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2875, pruned_loss=0.05553, over 3106743.19 frames. ], batch size: 246, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:19:49,532 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=259920.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:19:49,723 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4369, 3.4913, 2.7334, 2.1291, 2.3185, 2.4121, 3.6762, 3.1759], device='cuda:4'), covar=tensor([0.3177, 0.0655, 0.1880, 0.2864, 0.2592, 0.2171, 0.0509, 0.1357], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0274, 0.0310, 0.0322, 0.0303, 0.0271, 0.0302, 0.0347], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 05:20:00,757 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=259927.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:20:05,846 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=259930.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:20:14,224 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-02 05:20:42,382 INFO [train.py:904] (4/8) Epoch 26, batch 6200, loss[loss=0.2019, simple_loss=0.2728, pruned_loss=0.06554, over 11690.00 frames. ], tot_loss[loss=0.198, simple_loss=0.2856, pruned_loss=0.05522, over 3096442.61 frames. ], batch size: 247, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:20:44,631 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.064e+02 2.750e+02 3.295e+02 3.920e+02 8.789e+02, threshold=6.590e+02, percent-clipped=2.0 2023-05-02 05:21:04,922 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=259968.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:22:00,546 INFO [train.py:904] (4/8) Epoch 26, batch 6250, loss[loss=0.244, simple_loss=0.3233, pruned_loss=0.08231, over 16718.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2855, pruned_loss=0.05485, over 3117787.33 frames. ], batch size: 124, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:22:20,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0355, 2.1869, 2.2696, 3.5622, 2.1145, 2.4769, 2.2793, 2.3214], device='cuda:4'), covar=tensor([0.1462, 0.3686, 0.3102, 0.0634, 0.4131, 0.2555, 0.3645, 0.3276], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0467, 0.0381, 0.0334, 0.0444, 0.0535, 0.0439, 0.0546], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:22:27,893 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2223, 2.8551, 3.1220, 1.8043, 3.2511, 3.3018, 2.7082, 2.5833], device='cuda:4'), covar=tensor([0.0826, 0.0287, 0.0224, 0.1221, 0.0122, 0.0210, 0.0465, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0112, 0.0100, 0.0139, 0.0086, 0.0130, 0.0129, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 05:22:41,283 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=260030.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:23:15,623 INFO [train.py:904] (4/8) Epoch 26, batch 6300, loss[loss=0.1984, simple_loss=0.2736, pruned_loss=0.06162, over 11326.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2853, pruned_loss=0.05417, over 3123955.58 frames. ], batch size: 247, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:23:19,601 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.731e+02 2.817e+02 3.510e+02 4.108e+02 7.742e+02, threshold=7.020e+02, percent-clipped=4.0 2023-05-02 05:23:44,830 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=260070.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:24:08,684 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6775, 3.9834, 2.9826, 2.3798, 2.5952, 2.5706, 4.2668, 3.4539], device='cuda:4'), covar=tensor([0.2934, 0.0573, 0.1781, 0.2863, 0.2783, 0.2076, 0.0455, 0.1377], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0273, 0.0309, 0.0322, 0.0303, 0.0271, 0.0300, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 05:24:16,374 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9438, 2.2137, 2.4646, 3.1520, 2.2509, 2.4489, 2.3682, 2.3318], device='cuda:4'), covar=tensor([0.1481, 0.3479, 0.2366, 0.0751, 0.3913, 0.2185, 0.3420, 0.3016], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0467, 0.0381, 0.0334, 0.0444, 0.0534, 0.0438, 0.0545], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:24:17,937 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=260091.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:24:35,599 INFO [train.py:904] (4/8) Epoch 26, batch 6350, loss[loss=0.1879, simple_loss=0.281, pruned_loss=0.04738, over 16465.00 frames. ], tot_loss[loss=0.1981, simple_loss=0.2853, pruned_loss=0.05544, over 3106293.21 frames. ], batch size: 68, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:24:39,472 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=260105.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 05:24:59,054 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=260118.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:25:23,019 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1462, 2.2850, 2.3073, 3.8315, 2.0998, 2.6292, 2.3647, 2.4045], device='cuda:4'), covar=tensor([0.1465, 0.3611, 0.3066, 0.0584, 0.4434, 0.2478, 0.3523, 0.3307], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0465, 0.0380, 0.0332, 0.0443, 0.0532, 0.0436, 0.0544], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:25:30,299 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=260138.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 05:25:52,108 INFO [train.py:904] (4/8) Epoch 26, batch 6400, loss[loss=0.1819, simple_loss=0.2776, pruned_loss=0.04311, over 16825.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2867, pruned_loss=0.05708, over 3070353.38 frames. ], batch size: 102, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:25:54,618 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.816e+02 2.921e+02 3.424e+02 4.125e+02 9.297e+02, threshold=6.848e+02, percent-clipped=3.0 2023-05-02 05:25:56,226 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-02 05:27:08,957 INFO [train.py:904] (4/8) Epoch 26, batch 6450, loss[loss=0.1835, simple_loss=0.272, pruned_loss=0.04751, over 16696.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2864, pruned_loss=0.05658, over 3064600.48 frames. ], batch size: 124, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:27:35,048 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.04 vs. limit=5.0 2023-05-02 05:27:39,884 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=260222.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:27:43,908 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=260224.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 05:27:52,703 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-05-02 05:28:28,519 INFO [train.py:904] (4/8) Epoch 26, batch 6500, loss[loss=0.1922, simple_loss=0.2795, pruned_loss=0.05239, over 17219.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2841, pruned_loss=0.05573, over 3066204.59 frames. ], batch size: 45, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:28:31,533 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.633e+02 2.830e+02 3.427e+02 3.864e+02 5.518e+02, threshold=6.855e+02, percent-clipped=0.0 2023-05-02 05:29:17,807 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.14 vs. limit=5.0 2023-05-02 05:29:20,105 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=260285.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 05:29:49,648 INFO [train.py:904] (4/8) Epoch 26, batch 6550, loss[loss=0.2001, simple_loss=0.3013, pruned_loss=0.04945, over 16675.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2871, pruned_loss=0.05651, over 3066895.82 frames. ], batch size: 62, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:30:14,448 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-02 05:31:07,555 INFO [train.py:904] (4/8) Epoch 26, batch 6600, loss[loss=0.1878, simple_loss=0.2818, pruned_loss=0.04695, over 16856.00 frames. ], tot_loss[loss=0.2014, simple_loss=0.2891, pruned_loss=0.05682, over 3068394.00 frames. ], batch size: 96, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:31:09,946 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.723e+02 2.646e+02 3.224e+02 4.276e+02 9.076e+02, threshold=6.447e+02, percent-clipped=2.0 2023-05-02 05:31:11,283 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6135, 3.7164, 2.4032, 4.2535, 2.9473, 4.2124, 2.4176, 3.0172], device='cuda:4'), covar=tensor([0.0300, 0.0384, 0.1659, 0.0192, 0.0859, 0.0530, 0.1630, 0.0835], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0181, 0.0197, 0.0171, 0.0179, 0.0220, 0.0206, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 05:31:21,661 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-05-02 05:31:59,445 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=260386.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:32:00,058 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.85 vs. limit=5.0 2023-05-02 05:32:26,375 INFO [train.py:904] (4/8) Epoch 26, batch 6650, loss[loss=0.2081, simple_loss=0.2948, pruned_loss=0.06073, over 16690.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2894, pruned_loss=0.05733, over 3077485.53 frames. ], batch size: 89, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:32:30,389 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=260405.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 05:33:21,097 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=260438.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:33:43,217 INFO [train.py:904] (4/8) Epoch 26, batch 6700, loss[loss=0.2028, simple_loss=0.278, pruned_loss=0.06384, over 16626.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2881, pruned_loss=0.05702, over 3076964.53 frames. ], batch size: 35, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:33:43,549 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=260453.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 05:33:45,910 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.097e+02 2.665e+02 3.203e+02 3.728e+02 7.997e+02, threshold=6.406e+02, percent-clipped=3.0 2023-05-02 05:34:35,490 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=260486.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:35:00,760 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-05-02 05:35:01,130 INFO [train.py:904] (4/8) Epoch 26, batch 6750, loss[loss=0.187, simple_loss=0.2664, pruned_loss=0.05378, over 16967.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2875, pruned_loss=0.05724, over 3084287.45 frames. ], batch size: 55, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:35:31,761 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=260522.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:35:34,841 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2779, 5.2710, 5.0066, 4.3277, 5.1992, 1.7147, 4.8810, 4.7323], device='cuda:4'), covar=tensor([0.0086, 0.0073, 0.0219, 0.0420, 0.0078, 0.3082, 0.0130, 0.0239], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0172, 0.0210, 0.0185, 0.0186, 0.0217, 0.0198, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:36:19,268 INFO [train.py:904] (4/8) Epoch 26, batch 6800, loss[loss=0.1953, simple_loss=0.2887, pruned_loss=0.05098, over 16287.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2871, pruned_loss=0.05688, over 3098324.49 frames. ], batch size: 165, lr: 2.58e-03, grad_scale: 8.0 2023-05-02 05:36:21,473 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.811e+02 2.923e+02 3.478e+02 4.093e+02 6.667e+02, threshold=6.957e+02, percent-clipped=2.0 2023-05-02 05:36:45,310 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=260570.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:36:47,146 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-05-02 05:36:51,840 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8493, 2.8028, 2.8638, 2.1464, 2.6478, 2.1646, 2.7033, 2.9325], device='cuda:4'), covar=tensor([0.0265, 0.0740, 0.0501, 0.1687, 0.0798, 0.0900, 0.0604, 0.0747], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0170, 0.0172, 0.0157, 0.0149, 0.0133, 0.0147, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 05:37:01,724 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=260580.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 05:37:13,627 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-05-02 05:37:27,524 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2604, 3.6142, 3.8207, 2.1527, 3.1735, 2.4182, 3.8515, 3.8495], device='cuda:4'), covar=tensor([0.0223, 0.0743, 0.0618, 0.2134, 0.0819, 0.1076, 0.0483, 0.0780], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0170, 0.0172, 0.0157, 0.0149, 0.0133, 0.0147, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 05:37:35,988 INFO [train.py:904] (4/8) Epoch 26, batch 6850, loss[loss=0.1783, simple_loss=0.2781, pruned_loss=0.03923, over 16373.00 frames. ], tot_loss[loss=0.2013, simple_loss=0.2883, pruned_loss=0.05712, over 3093777.60 frames. ], batch size: 165, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:37:58,886 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.28 vs. limit=5.0 2023-05-02 05:38:50,099 INFO [train.py:904] (4/8) Epoch 26, batch 6900, loss[loss=0.2059, simple_loss=0.298, pruned_loss=0.05692, over 16709.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2908, pruned_loss=0.05683, over 3095318.33 frames. ], batch size: 89, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:38:53,859 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.801e+02 2.543e+02 3.098e+02 3.712e+02 7.299e+02, threshold=6.197e+02, percent-clipped=1.0 2023-05-02 05:38:55,120 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4072, 2.6028, 2.2289, 2.3129, 2.9781, 2.6192, 2.9652, 3.1586], device='cuda:4'), covar=tensor([0.0166, 0.0478, 0.0577, 0.0544, 0.0269, 0.0415, 0.0258, 0.0306], device='cuda:4'), in_proj_covar=tensor([0.0224, 0.0240, 0.0231, 0.0231, 0.0241, 0.0239, 0.0239, 0.0239], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:39:09,494 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7932, 1.3664, 1.7342, 1.6489, 1.7424, 1.9191, 1.6658, 1.7895], device='cuda:4'), covar=tensor([0.0282, 0.0423, 0.0221, 0.0333, 0.0291, 0.0204, 0.0437, 0.0153], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0195, 0.0183, 0.0188, 0.0203, 0.0161, 0.0200, 0.0161], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:39:26,971 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-05-02 05:39:40,138 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=260686.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:40:05,851 INFO [train.py:904] (4/8) Epoch 26, batch 6950, loss[loss=0.247, simple_loss=0.3127, pruned_loss=0.09067, over 11376.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.292, pruned_loss=0.05837, over 3085569.65 frames. ], batch size: 248, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:40:06,993 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2364, 4.2335, 4.1436, 3.3475, 4.2066, 1.8129, 3.9986, 3.7677], device='cuda:4'), covar=tensor([0.0129, 0.0116, 0.0196, 0.0317, 0.0100, 0.2895, 0.0143, 0.0273], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0171, 0.0209, 0.0184, 0.0184, 0.0215, 0.0197, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:40:48,872 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=260730.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:40:54,432 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=260734.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:41:21,382 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=260751.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:41:23,879 INFO [train.py:904] (4/8) Epoch 26, batch 7000, loss[loss=0.217, simple_loss=0.3065, pruned_loss=0.06378, over 16683.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2927, pruned_loss=0.05821, over 3083667.67 frames. ], batch size: 124, lr: 2.58e-03, grad_scale: 2.0 2023-05-02 05:41:29,417 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.929e+02 2.843e+02 3.539e+02 4.501e+02 8.180e+02, threshold=7.079e+02, percent-clipped=4.0 2023-05-02 05:41:47,220 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0964, 2.3816, 2.5839, 1.8793, 2.6918, 2.8004, 2.4550, 2.3830], device='cuda:4'), covar=tensor([0.0739, 0.0265, 0.0255, 0.1063, 0.0153, 0.0269, 0.0461, 0.0447], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0111, 0.0100, 0.0138, 0.0086, 0.0129, 0.0129, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 05:42:24,289 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=260791.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:42:41,479 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0887, 2.1323, 2.1231, 3.6940, 2.1034, 2.4534, 2.2296, 2.2834], device='cuda:4'), covar=tensor([0.1527, 0.3640, 0.3192, 0.0659, 0.4404, 0.2604, 0.3732, 0.3537], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0465, 0.0379, 0.0332, 0.0443, 0.0532, 0.0437, 0.0544], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:42:42,009 INFO [train.py:904] (4/8) Epoch 26, batch 7050, loss[loss=0.2121, simple_loss=0.2965, pruned_loss=0.06387, over 16923.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2928, pruned_loss=0.05812, over 3070889.50 frames. ], batch size: 109, lr: 2.58e-03, grad_scale: 2.0 2023-05-02 05:42:56,400 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=260812.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:43:00,785 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2779, 4.3429, 4.6392, 4.6009, 4.6264, 4.3470, 4.3227, 4.2624], device='cuda:4'), covar=tensor([0.0383, 0.0663, 0.0419, 0.0439, 0.0555, 0.0456, 0.0969, 0.0591], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0477, 0.0463, 0.0425, 0.0511, 0.0487, 0.0564, 0.0391], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 05:43:06,809 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8300, 2.6981, 2.5195, 4.6369, 3.4786, 4.0737, 1.5736, 2.9729], device='cuda:4'), covar=tensor([0.1343, 0.0853, 0.1309, 0.0174, 0.0299, 0.0428, 0.1719, 0.0839], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0180, 0.0199, 0.0198, 0.0208, 0.0218, 0.0208, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 05:43:59,434 INFO [train.py:904] (4/8) Epoch 26, batch 7100, loss[loss=0.2155, simple_loss=0.2969, pruned_loss=0.06702, over 15347.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2918, pruned_loss=0.05825, over 3051990.79 frames. ], batch size: 190, lr: 2.58e-03, grad_scale: 2.0 2023-05-02 05:44:05,368 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.003e+02 2.616e+02 2.972e+02 3.634e+02 7.546e+02, threshold=5.943e+02, percent-clipped=1.0 2023-05-02 05:44:40,124 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.11 vs. limit=2.0 2023-05-02 05:44:42,688 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=260880.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 05:44:44,309 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=260881.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:45:07,325 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6194, 2.2904, 1.8487, 2.0886, 2.5872, 2.2499, 2.3134, 2.6965], device='cuda:4'), covar=tensor([0.0234, 0.0457, 0.0634, 0.0525, 0.0295, 0.0425, 0.0250, 0.0292], device='cuda:4'), in_proj_covar=tensor([0.0222, 0.0238, 0.0229, 0.0230, 0.0239, 0.0237, 0.0236, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 05:45:17,429 INFO [train.py:904] (4/8) Epoch 26, batch 7150, loss[loss=0.2121, simple_loss=0.2966, pruned_loss=0.06376, over 16636.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2897, pruned_loss=0.05755, over 3076161.98 frames. ], batch size: 62, lr: 2.58e-03, grad_scale: 2.0 2023-05-02 05:45:48,427 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3389, 3.4465, 3.5965, 3.5756, 3.5908, 3.4119, 3.4682, 3.4852], device='cuda:4'), covar=tensor([0.0423, 0.0797, 0.0466, 0.0461, 0.0596, 0.0626, 0.0832, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0477, 0.0462, 0.0426, 0.0511, 0.0487, 0.0563, 0.0390], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 05:45:53,426 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=260928.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 05:46:13,537 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=260942.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:46:29,363 INFO [train.py:904] (4/8) Epoch 26, batch 7200, loss[loss=0.2052, simple_loss=0.282, pruned_loss=0.06419, over 11569.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2876, pruned_loss=0.05607, over 3075507.50 frames. ], batch size: 246, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:46:35,554 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.884e+02 2.782e+02 3.336e+02 4.472e+02 6.403e+02, threshold=6.673e+02, percent-clipped=2.0 2023-05-02 05:47:05,503 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6201, 4.2096, 4.1354, 2.7068, 3.6828, 4.1884, 3.6835, 2.3511], device='cuda:4'), covar=tensor([0.0533, 0.0054, 0.0061, 0.0436, 0.0118, 0.0134, 0.0120, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0087, 0.0088, 0.0134, 0.0100, 0.0113, 0.0097, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 05:47:53,028 INFO [train.py:904] (4/8) Epoch 26, batch 7250, loss[loss=0.1799, simple_loss=0.2644, pruned_loss=0.04774, over 16759.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2855, pruned_loss=0.05487, over 3081741.22 frames. ], batch size: 124, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:48:51,423 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-05-02 05:49:09,252 INFO [train.py:904] (4/8) Epoch 26, batch 7300, loss[loss=0.1949, simple_loss=0.2961, pruned_loss=0.04689, over 16733.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2855, pruned_loss=0.05503, over 3080258.48 frames. ], batch size: 89, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:49:15,973 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.944e+02 2.610e+02 3.123e+02 3.831e+02 8.496e+02, threshold=6.245e+02, percent-clipped=1.0 2023-05-02 05:50:00,870 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=261086.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:50:26,307 INFO [train.py:904] (4/8) Epoch 26, batch 7350, loss[loss=0.1952, simple_loss=0.284, pruned_loss=0.05325, over 16187.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2864, pruned_loss=0.0557, over 3074361.95 frames. ], batch size: 165, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:50:33,367 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=261107.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:51:18,119 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=261136.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 05:51:44,988 INFO [train.py:904] (4/8) Epoch 26, batch 7400, loss[loss=0.2059, simple_loss=0.2964, pruned_loss=0.05775, over 16773.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2874, pruned_loss=0.05629, over 3079905.61 frames. ], batch size: 83, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:51:50,753 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.938e+02 2.670e+02 3.232e+02 3.696e+02 7.635e+02, threshold=6.463e+02, percent-clipped=2.0 2023-05-02 05:52:39,425 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7407, 4.8069, 5.1208, 5.0882, 5.1253, 4.8379, 4.7913, 4.6274], device='cuda:4'), covar=tensor([0.0314, 0.0526, 0.0409, 0.0433, 0.0467, 0.0411, 0.0908, 0.0495], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0474, 0.0460, 0.0422, 0.0509, 0.0483, 0.0561, 0.0389], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 05:52:55,043 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=261197.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 05:53:04,388 INFO [train.py:904] (4/8) Epoch 26, batch 7450, loss[loss=0.222, simple_loss=0.3105, pruned_loss=0.06676, over 16487.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2871, pruned_loss=0.05621, over 3101895.50 frames. ], batch size: 75, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:54:01,620 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=261237.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:54:17,449 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-05-02 05:54:27,095 INFO [train.py:904] (4/8) Epoch 26, batch 7500, loss[loss=0.1873, simple_loss=0.2775, pruned_loss=0.04853, over 16744.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.287, pruned_loss=0.05616, over 3088297.30 frames. ], batch size: 83, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:54:33,503 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.912e+02 2.815e+02 3.421e+02 3.855e+02 8.023e+02, threshold=6.841e+02, percent-clipped=2.0 2023-05-02 05:55:45,643 INFO [train.py:904] (4/8) Epoch 26, batch 7550, loss[loss=0.2155, simple_loss=0.2881, pruned_loss=0.07145, over 11803.00 frames. ], tot_loss[loss=0.1994, simple_loss=0.2859, pruned_loss=0.05638, over 3074730.49 frames. ], batch size: 248, lr: 2.58e-03, grad_scale: 4.0 2023-05-02 05:57:01,728 INFO [train.py:904] (4/8) Epoch 26, batch 7600, loss[loss=0.2412, simple_loss=0.3116, pruned_loss=0.08541, over 11621.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2854, pruned_loss=0.05705, over 3051177.19 frames. ], batch size: 248, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 05:57:07,521 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.730e+02 2.718e+02 3.314e+02 3.844e+02 7.163e+02, threshold=6.628e+02, percent-clipped=1.0 2023-05-02 05:57:22,479 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6632, 4.7298, 5.0741, 5.0486, 5.0464, 4.7204, 4.7289, 4.5508], device='cuda:4'), covar=tensor([0.0350, 0.0603, 0.0342, 0.0352, 0.0452, 0.0419, 0.1018, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0471, 0.0456, 0.0419, 0.0505, 0.0482, 0.0557, 0.0386], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 05:57:54,565 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=261386.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:58:04,582 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5031, 4.6421, 4.7875, 4.5473, 4.6398, 5.1655, 4.7046, 4.4253], device='cuda:4'), covar=tensor([0.1392, 0.2058, 0.2674, 0.2042, 0.2447, 0.1064, 0.1758, 0.2590], device='cuda:4'), in_proj_covar=tensor([0.0423, 0.0629, 0.0692, 0.0511, 0.0679, 0.0718, 0.0540, 0.0687], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 05:58:20,639 INFO [train.py:904] (4/8) Epoch 26, batch 7650, loss[loss=0.2521, simple_loss=0.3197, pruned_loss=0.0922, over 11790.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2868, pruned_loss=0.05821, over 3051008.74 frames. ], batch size: 248, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 05:58:26,577 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=261407.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:59:08,082 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=261434.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:59:36,565 INFO [train.py:904] (4/8) Epoch 26, batch 7700, loss[loss=0.1877, simple_loss=0.2781, pruned_loss=0.04862, over 16735.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2874, pruned_loss=0.05865, over 3050019.43 frames. ], batch size: 89, lr: 2.57e-03, grad_scale: 2.0 2023-05-02 05:59:40,111 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=261455.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:59:42,157 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=261456.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 05:59:46,566 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.047e+02 2.911e+02 3.600e+02 4.177e+02 7.928e+02, threshold=7.200e+02, percent-clipped=2.0 2023-05-02 06:00:36,508 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=261492.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 06:00:52,206 INFO [train.py:904] (4/8) Epoch 26, batch 7750, loss[loss=0.2113, simple_loss=0.2964, pruned_loss=0.06307, over 16620.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.2877, pruned_loss=0.05825, over 3056154.76 frames. ], batch size: 57, lr: 2.57e-03, grad_scale: 2.0 2023-05-02 06:00:58,067 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7402, 4.8128, 5.1502, 5.1151, 5.1178, 4.8243, 4.7903, 4.6249], device='cuda:4'), covar=tensor([0.0332, 0.0538, 0.0363, 0.0400, 0.0488, 0.0408, 0.0985, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0473, 0.0457, 0.0422, 0.0506, 0.0483, 0.0559, 0.0387], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 06:01:13,788 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=261517.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:01:44,919 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=261537.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:01:54,851 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-05-02 06:02:09,780 INFO [train.py:904] (4/8) Epoch 26, batch 7800, loss[loss=0.2145, simple_loss=0.3014, pruned_loss=0.06383, over 16191.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2882, pruned_loss=0.05816, over 3073979.41 frames. ], batch size: 35, lr: 2.57e-03, grad_scale: 2.0 2023-05-02 06:02:19,354 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.114e+02 2.737e+02 3.262e+02 3.975e+02 6.664e+02, threshold=6.524e+02, percent-clipped=0.0 2023-05-02 06:03:00,430 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=261585.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:03:28,319 INFO [train.py:904] (4/8) Epoch 26, batch 7850, loss[loss=0.1816, simple_loss=0.2762, pruned_loss=0.04347, over 17123.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2883, pruned_loss=0.05736, over 3077236.69 frames. ], batch size: 48, lr: 2.57e-03, grad_scale: 2.0 2023-05-02 06:04:46,515 INFO [train.py:904] (4/8) Epoch 26, batch 7900, loss[loss=0.1969, simple_loss=0.2909, pruned_loss=0.05139, over 16406.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2874, pruned_loss=0.0568, over 3084957.48 frames. ], batch size: 146, lr: 2.57e-03, grad_scale: 2.0 2023-05-02 06:04:55,564 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.825e+02 2.587e+02 3.177e+02 3.708e+02 5.885e+02, threshold=6.353e+02, percent-clipped=0.0 2023-05-02 06:06:05,689 INFO [train.py:904] (4/8) Epoch 26, batch 7950, loss[loss=0.2144, simple_loss=0.2953, pruned_loss=0.06672, over 15277.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.288, pruned_loss=0.05753, over 3072194.71 frames. ], batch size: 191, lr: 2.57e-03, grad_scale: 2.0 2023-05-02 06:06:11,203 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-05-02 06:07:23,909 INFO [train.py:904] (4/8) Epoch 26, batch 8000, loss[loss=0.2082, simple_loss=0.3003, pruned_loss=0.058, over 16907.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2883, pruned_loss=0.05799, over 3070460.57 frames. ], batch size: 96, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 06:07:29,097 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6117, 3.7241, 2.7530, 2.2444, 2.3872, 2.4159, 4.0460, 3.3099], device='cuda:4'), covar=tensor([0.3240, 0.0678, 0.1991, 0.3051, 0.3073, 0.2315, 0.0446, 0.1449], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0274, 0.0310, 0.0323, 0.0305, 0.0273, 0.0302, 0.0349], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 06:07:32,670 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.840e+02 2.545e+02 3.130e+02 3.720e+02 6.505e+02, threshold=6.260e+02, percent-clipped=1.0 2023-05-02 06:07:37,805 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6425, 5.9602, 5.6937, 5.7365, 5.2780, 5.2652, 5.3845, 6.0857], device='cuda:4'), covar=tensor([0.1177, 0.0799, 0.0957, 0.0834, 0.0888, 0.0802, 0.1263, 0.0761], device='cuda:4'), in_proj_covar=tensor([0.0699, 0.0839, 0.0691, 0.0647, 0.0536, 0.0536, 0.0707, 0.0657], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 06:08:24,244 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=261792.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 06:08:40,303 INFO [train.py:904] (4/8) Epoch 26, batch 8050, loss[loss=0.1944, simple_loss=0.2846, pruned_loss=0.05215, over 16193.00 frames. ], tot_loss[loss=0.2019, simple_loss=0.2882, pruned_loss=0.0578, over 3078637.82 frames. ], batch size: 165, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 06:08:54,192 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=261812.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:09:12,987 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=261824.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:09:13,249 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-05-02 06:09:27,614 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=261833.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:09:37,846 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=261840.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 06:09:57,224 INFO [train.py:904] (4/8) Epoch 26, batch 8100, loss[loss=0.1957, simple_loss=0.2835, pruned_loss=0.05393, over 16905.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2876, pruned_loss=0.05699, over 3088417.64 frames. ], batch size: 109, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 06:10:06,903 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.718e+02 2.493e+02 2.989e+02 3.536e+02 5.601e+02, threshold=5.979e+02, percent-clipped=0.0 2023-05-02 06:10:46,778 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=261885.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:11:00,388 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=261894.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:11:14,754 INFO [train.py:904] (4/8) Epoch 26, batch 8150, loss[loss=0.1944, simple_loss=0.2687, pruned_loss=0.06007, over 11372.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2851, pruned_loss=0.05592, over 3089096.99 frames. ], batch size: 250, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 06:12:18,368 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5057, 3.5354, 2.6591, 2.1751, 2.3024, 2.3532, 3.6730, 3.1745], device='cuda:4'), covar=tensor([0.3217, 0.0736, 0.2055, 0.3006, 0.2763, 0.2327, 0.0596, 0.1441], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0274, 0.0310, 0.0323, 0.0305, 0.0273, 0.0302, 0.0348], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 06:12:24,807 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-05-02 06:12:33,065 INFO [train.py:904] (4/8) Epoch 26, batch 8200, loss[loss=0.2099, simple_loss=0.279, pruned_loss=0.07037, over 11133.00 frames. ], tot_loss[loss=0.1962, simple_loss=0.2825, pruned_loss=0.05494, over 3104454.21 frames. ], batch size: 246, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 06:12:43,206 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.047e+02 2.684e+02 3.208e+02 4.137e+02 6.714e+02, threshold=6.416e+02, percent-clipped=2.0 2023-05-02 06:13:12,968 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=261977.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:13:58,686 INFO [train.py:904] (4/8) Epoch 26, batch 8250, loss[loss=0.1698, simple_loss=0.2688, pruned_loss=0.03542, over 16792.00 frames. ], tot_loss[loss=0.1932, simple_loss=0.2815, pruned_loss=0.0525, over 3086270.28 frames. ], batch size: 102, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 06:14:56,996 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=262038.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:15:13,031 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4115, 4.6976, 4.5142, 4.5186, 4.2107, 4.2252, 4.1652, 4.7447], device='cuda:4'), covar=tensor([0.1112, 0.0885, 0.0918, 0.0860, 0.0862, 0.1321, 0.1140, 0.0829], device='cuda:4'), in_proj_covar=tensor([0.0690, 0.0829, 0.0683, 0.0638, 0.0529, 0.0530, 0.0697, 0.0649], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 06:15:22,696 INFO [train.py:904] (4/8) Epoch 26, batch 8300, loss[loss=0.1654, simple_loss=0.2514, pruned_loss=0.03971, over 12007.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.2784, pruned_loss=0.0495, over 3066018.06 frames. ], batch size: 248, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 06:15:32,830 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.532e+02 2.264e+02 2.806e+02 3.280e+02 5.175e+02, threshold=5.611e+02, percent-clipped=0.0 2023-05-02 06:16:44,224 INFO [train.py:904] (4/8) Epoch 26, batch 8350, loss[loss=0.1885, simple_loss=0.2903, pruned_loss=0.04337, over 16272.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2784, pruned_loss=0.04816, over 3057897.20 frames. ], batch size: 165, lr: 2.57e-03, grad_scale: 4.0 2023-05-02 06:16:59,060 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=262112.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:17:20,948 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7145, 2.6531, 1.9257, 2.8213, 2.1866, 2.8223, 2.1945, 2.4614], device='cuda:4'), covar=tensor([0.0297, 0.0361, 0.1212, 0.0361, 0.0647, 0.0515, 0.1223, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0178, 0.0195, 0.0168, 0.0177, 0.0216, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 06:18:04,036 INFO [train.py:904] (4/8) Epoch 26, batch 8400, loss[loss=0.1752, simple_loss=0.2722, pruned_loss=0.03914, over 15202.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2754, pruned_loss=0.04626, over 3033917.53 frames. ], batch size: 190, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:18:13,167 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.602e+02 2.393e+02 2.671e+02 3.015e+02 4.313e+02, threshold=5.342e+02, percent-clipped=0.0 2023-05-02 06:18:14,934 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=262160.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:18:47,968 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=262180.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:19:02,232 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=262189.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:19:24,257 INFO [train.py:904] (4/8) Epoch 26, batch 8450, loss[loss=0.1692, simple_loss=0.2693, pruned_loss=0.03453, over 16900.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.274, pruned_loss=0.04432, over 3059799.25 frames. ], batch size: 90, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:19:31,570 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5331, 4.6935, 4.8168, 4.6193, 4.6990, 5.1926, 4.6976, 4.4066], device='cuda:4'), covar=tensor([0.1345, 0.1981, 0.2134, 0.2141, 0.2442, 0.0992, 0.1642, 0.2571], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0612, 0.0677, 0.0501, 0.0664, 0.0704, 0.0531, 0.0672], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 06:19:50,517 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7371, 3.4034, 3.6864, 1.9870, 3.8534, 3.8938, 3.1406, 3.0606], device='cuda:4'), covar=tensor([0.0626, 0.0248, 0.0264, 0.1147, 0.0081, 0.0198, 0.0381, 0.0401], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0109, 0.0099, 0.0137, 0.0084, 0.0128, 0.0128, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 06:20:26,767 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9499, 1.8435, 1.6990, 1.4656, 1.9830, 1.6237, 1.5184, 1.9411], device='cuda:4'), covar=tensor([0.0225, 0.0320, 0.0467, 0.0421, 0.0242, 0.0327, 0.0199, 0.0253], device='cuda:4'), in_proj_covar=tensor([0.0218, 0.0235, 0.0227, 0.0227, 0.0237, 0.0235, 0.0235, 0.0233], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 06:20:44,662 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-05-02 06:20:48,803 INFO [train.py:904] (4/8) Epoch 26, batch 8500, loss[loss=0.1771, simple_loss=0.2705, pruned_loss=0.04182, over 16754.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2706, pruned_loss=0.04244, over 3065158.28 frames. ], batch size: 124, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:20:57,922 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 2.087e+02 2.550e+02 3.186e+02 5.088e+02, threshold=5.100e+02, percent-clipped=0.0 2023-05-02 06:21:44,930 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9725, 2.1673, 2.3243, 2.8486, 1.8365, 3.1895, 1.8009, 2.7696], device='cuda:4'), covar=tensor([0.1455, 0.0791, 0.1149, 0.0217, 0.0121, 0.0427, 0.1761, 0.0708], device='cuda:4'), in_proj_covar=tensor([0.0170, 0.0177, 0.0197, 0.0195, 0.0205, 0.0215, 0.0206, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 06:22:06,120 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-05-02 06:22:09,819 INFO [train.py:904] (4/8) Epoch 26, batch 8550, loss[loss=0.1901, simple_loss=0.2933, pruned_loss=0.0434, over 16397.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2685, pruned_loss=0.04169, over 3038377.11 frames. ], batch size: 146, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:22:40,683 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-05-02 06:22:55,587 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0611, 4.0244, 3.9284, 3.0667, 3.9855, 1.8321, 3.7682, 3.4784], device='cuda:4'), covar=tensor([0.0113, 0.0111, 0.0191, 0.0264, 0.0098, 0.2971, 0.0137, 0.0314], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0166, 0.0204, 0.0179, 0.0180, 0.0210, 0.0192, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 06:23:07,966 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=262333.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:23:47,966 INFO [train.py:904] (4/8) Epoch 26, batch 8600, loss[loss=0.1902, simple_loss=0.2703, pruned_loss=0.05507, over 12397.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2688, pruned_loss=0.04089, over 3024841.83 frames. ], batch size: 248, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:23:59,714 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.426e+02 2.293e+02 2.800e+02 3.467e+02 6.859e+02, threshold=5.600e+02, percent-clipped=2.0 2023-05-02 06:24:32,725 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-05-02 06:24:57,564 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.12 vs. limit=5.0 2023-05-02 06:25:12,454 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9635, 4.9483, 4.7123, 4.1439, 4.8055, 1.9969, 4.5485, 4.5562], device='cuda:4'), covar=tensor([0.0093, 0.0090, 0.0197, 0.0347, 0.0097, 0.2536, 0.0129, 0.0237], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0166, 0.0203, 0.0178, 0.0179, 0.0210, 0.0192, 0.0172], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 06:25:27,327 INFO [train.py:904] (4/8) Epoch 26, batch 8650, loss[loss=0.1565, simple_loss=0.255, pruned_loss=0.02905, over 16703.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2665, pruned_loss=0.03936, over 3014254.36 frames. ], batch size: 134, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:27:13,544 INFO [train.py:904] (4/8) Epoch 26, batch 8700, loss[loss=0.1761, simple_loss=0.2564, pruned_loss=0.04788, over 12255.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2639, pruned_loss=0.03836, over 3018876.37 frames. ], batch size: 250, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:27:25,317 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.287e+02 2.080e+02 2.583e+02 3.221e+02 6.009e+02, threshold=5.165e+02, percent-clipped=1.0 2023-05-02 06:28:02,740 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=262480.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:28:20,420 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=262489.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:28:48,924 INFO [train.py:904] (4/8) Epoch 26, batch 8750, loss[loss=0.1806, simple_loss=0.2852, pruned_loss=0.03797, over 16658.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2632, pruned_loss=0.03765, over 3021064.00 frames. ], batch size: 89, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:29:48,639 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=262528.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:30:09,753 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=262537.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:30:35,267 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2922, 3.3545, 1.9676, 3.7190, 2.5056, 3.6938, 2.1960, 2.7523], device='cuda:4'), covar=tensor([0.0317, 0.0403, 0.1741, 0.0265, 0.0946, 0.0540, 0.1626, 0.0874], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0175, 0.0192, 0.0165, 0.0175, 0.0212, 0.0201, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 06:30:41,532 INFO [train.py:904] (4/8) Epoch 26, batch 8800, loss[loss=0.1489, simple_loss=0.2417, pruned_loss=0.02802, over 17123.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2613, pruned_loss=0.03679, over 3018790.63 frames. ], batch size: 49, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:30:52,411 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.093e+02 2.466e+02 2.812e+02 4.660e+02, threshold=4.932e+02, percent-clipped=0.0 2023-05-02 06:32:25,983 INFO [train.py:904] (4/8) Epoch 26, batch 8850, loss[loss=0.1752, simple_loss=0.28, pruned_loss=0.03516, over 16566.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2645, pruned_loss=0.03621, over 3020289.41 frames. ], batch size: 62, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:33:32,244 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=262633.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:34:13,867 INFO [train.py:904] (4/8) Epoch 26, batch 8900, loss[loss=0.1781, simple_loss=0.2746, pruned_loss=0.04086, over 16307.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2652, pruned_loss=0.03578, over 3030442.82 frames. ], batch size: 165, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:34:26,825 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.369e+02 2.190e+02 2.649e+02 3.221e+02 8.286e+02, threshold=5.298e+02, percent-clipped=4.0 2023-05-02 06:35:20,124 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=262681.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:35:27,223 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-02 06:36:18,595 INFO [train.py:904] (4/8) Epoch 26, batch 8950, loss[loss=0.1543, simple_loss=0.2463, pruned_loss=0.03119, over 16988.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2645, pruned_loss=0.03579, over 3061430.38 frames. ], batch size: 109, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:38:00,859 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=262749.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:38:08,095 INFO [train.py:904] (4/8) Epoch 26, batch 9000, loss[loss=0.1437, simple_loss=0.2388, pruned_loss=0.02427, over 16511.00 frames. ], tot_loss[loss=0.165, simple_loss=0.261, pruned_loss=0.03452, over 3064695.31 frames. ], batch size: 68, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:38:08,096 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 06:38:18,516 INFO [train.py:938] (4/8) Epoch 26, validation: loss=0.1435, simple_loss=0.2475, pruned_loss=0.01976, over 944034.00 frames. 2023-05-02 06:38:18,517 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 06:38:30,742 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 2.045e+02 2.519e+02 3.077e+02 5.140e+02, threshold=5.037e+02, percent-clipped=0.0 2023-05-02 06:38:47,589 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-05-02 06:40:03,316 INFO [train.py:904] (4/8) Epoch 26, batch 9050, loss[loss=0.1841, simple_loss=0.2689, pruned_loss=0.04965, over 12370.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2622, pruned_loss=0.03511, over 3074475.31 frames. ], batch size: 248, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:40:20,429 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=262810.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:41:48,706 INFO [train.py:904] (4/8) Epoch 26, batch 9100, loss[loss=0.1669, simple_loss=0.2677, pruned_loss=0.03303, over 15510.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2621, pruned_loss=0.03577, over 3066809.89 frames. ], batch size: 194, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:42:01,223 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.532e+02 2.262e+02 2.654e+02 3.227e+02 6.240e+02, threshold=5.309e+02, percent-clipped=4.0 2023-05-02 06:43:45,862 INFO [train.py:904] (4/8) Epoch 26, batch 9150, loss[loss=0.1483, simple_loss=0.2437, pruned_loss=0.02647, over 16780.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2627, pruned_loss=0.0355, over 3067907.83 frames. ], batch size: 83, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:44:00,909 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4011, 3.1080, 2.6205, 2.3068, 2.1456, 2.1317, 3.0187, 2.7469], device='cuda:4'), covar=tensor([0.2685, 0.0776, 0.1850, 0.2792, 0.2853, 0.2583, 0.0528, 0.1485], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0269, 0.0304, 0.0317, 0.0296, 0.0268, 0.0296, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 06:45:27,072 INFO [train.py:904] (4/8) Epoch 26, batch 9200, loss[loss=0.1657, simple_loss=0.2619, pruned_loss=0.03479, over 16903.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2588, pruned_loss=0.03467, over 3072771.63 frames. ], batch size: 102, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:45:36,599 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.418e+02 2.229e+02 2.531e+02 3.050e+02 5.061e+02, threshold=5.062e+02, percent-clipped=0.0 2023-05-02 06:47:01,941 INFO [train.py:904] (4/8) Epoch 26, batch 9250, loss[loss=0.1597, simple_loss=0.2575, pruned_loss=0.03098, over 15543.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2581, pruned_loss=0.03493, over 3040914.61 frames. ], batch size: 194, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:47:11,369 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6108, 3.7870, 2.8079, 2.1891, 2.2820, 2.4029, 3.9917, 3.1915], device='cuda:4'), covar=tensor([0.2965, 0.0567, 0.1933, 0.3355, 0.3369, 0.2335, 0.0387, 0.1557], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0268, 0.0304, 0.0316, 0.0294, 0.0267, 0.0296, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 06:47:24,505 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2382, 2.9769, 3.1064, 1.8528, 3.2726, 3.3610, 2.7692, 2.6881], device='cuda:4'), covar=tensor([0.0780, 0.0292, 0.0198, 0.1174, 0.0105, 0.0182, 0.0457, 0.0457], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0107, 0.0095, 0.0135, 0.0082, 0.0124, 0.0125, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 06:48:49,724 INFO [train.py:904] (4/8) Epoch 26, batch 9300, loss[loss=0.1469, simple_loss=0.2407, pruned_loss=0.02661, over 15454.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2564, pruned_loss=0.03427, over 3032866.90 frames. ], batch size: 192, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:48:53,437 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4067, 3.4874, 3.6905, 3.6703, 3.6834, 3.5075, 3.5467, 3.5866], device='cuda:4'), covar=tensor([0.0486, 0.1374, 0.0686, 0.0638, 0.0736, 0.0752, 0.0923, 0.0581], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0464, 0.0451, 0.0414, 0.0499, 0.0475, 0.0545, 0.0379], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 06:48:55,439 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7639, 2.6163, 2.3727, 3.8042, 2.1977, 3.7507, 1.4478, 2.8760], device='cuda:4'), covar=tensor([0.1405, 0.0777, 0.1270, 0.0186, 0.0123, 0.0412, 0.1810, 0.0733], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0175, 0.0195, 0.0191, 0.0200, 0.0212, 0.0204, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 06:49:02,019 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.290e+02 2.028e+02 2.420e+02 3.075e+02 6.199e+02, threshold=4.840e+02, percent-clipped=1.0 2023-05-02 06:50:20,264 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=263095.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:50:33,213 INFO [train.py:904] (4/8) Epoch 26, batch 9350, loss[loss=0.1638, simple_loss=0.2487, pruned_loss=0.0394, over 12268.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2568, pruned_loss=0.03443, over 3037821.61 frames. ], batch size: 250, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:50:38,633 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=263105.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:50:51,063 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=263111.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:51:48,077 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=263140.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:51:53,915 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4569, 3.5733, 2.2632, 4.0838, 2.7212, 3.9544, 2.2561, 2.8931], device='cuda:4'), covar=tensor([0.0333, 0.0405, 0.1623, 0.0182, 0.0940, 0.0586, 0.1690, 0.0848], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0172, 0.0188, 0.0162, 0.0173, 0.0209, 0.0198, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 06:51:55,849 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5404, 3.4962, 3.5183, 2.6489, 3.2824, 2.0487, 2.9713, 2.8940], device='cuda:4'), covar=tensor([0.0147, 0.0137, 0.0179, 0.0182, 0.0100, 0.2569, 0.0143, 0.0275], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0165, 0.0201, 0.0176, 0.0179, 0.0209, 0.0190, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 06:52:14,167 INFO [train.py:904] (4/8) Epoch 26, batch 9400, loss[loss=0.1844, simple_loss=0.2842, pruned_loss=0.04227, over 16316.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2568, pruned_loss=0.03418, over 3045779.59 frames. ], batch size: 146, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:52:19,893 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=263156.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:52:25,026 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.316e+02 2.115e+02 2.425e+02 3.007e+02 4.996e+02, threshold=4.851e+02, percent-clipped=1.0 2023-05-02 06:52:52,579 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=263172.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:53:21,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4454, 2.4387, 2.3023, 4.2316, 2.3812, 2.8213, 2.4834, 2.6002], device='cuda:4'), covar=tensor([0.1249, 0.3577, 0.3209, 0.0482, 0.4061, 0.2468, 0.3525, 0.3518], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0457, 0.0376, 0.0324, 0.0436, 0.0521, 0.0429, 0.0534], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 06:53:46,406 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6059, 1.8218, 2.1737, 2.5299, 2.4504, 2.8795, 1.9607, 2.8833], device='cuda:4'), covar=tensor([0.0279, 0.0605, 0.0428, 0.0391, 0.0398, 0.0243, 0.0622, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0192, 0.0179, 0.0183, 0.0199, 0.0157, 0.0196, 0.0157], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 06:53:50,922 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=263201.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:53:53,489 INFO [train.py:904] (4/8) Epoch 26, batch 9450, loss[loss=0.1777, simple_loss=0.2764, pruned_loss=0.03952, over 16909.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2587, pruned_loss=0.03442, over 3051640.60 frames. ], batch size: 116, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:54:14,316 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=263214.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:55:19,881 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=263246.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:55:33,109 INFO [train.py:904] (4/8) Epoch 26, batch 9500, loss[loss=0.136, simple_loss=0.2235, pruned_loss=0.02426, over 12868.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2577, pruned_loss=0.034, over 3058938.34 frames. ], batch size: 248, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:55:47,501 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.223e+02 2.048e+02 2.318e+02 3.052e+02 5.561e+02, threshold=4.636e+02, percent-clipped=2.0 2023-05-02 06:56:07,094 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-05-02 06:56:17,500 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4274, 3.5104, 2.1499, 3.8861, 2.6433, 3.8435, 2.3520, 2.8619], device='cuda:4'), covar=tensor([0.0315, 0.0380, 0.1617, 0.0242, 0.0853, 0.0535, 0.1557, 0.0764], device='cuda:4'), in_proj_covar=tensor([0.0168, 0.0172, 0.0189, 0.0163, 0.0173, 0.0209, 0.0198, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 06:56:19,456 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=263275.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:57:17,615 INFO [train.py:904] (4/8) Epoch 26, batch 9550, loss[loss=0.2022, simple_loss=0.307, pruned_loss=0.04866, over 15348.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2581, pruned_loss=0.03407, over 3067853.32 frames. ], batch size: 191, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:57:27,263 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=263307.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 06:58:58,571 INFO [train.py:904] (4/8) Epoch 26, batch 9600, loss[loss=0.1977, simple_loss=0.2927, pruned_loss=0.05129, over 15256.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2594, pruned_loss=0.03476, over 3065702.25 frames. ], batch size: 191, lr: 2.57e-03, grad_scale: 8.0 2023-05-02 06:59:09,908 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.689e+02 2.249e+02 2.699e+02 3.019e+02 5.664e+02, threshold=5.399e+02, percent-clipped=4.0 2023-05-02 06:59:27,608 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 07:00:45,647 INFO [train.py:904] (4/8) Epoch 26, batch 9650, loss[loss=0.1657, simple_loss=0.2635, pruned_loss=0.03395, over 16104.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2618, pruned_loss=0.0354, over 3076627.50 frames. ], batch size: 165, lr: 2.56e-03, grad_scale: 8.0 2023-05-02 07:00:52,327 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=263405.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:02:29,091 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=263451.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:02:31,833 INFO [train.py:904] (4/8) Epoch 26, batch 9700, loss[loss=0.1621, simple_loss=0.2499, pruned_loss=0.03715, over 12168.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2604, pruned_loss=0.03499, over 3058809.18 frames. ], batch size: 248, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:02:33,072 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=263453.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:02:42,005 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.181e+02 2.600e+02 3.178e+02 8.821e+02, threshold=5.200e+02, percent-clipped=1.0 2023-05-02 07:02:58,809 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=263467.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:04:01,226 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=263496.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:04:13,669 INFO [train.py:904] (4/8) Epoch 26, batch 9750, loss[loss=0.1421, simple_loss=0.2398, pruned_loss=0.02221, over 16770.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.259, pruned_loss=0.03514, over 3046809.55 frames. ], batch size: 76, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:05:51,432 INFO [train.py:904] (4/8) Epoch 26, batch 9800, loss[loss=0.1653, simple_loss=0.2736, pruned_loss=0.02845, over 16228.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2588, pruned_loss=0.03411, over 3061680.97 frames. ], batch size: 165, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:06:03,262 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.369e+02 2.162e+02 2.515e+02 2.948e+02 4.356e+02, threshold=5.031e+02, percent-clipped=0.0 2023-05-02 07:06:23,918 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=263570.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:07:23,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5836, 3.3113, 3.5167, 1.6669, 3.6807, 3.8675, 3.0097, 2.8283], device='cuda:4'), covar=tensor([0.0742, 0.0298, 0.0265, 0.1474, 0.0117, 0.0155, 0.0456, 0.0530], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0106, 0.0094, 0.0134, 0.0082, 0.0123, 0.0125, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 07:07:23,338 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.46 vs. limit=2.0 2023-05-02 07:07:33,256 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=263602.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:07:34,239 INFO [train.py:904] (4/8) Epoch 26, batch 9850, loss[loss=0.1564, simple_loss=0.2561, pruned_loss=0.02834, over 16183.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.26, pruned_loss=0.03385, over 3065597.50 frames. ], batch size: 165, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:09:24,088 INFO [train.py:904] (4/8) Epoch 26, batch 9900, loss[loss=0.1694, simple_loss=0.2676, pruned_loss=0.03564, over 16715.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2608, pruned_loss=0.03357, over 3077079.80 frames. ], batch size: 134, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:09:36,804 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.367e+02 1.934e+02 2.301e+02 2.936e+02 6.933e+02, threshold=4.602e+02, percent-clipped=3.0 2023-05-02 07:09:41,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8643, 2.1538, 2.3576, 3.1941, 2.1677, 2.3189, 2.3095, 2.2329], device='cuda:4'), covar=tensor([0.1362, 0.3752, 0.2918, 0.0747, 0.4461, 0.2762, 0.3589, 0.3897], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0457, 0.0375, 0.0324, 0.0435, 0.0520, 0.0429, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 07:10:26,099 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4230, 3.0432, 2.6885, 2.3062, 2.2004, 2.3092, 3.0801, 2.8185], device='cuda:4'), covar=tensor([0.2678, 0.0713, 0.1697, 0.2695, 0.2841, 0.2274, 0.0447, 0.1542], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0268, 0.0304, 0.0317, 0.0293, 0.0267, 0.0296, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:11:20,747 INFO [train.py:904] (4/8) Epoch 26, batch 9950, loss[loss=0.1724, simple_loss=0.2706, pruned_loss=0.03714, over 16173.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2623, pruned_loss=0.03388, over 3059427.14 frames. ], batch size: 165, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:13:18,984 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=263751.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:13:21,139 INFO [train.py:904] (4/8) Epoch 26, batch 10000, loss[loss=0.1437, simple_loss=0.2482, pruned_loss=0.01958, over 16855.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.261, pruned_loss=0.03368, over 3067635.29 frames. ], batch size: 96, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:13:34,912 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.519e+02 2.135e+02 2.379e+02 2.761e+02 5.378e+02, threshold=4.758e+02, percent-clipped=3.0 2023-05-02 07:13:51,310 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=263767.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:14:48,922 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=263796.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:14:53,762 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=263799.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:15:00,032 INFO [train.py:904] (4/8) Epoch 26, batch 10050, loss[loss=0.183, simple_loss=0.2832, pruned_loss=0.04138, over 15396.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2619, pruned_loss=0.03411, over 3066314.08 frames. ], batch size: 191, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:15:23,653 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=263815.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:15:57,739 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=263833.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 07:16:17,279 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=263844.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:16:17,742 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 07:16:30,282 INFO [train.py:904] (4/8) Epoch 26, batch 10100, loss[loss=0.1605, simple_loss=0.2532, pruned_loss=0.03388, over 16225.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2618, pruned_loss=0.03419, over 3058278.37 frames. ], batch size: 165, lr: 2.56e-03, grad_scale: 16.0 2023-05-02 07:16:39,584 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.545e+02 2.128e+02 2.539e+02 3.188e+02 8.403e+02, threshold=5.079e+02, percent-clipped=3.0 2023-05-02 07:17:02,800 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=263870.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:17:36,274 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=263894.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 07:18:08,980 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=263902.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:18:09,862 INFO [train.py:904] (4/8) Epoch 27, batch 0, loss[loss=0.2284, simple_loss=0.297, pruned_loss=0.0799, over 16889.00 frames. ], tot_loss[loss=0.2284, simple_loss=0.297, pruned_loss=0.0799, over 16889.00 frames. ], batch size: 116, lr: 2.51e-03, grad_scale: 16.0 2023-05-02 07:18:09,862 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 07:18:17,113 INFO [train.py:938] (4/8) Epoch 27, validation: loss=0.1434, simple_loss=0.2467, pruned_loss=0.02006, over 944034.00 frames. 2023-05-02 07:18:17,114 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 07:18:38,687 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=263918.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:19:22,226 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=263950.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:19:26,179 INFO [train.py:904] (4/8) Epoch 27, batch 50, loss[loss=0.1974, simple_loss=0.2845, pruned_loss=0.05517, over 15628.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2634, pruned_loss=0.04445, over 749768.82 frames. ], batch size: 191, lr: 2.51e-03, grad_scale: 1.0 2023-05-02 07:19:36,272 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=263960.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:19:39,888 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.543e+02 2.464e+02 3.043e+02 3.721e+02 7.050e+02, threshold=6.086e+02, percent-clipped=6.0 2023-05-02 07:19:50,835 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9848, 4.6720, 3.2381, 2.4885, 2.7968, 2.7256, 4.9762, 3.6305], device='cuda:4'), covar=tensor([0.2886, 0.0530, 0.1845, 0.3134, 0.3001, 0.2203, 0.0335, 0.1533], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0268, 0.0306, 0.0317, 0.0293, 0.0268, 0.0297, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:20:36,408 INFO [train.py:904] (4/8) Epoch 27, batch 100, loss[loss=0.1912, simple_loss=0.2821, pruned_loss=0.05019, over 16550.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2614, pruned_loss=0.04268, over 1323133.78 frames. ], batch size: 68, lr: 2.51e-03, grad_scale: 1.0 2023-05-02 07:21:01,025 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=264021.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:21:10,183 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-05-02 07:21:36,110 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0623, 5.7137, 5.7931, 5.4739, 5.6551, 6.1721, 5.6046, 5.2538], device='cuda:4'), covar=tensor([0.0965, 0.1804, 0.2549, 0.2127, 0.2408, 0.0972, 0.1511, 0.2377], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0602, 0.0669, 0.0493, 0.0652, 0.0693, 0.0522, 0.0657], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:21:44,756 INFO [train.py:904] (4/8) Epoch 27, batch 150, loss[loss=0.1839, simple_loss=0.2757, pruned_loss=0.046, over 17039.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2598, pruned_loss=0.04142, over 1768621.87 frames. ], batch size: 53, lr: 2.51e-03, grad_scale: 1.0 2023-05-02 07:21:57,550 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.626e+02 2.228e+02 2.541e+02 3.123e+02 6.802e+02, threshold=5.083e+02, percent-clipped=1.0 2023-05-02 07:22:06,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7673, 4.5075, 4.8170, 4.9708, 5.1582, 4.5889, 5.1285, 5.1600], device='cuda:4'), covar=tensor([0.1983, 0.1557, 0.1834, 0.0898, 0.0674, 0.1038, 0.0727, 0.0682], device='cuda:4'), in_proj_covar=tensor([0.0645, 0.0797, 0.0913, 0.0804, 0.0613, 0.0637, 0.0670, 0.0777], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 07:22:25,152 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0983, 5.6582, 5.8214, 5.4782, 5.6411, 6.1993, 5.7321, 5.4393], device='cuda:4'), covar=tensor([0.0972, 0.2049, 0.2617, 0.2232, 0.2472, 0.0915, 0.1469, 0.2194], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0604, 0.0671, 0.0495, 0.0654, 0.0696, 0.0523, 0.0660], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:22:53,632 INFO [train.py:904] (4/8) Epoch 27, batch 200, loss[loss=0.1869, simple_loss=0.2708, pruned_loss=0.05152, over 16484.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2599, pruned_loss=0.04168, over 2114474.62 frames. ], batch size: 146, lr: 2.51e-03, grad_scale: 1.0 2023-05-02 07:23:48,485 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8009, 4.2394, 3.0287, 2.3819, 2.6338, 2.5889, 4.5909, 3.5243], device='cuda:4'), covar=tensor([0.3031, 0.0609, 0.1904, 0.3073, 0.2956, 0.2214, 0.0381, 0.1466], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0271, 0.0309, 0.0320, 0.0297, 0.0271, 0.0300, 0.0345], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:24:00,835 INFO [train.py:904] (4/8) Epoch 27, batch 250, loss[loss=0.167, simple_loss=0.2428, pruned_loss=0.0456, over 16896.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2585, pruned_loss=0.04268, over 2386227.13 frames. ], batch size: 109, lr: 2.51e-03, grad_scale: 1.0 2023-05-02 07:24:14,023 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.539e+02 2.137e+02 2.566e+02 3.283e+02 9.180e+02, threshold=5.132e+02, percent-clipped=3.0 2023-05-02 07:24:26,784 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-02 07:24:50,573 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=264189.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 07:25:10,467 INFO [train.py:904] (4/8) Epoch 27, batch 300, loss[loss=0.1792, simple_loss=0.2697, pruned_loss=0.04432, over 17228.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2569, pruned_loss=0.04113, over 2594589.10 frames. ], batch size: 43, lr: 2.51e-03, grad_scale: 1.0 2023-05-02 07:25:53,564 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=264234.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:26:19,419 INFO [train.py:904] (4/8) Epoch 27, batch 350, loss[loss=0.1415, simple_loss=0.2349, pruned_loss=0.02402, over 17142.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2541, pruned_loss=0.04062, over 2750946.92 frames. ], batch size: 48, lr: 2.51e-03, grad_scale: 1.0 2023-05-02 07:26:25,560 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.41 vs. limit=5.0 2023-05-02 07:26:34,329 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.322e+02 2.070e+02 2.475e+02 2.951e+02 5.112e+02, threshold=4.951e+02, percent-clipped=0.0 2023-05-02 07:27:17,847 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=264295.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:27:27,689 INFO [train.py:904] (4/8) Epoch 27, batch 400, loss[loss=0.1516, simple_loss=0.2514, pruned_loss=0.02594, over 16990.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2523, pruned_loss=0.04055, over 2876287.90 frames. ], batch size: 50, lr: 2.51e-03, grad_scale: 2.0 2023-05-02 07:27:38,128 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4155, 3.4392, 3.9793, 2.1799, 3.2223, 2.4338, 3.8300, 3.6446], device='cuda:4'), covar=tensor([0.0266, 0.1089, 0.0510, 0.2144, 0.0823, 0.1041, 0.0607, 0.1211], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0165, 0.0167, 0.0155, 0.0146, 0.0130, 0.0144, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 07:27:45,229 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=264316.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:27:59,270 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6974, 2.6926, 2.3848, 2.4615, 2.9714, 2.7671, 3.1848, 3.1894], device='cuda:4'), covar=tensor([0.0202, 0.0497, 0.0630, 0.0585, 0.0361, 0.0462, 0.0311, 0.0319], device='cuda:4'), in_proj_covar=tensor([0.0225, 0.0243, 0.0233, 0.0233, 0.0243, 0.0242, 0.0240, 0.0239], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 07:28:33,908 INFO [train.py:904] (4/8) Epoch 27, batch 450, loss[loss=0.1473, simple_loss=0.236, pruned_loss=0.02934, over 17008.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.251, pruned_loss=0.03979, over 2962879.96 frames. ], batch size: 41, lr: 2.51e-03, grad_scale: 2.0 2023-05-02 07:28:48,385 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.403e+02 2.223e+02 2.597e+02 3.088e+02 6.420e+02, threshold=5.195e+02, percent-clipped=1.0 2023-05-02 07:29:03,218 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-05-02 07:29:44,911 INFO [train.py:904] (4/8) Epoch 27, batch 500, loss[loss=0.1712, simple_loss=0.2709, pruned_loss=0.03578, over 17091.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2505, pruned_loss=0.03897, over 3039924.47 frames. ], batch size: 47, lr: 2.51e-03, grad_scale: 2.0 2023-05-02 07:30:40,849 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2645, 5.2617, 5.1430, 4.5749, 4.7497, 5.1538, 5.1383, 4.7957], device='cuda:4'), covar=tensor([0.0616, 0.0577, 0.0311, 0.0394, 0.1105, 0.0498, 0.0341, 0.0843], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0456, 0.0356, 0.0358, 0.0355, 0.0413, 0.0245, 0.0428], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 07:30:51,107 INFO [train.py:904] (4/8) Epoch 27, batch 550, loss[loss=0.1808, simple_loss=0.2783, pruned_loss=0.04164, over 16756.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2494, pruned_loss=0.03823, over 3100866.28 frames. ], batch size: 57, lr: 2.51e-03, grad_scale: 2.0 2023-05-02 07:31:04,228 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 2.059e+02 2.347e+02 2.787e+02 4.742e+02, threshold=4.693e+02, percent-clipped=0.0 2023-05-02 07:31:26,700 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=264479.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:31:41,321 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=264489.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 07:31:55,537 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=264500.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 07:31:59,347 INFO [train.py:904] (4/8) Epoch 27, batch 600, loss[loss=0.1612, simple_loss=0.2563, pruned_loss=0.03302, over 17034.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2492, pruned_loss=0.03828, over 3147169.47 frames. ], batch size: 53, lr: 2.51e-03, grad_scale: 2.0 2023-05-02 07:32:29,864 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-02 07:32:46,461 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=264537.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 07:32:50,470 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=264540.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:33:08,801 INFO [train.py:904] (4/8) Epoch 27, batch 650, loss[loss=0.1757, simple_loss=0.2511, pruned_loss=0.05019, over 16767.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.2472, pruned_loss=0.0373, over 3168806.47 frames. ], batch size: 124, lr: 2.51e-03, grad_scale: 2.0 2023-05-02 07:33:18,840 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=264561.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 07:33:20,702 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.585e+02 2.185e+02 2.561e+02 3.175e+02 5.971e+02, threshold=5.123e+02, percent-clipped=4.0 2023-05-02 07:33:58,956 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=264590.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:34:15,989 INFO [train.py:904] (4/8) Epoch 27, batch 700, loss[loss=0.1629, simple_loss=0.2574, pruned_loss=0.03415, over 17145.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2474, pruned_loss=0.03713, over 3207070.59 frames. ], batch size: 47, lr: 2.51e-03, grad_scale: 2.0 2023-05-02 07:34:31,885 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.08 vs. limit=5.0 2023-05-02 07:34:34,269 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=264616.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:34:40,653 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8886, 4.9338, 5.3733, 5.3551, 5.3372, 5.0044, 4.9769, 4.8403], device='cuda:4'), covar=tensor([0.0404, 0.0691, 0.0460, 0.0466, 0.0578, 0.0497, 0.0983, 0.0572], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0484, 0.0468, 0.0430, 0.0518, 0.0496, 0.0568, 0.0394], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 07:34:41,770 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1417, 5.1541, 5.6012, 5.5511, 5.5910, 5.2330, 5.1568, 5.0390], device='cuda:4'), covar=tensor([0.0373, 0.0701, 0.0406, 0.0482, 0.0495, 0.0414, 0.0967, 0.0504], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0484, 0.0468, 0.0430, 0.0518, 0.0496, 0.0568, 0.0394], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 07:34:42,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9749, 5.4870, 5.5501, 5.2706, 5.3597, 5.9700, 5.3850, 5.0498], device='cuda:4'), covar=tensor([0.1041, 0.1988, 0.2949, 0.2086, 0.2630, 0.0944, 0.1700, 0.2442], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0619, 0.0688, 0.0508, 0.0672, 0.0712, 0.0535, 0.0676], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:35:22,117 INFO [train.py:904] (4/8) Epoch 27, batch 750, loss[loss=0.1378, simple_loss=0.2215, pruned_loss=0.02701, over 16780.00 frames. ], tot_loss[loss=0.162, simple_loss=0.2477, pruned_loss=0.03817, over 3228849.30 frames. ], batch size: 39, lr: 2.51e-03, grad_scale: 2.0 2023-05-02 07:35:35,421 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.615e+02 2.184e+02 2.458e+02 2.836e+02 5.785e+02, threshold=4.916e+02, percent-clipped=2.0 2023-05-02 07:35:37,451 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=264664.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:35:53,549 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.94 vs. limit=5.0 2023-05-02 07:36:29,180 INFO [train.py:904] (4/8) Epoch 27, batch 800, loss[loss=0.1506, simple_loss=0.2452, pruned_loss=0.02798, over 17052.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2483, pruned_loss=0.03813, over 3245795.25 frames. ], batch size: 50, lr: 2.51e-03, grad_scale: 4.0 2023-05-02 07:37:36,931 INFO [train.py:904] (4/8) Epoch 27, batch 850, loss[loss=0.1717, simple_loss=0.2475, pruned_loss=0.0479, over 16869.00 frames. ], tot_loss[loss=0.1619, simple_loss=0.2483, pruned_loss=0.03777, over 3269601.57 frames. ], batch size: 116, lr: 2.51e-03, grad_scale: 4.0 2023-05-02 07:37:51,804 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.434e+02 2.032e+02 2.277e+02 2.694e+02 3.998e+02, threshold=4.553e+02, percent-clipped=0.0 2023-05-02 07:37:53,956 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2883, 4.2341, 4.1644, 3.4691, 4.2251, 1.7532, 3.9905, 3.6222], device='cuda:4'), covar=tensor([0.0240, 0.0170, 0.0247, 0.0352, 0.0156, 0.3315, 0.0201, 0.0387], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0169, 0.0206, 0.0179, 0.0183, 0.0214, 0.0196, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 07:38:44,877 INFO [train.py:904] (4/8) Epoch 27, batch 900, loss[loss=0.1515, simple_loss=0.2335, pruned_loss=0.03477, over 16440.00 frames. ], tot_loss[loss=0.1611, simple_loss=0.2473, pruned_loss=0.03748, over 3284817.14 frames. ], batch size: 75, lr: 2.51e-03, grad_scale: 4.0 2023-05-02 07:38:47,145 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=264804.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:39:02,512 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6568, 3.7927, 2.5170, 4.5038, 3.0403, 4.4109, 2.5936, 3.1897], device='cuda:4'), covar=tensor([0.0391, 0.0463, 0.1632, 0.0333, 0.0920, 0.0633, 0.1509, 0.0826], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0182, 0.0198, 0.0174, 0.0182, 0.0222, 0.0208, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 07:39:28,075 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=264835.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:39:51,349 INFO [train.py:904] (4/8) Epoch 27, batch 950, loss[loss=0.1799, simple_loss=0.2467, pruned_loss=0.05661, over 16909.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2477, pruned_loss=0.03768, over 3298896.56 frames. ], batch size: 109, lr: 2.51e-03, grad_scale: 4.0 2023-05-02 07:39:55,587 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=264856.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 07:40:04,148 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.346e+02 2.026e+02 2.339e+02 3.072e+02 8.203e+02, threshold=4.678e+02, percent-clipped=3.0 2023-05-02 07:40:07,741 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=264865.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:40:31,516 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9653, 5.0457, 5.4297, 5.3949, 5.4069, 5.1041, 5.0323, 4.9107], device='cuda:4'), covar=tensor([0.0376, 0.0577, 0.0432, 0.0478, 0.0505, 0.0490, 0.0959, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0433, 0.0489, 0.0472, 0.0433, 0.0524, 0.0501, 0.0575, 0.0399], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 07:40:41,671 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=264890.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:40:57,191 INFO [train.py:904] (4/8) Epoch 27, batch 1000, loss[loss=0.1579, simple_loss=0.2316, pruned_loss=0.04213, over 16892.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2466, pruned_loss=0.03753, over 3307299.29 frames. ], batch size: 90, lr: 2.51e-03, grad_scale: 4.0 2023-05-02 07:41:46,841 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=264938.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:42:07,407 INFO [train.py:904] (4/8) Epoch 27, batch 1050, loss[loss=0.1619, simple_loss=0.241, pruned_loss=0.04142, over 16890.00 frames. ], tot_loss[loss=0.1607, simple_loss=0.2463, pruned_loss=0.03758, over 3315591.04 frames. ], batch size: 96, lr: 2.51e-03, grad_scale: 4.0 2023-05-02 07:42:19,711 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.472e+02 2.065e+02 2.389e+02 2.901e+02 6.221e+02, threshold=4.777e+02, percent-clipped=3.0 2023-05-02 07:42:22,726 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.52 vs. limit=5.0 2023-05-02 07:42:44,597 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-02 07:43:16,444 INFO [train.py:904] (4/8) Epoch 27, batch 1100, loss[loss=0.1665, simple_loss=0.2527, pruned_loss=0.04012, over 16518.00 frames. ], tot_loss[loss=0.1595, simple_loss=0.2452, pruned_loss=0.03687, over 3322295.29 frames. ], batch size: 68, lr: 2.51e-03, grad_scale: 4.0 2023-05-02 07:43:52,953 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-02 07:44:10,171 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0776, 4.0929, 3.9614, 3.6457, 3.7441, 4.0477, 3.6816, 3.8739], device='cuda:4'), covar=tensor([0.0612, 0.0763, 0.0321, 0.0333, 0.0671, 0.0510, 0.1090, 0.0609], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0472, 0.0368, 0.0371, 0.0368, 0.0429, 0.0253, 0.0442], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:44:16,537 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5899, 3.6686, 2.3397, 3.9145, 2.9909, 3.8585, 2.4737, 3.0684], device='cuda:4'), covar=tensor([0.0283, 0.0429, 0.1517, 0.0386, 0.0775, 0.0739, 0.1373, 0.0665], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0182, 0.0197, 0.0174, 0.0181, 0.0222, 0.0207, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 07:44:25,473 INFO [train.py:904] (4/8) Epoch 27, batch 1150, loss[loss=0.1442, simple_loss=0.217, pruned_loss=0.03564, over 16720.00 frames. ], tot_loss[loss=0.1591, simple_loss=0.2454, pruned_loss=0.03642, over 3324122.31 frames. ], batch size: 83, lr: 2.51e-03, grad_scale: 4.0 2023-05-02 07:44:39,263 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.259e+02 1.949e+02 2.374e+02 2.990e+02 6.257e+02, threshold=4.748e+02, percent-clipped=3.0 2023-05-02 07:45:34,353 INFO [train.py:904] (4/8) Epoch 27, batch 1200, loss[loss=0.1579, simple_loss=0.2505, pruned_loss=0.03264, over 16616.00 frames. ], tot_loss[loss=0.1581, simple_loss=0.2441, pruned_loss=0.036, over 3313823.77 frames. ], batch size: 57, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:46:00,621 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6049, 3.7131, 2.5486, 3.9770, 3.0259, 3.9103, 2.5105, 3.0299], device='cuda:4'), covar=tensor([0.0305, 0.0490, 0.1522, 0.0440, 0.0795, 0.1009, 0.1486, 0.0748], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0183, 0.0198, 0.0175, 0.0182, 0.0223, 0.0208, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 07:46:19,385 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=265135.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:46:42,960 INFO [train.py:904] (4/8) Epoch 27, batch 1250, loss[loss=0.1401, simple_loss=0.2201, pruned_loss=0.03006, over 15851.00 frames. ], tot_loss[loss=0.1584, simple_loss=0.2439, pruned_loss=0.03645, over 3295844.71 frames. ], batch size: 35, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:46:48,320 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=265156.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 07:46:53,504 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=265160.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:46:57,310 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.528e+02 2.048e+02 2.480e+02 3.047e+02 6.939e+02, threshold=4.960e+02, percent-clipped=4.0 2023-05-02 07:47:01,712 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-02 07:47:25,262 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=265183.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:47:31,440 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1948, 5.1676, 5.0634, 4.5246, 4.6904, 5.0443, 5.0378, 4.7277], device='cuda:4'), covar=tensor([0.0639, 0.0464, 0.0364, 0.0435, 0.1139, 0.0545, 0.0328, 0.0844], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0474, 0.0370, 0.0373, 0.0368, 0.0430, 0.0254, 0.0444], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:47:53,012 INFO [train.py:904] (4/8) Epoch 27, batch 1300, loss[loss=0.185, simple_loss=0.2677, pruned_loss=0.05112, over 16905.00 frames. ], tot_loss[loss=0.1592, simple_loss=0.2443, pruned_loss=0.03705, over 3304672.70 frames. ], batch size: 116, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:47:54,313 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=265204.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 07:48:33,153 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8930, 2.5750, 2.4291, 4.1212, 3.3766, 4.0455, 1.6683, 2.9895], device='cuda:4'), covar=tensor([0.1372, 0.0750, 0.1265, 0.0182, 0.0137, 0.0390, 0.1581, 0.0823], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0179, 0.0198, 0.0198, 0.0204, 0.0218, 0.0207, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 07:49:00,603 INFO [train.py:904] (4/8) Epoch 27, batch 1350, loss[loss=0.1587, simple_loss=0.238, pruned_loss=0.03976, over 16712.00 frames. ], tot_loss[loss=0.1592, simple_loss=0.2444, pruned_loss=0.03699, over 3306627.13 frames. ], batch size: 134, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:49:14,442 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 2.174e+02 2.461e+02 3.019e+02 8.065e+02, threshold=4.923e+02, percent-clipped=2.0 2023-05-02 07:49:35,189 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=265278.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:49:47,291 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-02 07:50:07,766 INFO [train.py:904] (4/8) Epoch 27, batch 1400, loss[loss=0.1469, simple_loss=0.2421, pruned_loss=0.0258, over 17193.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2452, pruned_loss=0.03725, over 3315732.32 frames. ], batch size: 44, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:50:21,254 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6246, 2.8335, 3.0882, 2.0060, 2.7226, 2.1161, 3.1925, 3.1646], device='cuda:4'), covar=tensor([0.0290, 0.1062, 0.0635, 0.2136, 0.0951, 0.1104, 0.0669, 0.1208], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0170, 0.0171, 0.0158, 0.0148, 0.0133, 0.0147, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 07:50:47,463 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5549, 4.6705, 4.8118, 4.6296, 4.6648, 5.2507, 4.7781, 4.4368], device='cuda:4'), covar=tensor([0.1750, 0.2403, 0.2727, 0.2347, 0.3039, 0.1279, 0.1816, 0.2830], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0637, 0.0707, 0.0524, 0.0692, 0.0732, 0.0549, 0.0697], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:50:56,846 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=265339.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:51:15,555 INFO [train.py:904] (4/8) Epoch 27, batch 1450, loss[loss=0.1681, simple_loss=0.2563, pruned_loss=0.03994, over 16545.00 frames. ], tot_loss[loss=0.1601, simple_loss=0.2452, pruned_loss=0.03752, over 3318568.38 frames. ], batch size: 68, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:51:29,668 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.640e+02 2.111e+02 2.470e+02 3.011e+02 5.399e+02, threshold=4.940e+02, percent-clipped=1.0 2023-05-02 07:51:40,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6933, 2.0223, 2.3335, 2.5168, 2.6431, 2.5551, 1.9086, 2.7521], device='cuda:4'), covar=tensor([0.0188, 0.0492, 0.0332, 0.0295, 0.0354, 0.0386, 0.0588, 0.0222], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0196, 0.0184, 0.0190, 0.0205, 0.0164, 0.0202, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 07:52:24,302 INFO [train.py:904] (4/8) Epoch 27, batch 1500, loss[loss=0.1623, simple_loss=0.2375, pruned_loss=0.04349, over 16451.00 frames. ], tot_loss[loss=0.1603, simple_loss=0.2451, pruned_loss=0.03771, over 3313379.88 frames. ], batch size: 146, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:53:36,975 INFO [train.py:904] (4/8) Epoch 27, batch 1550, loss[loss=0.1693, simple_loss=0.2416, pruned_loss=0.0485, over 16848.00 frames. ], tot_loss[loss=0.1607, simple_loss=0.2454, pruned_loss=0.03796, over 3314479.17 frames. ], batch size: 90, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:53:45,900 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=265460.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:53:49,832 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.731e+02 2.243e+02 2.540e+02 3.021e+02 6.885e+02, threshold=5.080e+02, percent-clipped=3.0 2023-05-02 07:54:43,792 INFO [train.py:904] (4/8) Epoch 27, batch 1600, loss[loss=0.1779, simple_loss=0.2647, pruned_loss=0.0455, over 16743.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2474, pruned_loss=0.0388, over 3312070.38 frames. ], batch size: 134, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:54:51,446 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=265508.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:55:10,281 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-02 07:55:21,281 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=265530.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:55:51,085 INFO [train.py:904] (4/8) Epoch 27, batch 1650, loss[loss=0.1712, simple_loss=0.2631, pruned_loss=0.0396, over 17036.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2493, pruned_loss=0.03979, over 3311090.18 frames. ], batch size: 50, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:56:04,414 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.682e+02 2.309e+02 2.667e+02 3.273e+02 6.043e+02, threshold=5.334e+02, percent-clipped=4.0 2023-05-02 07:56:43,897 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=265591.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:56:49,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1001, 5.7237, 5.9001, 5.5805, 5.7407, 6.2665, 5.7694, 5.4152], device='cuda:4'), covar=tensor([0.0869, 0.2090, 0.2230, 0.1885, 0.2144, 0.0892, 0.1471, 0.2196], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0637, 0.0705, 0.0523, 0.0693, 0.0730, 0.0549, 0.0696], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 07:57:00,226 INFO [train.py:904] (4/8) Epoch 27, batch 1700, loss[loss=0.1818, simple_loss=0.2631, pruned_loss=0.05028, over 16730.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2507, pruned_loss=0.04003, over 3311674.50 frames. ], batch size: 124, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:57:02,148 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.53 vs. limit=5.0 2023-05-02 07:57:15,986 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=265614.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:57:34,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6418, 3.7339, 2.1781, 4.0456, 2.9765, 3.9379, 2.2207, 2.9683], device='cuda:4'), covar=tensor([0.0284, 0.0396, 0.1742, 0.0377, 0.0736, 0.0681, 0.1769, 0.0761], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0183, 0.0198, 0.0175, 0.0182, 0.0223, 0.0207, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 07:57:42,660 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=265634.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 07:58:07,539 INFO [train.py:904] (4/8) Epoch 27, batch 1750, loss[loss=0.138, simple_loss=0.2209, pruned_loss=0.02753, over 16766.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2517, pruned_loss=0.03966, over 3323703.87 frames. ], batch size: 39, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:58:22,124 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.457e+02 2.193e+02 2.494e+02 2.973e+02 6.444e+02, threshold=4.988e+02, percent-clipped=1.0 2023-05-02 07:58:38,397 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=265675.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 07:58:39,471 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3396, 4.2050, 4.4194, 4.5442, 4.6100, 4.2474, 4.4656, 4.6294], device='cuda:4'), covar=tensor([0.1720, 0.1241, 0.1340, 0.0718, 0.0599, 0.1231, 0.2689, 0.0773], device='cuda:4'), in_proj_covar=tensor([0.0690, 0.0849, 0.0980, 0.0860, 0.0652, 0.0679, 0.0713, 0.0830], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 07:59:15,656 INFO [train.py:904] (4/8) Epoch 27, batch 1800, loss[loss=0.1379, simple_loss=0.2272, pruned_loss=0.02427, over 16723.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2515, pruned_loss=0.03875, over 3328171.34 frames. ], batch size: 39, lr: 2.51e-03, grad_scale: 8.0 2023-05-02 07:59:32,450 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=265715.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:00:23,854 INFO [train.py:904] (4/8) Epoch 27, batch 1850, loss[loss=0.156, simple_loss=0.2499, pruned_loss=0.03106, over 17134.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2534, pruned_loss=0.03904, over 3309587.94 frames. ], batch size: 48, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:00:37,844 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.494e+02 2.122e+02 2.552e+02 2.913e+02 6.429e+02, threshold=5.105e+02, percent-clipped=1.0 2023-05-02 08:00:55,762 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=265776.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:01:19,747 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9936, 4.7326, 5.0470, 5.2469, 5.4269, 4.7768, 5.4484, 5.4416], device='cuda:4'), covar=tensor([0.2143, 0.1507, 0.1918, 0.0840, 0.0559, 0.1037, 0.0509, 0.0667], device='cuda:4'), in_proj_covar=tensor([0.0692, 0.0851, 0.0983, 0.0862, 0.0653, 0.0681, 0.0715, 0.0832], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:01:33,281 INFO [train.py:904] (4/8) Epoch 27, batch 1900, loss[loss=0.1598, simple_loss=0.2574, pruned_loss=0.03107, over 16490.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2518, pruned_loss=0.03847, over 3314374.51 frames. ], batch size: 62, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:02:13,068 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9699, 3.6366, 4.2104, 2.0677, 4.3947, 4.5137, 3.2959, 3.3767], device='cuda:4'), covar=tensor([0.0770, 0.0317, 0.0269, 0.1283, 0.0107, 0.0221, 0.0459, 0.0468], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0101, 0.0139, 0.0086, 0.0131, 0.0130, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:02:42,098 INFO [train.py:904] (4/8) Epoch 27, batch 1950, loss[loss=0.176, simple_loss=0.2627, pruned_loss=0.04466, over 16289.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2519, pruned_loss=0.03793, over 3318904.30 frames. ], batch size: 165, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:02:54,907 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.539e+02 2.085e+02 2.462e+02 3.209e+02 6.071e+02, threshold=4.924e+02, percent-clipped=6.0 2023-05-02 08:03:01,145 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-02 08:03:08,288 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=265872.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:03:26,219 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=265886.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:03:45,858 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-05-02 08:03:50,711 INFO [train.py:904] (4/8) Epoch 27, batch 2000, loss[loss=0.1586, simple_loss=0.2565, pruned_loss=0.03031, over 16762.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2516, pruned_loss=0.03756, over 3319603.31 frames. ], batch size: 57, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:04:10,496 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=265917.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:04:33,666 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=265933.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:04:34,770 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=265934.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:05:01,213 INFO [train.py:904] (4/8) Epoch 27, batch 2050, loss[loss=0.1801, simple_loss=0.274, pruned_loss=0.04306, over 16512.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2517, pruned_loss=0.03826, over 3301492.39 frames. ], batch size: 68, lr: 2.50e-03, grad_scale: 16.0 2023-05-02 08:05:14,206 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.410e+02 2.179e+02 2.495e+02 3.022e+02 6.334e+02, threshold=4.990e+02, percent-clipped=2.0 2023-05-02 08:05:25,358 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=265970.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 08:05:36,779 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=265978.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:05:41,864 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=265982.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:05:44,584 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0045, 4.5031, 3.2084, 2.3771, 2.7228, 2.6943, 4.8879, 3.6689], device='cuda:4'), covar=tensor([0.2856, 0.0551, 0.1816, 0.3210, 0.3053, 0.2148, 0.0363, 0.1538], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0276, 0.0312, 0.0325, 0.0304, 0.0275, 0.0304, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 08:06:14,060 INFO [train.py:904] (4/8) Epoch 27, batch 2100, loss[loss=0.2388, simple_loss=0.3059, pruned_loss=0.08583, over 12184.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2529, pruned_loss=0.039, over 3295140.05 frames. ], batch size: 247, lr: 2.50e-03, grad_scale: 16.0 2023-05-02 08:06:43,920 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-05-02 08:06:48,588 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6527, 3.7111, 2.1418, 4.2708, 2.8559, 4.1530, 2.2633, 3.0749], device='cuda:4'), covar=tensor([0.0347, 0.0405, 0.2078, 0.0333, 0.0914, 0.0550, 0.2035, 0.0895], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0183, 0.0199, 0.0176, 0.0183, 0.0224, 0.0208, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:07:22,091 INFO [train.py:904] (4/8) Epoch 27, batch 2150, loss[loss=0.1862, simple_loss=0.2744, pruned_loss=0.04904, over 16803.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2535, pruned_loss=0.03927, over 3308653.06 frames. ], batch size: 124, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:07:23,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3501, 2.4226, 2.4061, 4.2350, 2.2971, 2.8263, 2.4919, 2.5690], device='cuda:4'), covar=tensor([0.1451, 0.3808, 0.3297, 0.0566, 0.4423, 0.2645, 0.3676, 0.3851], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0473, 0.0387, 0.0338, 0.0446, 0.0541, 0.0444, 0.0553], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:07:37,031 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.561e+02 2.274e+02 2.676e+02 3.152e+02 5.314e+02, threshold=5.352e+02, percent-clipped=2.0 2023-05-02 08:07:46,334 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=266071.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:07:54,359 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-02 08:08:32,964 INFO [train.py:904] (4/8) Epoch 27, batch 2200, loss[loss=0.1465, simple_loss=0.2352, pruned_loss=0.0289, over 17184.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2545, pruned_loss=0.04045, over 3303379.45 frames. ], batch size: 46, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:09:34,140 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8531, 2.8604, 2.6212, 4.8739, 3.6908, 4.2582, 1.5667, 3.0222], device='cuda:4'), covar=tensor([0.1463, 0.0904, 0.1335, 0.0263, 0.0263, 0.0441, 0.1820, 0.0870], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0180, 0.0199, 0.0200, 0.0205, 0.0219, 0.0208, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:09:41,315 INFO [train.py:904] (4/8) Epoch 27, batch 2250, loss[loss=0.1622, simple_loss=0.2505, pruned_loss=0.03693, over 16496.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2558, pruned_loss=0.04123, over 3291137.61 frames. ], batch size: 68, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:09:56,594 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.643e+02 2.173e+02 2.503e+02 2.978e+02 4.988e+02, threshold=5.005e+02, percent-clipped=0.0 2023-05-02 08:10:12,813 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-05-02 08:10:24,525 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6589, 6.0434, 5.7672, 5.9009, 5.4625, 5.4846, 5.4179, 6.1495], device='cuda:4'), covar=tensor([0.1381, 0.0902, 0.0969, 0.0851, 0.0855, 0.0634, 0.1266, 0.0864], device='cuda:4'), in_proj_covar=tensor([0.0721, 0.0871, 0.0711, 0.0672, 0.0553, 0.0549, 0.0737, 0.0684], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:10:28,099 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=266186.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:10:51,414 INFO [train.py:904] (4/8) Epoch 27, batch 2300, loss[loss=0.1664, simple_loss=0.252, pruned_loss=0.04038, over 16737.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2555, pruned_loss=0.04131, over 3294725.15 frames. ], batch size: 83, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:11:06,868 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=266214.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:11:23,572 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=266226.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 08:11:25,873 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=266228.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:11:34,294 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=266234.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:11:59,648 INFO [train.py:904] (4/8) Epoch 27, batch 2350, loss[loss=0.1815, simple_loss=0.2597, pruned_loss=0.05168, over 16776.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2559, pruned_loss=0.04079, over 3293686.60 frames. ], batch size: 102, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:12:14,159 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.417e+02 2.140e+02 2.442e+02 2.999e+02 5.112e+02, threshold=4.884e+02, percent-clipped=1.0 2023-05-02 08:12:22,934 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=266270.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 08:12:27,931 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=266273.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:12:30,388 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=266275.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:12:46,330 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=266287.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 08:13:07,763 INFO [train.py:904] (4/8) Epoch 27, batch 2400, loss[loss=0.195, simple_loss=0.2759, pruned_loss=0.05703, over 16654.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2554, pruned_loss=0.04047, over 3297987.54 frames. ], batch size: 134, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:13:28,958 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=266318.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:14:07,689 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=266346.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:14:16,214 INFO [train.py:904] (4/8) Epoch 27, batch 2450, loss[loss=0.177, simple_loss=0.26, pruned_loss=0.04702, over 16884.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2565, pruned_loss=0.04061, over 3298259.67 frames. ], batch size: 116, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:14:31,082 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.587e+02 2.203e+02 2.440e+02 2.935e+02 4.080e+02, threshold=4.880e+02, percent-clipped=0.0 2023-05-02 08:14:41,915 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=266371.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:15:26,443 INFO [train.py:904] (4/8) Epoch 27, batch 2500, loss[loss=0.1852, simple_loss=0.2769, pruned_loss=0.04675, over 16682.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2564, pruned_loss=0.04034, over 3305424.61 frames. ], batch size: 57, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:15:31,453 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1036, 4.1270, 4.4091, 4.3907, 4.4366, 4.1721, 4.1907, 4.1202], device='cuda:4'), covar=tensor([0.0382, 0.0747, 0.0448, 0.0447, 0.0556, 0.0478, 0.0883, 0.0635], device='cuda:4'), in_proj_covar=tensor([0.0440, 0.0496, 0.0478, 0.0441, 0.0530, 0.0508, 0.0583, 0.0405], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 08:15:31,575 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=266407.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:15:48,870 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=266419.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:16:35,710 INFO [train.py:904] (4/8) Epoch 27, batch 2550, loss[loss=0.1903, simple_loss=0.2714, pruned_loss=0.05456, over 12195.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2567, pruned_loss=0.04022, over 3306010.20 frames. ], batch size: 246, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:16:51,126 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.440e+02 2.045e+02 2.342e+02 2.686e+02 5.745e+02, threshold=4.685e+02, percent-clipped=2.0 2023-05-02 08:16:51,500 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=266464.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:17:45,262 INFO [train.py:904] (4/8) Epoch 27, batch 2600, loss[loss=0.1599, simple_loss=0.2538, pruned_loss=0.03297, over 17128.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2563, pruned_loss=0.03929, over 3307769.45 frames. ], batch size: 49, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:17:49,168 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1352, 5.4638, 5.2939, 5.3231, 4.9789, 4.8895, 4.9479, 5.6257], device='cuda:4'), covar=tensor([0.1445, 0.0998, 0.0980, 0.0847, 0.0839, 0.0882, 0.1283, 0.0893], device='cuda:4'), in_proj_covar=tensor([0.0724, 0.0876, 0.0713, 0.0674, 0.0555, 0.0550, 0.0740, 0.0686], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:18:07,376 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-05-02 08:18:17,148 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=266525.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:18:20,981 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=266528.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:18:55,437 INFO [train.py:904] (4/8) Epoch 27, batch 2650, loss[loss=0.1782, simple_loss=0.2603, pruned_loss=0.04805, over 16719.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2566, pruned_loss=0.03929, over 3315723.35 frames. ], batch size: 124, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:19:08,900 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6946, 2.9208, 2.8781, 5.0488, 4.0277, 4.4026, 1.8210, 3.1185], device='cuda:4'), covar=tensor([0.1483, 0.0830, 0.1158, 0.0185, 0.0221, 0.0430, 0.1612, 0.0842], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0180, 0.0199, 0.0200, 0.0206, 0.0219, 0.0208, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:19:11,569 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 2.008e+02 2.392e+02 2.897e+02 5.210e+02, threshold=4.785e+02, percent-clipped=1.0 2023-05-02 08:19:20,871 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=266570.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:19:21,057 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0945, 4.7630, 4.5513, 3.2550, 3.9205, 4.5507, 4.0207, 2.8338], device='cuda:4'), covar=tensor([0.0485, 0.0076, 0.0054, 0.0380, 0.0140, 0.0105, 0.0116, 0.0442], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0090, 0.0091, 0.0136, 0.0103, 0.0116, 0.0099, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 08:19:24,563 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=266573.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:19:29,419 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=266576.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:19:37,429 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=266582.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 08:20:05,798 INFO [train.py:904] (4/8) Epoch 27, batch 2700, loss[loss=0.1613, simple_loss=0.2486, pruned_loss=0.03698, over 16687.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2558, pruned_loss=0.03864, over 3323584.94 frames. ], batch size: 134, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:20:31,906 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=266621.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:21:15,401 INFO [train.py:904] (4/8) Epoch 27, batch 2750, loss[loss=0.164, simple_loss=0.2607, pruned_loss=0.03365, over 17064.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2561, pruned_loss=0.03855, over 3322867.90 frames. ], batch size: 55, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:21:29,196 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.625e+02 2.109e+02 2.416e+02 2.772e+02 5.063e+02, threshold=4.832e+02, percent-clipped=2.0 2023-05-02 08:21:51,264 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=266679.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:22:22,656 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=266702.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:22:23,526 INFO [train.py:904] (4/8) Epoch 27, batch 2800, loss[loss=0.1759, simple_loss=0.2567, pruned_loss=0.04755, over 16875.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2557, pruned_loss=0.03882, over 3326086.92 frames. ], batch size: 116, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:22:25,103 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=266704.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:22:32,454 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6996, 3.3539, 3.8030, 1.9602, 3.8455, 3.9073, 3.2079, 2.9629], device='cuda:4'), covar=tensor([0.0763, 0.0277, 0.0173, 0.1252, 0.0114, 0.0199, 0.0372, 0.0442], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0101, 0.0139, 0.0086, 0.0132, 0.0130, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:22:38,188 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0712, 3.8303, 4.3763, 2.3278, 4.5411, 4.7138, 3.3436, 3.6540], device='cuda:4'), covar=tensor([0.0698, 0.0287, 0.0229, 0.1119, 0.0081, 0.0171, 0.0432, 0.0373], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0101, 0.0139, 0.0086, 0.0132, 0.0131, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:23:15,439 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=266740.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:23:31,358 INFO [train.py:904] (4/8) Epoch 27, batch 2850, loss[loss=0.1809, simple_loss=0.261, pruned_loss=0.05045, over 16731.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2541, pruned_loss=0.03864, over 3321684.72 frames. ], batch size: 134, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:23:48,195 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.557e+02 2.101e+02 2.361e+02 2.749e+02 4.923e+02, threshold=4.721e+02, percent-clipped=1.0 2023-05-02 08:23:49,872 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=266765.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:24:13,899 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8266, 3.7751, 2.9365, 2.3463, 2.4392, 2.4294, 4.0088, 3.3398], device='cuda:4'), covar=tensor([0.2832, 0.0651, 0.1773, 0.3223, 0.3105, 0.2272, 0.0565, 0.1579], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0277, 0.0314, 0.0327, 0.0306, 0.0276, 0.0305, 0.0354], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 08:24:40,561 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0051, 3.7250, 4.3139, 2.1680, 4.4526, 4.6089, 3.3267, 3.5916], device='cuda:4'), covar=tensor([0.0761, 0.0308, 0.0259, 0.1230, 0.0090, 0.0217, 0.0425, 0.0393], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0102, 0.0139, 0.0086, 0.0132, 0.0131, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:24:41,244 INFO [train.py:904] (4/8) Epoch 27, batch 2900, loss[loss=0.171, simple_loss=0.2669, pruned_loss=0.03751, over 17051.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2534, pruned_loss=0.03864, over 3319031.32 frames. ], batch size: 50, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:25:04,248 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=266820.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:25:49,402 INFO [train.py:904] (4/8) Epoch 27, batch 2950, loss[loss=0.1762, simple_loss=0.2487, pruned_loss=0.05189, over 15449.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2528, pruned_loss=0.03918, over 3316986.64 frames. ], batch size: 190, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:26:04,411 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.609e+02 2.163e+02 2.599e+02 3.216e+02 6.007e+02, threshold=5.198e+02, percent-clipped=1.0 2023-05-02 08:26:14,739 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=266870.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:26:29,625 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=266882.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 08:26:58,264 INFO [train.py:904] (4/8) Epoch 27, batch 3000, loss[loss=0.1673, simple_loss=0.267, pruned_loss=0.03379, over 16647.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.253, pruned_loss=0.03972, over 3323067.31 frames. ], batch size: 62, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:26:58,265 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 08:27:07,044 INFO [train.py:938] (4/8) Epoch 27, validation: loss=0.1336, simple_loss=0.2386, pruned_loss=0.01429, over 944034.00 frames. 2023-05-02 08:27:07,045 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 08:27:27,849 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=266918.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:27:30,368 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=266920.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:27:45,280 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=266930.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 08:27:53,098 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0047, 2.1962, 2.2796, 3.6705, 2.1853, 2.4473, 2.2581, 2.3266], device='cuda:4'), covar=tensor([0.1692, 0.3736, 0.3094, 0.0746, 0.4019, 0.2651, 0.3910, 0.3244], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0472, 0.0386, 0.0337, 0.0444, 0.0540, 0.0443, 0.0552], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:28:02,416 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8897, 4.1005, 2.9217, 4.7589, 3.2924, 4.6505, 2.8395, 3.5173], device='cuda:4'), covar=tensor([0.0360, 0.0387, 0.1391, 0.0258, 0.0768, 0.0499, 0.1398, 0.0681], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0183, 0.0198, 0.0177, 0.0183, 0.0224, 0.0207, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:28:16,681 INFO [train.py:904] (4/8) Epoch 27, batch 3050, loss[loss=0.1438, simple_loss=0.2304, pruned_loss=0.02864, over 16812.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2532, pruned_loss=0.03993, over 3326008.89 frames. ], batch size: 39, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:28:26,993 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 08:28:30,545 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.662e+02 2.121e+02 2.455e+02 2.792e+02 5.832e+02, threshold=4.910e+02, percent-clipped=2.0 2023-05-02 08:28:39,963 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2269, 2.4328, 2.5694, 4.0816, 2.3344, 2.6905, 2.4812, 2.5467], device='cuda:4'), covar=tensor([0.1619, 0.3647, 0.2992, 0.0662, 0.3925, 0.2660, 0.3779, 0.3253], device='cuda:4'), in_proj_covar=tensor([0.0423, 0.0475, 0.0388, 0.0340, 0.0447, 0.0544, 0.0446, 0.0555], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:28:55,840 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=266981.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:29:24,850 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=267002.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:29:25,704 INFO [train.py:904] (4/8) Epoch 27, batch 3100, loss[loss=0.1609, simple_loss=0.2335, pruned_loss=0.04418, over 16392.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2532, pruned_loss=0.03987, over 3327048.95 frames. ], batch size: 146, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:29:54,110 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6543, 3.7457, 2.4181, 4.0806, 2.9754, 4.0109, 2.4368, 3.0119], device='cuda:4'), covar=tensor([0.0297, 0.0378, 0.1521, 0.0329, 0.0744, 0.0697, 0.1454, 0.0741], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0183, 0.0198, 0.0176, 0.0182, 0.0224, 0.0206, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:30:11,821 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=267035.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:30:33,230 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=267050.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:30:36,918 INFO [train.py:904] (4/8) Epoch 27, batch 3150, loss[loss=0.1542, simple_loss=0.2364, pruned_loss=0.03603, over 16446.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2523, pruned_loss=0.03978, over 3328751.21 frames. ], batch size: 146, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:30:45,722 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=267060.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:30:50,834 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.616e+02 2.137e+02 2.489e+02 2.954e+02 4.893e+02, threshold=4.979e+02, percent-clipped=1.0 2023-05-02 08:31:15,376 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=267082.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:31:32,681 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-02 08:31:37,542 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=267098.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:31:43,509 INFO [train.py:904] (4/8) Epoch 27, batch 3200, loss[loss=0.2004, simple_loss=0.2775, pruned_loss=0.06167, over 11973.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2523, pruned_loss=0.03957, over 3325332.85 frames. ], batch size: 246, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:32:07,958 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=267120.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:32:28,761 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8088, 1.9878, 2.5303, 2.8471, 2.6634, 3.2590, 2.3172, 3.2889], device='cuda:4'), covar=tensor([0.0339, 0.0644, 0.0395, 0.0442, 0.0463, 0.0287, 0.0587, 0.0234], device='cuda:4'), in_proj_covar=tensor([0.0203, 0.0202, 0.0190, 0.0197, 0.0211, 0.0169, 0.0208, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 08:32:39,052 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=267143.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:32:50,807 INFO [train.py:904] (4/8) Epoch 27, batch 3250, loss[loss=0.1574, simple_loss=0.2418, pruned_loss=0.03645, over 16902.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2525, pruned_loss=0.03968, over 3332451.06 frames. ], batch size: 90, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:33:01,489 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=267159.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:33:07,328 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.439e+02 2.100e+02 2.558e+02 2.992e+02 5.304e+02, threshold=5.115e+02, percent-clipped=1.0 2023-05-02 08:33:08,859 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9847, 4.2390, 4.1004, 4.1256, 3.8107, 3.8393, 3.9191, 4.2333], device='cuda:4'), covar=tensor([0.1216, 0.0941, 0.0910, 0.0891, 0.0806, 0.1858, 0.0953, 0.1001], device='cuda:4'), in_proj_covar=tensor([0.0735, 0.0887, 0.0724, 0.0686, 0.0564, 0.0558, 0.0751, 0.0695], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:33:12,826 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=267168.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:33:48,064 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6322, 4.5971, 4.5445, 4.2132, 4.2790, 4.5806, 4.3831, 4.3484], device='cuda:4'), covar=tensor([0.0650, 0.0755, 0.0308, 0.0329, 0.0794, 0.0523, 0.0496, 0.0714], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0483, 0.0376, 0.0381, 0.0376, 0.0438, 0.0258, 0.0453], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 08:34:00,064 INFO [train.py:904] (4/8) Epoch 27, batch 3300, loss[loss=0.1719, simple_loss=0.2701, pruned_loss=0.03686, over 16739.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2538, pruned_loss=0.04046, over 3314950.74 frames. ], batch size: 62, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:34:11,117 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9263, 2.1785, 2.3096, 3.4382, 2.1274, 2.4126, 2.2766, 2.2603], device='cuda:4'), covar=tensor([0.1597, 0.3607, 0.3194, 0.0832, 0.4223, 0.2673, 0.3746, 0.3424], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0475, 0.0387, 0.0339, 0.0447, 0.0543, 0.0446, 0.0555], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:35:07,427 INFO [train.py:904] (4/8) Epoch 27, batch 3350, loss[loss=0.151, simple_loss=0.244, pruned_loss=0.02905, over 17262.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2531, pruned_loss=0.03967, over 3314738.57 frames. ], batch size: 45, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:35:22,678 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.323e+02 1.912e+02 2.198e+02 2.588e+02 3.700e+02, threshold=4.395e+02, percent-clipped=0.0 2023-05-02 08:35:34,250 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5083, 3.2638, 3.6604, 2.0037, 3.6802, 3.7190, 3.1245, 2.9156], device='cuda:4'), covar=tensor([0.0846, 0.0274, 0.0203, 0.1204, 0.0136, 0.0225, 0.0417, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0111, 0.0102, 0.0140, 0.0087, 0.0132, 0.0131, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:35:38,160 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=267276.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:36:14,881 INFO [train.py:904] (4/8) Epoch 27, batch 3400, loss[loss=0.1462, simple_loss=0.2426, pruned_loss=0.02485, over 17120.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2537, pruned_loss=0.03977, over 3314255.71 frames. ], batch size: 49, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:36:28,481 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3235, 4.6914, 4.6388, 3.4868, 3.9008, 4.6294, 4.1335, 2.7891], device='cuda:4'), covar=tensor([0.0455, 0.0068, 0.0046, 0.0348, 0.0134, 0.0095, 0.0103, 0.0478], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0091, 0.0092, 0.0137, 0.0104, 0.0117, 0.0100, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 08:36:58,243 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=267335.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:37:22,625 INFO [train.py:904] (4/8) Epoch 27, batch 3450, loss[loss=0.1766, simple_loss=0.2569, pruned_loss=0.04817, over 16423.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2525, pruned_loss=0.039, over 3325998.87 frames. ], batch size: 146, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:37:31,857 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=267360.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:37:36,672 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.354e+02 2.041e+02 2.356e+02 2.660e+02 5.077e+02, threshold=4.713e+02, percent-clipped=1.0 2023-05-02 08:38:03,485 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=267383.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:38:20,854 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4957, 3.5980, 4.0029, 2.3304, 3.2888, 2.4668, 3.7778, 3.7433], device='cuda:4'), covar=tensor([0.0267, 0.0947, 0.0491, 0.2059, 0.0800, 0.1045, 0.0643, 0.1120], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0173, 0.0171, 0.0158, 0.0148, 0.0133, 0.0147, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 08:38:30,923 INFO [train.py:904] (4/8) Epoch 27, batch 3500, loss[loss=0.1655, simple_loss=0.2524, pruned_loss=0.03937, over 16459.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2511, pruned_loss=0.03866, over 3316296.53 frames. ], batch size: 68, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:38:39,646 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=267408.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:39:19,870 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=267438.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:39:40,533 INFO [train.py:904] (4/8) Epoch 27, batch 3550, loss[loss=0.1682, simple_loss=0.2645, pruned_loss=0.0359, over 17104.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.25, pruned_loss=0.03776, over 3318750.62 frames. ], batch size: 55, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:39:41,926 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=267454.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:39:53,961 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.045e+02 2.366e+02 2.928e+02 5.405e+02, threshold=4.732e+02, percent-clipped=1.0 2023-05-02 08:40:36,697 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=267494.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:40:48,161 INFO [train.py:904] (4/8) Epoch 27, batch 3600, loss[loss=0.1621, simple_loss=0.2507, pruned_loss=0.03671, over 17203.00 frames. ], tot_loss[loss=0.1616, simple_loss=0.2487, pruned_loss=0.03719, over 3302903.80 frames. ], batch size: 44, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:42:00,665 INFO [train.py:904] (4/8) Epoch 27, batch 3650, loss[loss=0.1607, simple_loss=0.2406, pruned_loss=0.0404, over 16227.00 frames. ], tot_loss[loss=0.1619, simple_loss=0.2479, pruned_loss=0.03797, over 3300952.92 frames. ], batch size: 165, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:42:03,426 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=267555.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:42:16,686 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.290e+02 2.009e+02 2.392e+02 2.845e+02 4.698e+02, threshold=4.785e+02, percent-clipped=1.0 2023-05-02 08:42:36,689 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=267576.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:43:14,274 INFO [train.py:904] (4/8) Epoch 27, batch 3700, loss[loss=0.1675, simple_loss=0.246, pruned_loss=0.0445, over 16717.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2461, pruned_loss=0.03951, over 3291669.44 frames. ], batch size: 134, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:43:45,314 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=267624.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:43:57,339 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5976, 4.6629, 4.9322, 4.9214, 4.9791, 4.6574, 4.6443, 4.4996], device='cuda:4'), covar=tensor([0.0363, 0.0598, 0.0431, 0.0451, 0.0503, 0.0482, 0.0839, 0.0676], device='cuda:4'), in_proj_covar=tensor([0.0447, 0.0503, 0.0484, 0.0447, 0.0535, 0.0512, 0.0590, 0.0408], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 08:44:27,013 INFO [train.py:904] (4/8) Epoch 27, batch 3750, loss[loss=0.1787, simple_loss=0.2549, pruned_loss=0.05128, over 16173.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2464, pruned_loss=0.04054, over 3289684.09 frames. ], batch size: 165, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:44:28,613 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0463, 3.9698, 4.1076, 4.2136, 4.2843, 3.9273, 4.1192, 4.3188], device='cuda:4'), covar=tensor([0.1638, 0.1062, 0.1190, 0.0660, 0.0647, 0.1419, 0.2380, 0.0700], device='cuda:4'), in_proj_covar=tensor([0.0699, 0.0859, 0.0999, 0.0869, 0.0661, 0.0690, 0.0721, 0.0838], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:44:42,769 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.122e+02 2.445e+02 2.969e+02 5.138e+02, threshold=4.890e+02, percent-clipped=1.0 2023-05-02 08:45:40,198 INFO [train.py:904] (4/8) Epoch 27, batch 3800, loss[loss=0.169, simple_loss=0.2462, pruned_loss=0.04589, over 16784.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.248, pruned_loss=0.04177, over 3283518.90 frames. ], batch size: 124, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:45:41,321 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1096, 4.1653, 4.4243, 4.4130, 4.4557, 4.1807, 4.2075, 4.1801], device='cuda:4'), covar=tensor([0.0395, 0.0702, 0.0454, 0.0450, 0.0514, 0.0501, 0.0753, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0506, 0.0486, 0.0449, 0.0536, 0.0515, 0.0592, 0.0408], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 08:45:46,082 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3988, 3.5475, 3.7339, 2.5257, 3.3811, 3.8185, 3.5196, 2.2303], device='cuda:4'), covar=tensor([0.0494, 0.0158, 0.0063, 0.0422, 0.0119, 0.0100, 0.0102, 0.0475], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0090, 0.0092, 0.0137, 0.0103, 0.0117, 0.0100, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 08:46:31,902 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=267738.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:46:52,904 INFO [train.py:904] (4/8) Epoch 27, batch 3850, loss[loss=0.1874, simple_loss=0.2673, pruned_loss=0.05373, over 12506.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2482, pruned_loss=0.04245, over 3280805.67 frames. ], batch size: 246, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:46:54,235 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=267754.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:47:08,875 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.727e+02 2.090e+02 2.451e+02 2.947e+02 6.738e+02, threshold=4.901e+02, percent-clipped=4.0 2023-05-02 08:47:09,323 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4831, 5.5593, 5.4061, 4.9520, 5.0080, 5.4610, 5.3340, 5.2004], device='cuda:4'), covar=tensor([0.0619, 0.0344, 0.0262, 0.0304, 0.1063, 0.0326, 0.0246, 0.0531], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0482, 0.0374, 0.0380, 0.0375, 0.0436, 0.0256, 0.0451], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 08:47:39,701 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=267786.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:48:02,755 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=267802.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:48:03,542 INFO [train.py:904] (4/8) Epoch 27, batch 3900, loss[loss=0.1666, simple_loss=0.2422, pruned_loss=0.04553, over 16775.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2486, pruned_loss=0.04316, over 3292229.85 frames. ], batch size: 83, lr: 2.50e-03, grad_scale: 8.0 2023-05-02 08:48:10,172 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2839, 3.3060, 2.1604, 3.4320, 2.6240, 3.4675, 2.3134, 2.7631], device='cuda:4'), covar=tensor([0.0311, 0.0508, 0.1508, 0.0353, 0.0767, 0.0751, 0.1402, 0.0753], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0183, 0.0198, 0.0177, 0.0182, 0.0224, 0.0206, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 08:49:11,734 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=267850.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:49:12,112 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-02 08:49:14,858 INFO [train.py:904] (4/8) Epoch 27, batch 3950, loss[loss=0.1641, simple_loss=0.2441, pruned_loss=0.04207, over 16725.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2488, pruned_loss=0.04352, over 3300487.30 frames. ], batch size: 124, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 08:49:32,326 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.702e+02 2.233e+02 2.530e+02 3.152e+02 5.510e+02, threshold=5.059e+02, percent-clipped=1.0 2023-05-02 08:50:11,200 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.62 vs. limit=5.0 2023-05-02 08:50:28,771 INFO [train.py:904] (4/8) Epoch 27, batch 4000, loss[loss=0.2015, simple_loss=0.2794, pruned_loss=0.06178, over 12760.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2488, pruned_loss=0.04393, over 3296514.38 frames. ], batch size: 246, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 08:51:24,488 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-05-02 08:51:42,553 INFO [train.py:904] (4/8) Epoch 27, batch 4050, loss[loss=0.1678, simple_loss=0.2576, pruned_loss=0.03905, over 16654.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2491, pruned_loss=0.04307, over 3304367.95 frames. ], batch size: 89, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 08:51:58,300 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 1.841e+02 2.143e+02 2.539e+02 4.386e+02, threshold=4.286e+02, percent-clipped=0.0 2023-05-02 08:52:59,882 INFO [train.py:904] (4/8) Epoch 27, batch 4100, loss[loss=0.2093, simple_loss=0.2892, pruned_loss=0.06473, over 12039.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2511, pruned_loss=0.0427, over 3303528.84 frames. ], batch size: 246, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 08:53:18,334 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-05-02 08:54:16,165 INFO [train.py:904] (4/8) Epoch 27, batch 4150, loss[loss=0.2031, simple_loss=0.2863, pruned_loss=0.05997, over 16841.00 frames. ], tot_loss[loss=0.174, simple_loss=0.258, pruned_loss=0.04505, over 3244078.58 frames. ], batch size: 109, lr: 2.49e-03, grad_scale: 16.0 2023-05-02 08:54:33,171 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.413e+02 2.031e+02 2.466e+02 3.151e+02 6.245e+02, threshold=4.932e+02, percent-clipped=8.0 2023-05-02 08:55:32,120 INFO [train.py:904] (4/8) Epoch 27, batch 4200, loss[loss=0.1817, simple_loss=0.2792, pruned_loss=0.04207, over 16811.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2649, pruned_loss=0.0469, over 3195210.43 frames. ], batch size: 102, lr: 2.49e-03, grad_scale: 16.0 2023-05-02 08:55:46,801 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3043, 2.5566, 2.3554, 4.0995, 2.3249, 2.8881, 2.5118, 2.6558], device='cuda:4'), covar=tensor([0.1458, 0.3326, 0.2957, 0.0557, 0.3920, 0.2302, 0.3353, 0.3180], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0477, 0.0388, 0.0339, 0.0447, 0.0545, 0.0446, 0.0557], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:56:34,713 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-05-02 08:56:40,950 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=268150.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:56:44,662 INFO [train.py:904] (4/8) Epoch 27, batch 4250, loss[loss=0.1736, simple_loss=0.2736, pruned_loss=0.03683, over 16746.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2691, pruned_loss=0.04714, over 3183698.13 frames. ], batch size: 83, lr: 2.49e-03, grad_scale: 16.0 2023-05-02 08:57:00,876 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.686e+02 2.148e+02 2.458e+02 3.113e+02 4.874e+02, threshold=4.916e+02, percent-clipped=0.0 2023-05-02 08:57:45,678 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8838, 1.3952, 1.7194, 1.7190, 1.8061, 1.9332, 1.6726, 1.8069], device='cuda:4'), covar=tensor([0.0267, 0.0431, 0.0251, 0.0333, 0.0295, 0.0229, 0.0496, 0.0168], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0198, 0.0187, 0.0193, 0.0208, 0.0167, 0.0204, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 08:57:50,362 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=268198.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:57:57,371 INFO [train.py:904] (4/8) Epoch 27, batch 4300, loss[loss=0.2083, simple_loss=0.3035, pruned_loss=0.05656, over 16824.00 frames. ], tot_loss[loss=0.182, simple_loss=0.2707, pruned_loss=0.04662, over 3185696.07 frames. ], batch size: 116, lr: 2.49e-03, grad_scale: 16.0 2023-05-02 08:58:05,773 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-05-02 08:58:10,874 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=268212.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 08:59:12,216 INFO [train.py:904] (4/8) Epoch 27, batch 4350, loss[loss=0.1921, simple_loss=0.287, pruned_loss=0.04864, over 16707.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2739, pruned_loss=0.04716, over 3187602.62 frames. ], batch size: 124, lr: 2.49e-03, grad_scale: 16.0 2023-05-02 08:59:27,932 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.498e+02 2.120e+02 2.432e+02 2.957e+02 3.904e+02, threshold=4.863e+02, percent-clipped=0.0 2023-05-02 08:59:43,221 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=268273.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 08:59:45,539 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=268275.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:00:26,994 INFO [train.py:904] (4/8) Epoch 27, batch 4400, loss[loss=0.2058, simple_loss=0.2921, pruned_loss=0.05971, over 17051.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2756, pruned_loss=0.0482, over 3179287.29 frames. ], batch size: 53, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:00:41,115 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.8970, 6.1484, 5.8889, 6.0618, 5.6978, 5.3281, 5.6631, 6.3053], device='cuda:4'), covar=tensor([0.1200, 0.0801, 0.1207, 0.0761, 0.0733, 0.0667, 0.1123, 0.0810], device='cuda:4'), in_proj_covar=tensor([0.0718, 0.0871, 0.0708, 0.0670, 0.0553, 0.0549, 0.0733, 0.0684], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:01:03,898 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-05-02 09:01:14,951 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=268336.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 09:01:37,963 INFO [train.py:904] (4/8) Epoch 27, batch 4450, loss[loss=0.1843, simple_loss=0.2838, pruned_loss=0.04238, over 16725.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2789, pruned_loss=0.04914, over 3196178.89 frames. ], batch size: 89, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:01:55,122 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.500e+02 1.898e+02 2.329e+02 2.751e+02 5.089e+02, threshold=4.658e+02, percent-clipped=1.0 2023-05-02 09:02:50,811 INFO [train.py:904] (4/8) Epoch 27, batch 4500, loss[loss=0.2104, simple_loss=0.2947, pruned_loss=0.06308, over 15487.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2793, pruned_loss=0.04993, over 3213914.42 frames. ], batch size: 190, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:04:03,668 INFO [train.py:904] (4/8) Epoch 27, batch 4550, loss[loss=0.2116, simple_loss=0.3016, pruned_loss=0.06076, over 16489.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2797, pruned_loss=0.05072, over 3213387.00 frames. ], batch size: 68, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:04:20,725 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.307e+02 1.908e+02 2.230e+02 2.489e+02 1.268e+03, threshold=4.461e+02, percent-clipped=3.0 2023-05-02 09:04:52,638 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9870, 5.3900, 5.5966, 5.2152, 5.3305, 5.9397, 5.3209, 5.0141], device='cuda:4'), covar=tensor([0.0887, 0.1768, 0.1789, 0.1972, 0.2416, 0.0775, 0.1348, 0.2274], device='cuda:4'), in_proj_covar=tensor([0.0431, 0.0637, 0.0698, 0.0519, 0.0692, 0.0728, 0.0544, 0.0693], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 09:05:15,491 INFO [train.py:904] (4/8) Epoch 27, batch 4600, loss[loss=0.2131, simple_loss=0.2922, pruned_loss=0.06701, over 11595.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2813, pruned_loss=0.05141, over 3215639.73 frames. ], batch size: 246, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:05:20,703 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6516, 1.8716, 2.2761, 2.6748, 2.5934, 2.9816, 2.0186, 2.8707], device='cuda:4'), covar=tensor([0.0230, 0.0551, 0.0337, 0.0319, 0.0353, 0.0196, 0.0574, 0.0165], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0199, 0.0187, 0.0194, 0.0209, 0.0167, 0.0205, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 09:05:26,986 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 09:06:23,484 INFO [train.py:904] (4/8) Epoch 27, batch 4650, loss[loss=0.1762, simple_loss=0.2667, pruned_loss=0.04289, over 16498.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2805, pruned_loss=0.05155, over 3226537.93 frames. ], batch size: 146, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:06:40,832 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.363e+02 1.843e+02 2.064e+02 2.500e+02 6.008e+02, threshold=4.128e+02, percent-clipped=2.0 2023-05-02 09:06:45,258 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=268568.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 09:07:32,852 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=268601.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:07:34,884 INFO [train.py:904] (4/8) Epoch 27, batch 4700, loss[loss=0.2009, simple_loss=0.2932, pruned_loss=0.05423, over 16985.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2776, pruned_loss=0.05009, over 3231468.53 frames. ], batch size: 41, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:08:15,780 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=268631.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 09:08:17,698 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8464, 4.9106, 4.7310, 4.3641, 4.3564, 4.8080, 4.5922, 4.5320], device='cuda:4'), covar=tensor([0.0614, 0.0589, 0.0327, 0.0321, 0.1116, 0.0637, 0.0462, 0.0665], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0465, 0.0362, 0.0368, 0.0364, 0.0420, 0.0248, 0.0434], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:08:45,873 INFO [train.py:904] (4/8) Epoch 27, batch 4750, loss[loss=0.1667, simple_loss=0.2536, pruned_loss=0.03987, over 16792.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2735, pruned_loss=0.04814, over 3241333.90 frames. ], batch size: 39, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:09:00,243 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=268662.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:09:04,566 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 1.855e+02 2.157e+02 2.478e+02 7.312e+02, threshold=4.313e+02, percent-clipped=3.0 2023-05-02 09:09:21,978 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9359, 4.8923, 4.7627, 3.7112, 4.8828, 1.7418, 4.5377, 4.3540], device='cuda:4'), covar=tensor([0.0125, 0.0141, 0.0202, 0.0692, 0.0118, 0.3329, 0.0174, 0.0365], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0175, 0.0212, 0.0187, 0.0189, 0.0219, 0.0202, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:09:59,302 INFO [train.py:904] (4/8) Epoch 27, batch 4800, loss[loss=0.1589, simple_loss=0.2556, pruned_loss=0.03112, over 16744.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2697, pruned_loss=0.04623, over 3226200.45 frames. ], batch size: 83, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:10:38,839 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0293, 4.0214, 3.9450, 3.0753, 3.8743, 1.7066, 3.6953, 3.4435], device='cuda:4'), covar=tensor([0.0161, 0.0165, 0.0187, 0.0474, 0.0118, 0.3251, 0.0167, 0.0334], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0175, 0.0212, 0.0187, 0.0188, 0.0219, 0.0202, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:10:53,654 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.92 vs. limit=5.0 2023-05-02 09:11:14,169 INFO [train.py:904] (4/8) Epoch 27, batch 4850, loss[loss=0.2008, simple_loss=0.2808, pruned_loss=0.0604, over 11832.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2702, pruned_loss=0.04549, over 3203325.33 frames. ], batch size: 246, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:11:22,751 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6486, 4.6938, 4.9838, 4.9438, 4.9654, 4.7138, 4.6307, 4.5254], device='cuda:4'), covar=tensor([0.0283, 0.0478, 0.0343, 0.0389, 0.0426, 0.0332, 0.0888, 0.0451], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0481, 0.0466, 0.0429, 0.0513, 0.0492, 0.0568, 0.0392], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 09:11:31,499 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.342e+02 1.812e+02 2.075e+02 2.444e+02 6.308e+02, threshold=4.151e+02, percent-clipped=1.0 2023-05-02 09:11:46,636 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0922, 4.8954, 5.1268, 5.3261, 5.5262, 4.9289, 5.4990, 5.5129], device='cuda:4'), covar=tensor([0.1933, 0.1388, 0.1702, 0.0707, 0.0497, 0.0829, 0.0531, 0.0641], device='cuda:4'), in_proj_covar=tensor([0.0664, 0.0814, 0.0947, 0.0829, 0.0627, 0.0658, 0.0687, 0.0799], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:12:27,703 INFO [train.py:904] (4/8) Epoch 27, batch 4900, loss[loss=0.1788, simple_loss=0.2685, pruned_loss=0.04459, over 11985.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2693, pruned_loss=0.04447, over 3188941.73 frames. ], batch size: 246, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:12:38,288 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5655, 5.5881, 5.4512, 5.0224, 5.0693, 5.4982, 5.3707, 5.2509], device='cuda:4'), covar=tensor([0.0661, 0.0482, 0.0316, 0.0308, 0.1113, 0.0470, 0.0284, 0.0630], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0462, 0.0359, 0.0365, 0.0360, 0.0417, 0.0247, 0.0430], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:12:46,341 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.04 vs. limit=5.0 2023-05-02 09:13:37,968 INFO [train.py:904] (4/8) Epoch 27, batch 4950, loss[loss=0.167, simple_loss=0.2625, pruned_loss=0.03576, over 17147.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2691, pruned_loss=0.04388, over 3198117.23 frames. ], batch size: 47, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:13:54,418 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 1.970e+02 2.304e+02 2.682e+02 4.795e+02, threshold=4.609e+02, percent-clipped=2.0 2023-05-02 09:13:59,072 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=268868.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 09:14:36,338 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=268894.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:14:50,662 INFO [train.py:904] (4/8) Epoch 27, batch 5000, loss[loss=0.2124, simple_loss=0.2937, pruned_loss=0.06557, over 12129.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2708, pruned_loss=0.04394, over 3199538.70 frames. ], batch size: 246, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:15:04,676 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8102, 2.1568, 2.4196, 3.0617, 2.1717, 2.2979, 2.2816, 2.2478], device='cuda:4'), covar=tensor([0.1595, 0.3331, 0.2540, 0.0769, 0.4041, 0.2503, 0.3350, 0.3107], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0472, 0.0383, 0.0336, 0.0445, 0.0539, 0.0442, 0.0553], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:15:09,314 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=268916.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:15:29,372 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=268931.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:15:31,465 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9638, 2.0474, 2.6079, 2.9672, 2.8573, 3.4330, 2.2206, 3.4099], device='cuda:4'), covar=tensor([0.0280, 0.0582, 0.0372, 0.0385, 0.0359, 0.0170, 0.0621, 0.0147], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0200, 0.0187, 0.0194, 0.0208, 0.0166, 0.0205, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:16:00,154 INFO [train.py:904] (4/8) Epoch 27, batch 5050, loss[loss=0.1773, simple_loss=0.2702, pruned_loss=0.04222, over 16507.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2724, pruned_loss=0.04417, over 3177141.57 frames. ], batch size: 68, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:16:03,776 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5455, 1.7221, 2.2005, 2.4778, 2.5189, 2.8121, 1.8308, 2.7391], device='cuda:4'), covar=tensor([0.0252, 0.0615, 0.0365, 0.0422, 0.0368, 0.0220, 0.0669, 0.0166], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0199, 0.0187, 0.0194, 0.0208, 0.0166, 0.0204, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:16:03,786 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=268955.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:16:05,781 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=268957.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:16:14,597 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-05-02 09:16:17,236 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.360e+02 1.970e+02 2.360e+02 2.844e+02 4.584e+02, threshold=4.719e+02, percent-clipped=0.0 2023-05-02 09:16:37,071 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=268979.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:16:54,168 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=268991.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:17:11,192 INFO [train.py:904] (4/8) Epoch 27, batch 5100, loss[loss=0.1562, simple_loss=0.2486, pruned_loss=0.03189, over 16822.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2707, pruned_loss=0.04373, over 3186759.66 frames. ], batch size: 83, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:17:13,388 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5848, 4.5668, 4.5282, 3.6126, 4.5535, 1.6030, 4.2673, 4.1510], device='cuda:4'), covar=tensor([0.0126, 0.0119, 0.0173, 0.0571, 0.0114, 0.3165, 0.0152, 0.0297], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0173, 0.0210, 0.0185, 0.0187, 0.0217, 0.0200, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:18:13,212 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.84 vs. limit=2.0 2023-05-02 09:18:22,989 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=269052.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:18:23,595 INFO [train.py:904] (4/8) Epoch 27, batch 5150, loss[loss=0.1862, simple_loss=0.2848, pruned_loss=0.04379, over 16386.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2702, pruned_loss=0.04283, over 3188241.98 frames. ], batch size: 146, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:18:41,483 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.484e+02 1.963e+02 2.152e+02 2.508e+02 3.926e+02, threshold=4.304e+02, percent-clipped=0.0 2023-05-02 09:18:58,309 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-02 09:19:26,982 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3107, 3.3387, 1.7669, 3.6242, 2.4745, 3.5817, 1.8898, 2.6749], device='cuda:4'), covar=tensor([0.0273, 0.0356, 0.1983, 0.0207, 0.0866, 0.0506, 0.2128, 0.0868], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0180, 0.0196, 0.0171, 0.0179, 0.0219, 0.0203, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 09:19:36,079 INFO [train.py:904] (4/8) Epoch 27, batch 5200, loss[loss=0.1544, simple_loss=0.2452, pruned_loss=0.03183, over 16823.00 frames. ], tot_loss[loss=0.1764, simple_loss=0.2685, pruned_loss=0.04217, over 3200005.28 frames. ], batch size: 83, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:20:20,490 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8641, 5.1416, 4.9126, 4.9596, 4.7168, 4.6878, 4.5014, 5.2039], device='cuda:4'), covar=tensor([0.1233, 0.0809, 0.0967, 0.0796, 0.0869, 0.1066, 0.1282, 0.0892], device='cuda:4'), in_proj_covar=tensor([0.0711, 0.0863, 0.0703, 0.0662, 0.0547, 0.0543, 0.0726, 0.0677], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:20:47,202 INFO [train.py:904] (4/8) Epoch 27, batch 5250, loss[loss=0.1768, simple_loss=0.2682, pruned_loss=0.04273, over 16216.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2659, pruned_loss=0.04177, over 3209356.42 frames. ], batch size: 165, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:21:04,386 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.499e+02 1.922e+02 2.287e+02 2.667e+02 4.356e+02, threshold=4.574e+02, percent-clipped=2.0 2023-05-02 09:22:00,557 INFO [train.py:904] (4/8) Epoch 27, batch 5300, loss[loss=0.1633, simple_loss=0.2529, pruned_loss=0.03683, over 16347.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2621, pruned_loss=0.04054, over 3210541.00 frames. ], batch size: 165, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:22:16,973 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9705, 2.2133, 2.3077, 2.6800, 1.8961, 3.2221, 1.7567, 2.6817], device='cuda:4'), covar=tensor([0.1182, 0.0740, 0.1080, 0.0176, 0.0105, 0.0337, 0.1556, 0.0758], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0181, 0.0201, 0.0201, 0.0208, 0.0219, 0.0210, 0.0200], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 09:22:55,455 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=269241.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:22:59,855 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-05-02 09:23:08,880 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=269250.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:23:12,961 INFO [train.py:904] (4/8) Epoch 27, batch 5350, loss[loss=0.161, simple_loss=0.2632, pruned_loss=0.02944, over 16839.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.261, pruned_loss=0.04013, over 3214235.21 frames. ], batch size: 96, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:23:18,868 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=269257.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:23:29,780 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.403e+02 2.050e+02 2.380e+02 2.740e+02 5.067e+02, threshold=4.761e+02, percent-clipped=1.0 2023-05-02 09:24:24,473 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=269302.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:24:25,134 INFO [train.py:904] (4/8) Epoch 27, batch 5400, loss[loss=0.1674, simple_loss=0.2661, pruned_loss=0.03436, over 16891.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2638, pruned_loss=0.04078, over 3211746.61 frames. ], batch size: 109, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:24:28,350 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=269305.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:24:36,387 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-02 09:25:05,635 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-05-02 09:25:31,283 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=269347.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:25:40,788 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=269352.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:25:42,376 INFO [train.py:904] (4/8) Epoch 27, batch 5450, loss[loss=0.1829, simple_loss=0.2759, pruned_loss=0.04496, over 16743.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2672, pruned_loss=0.04258, over 3203669.93 frames. ], batch size: 124, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:26:01,149 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.423e+02 2.011e+02 2.409e+02 3.005e+02 5.920e+02, threshold=4.819e+02, percent-clipped=1.0 2023-05-02 09:26:46,252 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3763, 3.3286, 3.3863, 3.4452, 3.5122, 3.2874, 3.4780, 3.5427], device='cuda:4'), covar=tensor([0.1190, 0.0922, 0.1055, 0.0685, 0.0672, 0.2368, 0.1186, 0.0882], device='cuda:4'), in_proj_covar=tensor([0.0664, 0.0813, 0.0946, 0.0827, 0.0627, 0.0658, 0.0684, 0.0799], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:27:00,964 INFO [train.py:904] (4/8) Epoch 27, batch 5500, loss[loss=0.2494, simple_loss=0.3197, pruned_loss=0.08953, over 11391.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2743, pruned_loss=0.04676, over 3170451.51 frames. ], batch size: 246, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:27:14,299 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3877, 2.9655, 2.6792, 2.3486, 2.3300, 2.3247, 2.9541, 2.8916], device='cuda:4'), covar=tensor([0.2277, 0.0590, 0.1419, 0.2438, 0.2228, 0.2119, 0.0491, 0.1190], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0276, 0.0313, 0.0326, 0.0305, 0.0275, 0.0305, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 09:27:16,671 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=269413.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:28:00,716 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=269441.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:28:19,438 INFO [train.py:904] (4/8) Epoch 27, batch 5550, loss[loss=0.193, simple_loss=0.2799, pruned_loss=0.05303, over 16394.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2803, pruned_loss=0.05093, over 3149108.68 frames. ], batch size: 146, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:28:38,497 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.260e+02 3.053e+02 3.491e+02 4.207e+02 9.161e+02, threshold=6.983e+02, percent-clipped=6.0 2023-05-02 09:29:33,081 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 09:29:36,057 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6930, 4.5210, 4.3898, 2.8406, 3.7950, 4.3969, 3.8388, 2.5765], device='cuda:4'), covar=tensor([0.0597, 0.0044, 0.0055, 0.0477, 0.0114, 0.0109, 0.0102, 0.0482], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0089, 0.0135, 0.0101, 0.0114, 0.0097, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 09:29:37,456 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=269502.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 09:29:38,092 INFO [train.py:904] (4/8) Epoch 27, batch 5600, loss[loss=0.2283, simple_loss=0.3045, pruned_loss=0.0761, over 17079.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2847, pruned_loss=0.05442, over 3120771.87 frames. ], batch size: 53, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:30:40,509 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4129, 1.7305, 2.1679, 2.3956, 2.4949, 2.7027, 1.8687, 2.5910], device='cuda:4'), covar=tensor([0.0248, 0.0564, 0.0310, 0.0361, 0.0338, 0.0223, 0.0596, 0.0164], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0198, 0.0185, 0.0192, 0.0206, 0.0165, 0.0203, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:30:46,870 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3306, 4.3204, 4.2283, 3.4104, 4.3000, 1.6705, 4.0849, 3.7706], device='cuda:4'), covar=tensor([0.0124, 0.0111, 0.0201, 0.0343, 0.0099, 0.2963, 0.0137, 0.0325], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0173, 0.0209, 0.0185, 0.0186, 0.0215, 0.0199, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:30:56,898 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=269550.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:31:01,866 INFO [train.py:904] (4/8) Epoch 27, batch 5650, loss[loss=0.2103, simple_loss=0.3035, pruned_loss=0.05854, over 17200.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2898, pruned_loss=0.05874, over 3085633.83 frames. ], batch size: 44, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:31:20,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.426e+02 3.478e+02 4.295e+02 5.160e+02 1.255e+03, threshold=8.591e+02, percent-clipped=5.0 2023-05-02 09:32:11,860 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=269597.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:32:13,054 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=269598.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:32:20,977 INFO [train.py:904] (4/8) Epoch 27, batch 5700, loss[loss=0.2232, simple_loss=0.3202, pruned_loss=0.06315, over 16389.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2916, pruned_loss=0.06042, over 3073847.25 frames. ], batch size: 35, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:32:51,166 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.50 vs. limit=5.0 2023-05-02 09:33:31,912 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=269647.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:33:41,283 INFO [train.py:904] (4/8) Epoch 27, batch 5750, loss[loss=0.2177, simple_loss=0.299, pruned_loss=0.06823, over 16877.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2944, pruned_loss=0.06225, over 3052766.20 frames. ], batch size: 109, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:33:59,183 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.119e+02 2.911e+02 3.444e+02 4.109e+02 8.393e+02, threshold=6.889e+02, percent-clipped=0.0 2023-05-02 09:34:17,097 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-05-02 09:34:24,875 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5062, 3.5800, 3.3234, 2.9329, 3.1820, 3.4766, 3.2943, 3.2844], device='cuda:4'), covar=tensor([0.0569, 0.0642, 0.0291, 0.0298, 0.0461, 0.0501, 0.1542, 0.0476], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0462, 0.0359, 0.0363, 0.0359, 0.0418, 0.0246, 0.0430], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:34:50,303 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=269695.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:35:02,096 INFO [train.py:904] (4/8) Epoch 27, batch 5800, loss[loss=0.1823, simple_loss=0.2743, pruned_loss=0.04519, over 17058.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2937, pruned_loss=0.06114, over 3044802.59 frames. ], batch size: 55, lr: 2.49e-03, grad_scale: 4.0 2023-05-02 09:35:11,564 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=269708.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:35:15,054 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-05-02 09:36:21,257 INFO [train.py:904] (4/8) Epoch 27, batch 5850, loss[loss=0.2219, simple_loss=0.3007, pruned_loss=0.0716, over 16713.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2913, pruned_loss=0.05979, over 3050966.91 frames. ], batch size: 62, lr: 2.49e-03, grad_scale: 4.0 2023-05-02 09:36:40,939 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.068e+02 3.085e+02 3.568e+02 4.510e+02 7.349e+02, threshold=7.136e+02, percent-clipped=2.0 2023-05-02 09:36:53,558 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=269773.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:37:34,885 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=269797.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 09:37:45,800 INFO [train.py:904] (4/8) Epoch 27, batch 5900, loss[loss=0.2386, simple_loss=0.3112, pruned_loss=0.08301, over 11288.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2913, pruned_loss=0.05951, over 3071077.64 frames. ], batch size: 248, lr: 2.49e-03, grad_scale: 4.0 2023-05-02 09:38:38,117 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=269834.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 09:38:40,690 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=269836.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:39:06,822 INFO [train.py:904] (4/8) Epoch 27, batch 5950, loss[loss=0.2037, simple_loss=0.2966, pruned_loss=0.05543, over 16191.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2919, pruned_loss=0.05801, over 3100655.65 frames. ], batch size: 35, lr: 2.49e-03, grad_scale: 4.0 2023-05-02 09:39:23,479 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3600, 3.3975, 2.0127, 3.7509, 2.5559, 3.7541, 2.1399, 2.7039], device='cuda:4'), covar=tensor([0.0319, 0.0415, 0.1774, 0.0277, 0.0906, 0.0622, 0.1646, 0.0910], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0180, 0.0196, 0.0171, 0.0179, 0.0219, 0.0202, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 09:39:27,599 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.538e+02 2.704e+02 3.273e+02 4.077e+02 6.463e+02, threshold=6.547e+02, percent-clipped=0.0 2023-05-02 09:39:30,421 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3581, 2.4775, 2.4520, 4.0011, 2.2683, 2.8330, 2.5119, 2.5933], device='cuda:4'), covar=tensor([0.1392, 0.3409, 0.2914, 0.0586, 0.4241, 0.2378, 0.3500, 0.3167], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0466, 0.0378, 0.0332, 0.0440, 0.0534, 0.0437, 0.0545], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:39:41,563 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7997, 1.8947, 2.3812, 2.7042, 2.6244, 3.1487, 1.9915, 3.1398], device='cuda:4'), covar=tensor([0.0252, 0.0625, 0.0411, 0.0388, 0.0389, 0.0220, 0.0675, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0197, 0.0185, 0.0191, 0.0205, 0.0164, 0.0202, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:40:17,556 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=269897.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:40:17,591 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=269897.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:40:24,914 INFO [train.py:904] (4/8) Epoch 27, batch 6000, loss[loss=0.1784, simple_loss=0.2671, pruned_loss=0.04486, over 16750.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.291, pruned_loss=0.05758, over 3118390.22 frames. ], batch size: 124, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:40:24,914 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 09:40:35,107 INFO [train.py:938] (4/8) Epoch 27, validation: loss=0.148, simple_loss=0.2603, pruned_loss=0.01783, over 944034.00 frames. 2023-05-02 09:40:35,108 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 09:41:15,678 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2388, 4.3321, 4.4828, 4.2748, 4.3866, 4.8507, 4.3808, 4.1246], device='cuda:4'), covar=tensor([0.1702, 0.2020, 0.2288, 0.1971, 0.2456, 0.1075, 0.1705, 0.2397], device='cuda:4'), in_proj_covar=tensor([0.0424, 0.0627, 0.0688, 0.0510, 0.0682, 0.0715, 0.0538, 0.0685], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 09:41:37,530 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=269945.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:41:52,113 INFO [train.py:904] (4/8) Epoch 27, batch 6050, loss[loss=0.2075, simple_loss=0.2988, pruned_loss=0.0581, over 15427.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2897, pruned_loss=0.05694, over 3111203.26 frames. ], batch size: 190, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:41:54,351 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8763, 5.0305, 4.7912, 4.3730, 4.2591, 4.8894, 4.8455, 4.4888], device='cuda:4'), covar=tensor([0.1166, 0.1584, 0.0576, 0.0637, 0.1392, 0.1218, 0.0806, 0.1312], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0461, 0.0358, 0.0362, 0.0358, 0.0416, 0.0246, 0.0429], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:42:12,241 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.929e+02 2.707e+02 3.068e+02 3.739e+02 6.756e+02, threshold=6.136e+02, percent-clipped=2.0 2023-05-02 09:42:35,796 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6755, 4.4714, 4.6917, 4.8543, 5.0191, 4.5709, 5.0056, 5.0204], device='cuda:4'), covar=tensor([0.1741, 0.1317, 0.1522, 0.0750, 0.0645, 0.1006, 0.0689, 0.0707], device='cuda:4'), in_proj_covar=tensor([0.0662, 0.0815, 0.0944, 0.0824, 0.0627, 0.0658, 0.0684, 0.0799], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:43:13,684 INFO [train.py:904] (4/8) Epoch 27, batch 6100, loss[loss=0.1886, simple_loss=0.2824, pruned_loss=0.04737, over 16790.00 frames. ], tot_loss[loss=0.1994, simple_loss=0.2884, pruned_loss=0.0552, over 3133634.03 frames. ], batch size: 124, lr: 2.49e-03, grad_scale: 8.0 2023-05-02 09:43:22,363 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=270008.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:43:42,148 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9731, 3.1896, 3.4260, 2.1595, 2.9343, 2.1656, 3.4702, 3.5064], device='cuda:4'), covar=tensor([0.0259, 0.0855, 0.0611, 0.2096, 0.0856, 0.1053, 0.0648, 0.0898], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0171, 0.0170, 0.0157, 0.0149, 0.0133, 0.0147, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 09:44:33,190 INFO [train.py:904] (4/8) Epoch 27, batch 6150, loss[loss=0.185, simple_loss=0.2729, pruned_loss=0.04855, over 16687.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.2859, pruned_loss=0.05426, over 3147004.79 frames. ], batch size: 134, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:44:36,464 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5922, 1.8231, 2.2036, 2.5225, 2.4901, 2.8318, 1.8880, 2.7509], device='cuda:4'), covar=tensor([0.0249, 0.0548, 0.0356, 0.0342, 0.0369, 0.0208, 0.0603, 0.0143], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0197, 0.0185, 0.0190, 0.0206, 0.0164, 0.0202, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:44:37,805 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=270056.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:44:53,892 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.916e+02 2.602e+02 3.189e+02 4.066e+02 6.857e+02, threshold=6.378e+02, percent-clipped=4.0 2023-05-02 09:45:40,116 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-02 09:45:42,210 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=270097.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 09:45:51,121 INFO [train.py:904] (4/8) Epoch 27, batch 6200, loss[loss=0.2186, simple_loss=0.2849, pruned_loss=0.07618, over 11157.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.2846, pruned_loss=0.05458, over 3130041.91 frames. ], batch size: 248, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:45:54,416 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-05-02 09:46:33,638 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=270129.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 09:46:52,596 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3348, 5.6640, 5.4180, 5.4344, 5.1336, 5.0980, 5.0412, 5.7576], device='cuda:4'), covar=tensor([0.1286, 0.0893, 0.0986, 0.0842, 0.0803, 0.0772, 0.1287, 0.0840], device='cuda:4'), in_proj_covar=tensor([0.0706, 0.0855, 0.0697, 0.0658, 0.0539, 0.0538, 0.0718, 0.0669], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:46:57,925 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=270145.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:47:09,692 INFO [train.py:904] (4/8) Epoch 27, batch 6250, loss[loss=0.215, simple_loss=0.2882, pruned_loss=0.07091, over 11467.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.284, pruned_loss=0.05452, over 3117307.08 frames. ], batch size: 248, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:47:15,260 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-05-02 09:47:29,202 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.849e+02 2.710e+02 3.440e+02 4.218e+02 8.566e+02, threshold=6.880e+02, percent-clipped=4.0 2023-05-02 09:48:08,658 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=270192.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:48:25,165 INFO [train.py:904] (4/8) Epoch 27, batch 6300, loss[loss=0.1786, simple_loss=0.273, pruned_loss=0.04217, over 16544.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2837, pruned_loss=0.05342, over 3136964.51 frames. ], batch size: 75, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:49:18,775 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1122, 2.1042, 2.6401, 3.0718, 2.9081, 3.5319, 2.2624, 3.5211], device='cuda:4'), covar=tensor([0.0260, 0.0572, 0.0385, 0.0342, 0.0371, 0.0199, 0.0626, 0.0161], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0197, 0.0185, 0.0191, 0.0206, 0.0165, 0.0203, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:49:45,985 INFO [train.py:904] (4/8) Epoch 27, batch 6350, loss[loss=0.1933, simple_loss=0.2767, pruned_loss=0.05497, over 16606.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.285, pruned_loss=0.05479, over 3115584.34 frames. ], batch size: 57, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:49:50,879 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7227, 1.7739, 2.2585, 2.6428, 2.5887, 2.9691, 1.8943, 2.9868], device='cuda:4'), covar=tensor([0.0243, 0.0646, 0.0414, 0.0373, 0.0379, 0.0231, 0.0716, 0.0169], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0198, 0.0185, 0.0191, 0.0206, 0.0165, 0.0203, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:50:05,608 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.880e+02 2.767e+02 3.274e+02 3.776e+02 6.685e+02, threshold=6.549e+02, percent-clipped=0.0 2023-05-02 09:50:19,898 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2391, 4.2102, 4.1318, 3.2825, 4.1908, 1.7654, 3.9927, 3.6798], device='cuda:4'), covar=tensor([0.0123, 0.0123, 0.0197, 0.0330, 0.0096, 0.2962, 0.0130, 0.0300], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0173, 0.0210, 0.0185, 0.0185, 0.0215, 0.0199, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:50:28,138 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=270280.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 09:51:02,482 INFO [train.py:904] (4/8) Epoch 27, batch 6400, loss[loss=0.1686, simple_loss=0.2642, pruned_loss=0.03648, over 16792.00 frames. ], tot_loss[loss=0.1979, simple_loss=0.2848, pruned_loss=0.05549, over 3110541.12 frames. ], batch size: 102, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:51:08,861 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7375, 2.8946, 2.5960, 4.3884, 3.1472, 4.0389, 1.4923, 2.9767], device='cuda:4'), covar=tensor([0.1405, 0.0757, 0.1230, 0.0183, 0.0272, 0.0408, 0.1842, 0.0836], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0180, 0.0201, 0.0202, 0.0208, 0.0219, 0.0210, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 09:51:29,411 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1672, 2.3605, 2.2843, 3.8717, 2.2255, 2.7214, 2.3722, 2.5019], device='cuda:4'), covar=tensor([0.1516, 0.3636, 0.3195, 0.0569, 0.4166, 0.2510, 0.3919, 0.3198], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0467, 0.0379, 0.0332, 0.0441, 0.0535, 0.0438, 0.0547], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:52:01,055 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=270341.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 09:52:19,089 INFO [train.py:904] (4/8) Epoch 27, batch 6450, loss[loss=0.1852, simple_loss=0.2703, pruned_loss=0.05004, over 15374.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2849, pruned_loss=0.05494, over 3105633.48 frames. ], batch size: 190, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:52:39,047 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.936e+02 2.688e+02 3.236e+02 4.371e+02 9.664e+02, threshold=6.472e+02, percent-clipped=9.0 2023-05-02 09:53:27,570 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=270396.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:53:37,244 INFO [train.py:904] (4/8) Epoch 27, batch 6500, loss[loss=0.2297, simple_loss=0.3005, pruned_loss=0.07947, over 11805.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2833, pruned_loss=0.05427, over 3118331.01 frames. ], batch size: 247, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:54:12,483 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9418, 4.1840, 4.0240, 4.0521, 3.7693, 3.7914, 3.8336, 4.1793], device='cuda:4'), covar=tensor([0.1166, 0.0869, 0.1001, 0.0851, 0.0812, 0.1805, 0.1008, 0.0984], device='cuda:4'), in_proj_covar=tensor([0.0706, 0.0853, 0.0696, 0.0658, 0.0539, 0.0538, 0.0717, 0.0666], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:54:17,014 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=270429.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 09:54:33,593 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=270439.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:54:42,986 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8045, 1.4423, 1.7370, 1.6629, 1.7766, 1.8776, 1.6623, 1.7541], device='cuda:4'), covar=tensor([0.0269, 0.0420, 0.0225, 0.0316, 0.0277, 0.0191, 0.0443, 0.0149], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0198, 0.0185, 0.0191, 0.0206, 0.0165, 0.0203, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 09:54:58,115 INFO [train.py:904] (4/8) Epoch 27, batch 6550, loss[loss=0.1953, simple_loss=0.2968, pruned_loss=0.04691, over 16643.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2864, pruned_loss=0.05517, over 3116973.60 frames. ], batch size: 134, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:55:04,745 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=270457.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:55:17,872 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.942e+02 2.679e+02 3.213e+02 3.936e+02 7.742e+02, threshold=6.427e+02, percent-clipped=6.0 2023-05-02 09:55:34,671 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=270477.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:55:58,276 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=270492.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:56:10,607 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=270500.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:56:14,142 INFO [train.py:904] (4/8) Epoch 27, batch 6600, loss[loss=0.2442, simple_loss=0.3149, pruned_loss=0.08674, over 11600.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2882, pruned_loss=0.05551, over 3109636.10 frames. ], batch size: 248, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:57:09,703 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=270540.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:57:18,680 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-05-02 09:57:30,476 INFO [train.py:904] (4/8) Epoch 27, batch 6650, loss[loss=0.1903, simple_loss=0.2859, pruned_loss=0.04732, over 16819.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2881, pruned_loss=0.05617, over 3109551.27 frames. ], batch size: 102, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:57:50,322 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.544e+02 2.739e+02 3.273e+02 3.972e+02 7.700e+02, threshold=6.545e+02, percent-clipped=3.0 2023-05-02 09:58:46,109 INFO [train.py:904] (4/8) Epoch 27, batch 6700, loss[loss=0.192, simple_loss=0.2783, pruned_loss=0.05287, over 16731.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2879, pruned_loss=0.05695, over 3114113.48 frames. ], batch size: 124, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 09:59:01,553 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=270613.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 09:59:35,300 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=270636.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 10:00:01,086 INFO [train.py:904] (4/8) Epoch 27, batch 6750, loss[loss=0.189, simple_loss=0.2894, pruned_loss=0.04431, over 16814.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.2874, pruned_loss=0.05749, over 3096124.03 frames. ], batch size: 102, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:00:11,608 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6244, 1.7805, 1.6845, 1.5324, 1.9638, 1.5525, 1.6402, 1.9230], device='cuda:4'), covar=tensor([0.0212, 0.0314, 0.0424, 0.0362, 0.0213, 0.0268, 0.0199, 0.0222], device='cuda:4'), in_proj_covar=tensor([0.0223, 0.0238, 0.0228, 0.0230, 0.0240, 0.0238, 0.0237, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:00:20,167 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.179e+02 2.756e+02 3.273e+02 4.054e+02 1.247e+03, threshold=6.545e+02, percent-clipped=2.0 2023-05-02 10:00:32,946 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=270674.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:01:15,215 INFO [train.py:904] (4/8) Epoch 27, batch 6800, loss[loss=0.1946, simple_loss=0.2895, pruned_loss=0.0499, over 16897.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2872, pruned_loss=0.05702, over 3111068.16 frames. ], batch size: 96, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:01:53,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6390, 2.5669, 1.9427, 2.7357, 2.1629, 2.8012, 2.1717, 2.4026], device='cuda:4'), covar=tensor([0.0333, 0.0361, 0.1274, 0.0273, 0.0708, 0.0480, 0.1290, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0180, 0.0195, 0.0171, 0.0179, 0.0219, 0.0202, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 10:01:53,384 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=270726.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:02:33,337 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=270752.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:02:34,157 INFO [train.py:904] (4/8) Epoch 27, batch 6850, loss[loss=0.2359, simple_loss=0.3041, pruned_loss=0.08379, over 11610.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2884, pruned_loss=0.05756, over 3101964.64 frames. ], batch size: 247, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:02:53,223 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.651e+02 2.704e+02 3.220e+02 3.695e+02 5.610e+02, threshold=6.440e+02, percent-clipped=0.0 2023-05-02 10:03:25,311 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=270787.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:03:26,463 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7393, 1.7867, 2.2455, 2.5826, 2.5373, 2.8694, 1.9011, 2.9468], device='cuda:4'), covar=tensor([0.0220, 0.0639, 0.0413, 0.0400, 0.0400, 0.0253, 0.0715, 0.0161], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0197, 0.0185, 0.0190, 0.0206, 0.0165, 0.0202, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:03:37,581 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=270795.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:03:49,486 INFO [train.py:904] (4/8) Epoch 27, batch 6900, loss[loss=0.1968, simple_loss=0.2904, pruned_loss=0.05164, over 16191.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2902, pruned_loss=0.05591, over 3132283.08 frames. ], batch size: 165, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:04:59,212 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4505, 3.4979, 2.8079, 2.1896, 2.3034, 2.3922, 3.6430, 3.1204], device='cuda:4'), covar=tensor([0.3139, 0.0609, 0.1802, 0.2930, 0.2681, 0.2269, 0.0507, 0.1392], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0276, 0.0313, 0.0328, 0.0306, 0.0276, 0.0305, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 10:05:07,698 INFO [train.py:904] (4/8) Epoch 27, batch 6950, loss[loss=0.2071, simple_loss=0.2939, pruned_loss=0.0601, over 16267.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2914, pruned_loss=0.05733, over 3130297.75 frames. ], batch size: 165, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:05:28,446 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.875e+02 2.915e+02 3.513e+02 4.432e+02 8.204e+02, threshold=7.026e+02, percent-clipped=5.0 2023-05-02 10:05:31,802 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2367, 5.2186, 5.0169, 4.2733, 5.1164, 1.8422, 4.8212, 4.6569], device='cuda:4'), covar=tensor([0.0119, 0.0115, 0.0213, 0.0437, 0.0102, 0.2959, 0.0152, 0.0283], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0173, 0.0211, 0.0185, 0.0186, 0.0216, 0.0199, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:05:58,213 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 10:05:59,367 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.86 vs. limit=5.0 2023-05-02 10:06:05,077 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.32 vs. limit=5.0 2023-05-02 10:06:08,463 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6300, 4.7116, 4.4943, 4.1518, 4.1445, 4.5917, 4.4100, 4.2858], device='cuda:4'), covar=tensor([0.0840, 0.1110, 0.0418, 0.0460, 0.1074, 0.0851, 0.0766, 0.0998], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0461, 0.0357, 0.0361, 0.0357, 0.0415, 0.0245, 0.0429], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:06:23,848 INFO [train.py:904] (4/8) Epoch 27, batch 7000, loss[loss=0.2114, simple_loss=0.2848, pruned_loss=0.06902, over 11752.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.291, pruned_loss=0.05702, over 3107848.68 frames. ], batch size: 247, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:06:46,508 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=270917.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:06:58,421 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 10:07:15,590 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=270936.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 10:07:40,697 INFO [train.py:904] (4/8) Epoch 27, batch 7050, loss[loss=0.197, simple_loss=0.2877, pruned_loss=0.05311, over 16895.00 frames. ], tot_loss[loss=0.2019, simple_loss=0.2913, pruned_loss=0.05629, over 3116874.99 frames. ], batch size: 109, lr: 2.48e-03, grad_scale: 4.0 2023-05-02 10:08:01,040 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.024e+02 2.738e+02 3.222e+02 3.975e+02 6.307e+02, threshold=6.443e+02, percent-clipped=0.0 2023-05-02 10:08:04,268 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-05-02 10:08:05,077 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=270969.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:08:18,503 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=270978.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:08:27,529 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=270984.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 10:08:55,738 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.65 vs. limit=5.0 2023-05-02 10:08:58,235 INFO [train.py:904] (4/8) Epoch 27, batch 7100, loss[loss=0.2234, simple_loss=0.3001, pruned_loss=0.07328, over 11312.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2902, pruned_loss=0.05649, over 3102453.11 frames. ], batch size: 248, lr: 2.48e-03, grad_scale: 4.0 2023-05-02 10:10:15,296 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=271052.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:10:16,497 INFO [train.py:904] (4/8) Epoch 27, batch 7150, loss[loss=0.226, simple_loss=0.2954, pruned_loss=0.07829, over 11318.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2884, pruned_loss=0.05631, over 3115799.44 frames. ], batch size: 247, lr: 2.48e-03, grad_scale: 4.0 2023-05-02 10:10:18,565 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9321, 3.1737, 3.4835, 2.0167, 2.9559, 2.2308, 3.4071, 3.4946], device='cuda:4'), covar=tensor([0.0296, 0.0857, 0.0586, 0.2236, 0.0865, 0.1051, 0.0688, 0.0967], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0172, 0.0171, 0.0157, 0.0149, 0.0133, 0.0147, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 10:10:37,964 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.119e+02 2.616e+02 3.372e+02 3.961e+02 8.125e+02, threshold=6.744e+02, percent-clipped=3.0 2023-05-02 10:10:59,691 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=271082.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:11:18,924 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=271095.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:11:28,070 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=271100.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:11:31,639 INFO [train.py:904] (4/8) Epoch 27, batch 7200, loss[loss=0.1623, simple_loss=0.2689, pruned_loss=0.0278, over 16876.00 frames. ], tot_loss[loss=0.199, simple_loss=0.287, pruned_loss=0.05549, over 3090809.96 frames. ], batch size: 96, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:11:42,792 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3227, 5.1407, 5.3594, 5.5307, 5.7252, 5.0779, 5.6775, 5.7151], device='cuda:4'), covar=tensor([0.2056, 0.1295, 0.1517, 0.0645, 0.0498, 0.0868, 0.0536, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0660, 0.0806, 0.0939, 0.0816, 0.0624, 0.0653, 0.0683, 0.0794], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:12:37,999 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=271143.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:12:43,379 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0417, 3.0893, 2.7549, 2.8532, 3.3739, 2.9835, 3.5385, 3.5656], device='cuda:4'), covar=tensor([0.0104, 0.0426, 0.0534, 0.0449, 0.0266, 0.0408, 0.0279, 0.0275], device='cuda:4'), in_proj_covar=tensor([0.0224, 0.0241, 0.0230, 0.0232, 0.0242, 0.0240, 0.0239, 0.0239], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:12:53,488 INFO [train.py:904] (4/8) Epoch 27, batch 7250, loss[loss=0.1741, simple_loss=0.2659, pruned_loss=0.0411, over 16826.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2846, pruned_loss=0.05442, over 3079439.48 frames. ], batch size: 102, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:13:16,038 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.855e+02 2.666e+02 3.085e+02 3.699e+02 8.603e+02, threshold=6.169e+02, percent-clipped=3.0 2023-05-02 10:13:17,272 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8166, 2.6948, 2.5587, 1.9151, 2.5717, 2.6625, 2.5498, 1.9380], device='cuda:4'), covar=tensor([0.0466, 0.0107, 0.0108, 0.0404, 0.0158, 0.0169, 0.0152, 0.0415], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0089, 0.0090, 0.0136, 0.0101, 0.0115, 0.0098, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 10:14:10,773 INFO [train.py:904] (4/8) Epoch 27, batch 7300, loss[loss=0.1795, simple_loss=0.2709, pruned_loss=0.04407, over 17270.00 frames. ], tot_loss[loss=0.196, simple_loss=0.2837, pruned_loss=0.05413, over 3083177.45 frames. ], batch size: 52, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:14:18,943 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9329, 2.7518, 2.6271, 4.8711, 3.4634, 3.9664, 1.8011, 2.9083], device='cuda:4'), covar=tensor([0.1379, 0.0982, 0.1441, 0.0211, 0.0478, 0.0586, 0.1716, 0.1061], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0179, 0.0199, 0.0200, 0.0207, 0.0218, 0.0208, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 10:15:05,550 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9154, 3.2293, 3.1299, 2.0615, 2.9877, 3.2014, 2.9831, 1.8834], device='cuda:4'), covar=tensor([0.0603, 0.0080, 0.0096, 0.0506, 0.0135, 0.0141, 0.0146, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0089, 0.0090, 0.0136, 0.0101, 0.0115, 0.0098, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 10:15:29,743 INFO [train.py:904] (4/8) Epoch 27, batch 7350, loss[loss=0.1803, simple_loss=0.271, pruned_loss=0.04486, over 16692.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2846, pruned_loss=0.05472, over 3081415.50 frames. ], batch size: 83, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:15:42,267 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0805, 4.0684, 3.9544, 3.1088, 4.0120, 1.8342, 3.8040, 3.4237], device='cuda:4'), covar=tensor([0.0138, 0.0116, 0.0202, 0.0332, 0.0096, 0.2998, 0.0135, 0.0313], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0170, 0.0209, 0.0183, 0.0183, 0.0213, 0.0197, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:15:42,284 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=271261.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:15:50,537 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.854e+02 2.875e+02 3.303e+02 4.103e+02 7.171e+02, threshold=6.606e+02, percent-clipped=3.0 2023-05-02 10:15:54,731 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=271269.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:16:01,242 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=271273.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:16:39,430 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-05-02 10:16:48,723 INFO [train.py:904] (4/8) Epoch 27, batch 7400, loss[loss=0.2044, simple_loss=0.2933, pruned_loss=0.05774, over 16782.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2862, pruned_loss=0.05551, over 3084524.83 frames. ], batch size: 76, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:16:58,698 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4406, 4.4441, 4.3005, 3.5422, 4.3872, 1.6801, 4.1024, 3.8849], device='cuda:4'), covar=tensor([0.0129, 0.0119, 0.0216, 0.0392, 0.0107, 0.3082, 0.0154, 0.0310], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0170, 0.0209, 0.0183, 0.0184, 0.0214, 0.0196, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:17:11,160 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=271317.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:17:18,925 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9918, 2.0329, 2.6112, 2.9648, 2.8700, 3.4031, 2.1782, 3.3673], device='cuda:4'), covar=tensor([0.0239, 0.0587, 0.0352, 0.0364, 0.0306, 0.0177, 0.0608, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0195, 0.0183, 0.0188, 0.0203, 0.0163, 0.0200, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:17:18,934 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=271322.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:17:28,769 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5208, 3.5778, 3.3616, 3.0344, 3.2141, 3.5064, 3.2992, 3.3546], device='cuda:4'), covar=tensor([0.0601, 0.0826, 0.0321, 0.0296, 0.0505, 0.0573, 0.1526, 0.0512], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0455, 0.0354, 0.0356, 0.0352, 0.0409, 0.0243, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:18:06,473 INFO [train.py:904] (4/8) Epoch 27, batch 7450, loss[loss=0.2101, simple_loss=0.2961, pruned_loss=0.06211, over 16862.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2872, pruned_loss=0.05637, over 3096022.16 frames. ], batch size: 42, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:18:30,891 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.940e+02 2.819e+02 3.560e+02 4.384e+02 9.484e+02, threshold=7.119e+02, percent-clipped=1.0 2023-05-02 10:18:45,269 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=271375.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:18:55,894 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=271382.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:19:30,450 INFO [train.py:904] (4/8) Epoch 27, batch 7500, loss[loss=0.1798, simple_loss=0.2723, pruned_loss=0.04368, over 16486.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2866, pruned_loss=0.05529, over 3095341.20 frames. ], batch size: 68, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:19:34,025 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7602, 4.9943, 5.1513, 4.9044, 4.9490, 5.5001, 4.9851, 4.7371], device='cuda:4'), covar=tensor([0.1074, 0.1838, 0.2276, 0.1935, 0.2359, 0.0999, 0.1602, 0.2365], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0632, 0.0694, 0.0514, 0.0687, 0.0724, 0.0545, 0.0692], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 10:19:35,547 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-05-02 10:20:08,115 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=271426.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:20:14,059 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=271430.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:20:22,931 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=271436.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:20:49,667 INFO [train.py:904] (4/8) Epoch 27, batch 7550, loss[loss=0.2363, simple_loss=0.3018, pruned_loss=0.08546, over 11255.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2862, pruned_loss=0.0561, over 3092064.24 frames. ], batch size: 246, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:21:11,189 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.886e+02 2.650e+02 3.300e+02 4.210e+02 7.541e+02, threshold=6.599e+02, percent-clipped=2.0 2023-05-02 10:21:41,027 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=271487.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:22:05,495 INFO [train.py:904] (4/8) Epoch 27, batch 7600, loss[loss=0.1878, simple_loss=0.2861, pruned_loss=0.04478, over 16394.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2849, pruned_loss=0.05522, over 3116437.93 frames. ], batch size: 146, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:23:22,937 INFO [train.py:904] (4/8) Epoch 27, batch 7650, loss[loss=0.1907, simple_loss=0.2861, pruned_loss=0.04767, over 16177.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2855, pruned_loss=0.05607, over 3103524.10 frames. ], batch size: 165, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:23:45,488 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.699e+02 2.765e+02 3.286e+02 4.189e+02 6.927e+02, threshold=6.573e+02, percent-clipped=1.0 2023-05-02 10:23:55,622 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=271573.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:24:43,692 INFO [train.py:904] (4/8) Epoch 27, batch 7700, loss[loss=0.1916, simple_loss=0.2772, pruned_loss=0.05304, over 16477.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2858, pruned_loss=0.05689, over 3082302.29 frames. ], batch size: 146, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:25:06,078 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=271617.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:25:12,495 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=271621.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:26:02,413 INFO [train.py:904] (4/8) Epoch 27, batch 7750, loss[loss=0.181, simple_loss=0.2736, pruned_loss=0.04416, over 16741.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.286, pruned_loss=0.05683, over 3078171.40 frames. ], batch size: 89, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:26:24,501 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.915e+02 2.896e+02 3.348e+02 4.032e+02 8.871e+02, threshold=6.696e+02, percent-clipped=2.0 2023-05-02 10:27:20,090 INFO [train.py:904] (4/8) Epoch 27, batch 7800, loss[loss=0.2116, simple_loss=0.3004, pruned_loss=0.06144, over 16603.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2875, pruned_loss=0.05824, over 3060113.82 frames. ], batch size: 62, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:28:04,771 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=271731.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:28:36,664 INFO [train.py:904] (4/8) Epoch 27, batch 7850, loss[loss=0.2202, simple_loss=0.3056, pruned_loss=0.06745, over 15424.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2886, pruned_loss=0.05802, over 3069843.75 frames. ], batch size: 191, lr: 2.48e-03, grad_scale: 4.0 2023-05-02 10:28:57,993 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.815e+02 2.838e+02 3.379e+02 4.179e+02 9.334e+02, threshold=6.757e+02, percent-clipped=5.0 2023-05-02 10:29:20,881 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=271782.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:29:52,539 INFO [train.py:904] (4/8) Epoch 27, batch 7900, loss[loss=0.1861, simple_loss=0.2798, pruned_loss=0.0462, over 16729.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2876, pruned_loss=0.05733, over 3082057.49 frames. ], batch size: 89, lr: 2.48e-03, grad_scale: 4.0 2023-05-02 10:31:12,165 INFO [train.py:904] (4/8) Epoch 27, batch 7950, loss[loss=0.2384, simple_loss=0.309, pruned_loss=0.08385, over 11785.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2877, pruned_loss=0.05706, over 3090992.35 frames. ], batch size: 246, lr: 2.48e-03, grad_scale: 4.0 2023-05-02 10:31:15,288 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=271854.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:31:36,730 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.948e+02 2.764e+02 3.233e+02 3.778e+02 5.723e+02, threshold=6.466e+02, percent-clipped=0.0 2023-05-02 10:32:02,786 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-05-02 10:32:21,164 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-05-02 10:32:29,919 INFO [train.py:904] (4/8) Epoch 27, batch 8000, loss[loss=0.1941, simple_loss=0.2854, pruned_loss=0.05141, over 16694.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2872, pruned_loss=0.05693, over 3098999.34 frames. ], batch size: 89, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:32:49,709 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=271915.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:32:52,894 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=271917.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:33:19,530 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=271934.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:33:47,183 INFO [train.py:904] (4/8) Epoch 27, batch 8050, loss[loss=0.2162, simple_loss=0.3046, pruned_loss=0.06394, over 16693.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2877, pruned_loss=0.05716, over 3089342.58 frames. ], batch size: 134, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:34:05,653 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=271965.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:34:09,131 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.787e+02 2.577e+02 3.144e+02 3.701e+02 6.468e+02, threshold=6.287e+02, percent-clipped=1.0 2023-05-02 10:34:41,542 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5556, 3.6329, 3.3525, 3.0065, 3.2052, 3.5318, 3.3364, 3.3106], device='cuda:4'), covar=tensor([0.0598, 0.0714, 0.0304, 0.0280, 0.0526, 0.0502, 0.1210, 0.0510], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0451, 0.0351, 0.0353, 0.0351, 0.0406, 0.0241, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:34:49,598 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=271995.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:35:01,726 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1701, 4.2700, 4.5766, 4.5003, 4.5536, 4.2709, 4.1768, 4.1657], device='cuda:4'), covar=tensor([0.0500, 0.0738, 0.0423, 0.0561, 0.0616, 0.0553, 0.1303, 0.0627], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0480, 0.0464, 0.0426, 0.0512, 0.0490, 0.0563, 0.0392], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 10:35:04,671 INFO [train.py:904] (4/8) Epoch 27, batch 8100, loss[loss=0.2043, simple_loss=0.2925, pruned_loss=0.0581, over 16821.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2871, pruned_loss=0.05659, over 3098389.65 frames. ], batch size: 102, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:35:45,771 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=272031.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:36:17,758 INFO [train.py:904] (4/8) Epoch 27, batch 8150, loss[loss=0.1953, simple_loss=0.2766, pruned_loss=0.05705, over 15404.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2846, pruned_loss=0.05587, over 3088130.65 frames. ], batch size: 190, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:36:39,765 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.895e+02 2.743e+02 3.279e+02 3.939e+02 6.192e+02, threshold=6.559e+02, percent-clipped=0.0 2023-05-02 10:36:57,159 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=272079.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:37:01,603 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=272082.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:37:32,133 INFO [train.py:904] (4/8) Epoch 27, batch 8200, loss[loss=0.2088, simple_loss=0.2845, pruned_loss=0.06652, over 11766.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.2828, pruned_loss=0.05547, over 3080977.48 frames. ], batch size: 248, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:38:12,508 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=272127.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:38:16,441 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=272130.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:38:34,487 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7856, 4.8493, 4.6532, 4.2706, 4.3283, 4.7624, 4.5769, 4.4133], device='cuda:4'), covar=tensor([0.0570, 0.0501, 0.0345, 0.0361, 0.0982, 0.0470, 0.0410, 0.0797], device='cuda:4'), in_proj_covar=tensor([0.0300, 0.0454, 0.0353, 0.0354, 0.0352, 0.0407, 0.0243, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:38:53,592 INFO [train.py:904] (4/8) Epoch 27, batch 8250, loss[loss=0.1695, simple_loss=0.2698, pruned_loss=0.03462, over 16209.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2814, pruned_loss=0.05265, over 3082324.75 frames. ], batch size: 165, lr: 2.48e-03, grad_scale: 8.0 2023-05-02 10:39:09,989 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8248, 1.3761, 1.8070, 1.6613, 1.8191, 1.8689, 1.6383, 1.8374], device='cuda:4'), covar=tensor([0.0291, 0.0466, 0.0241, 0.0359, 0.0336, 0.0232, 0.0501, 0.0180], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0198, 0.0185, 0.0190, 0.0205, 0.0164, 0.0201, 0.0164], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:39:19,036 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.616e+02 2.668e+02 3.172e+02 3.719e+02 6.913e+02, threshold=6.344e+02, percent-clipped=1.0 2023-05-02 10:39:40,153 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=272181.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 10:39:50,747 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=272188.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:40:14,329 INFO [train.py:904] (4/8) Epoch 27, batch 8300, loss[loss=0.1698, simple_loss=0.2748, pruned_loss=0.03238, over 16332.00 frames. ], tot_loss[loss=0.189, simple_loss=0.2785, pruned_loss=0.04977, over 3072189.45 frames. ], batch size: 165, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:40:26,347 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=272210.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:40:46,076 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=272223.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:40:54,413 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-02 10:41:04,781 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-02 10:41:15,138 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=272242.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 10:41:22,858 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=272247.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:41:31,413 INFO [train.py:904] (4/8) Epoch 27, batch 8350, loss[loss=0.1976, simple_loss=0.2904, pruned_loss=0.05245, over 15216.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2777, pruned_loss=0.04819, over 3061192.91 frames. ], batch size: 191, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:41:54,864 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.512e+02 2.144e+02 2.517e+02 3.037e+02 5.148e+02, threshold=5.034e+02, percent-clipped=0.0 2023-05-02 10:42:20,866 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=272284.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:42:30,780 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=272290.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:42:43,542 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8000, 4.8597, 4.6673, 4.3052, 4.3542, 4.7677, 4.5960, 4.4579], device='cuda:4'), covar=tensor([0.0649, 0.0705, 0.0385, 0.0347, 0.1038, 0.0588, 0.0450, 0.0743], device='cuda:4'), in_proj_covar=tensor([0.0300, 0.0456, 0.0353, 0.0355, 0.0351, 0.0408, 0.0243, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:42:50,436 INFO [train.py:904] (4/8) Epoch 27, batch 8400, loss[loss=0.1664, simple_loss=0.2616, pruned_loss=0.03558, over 16723.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.275, pruned_loss=0.04613, over 3059712.92 frames. ], batch size: 124, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:42:59,231 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=272308.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:43:42,971 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-02 10:44:09,888 INFO [train.py:904] (4/8) Epoch 27, batch 8450, loss[loss=0.1739, simple_loss=0.27, pruned_loss=0.0389, over 16906.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2733, pruned_loss=0.04455, over 3055459.32 frames. ], batch size: 116, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:44:34,139 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.487e+02 2.148e+02 2.573e+02 3.182e+02 7.232e+02, threshold=5.146e+02, percent-clipped=1.0 2023-05-02 10:44:42,777 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=272373.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:45:32,408 INFO [train.py:904] (4/8) Epoch 27, batch 8500, loss[loss=0.1493, simple_loss=0.2502, pruned_loss=0.02418, over 16804.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2698, pruned_loss=0.04245, over 3051511.60 frames. ], batch size: 102, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:46:24,098 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=272434.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:46:57,058 INFO [train.py:904] (4/8) Epoch 27, batch 8550, loss[loss=0.1635, simple_loss=0.2648, pruned_loss=0.03113, over 16873.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.267, pruned_loss=0.04114, over 3041167.13 frames. ], batch size: 96, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:47:25,002 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 2.247e+02 2.591e+02 3.219e+02 6.393e+02, threshold=5.181e+02, percent-clipped=2.0 2023-05-02 10:47:55,106 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=272483.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:48:35,726 INFO [train.py:904] (4/8) Epoch 27, batch 8600, loss[loss=0.1823, simple_loss=0.2781, pruned_loss=0.04323, over 16380.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2676, pruned_loss=0.04015, over 3046043.56 frames. ], batch size: 146, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:48:50,699 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=272510.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:48:55,558 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5775, 2.1333, 1.8574, 1.9076, 2.4007, 2.1139, 1.9475, 2.5404], device='cuda:4'), covar=tensor([0.0198, 0.0492, 0.0565, 0.0568, 0.0314, 0.0437, 0.0190, 0.0311], device='cuda:4'), in_proj_covar=tensor([0.0220, 0.0239, 0.0227, 0.0229, 0.0239, 0.0237, 0.0236, 0.0235], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:49:22,588 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-05-02 10:49:43,698 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=272537.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 10:49:50,986 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2083, 1.5577, 1.9470, 2.2138, 2.2965, 2.4932, 1.8439, 2.4263], device='cuda:4'), covar=tensor([0.0290, 0.0633, 0.0399, 0.0399, 0.0414, 0.0253, 0.0560, 0.0197], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0197, 0.0183, 0.0188, 0.0204, 0.0163, 0.0200, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:50:13,935 INFO [train.py:904] (4/8) Epoch 27, batch 8650, loss[loss=0.1631, simple_loss=0.261, pruned_loss=0.03266, over 16680.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2663, pruned_loss=0.03858, over 3064330.76 frames. ], batch size: 134, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:50:26,981 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=272558.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:50:47,047 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-02 10:50:48,990 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 2.181e+02 2.564e+02 3.046e+02 5.835e+02, threshold=5.129e+02, percent-clipped=3.0 2023-05-02 10:51:15,177 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=272579.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:51:36,584 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=272590.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:51:44,322 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8740, 2.7780, 2.5777, 1.9895, 2.5540, 2.7784, 2.6764, 1.9258], device='cuda:4'), covar=tensor([0.0446, 0.0090, 0.0078, 0.0370, 0.0149, 0.0119, 0.0109, 0.0491], device='cuda:4'), in_proj_covar=tensor([0.0134, 0.0087, 0.0088, 0.0133, 0.0098, 0.0111, 0.0095, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 10:52:00,241 INFO [train.py:904] (4/8) Epoch 27, batch 8700, loss[loss=0.1491, simple_loss=0.2385, pruned_loss=0.02988, over 12358.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2636, pruned_loss=0.03751, over 3058224.95 frames. ], batch size: 248, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:52:01,200 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=272603.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:52:21,599 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=272613.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:53:06,582 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=272638.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:53:35,410 INFO [train.py:904] (4/8) Epoch 27, batch 8750, loss[loss=0.1889, simple_loss=0.2858, pruned_loss=0.04596, over 16715.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2642, pruned_loss=0.03732, over 3068822.13 frames. ], batch size: 134, lr: 2.47e-03, grad_scale: 4.0 2023-05-02 10:54:15,876 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.582e+02 2.360e+02 2.765e+02 3.360e+02 6.939e+02, threshold=5.529e+02, percent-clipped=4.0 2023-05-02 10:54:28,887 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=272674.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:55:27,068 INFO [train.py:904] (4/8) Epoch 27, batch 8800, loss[loss=0.1765, simple_loss=0.2677, pruned_loss=0.04261, over 12536.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2632, pruned_loss=0.03681, over 3065787.48 frames. ], batch size: 248, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:56:23,109 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=272729.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:57:03,354 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5062, 3.4493, 3.5477, 3.5993, 3.6349, 3.3362, 3.6312, 3.6930], device='cuda:4'), covar=tensor([0.1136, 0.0868, 0.0863, 0.0572, 0.0575, 0.2435, 0.0791, 0.0745], device='cuda:4'), in_proj_covar=tensor([0.0635, 0.0778, 0.0904, 0.0789, 0.0606, 0.0632, 0.0661, 0.0772], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 10:57:12,470 INFO [train.py:904] (4/8) Epoch 27, batch 8850, loss[loss=0.1541, simple_loss=0.2495, pruned_loss=0.02934, over 12687.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2653, pruned_loss=0.03622, over 3053870.55 frames. ], batch size: 248, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 10:57:46,549 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.637e+02 2.085e+02 2.509e+02 2.907e+02 5.151e+02, threshold=5.018e+02, percent-clipped=0.0 2023-05-02 10:58:17,210 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=272783.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 10:58:57,762 INFO [train.py:904] (4/8) Epoch 27, batch 8900, loss[loss=0.1679, simple_loss=0.2692, pruned_loss=0.03331, over 15461.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2653, pruned_loss=0.0355, over 3059608.60 frames. ], batch size: 192, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:00:01,529 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=272831.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:00:20,771 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=272837.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 11:00:45,759 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5708, 4.5967, 4.8981, 4.8991, 4.9024, 4.6394, 4.5758, 4.5220], device='cuda:4'), covar=tensor([0.0368, 0.0657, 0.0435, 0.0407, 0.0455, 0.0426, 0.0914, 0.0445], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0472, 0.0458, 0.0421, 0.0505, 0.0483, 0.0553, 0.0387], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 11:01:00,398 INFO [train.py:904] (4/8) Epoch 27, batch 8950, loss[loss=0.1571, simple_loss=0.2475, pruned_loss=0.0334, over 12868.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2648, pruned_loss=0.0356, over 3068298.68 frames. ], batch size: 250, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:01:35,337 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.412e+02 1.963e+02 2.469e+02 3.081e+02 5.293e+02, threshold=4.938e+02, percent-clipped=1.0 2023-05-02 11:01:57,000 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=272879.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:02:12,418 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=272885.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 11:02:49,712 INFO [train.py:904] (4/8) Epoch 27, batch 9000, loss[loss=0.1425, simple_loss=0.2345, pruned_loss=0.0253, over 16547.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.262, pruned_loss=0.03427, over 3087905.40 frames. ], batch size: 75, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:02:49,712 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 11:02:57,346 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3314, 1.7840, 2.2056, 2.2861, 2.3632, 2.5955, 1.9945, 2.5953], device='cuda:4'), covar=tensor([0.0314, 0.0556, 0.0324, 0.0376, 0.0387, 0.0277, 0.0532, 0.0188], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0196, 0.0183, 0.0187, 0.0204, 0.0162, 0.0199, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:02:59,742 INFO [train.py:938] (4/8) Epoch 27, validation: loss=0.1436, simple_loss=0.2474, pruned_loss=0.01989, over 944034.00 frames. 2023-05-02 11:02:59,742 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 11:03:00,818 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=272903.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:03:26,180 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7782, 5.0733, 4.8990, 4.8994, 4.6170, 4.5388, 4.5912, 5.1518], device='cuda:4'), covar=tensor([0.1231, 0.0908, 0.0957, 0.0861, 0.0854, 0.1220, 0.1215, 0.0867], device='cuda:4'), in_proj_covar=tensor([0.0691, 0.0837, 0.0686, 0.0643, 0.0527, 0.0529, 0.0699, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:03:51,478 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=272927.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:04:42,535 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=272951.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:04:44,610 INFO [train.py:904] (4/8) Epoch 27, batch 9050, loss[loss=0.1741, simple_loss=0.2678, pruned_loss=0.04023, over 16338.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2622, pruned_loss=0.03464, over 3085434.07 frames. ], batch size: 146, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:04:56,388 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9822, 2.3465, 2.3161, 3.0350, 1.7632, 3.2918, 1.7870, 2.8079], device='cuda:4'), covar=tensor([0.1309, 0.0718, 0.1147, 0.0175, 0.0091, 0.0372, 0.1639, 0.0711], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0177, 0.0197, 0.0195, 0.0202, 0.0214, 0.0207, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 11:05:18,776 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.343e+02 2.208e+02 2.579e+02 3.072e+02 5.940e+02, threshold=5.159e+02, percent-clipped=5.0 2023-05-02 11:05:19,929 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=272969.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:05:30,371 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5507, 3.4967, 2.7552, 2.2061, 2.2720, 2.3746, 3.6615, 3.1350], device='cuda:4'), covar=tensor([0.3002, 0.0628, 0.1915, 0.3285, 0.2901, 0.2387, 0.0491, 0.1487], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0268, 0.0306, 0.0319, 0.0296, 0.0270, 0.0297, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 11:06:32,230 INFO [train.py:904] (4/8) Epoch 27, batch 9100, loss[loss=0.1426, simple_loss=0.2416, pruned_loss=0.02178, over 17174.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2611, pruned_loss=0.03477, over 3078452.48 frames. ], batch size: 46, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:07:36,075 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=273029.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:08:32,208 INFO [train.py:904] (4/8) Epoch 27, batch 9150, loss[loss=0.1622, simple_loss=0.2553, pruned_loss=0.03456, over 15409.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2619, pruned_loss=0.03465, over 3075778.35 frames. ], batch size: 190, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:09:08,127 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.550e+02 2.260e+02 2.634e+02 3.413e+02 5.711e+02, threshold=5.268e+02, percent-clipped=2.0 2023-05-02 11:09:27,707 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=273077.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:10:18,352 INFO [train.py:904] (4/8) Epoch 27, batch 9200, loss[loss=0.1459, simple_loss=0.2386, pruned_loss=0.02663, over 17111.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.258, pruned_loss=0.0339, over 3083649.70 frames. ], batch size: 47, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:11:53,369 INFO [train.py:904] (4/8) Epoch 27, batch 9250, loss[loss=0.1677, simple_loss=0.2502, pruned_loss=0.04263, over 12483.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2575, pruned_loss=0.03386, over 3077071.47 frames. ], batch size: 247, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:12:10,516 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1611, 3.2444, 2.1580, 3.5439, 2.5064, 3.5052, 2.2652, 2.7153], device='cuda:4'), covar=tensor([0.0364, 0.0452, 0.1590, 0.0219, 0.0894, 0.0567, 0.1520, 0.0787], device='cuda:4'), in_proj_covar=tensor([0.0169, 0.0175, 0.0191, 0.0164, 0.0175, 0.0211, 0.0199, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 11:12:25,534 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.747e+02 2.225e+02 2.595e+02 3.229e+02 5.341e+02, threshold=5.191e+02, percent-clipped=1.0 2023-05-02 11:13:07,291 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-05-02 11:13:17,671 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1921, 1.5800, 1.9927, 2.1705, 2.2642, 2.3999, 1.7633, 2.3019], device='cuda:4'), covar=tensor([0.0271, 0.0585, 0.0347, 0.0373, 0.0380, 0.0232, 0.0609, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0193, 0.0181, 0.0184, 0.0200, 0.0159, 0.0196, 0.0159], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:13:43,821 INFO [train.py:904] (4/8) Epoch 27, batch 9300, loss[loss=0.1486, simple_loss=0.2397, pruned_loss=0.0288, over 12641.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2559, pruned_loss=0.03362, over 3060780.92 frames. ], batch size: 248, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:14:02,949 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-02 11:14:59,129 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6046, 3.6752, 3.4607, 3.1482, 3.2939, 3.5650, 3.3818, 3.3896], device='cuda:4'), covar=tensor([0.0583, 0.0713, 0.0319, 0.0277, 0.0509, 0.0556, 0.1214, 0.0493], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0449, 0.0349, 0.0351, 0.0346, 0.0405, 0.0241, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:15:28,958 INFO [train.py:904] (4/8) Epoch 27, batch 9350, loss[loss=0.1535, simple_loss=0.2397, pruned_loss=0.03361, over 12201.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.2552, pruned_loss=0.03336, over 3047848.16 frames. ], batch size: 246, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:16:02,980 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.192e+02 2.614e+02 3.426e+02 6.584e+02, threshold=5.229e+02, percent-clipped=2.0 2023-05-02 11:16:03,974 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=273269.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:16:30,841 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.55 vs. limit=5.0 2023-05-02 11:17:04,123 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4118, 4.5063, 4.3057, 3.9976, 4.0015, 4.4054, 4.1705, 4.1473], device='cuda:4'), covar=tensor([0.0714, 0.0884, 0.0389, 0.0379, 0.0940, 0.0771, 0.0710, 0.0709], device='cuda:4'), in_proj_covar=tensor([0.0299, 0.0447, 0.0349, 0.0351, 0.0345, 0.0404, 0.0241, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:17:10,294 INFO [train.py:904] (4/8) Epoch 27, batch 9400, loss[loss=0.1727, simple_loss=0.2804, pruned_loss=0.0325, over 16750.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.2556, pruned_loss=0.03314, over 3045572.92 frames. ], batch size: 83, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:17:39,650 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=273317.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:17:41,186 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3297, 1.6953, 2.0618, 2.3545, 2.3446, 2.6832, 1.9060, 2.5558], device='cuda:4'), covar=tensor([0.0305, 0.0624, 0.0411, 0.0396, 0.0432, 0.0234, 0.0629, 0.0204], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0195, 0.0183, 0.0185, 0.0202, 0.0160, 0.0198, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:17:55,114 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6937, 6.0717, 5.7811, 5.8802, 5.4925, 5.4857, 5.4756, 6.1723], device='cuda:4'), covar=tensor([0.1334, 0.0912, 0.1031, 0.0850, 0.0787, 0.0550, 0.1169, 0.0838], device='cuda:4'), in_proj_covar=tensor([0.0690, 0.0837, 0.0684, 0.0643, 0.0526, 0.0530, 0.0699, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:18:39,395 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2274, 4.3086, 4.1537, 3.8369, 3.8576, 4.2305, 3.9634, 3.9742], device='cuda:4'), covar=tensor([0.0582, 0.0593, 0.0325, 0.0321, 0.0763, 0.0577, 0.0777, 0.0681], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0445, 0.0347, 0.0349, 0.0344, 0.0401, 0.0239, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:18:52,055 INFO [train.py:904] (4/8) Epoch 27, batch 9450, loss[loss=0.1694, simple_loss=0.2678, pruned_loss=0.03546, over 15290.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2581, pruned_loss=0.03357, over 3050784.12 frames. ], batch size: 190, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:19:21,969 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 2.197e+02 2.626e+02 3.206e+02 9.861e+02, threshold=5.251e+02, percent-clipped=1.0 2023-05-02 11:19:40,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4440, 3.3264, 2.7466, 2.2171, 2.1475, 2.3895, 3.5004, 2.9339], device='cuda:4'), covar=tensor([0.3008, 0.0635, 0.1866, 0.3104, 0.2945, 0.2204, 0.0446, 0.1510], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0266, 0.0304, 0.0317, 0.0293, 0.0267, 0.0295, 0.0338], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 11:20:32,362 INFO [train.py:904] (4/8) Epoch 27, batch 9500, loss[loss=0.152, simple_loss=0.25, pruned_loss=0.02703, over 16766.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2576, pruned_loss=0.03332, over 3072640.79 frames. ], batch size: 83, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:20:49,240 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=273409.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:21:38,345 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=273435.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:22:18,626 INFO [train.py:904] (4/8) Epoch 27, batch 9550, loss[loss=0.1816, simple_loss=0.2754, pruned_loss=0.04388, over 16937.00 frames. ], tot_loss[loss=0.162, simple_loss=0.2573, pruned_loss=0.03334, over 3088186.13 frames. ], batch size: 109, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:22:51,234 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 2.156e+02 2.505e+02 2.980e+02 5.260e+02, threshold=5.009e+02, percent-clipped=1.0 2023-05-02 11:22:54,287 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=273470.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:23:35,300 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0896, 3.4587, 3.5092, 2.2966, 3.1484, 3.5492, 3.3681, 2.0035], device='cuda:4'), covar=tensor([0.0599, 0.0062, 0.0063, 0.0467, 0.0138, 0.0100, 0.0091, 0.0575], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0087, 0.0088, 0.0133, 0.0099, 0.0111, 0.0095, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 11:23:46,445 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=273496.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:23:51,495 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3106, 2.4452, 2.0888, 2.2084, 2.7846, 2.4680, 2.7254, 2.9748], device='cuda:4'), covar=tensor([0.0156, 0.0549, 0.0642, 0.0605, 0.0364, 0.0496, 0.0259, 0.0275], device='cuda:4'), in_proj_covar=tensor([0.0215, 0.0235, 0.0224, 0.0226, 0.0236, 0.0233, 0.0230, 0.0230], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:23:57,853 INFO [train.py:904] (4/8) Epoch 27, batch 9600, loss[loss=0.1553, simple_loss=0.2555, pruned_loss=0.02751, over 16800.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2584, pruned_loss=0.03382, over 3077511.81 frames. ], batch size: 83, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:24:28,909 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-02 11:24:34,276 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 11:25:44,212 INFO [train.py:904] (4/8) Epoch 27, batch 9650, loss[loss=0.1648, simple_loss=0.2687, pruned_loss=0.0305, over 16736.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2603, pruned_loss=0.0339, over 3080386.94 frames. ], batch size: 76, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:26:23,246 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.421e+02 2.167e+02 2.602e+02 2.993e+02 6.217e+02, threshold=5.204e+02, percent-clipped=3.0 2023-05-02 11:26:51,013 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4002, 3.2133, 3.5790, 1.8844, 3.6442, 3.7320, 2.9414, 2.8234], device='cuda:4'), covar=tensor([0.0795, 0.0295, 0.0220, 0.1250, 0.0095, 0.0194, 0.0465, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0143, 0.0106, 0.0097, 0.0134, 0.0082, 0.0124, 0.0125, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 11:27:19,646 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9911, 4.2486, 4.1066, 4.1379, 3.8037, 3.8074, 3.8766, 4.2451], device='cuda:4'), covar=tensor([0.1093, 0.0921, 0.0964, 0.0769, 0.0739, 0.1826, 0.0988, 0.0940], device='cuda:4'), in_proj_covar=tensor([0.0684, 0.0829, 0.0678, 0.0638, 0.0522, 0.0524, 0.0694, 0.0647], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:27:30,800 INFO [train.py:904] (4/8) Epoch 27, batch 9700, loss[loss=0.1809, simple_loss=0.2704, pruned_loss=0.04573, over 16146.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2591, pruned_loss=0.03397, over 3055068.30 frames. ], batch size: 165, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:27:35,257 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4798, 3.4256, 3.5230, 3.5891, 3.6428, 3.3337, 3.6114, 3.6770], device='cuda:4'), covar=tensor([0.1285, 0.0899, 0.1030, 0.0659, 0.0614, 0.2398, 0.0814, 0.0804], device='cuda:4'), in_proj_covar=tensor([0.0639, 0.0780, 0.0903, 0.0794, 0.0608, 0.0633, 0.0665, 0.0774], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:29:12,471 INFO [train.py:904] (4/8) Epoch 27, batch 9750, loss[loss=0.177, simple_loss=0.2707, pruned_loss=0.04167, over 16970.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2577, pruned_loss=0.03416, over 3043241.32 frames. ], batch size: 109, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:29:42,115 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.247e+02 2.115e+02 2.683e+02 3.309e+02 9.068e+02, threshold=5.365e+02, percent-clipped=4.0 2023-05-02 11:30:51,289 INFO [train.py:904] (4/8) Epoch 27, batch 9800, loss[loss=0.1704, simple_loss=0.2713, pruned_loss=0.0347, over 16751.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2583, pruned_loss=0.03352, over 3060339.41 frames. ], batch size: 134, lr: 2.47e-03, grad_scale: 4.0 2023-05-02 11:31:16,905 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-05-02 11:32:36,071 INFO [train.py:904] (4/8) Epoch 27, batch 9850, loss[loss=0.1835, simple_loss=0.2749, pruned_loss=0.04601, over 16647.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2596, pruned_loss=0.03327, over 3070705.13 frames. ], batch size: 134, lr: 2.47e-03, grad_scale: 4.0 2023-05-02 11:32:43,963 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-05-02 11:33:00,285 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=273765.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:33:11,080 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.328e+02 2.029e+02 2.351e+02 3.028e+02 6.435e+02, threshold=4.703e+02, percent-clipped=1.0 2023-05-02 11:33:36,792 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.16 vs. limit=5.0 2023-05-02 11:34:02,560 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=273791.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:34:25,790 INFO [train.py:904] (4/8) Epoch 27, batch 9900, loss[loss=0.1581, simple_loss=0.2598, pruned_loss=0.02816, over 16844.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2593, pruned_loss=0.03281, over 3071664.09 frames. ], batch size: 96, lr: 2.47e-03, grad_scale: 4.0 2023-05-02 11:35:12,355 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8481, 5.0974, 5.2286, 5.0146, 5.1181, 5.5852, 5.1595, 4.9054], device='cuda:4'), covar=tensor([0.1004, 0.1849, 0.2011, 0.2027, 0.2257, 0.0880, 0.1513, 0.2327], device='cuda:4'), in_proj_covar=tensor([0.0402, 0.0601, 0.0658, 0.0488, 0.0651, 0.0692, 0.0518, 0.0649], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 11:36:11,395 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6368, 4.6687, 4.9909, 4.9788, 4.9826, 4.7182, 4.6648, 4.6169], device='cuda:4'), covar=tensor([0.0456, 0.0817, 0.0533, 0.0508, 0.0537, 0.0581, 0.1046, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0463, 0.0452, 0.0414, 0.0498, 0.0476, 0.0544, 0.0381], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 11:36:22,572 INFO [train.py:904] (4/8) Epoch 27, batch 9950, loss[loss=0.1514, simple_loss=0.2512, pruned_loss=0.02585, over 17166.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2617, pruned_loss=0.0332, over 3078301.86 frames. ], batch size: 44, lr: 2.47e-03, grad_scale: 4.0 2023-05-02 11:37:03,258 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.425e+02 1.975e+02 2.538e+02 3.032e+02 5.048e+02, threshold=5.077e+02, percent-clipped=2.0 2023-05-02 11:38:25,116 INFO [train.py:904] (4/8) Epoch 27, batch 10000, loss[loss=0.1695, simple_loss=0.2683, pruned_loss=0.03533, over 15361.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2607, pruned_loss=0.03302, over 3107021.20 frames. ], batch size: 192, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:40:06,628 INFO [train.py:904] (4/8) Epoch 27, batch 10050, loss[loss=0.1778, simple_loss=0.2662, pruned_loss=0.04469, over 11710.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2611, pruned_loss=0.03322, over 3109155.70 frames. ], batch size: 248, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:40:24,800 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0636, 5.3618, 5.5356, 5.3065, 5.3957, 5.9113, 5.4240, 5.1259], device='cuda:4'), covar=tensor([0.0893, 0.1890, 0.2196, 0.2095, 0.2348, 0.0826, 0.1726, 0.2509], device='cuda:4'), in_proj_covar=tensor([0.0404, 0.0604, 0.0663, 0.0492, 0.0654, 0.0697, 0.0523, 0.0654], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 11:40:39,173 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.416e+02 2.137e+02 2.431e+02 3.038e+02 5.006e+02, threshold=4.862e+02, percent-clipped=0.0 2023-05-02 11:41:39,226 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3627, 3.5191, 4.0694, 2.2820, 3.2872, 2.5648, 3.7510, 3.6671], device='cuda:4'), covar=tensor([0.0271, 0.0887, 0.0419, 0.2115, 0.0783, 0.0970, 0.0635, 0.1087], device='cuda:4'), in_proj_covar=tensor([0.0155, 0.0162, 0.0163, 0.0152, 0.0143, 0.0128, 0.0140, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 11:41:41,043 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=274001.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:41:43,715 INFO [train.py:904] (4/8) Epoch 27, batch 10100, loss[loss=0.1755, simple_loss=0.2548, pruned_loss=0.04805, over 12622.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2614, pruned_loss=0.03371, over 3084259.62 frames. ], batch size: 248, lr: 2.47e-03, grad_scale: 8.0 2023-05-02 11:42:12,028 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.79 vs. limit=2.0 2023-05-02 11:43:27,165 INFO [train.py:904] (4/8) Epoch 28, batch 0, loss[loss=0.2068, simple_loss=0.2782, pruned_loss=0.06772, over 16828.00 frames. ], tot_loss[loss=0.2068, simple_loss=0.2782, pruned_loss=0.06772, over 16828.00 frames. ], batch size: 96, lr: 2.42e-03, grad_scale: 8.0 2023-05-02 11:43:27,166 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 11:43:34,594 INFO [train.py:938] (4/8) Epoch 28, validation: loss=0.1434, simple_loss=0.2464, pruned_loss=0.02022, over 944034.00 frames. 2023-05-02 11:43:34,594 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 11:43:48,018 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=274062.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:43:52,218 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=274065.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:44:01,204 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.517e+02 2.305e+02 2.657e+02 3.293e+02 5.505e+02, threshold=5.314e+02, percent-clipped=2.0 2023-05-02 11:44:27,473 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=274091.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:44:43,857 INFO [train.py:904] (4/8) Epoch 28, batch 50, loss[loss=0.1882, simple_loss=0.268, pruned_loss=0.05423, over 16444.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2662, pruned_loss=0.04785, over 747924.71 frames. ], batch size: 146, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 11:44:55,636 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.61 vs. limit=5.0 2023-05-02 11:44:59,064 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=274113.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:45:24,147 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=274131.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:45:34,056 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=274139.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:45:53,997 INFO [train.py:904] (4/8) Epoch 28, batch 100, loss[loss=0.1766, simple_loss=0.2583, pruned_loss=0.04744, over 16717.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2614, pruned_loss=0.04394, over 1321302.43 frames. ], batch size: 134, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 11:46:20,296 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.535e+02 2.302e+02 2.829e+02 3.561e+02 6.259e+02, threshold=5.657e+02, percent-clipped=1.0 2023-05-02 11:46:23,012 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3112, 4.1030, 4.3523, 4.4828, 4.5823, 4.1476, 4.4284, 4.5674], device='cuda:4'), covar=tensor([0.1680, 0.1279, 0.1320, 0.0715, 0.0579, 0.1283, 0.2459, 0.0713], device='cuda:4'), in_proj_covar=tensor([0.0640, 0.0782, 0.0904, 0.0796, 0.0609, 0.0634, 0.0666, 0.0776], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:46:47,961 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=274192.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:47:02,021 INFO [train.py:904] (4/8) Epoch 28, batch 150, loss[loss=0.1468, simple_loss=0.2317, pruned_loss=0.0309, over 16967.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2612, pruned_loss=0.04383, over 1746182.48 frames. ], batch size: 41, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 11:48:08,590 INFO [train.py:904] (4/8) Epoch 28, batch 200, loss[loss=0.15, simple_loss=0.2359, pruned_loss=0.03204, over 16735.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2601, pruned_loss=0.04257, over 2095837.79 frames. ], batch size: 39, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 11:48:30,326 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=274269.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:48:33,858 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.454e+02 2.228e+02 2.594e+02 3.067e+02 5.865e+02, threshold=5.188e+02, percent-clipped=1.0 2023-05-02 11:48:51,886 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9398, 4.6855, 4.9453, 5.1406, 5.3237, 4.7352, 5.2909, 5.3172], device='cuda:4'), covar=tensor([0.1913, 0.1360, 0.1769, 0.0813, 0.0613, 0.0904, 0.0603, 0.0687], device='cuda:4'), in_proj_covar=tensor([0.0643, 0.0786, 0.0909, 0.0798, 0.0611, 0.0636, 0.0670, 0.0778], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 11:48:52,327 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-05-02 11:49:16,127 INFO [train.py:904] (4/8) Epoch 28, batch 250, loss[loss=0.1515, simple_loss=0.2402, pruned_loss=0.03146, over 17165.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.257, pruned_loss=0.0411, over 2370536.88 frames. ], batch size: 46, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 11:49:16,573 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=274303.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:49:43,375 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7527, 3.8973, 2.6394, 4.4580, 3.1287, 4.4022, 2.5522, 3.2128], device='cuda:4'), covar=tensor([0.0417, 0.0455, 0.1649, 0.0420, 0.0865, 0.0612, 0.1696, 0.0853], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0178, 0.0194, 0.0168, 0.0177, 0.0216, 0.0203, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 11:49:51,989 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=274330.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:50:23,017 INFO [train.py:904] (4/8) Epoch 28, batch 300, loss[loss=0.1647, simple_loss=0.242, pruned_loss=0.04369, over 16878.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2545, pruned_loss=0.04013, over 2579719.61 frames. ], batch size: 90, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 11:50:30,575 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=274357.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:50:39,282 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=274364.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:50:49,959 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.456e+02 2.173e+02 2.525e+02 3.031e+02 5.896e+02, threshold=5.050e+02, percent-clipped=1.0 2023-05-02 11:50:53,478 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7977, 3.9362, 2.7235, 4.6128, 3.1466, 4.4652, 2.6887, 3.2064], device='cuda:4'), covar=tensor([0.0379, 0.0450, 0.1551, 0.0330, 0.0908, 0.0711, 0.1626, 0.0878], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0178, 0.0195, 0.0169, 0.0177, 0.0217, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 11:51:25,352 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7885, 2.6129, 2.6787, 4.9983, 3.9187, 4.3628, 1.6551, 3.1662], device='cuda:4'), covar=tensor([0.1572, 0.0986, 0.1385, 0.0211, 0.0247, 0.0428, 0.1897, 0.0867], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0179, 0.0200, 0.0199, 0.0203, 0.0218, 0.0209, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 11:51:31,008 INFO [train.py:904] (4/8) Epoch 28, batch 350, loss[loss=0.1681, simple_loss=0.2456, pruned_loss=0.04529, over 16805.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.253, pruned_loss=0.03932, over 2752098.57 frames. ], batch size: 83, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 11:52:07,035 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4929, 4.5653, 4.6984, 4.5079, 4.5352, 5.1429, 4.6170, 4.3026], device='cuda:4'), covar=tensor([0.1581, 0.2178, 0.2533, 0.2322, 0.2776, 0.1144, 0.1810, 0.2611], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0622, 0.0683, 0.0505, 0.0674, 0.0714, 0.0536, 0.0670], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 11:52:37,185 INFO [train.py:904] (4/8) Epoch 28, batch 400, loss[loss=0.1862, simple_loss=0.2717, pruned_loss=0.05033, over 16666.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2519, pruned_loss=0.03961, over 2870319.70 frames. ], batch size: 134, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 11:53:03,891 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.265e+02 2.649e+02 3.069e+02 5.270e+02, threshold=5.298e+02, percent-clipped=3.0 2023-05-02 11:53:24,169 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=274487.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:53:44,149 INFO [train.py:904] (4/8) Epoch 28, batch 450, loss[loss=0.1568, simple_loss=0.2529, pruned_loss=0.03035, over 17145.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2507, pruned_loss=0.03908, over 2966888.87 frames. ], batch size: 47, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 11:54:17,197 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-05-02 11:54:35,039 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=274540.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:54:53,415 INFO [train.py:904] (4/8) Epoch 28, batch 500, loss[loss=0.1445, simple_loss=0.2264, pruned_loss=0.03125, over 17089.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2494, pruned_loss=0.03876, over 3046778.08 frames. ], batch size: 41, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 11:55:01,587 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4479, 3.4232, 2.2466, 3.6367, 2.8074, 3.6027, 2.3083, 2.9036], device='cuda:4'), covar=tensor([0.0302, 0.0503, 0.1451, 0.0413, 0.0706, 0.0919, 0.1421, 0.0670], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0179, 0.0195, 0.0170, 0.0178, 0.0217, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 11:55:21,737 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.414e+02 2.207e+02 2.697e+02 3.239e+02 1.435e+03, threshold=5.394e+02, percent-clipped=3.0 2023-05-02 11:55:58,922 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=274601.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:56:00,838 INFO [train.py:904] (4/8) Epoch 28, batch 550, loss[loss=0.1661, simple_loss=0.2463, pruned_loss=0.04292, over 16883.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.249, pruned_loss=0.0384, over 3099267.33 frames. ], batch size: 90, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 11:56:21,047 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 11:56:31,527 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=274625.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:56:41,933 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5381, 3.5994, 2.3245, 3.8270, 2.9110, 3.7597, 2.3726, 2.9803], device='cuda:4'), covar=tensor([0.0312, 0.0486, 0.1510, 0.0409, 0.0735, 0.0796, 0.1439, 0.0668], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0180, 0.0197, 0.0172, 0.0179, 0.0219, 0.0206, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 11:56:56,841 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-02 11:57:09,862 INFO [train.py:904] (4/8) Epoch 28, batch 600, loss[loss=0.1481, simple_loss=0.2253, pruned_loss=0.03543, over 16352.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2485, pruned_loss=0.03881, over 3144831.26 frames. ], batch size: 165, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 11:57:15,304 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=274657.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:57:17,483 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=274659.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:57:36,633 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.424e+02 2.130e+02 2.496e+02 2.954e+02 2.616e+03, threshold=4.992e+02, percent-clipped=3.0 2023-05-02 11:58:16,075 INFO [train.py:904] (4/8) Epoch 28, batch 650, loss[loss=0.1325, simple_loss=0.2148, pruned_loss=0.02514, over 17017.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2472, pruned_loss=0.03861, over 3175799.01 frames. ], batch size: 41, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 11:58:18,639 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=274705.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 11:58:57,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9158, 2.0865, 2.5957, 2.9134, 2.7722, 3.3590, 2.3289, 3.2878], device='cuda:4'), covar=tensor([0.0318, 0.0610, 0.0374, 0.0401, 0.0419, 0.0243, 0.0596, 0.0193], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0200, 0.0188, 0.0193, 0.0208, 0.0167, 0.0206, 0.0166], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 11:59:25,785 INFO [train.py:904] (4/8) Epoch 28, batch 700, loss[loss=0.1559, simple_loss=0.2479, pruned_loss=0.03189, over 17065.00 frames. ], tot_loss[loss=0.1612, simple_loss=0.2467, pruned_loss=0.03786, over 3207271.55 frames. ], batch size: 50, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 11:59:54,341 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.355e+02 2.008e+02 2.338e+02 2.859e+02 5.050e+02, threshold=4.676e+02, percent-clipped=1.0 2023-05-02 12:00:13,349 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=274787.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:00:34,708 INFO [train.py:904] (4/8) Epoch 28, batch 750, loss[loss=0.1746, simple_loss=0.2482, pruned_loss=0.05052, over 16517.00 frames. ], tot_loss[loss=0.1611, simple_loss=0.2469, pruned_loss=0.03759, over 3239088.57 frames. ], batch size: 75, lr: 2.42e-03, grad_scale: 1.0 2023-05-02 12:01:19,982 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=274835.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:01:22,561 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0001, 2.7077, 2.6708, 4.4510, 3.6125, 4.2034, 1.7481, 3.1760], device='cuda:4'), covar=tensor([0.1319, 0.0773, 0.1172, 0.0182, 0.0186, 0.0371, 0.1584, 0.0795], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0179, 0.0200, 0.0200, 0.0204, 0.0218, 0.0209, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 12:01:45,403 INFO [train.py:904] (4/8) Epoch 28, batch 800, loss[loss=0.1627, simple_loss=0.2594, pruned_loss=0.03303, over 17257.00 frames. ], tot_loss[loss=0.1618, simple_loss=0.2475, pruned_loss=0.03803, over 3265229.77 frames. ], batch size: 52, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 12:02:15,002 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.216e+02 2.000e+02 2.423e+02 2.829e+02 5.216e+02, threshold=4.846e+02, percent-clipped=1.0 2023-05-02 12:02:43,467 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=274894.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:02:46,941 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=274896.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:02:57,015 INFO [train.py:904] (4/8) Epoch 28, batch 850, loss[loss=0.1533, simple_loss=0.2481, pruned_loss=0.02925, over 17199.00 frames. ], tot_loss[loss=0.1611, simple_loss=0.2473, pruned_loss=0.0375, over 3271396.66 frames. ], batch size: 46, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 12:03:26,692 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=274925.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:04:05,741 INFO [train.py:904] (4/8) Epoch 28, batch 900, loss[loss=0.1523, simple_loss=0.2413, pruned_loss=0.03163, over 17202.00 frames. ], tot_loss[loss=0.1603, simple_loss=0.2463, pruned_loss=0.03719, over 3280134.88 frames. ], batch size: 45, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 12:04:09,205 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=274955.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:04:14,411 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=274959.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:04:33,593 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=274973.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:04:34,535 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.476e+02 2.075e+02 2.378e+02 2.870e+02 6.521e+02, threshold=4.756e+02, percent-clipped=4.0 2023-05-02 12:05:14,546 INFO [train.py:904] (4/8) Epoch 28, batch 950, loss[loss=0.1555, simple_loss=0.2543, pruned_loss=0.02832, over 17084.00 frames. ], tot_loss[loss=0.1607, simple_loss=0.2465, pruned_loss=0.03743, over 3293462.71 frames. ], batch size: 53, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 12:05:20,983 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=275007.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:06:23,168 INFO [train.py:904] (4/8) Epoch 28, batch 1000, loss[loss=0.1571, simple_loss=0.2436, pruned_loss=0.03532, over 12114.00 frames. ], tot_loss[loss=0.1602, simple_loss=0.246, pruned_loss=0.03721, over 3295836.13 frames. ], batch size: 247, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 12:06:52,347 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.382e+02 2.067e+02 2.369e+02 2.767e+02 9.232e+02, threshold=4.738e+02, percent-clipped=2.0 2023-05-02 12:07:31,780 INFO [train.py:904] (4/8) Epoch 28, batch 1050, loss[loss=0.1635, simple_loss=0.2427, pruned_loss=0.04211, over 15462.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2455, pruned_loss=0.03715, over 3297327.89 frames. ], batch size: 191, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 12:08:11,505 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.06 vs. limit=5.0 2023-05-02 12:08:17,044 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-02 12:08:39,619 INFO [train.py:904] (4/8) Epoch 28, batch 1100, loss[loss=0.1467, simple_loss=0.2285, pruned_loss=0.03241, over 16831.00 frames. ], tot_loss[loss=0.1593, simple_loss=0.2448, pruned_loss=0.03695, over 3303217.55 frames. ], batch size: 83, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 12:08:47,398 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=275158.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:09:07,965 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.388e+02 1.984e+02 2.393e+02 2.792e+02 5.361e+02, threshold=4.787e+02, percent-clipped=2.0 2023-05-02 12:09:39,593 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=275196.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:09:39,932 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.12 vs. limit=5.0 2023-05-02 12:09:49,907 INFO [train.py:904] (4/8) Epoch 28, batch 1150, loss[loss=0.1497, simple_loss=0.2441, pruned_loss=0.02762, over 17188.00 frames. ], tot_loss[loss=0.1586, simple_loss=0.2447, pruned_loss=0.03625, over 3308595.42 frames. ], batch size: 44, lr: 2.42e-03, grad_scale: 2.0 2023-05-02 12:09:50,696 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-02 12:09:58,810 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.60 vs. limit=2.0 2023-05-02 12:10:11,490 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=275219.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:10:22,093 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4609, 5.7940, 5.6031, 5.6659, 5.2680, 5.2392, 5.1948, 5.9334], device='cuda:4'), covar=tensor([0.1471, 0.1161, 0.1086, 0.0952, 0.0914, 0.0798, 0.1395, 0.0961], device='cuda:4'), in_proj_covar=tensor([0.0724, 0.0884, 0.0717, 0.0679, 0.0557, 0.0553, 0.0741, 0.0687], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:10:44,955 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=275244.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:10:53,442 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=275250.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:10:57,364 INFO [train.py:904] (4/8) Epoch 28, batch 1200, loss[loss=0.1457, simple_loss=0.2273, pruned_loss=0.03207, over 17201.00 frames. ], tot_loss[loss=0.1576, simple_loss=0.2441, pruned_loss=0.03549, over 3312743.94 frames. ], batch size: 44, lr: 2.42e-03, grad_scale: 4.0 2023-05-02 12:11:20,279 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5701, 3.7286, 3.9102, 2.8464, 3.5377, 4.0324, 3.6832, 2.3810], device='cuda:4'), covar=tensor([0.0548, 0.0328, 0.0074, 0.0381, 0.0149, 0.0112, 0.0111, 0.0529], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0090, 0.0091, 0.0135, 0.0102, 0.0114, 0.0098, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 12:11:21,666 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-02 12:11:26,948 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.367e+02 2.036e+02 2.270e+02 2.715e+02 4.480e+02, threshold=4.540e+02, percent-clipped=0.0 2023-05-02 12:12:06,865 INFO [train.py:904] (4/8) Epoch 28, batch 1250, loss[loss=0.1865, simple_loss=0.2536, pruned_loss=0.05974, over 16732.00 frames. ], tot_loss[loss=0.1584, simple_loss=0.2442, pruned_loss=0.0363, over 3324310.14 frames. ], batch size: 124, lr: 2.42e-03, grad_scale: 4.0 2023-05-02 12:12:12,681 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=275307.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:12:56,798 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9428, 2.6512, 2.6115, 4.4603, 3.5107, 4.2358, 1.7285, 3.1176], device='cuda:4'), covar=tensor([0.1367, 0.0816, 0.1238, 0.0191, 0.0188, 0.0403, 0.1611, 0.0820], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0181, 0.0201, 0.0202, 0.0206, 0.0220, 0.0210, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 12:13:08,361 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=275347.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 12:13:17,199 INFO [train.py:904] (4/8) Epoch 28, batch 1300, loss[loss=0.1501, simple_loss=0.2474, pruned_loss=0.02639, over 17038.00 frames. ], tot_loss[loss=0.1576, simple_loss=0.2435, pruned_loss=0.03591, over 3321527.71 frames. ], batch size: 50, lr: 2.42e-03, grad_scale: 4.0 2023-05-02 12:13:25,927 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-05-02 12:13:39,184 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=275368.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:13:46,872 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.360e+02 1.989e+02 2.411e+02 2.818e+02 5.022e+02, threshold=4.822e+02, percent-clipped=1.0 2023-05-02 12:14:02,916 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=275385.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:14:12,678 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9655, 4.8851, 4.8293, 4.3938, 4.4992, 4.8700, 4.7095, 4.5647], device='cuda:4'), covar=tensor([0.0580, 0.0743, 0.0322, 0.0357, 0.0982, 0.0526, 0.0441, 0.0741], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0475, 0.0369, 0.0372, 0.0367, 0.0426, 0.0254, 0.0443], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:14:26,876 INFO [train.py:904] (4/8) Epoch 28, batch 1350, loss[loss=0.1618, simple_loss=0.2384, pruned_loss=0.04258, over 16750.00 frames. ], tot_loss[loss=0.1574, simple_loss=0.2437, pruned_loss=0.03561, over 3331254.67 frames. ], batch size: 89, lr: 2.42e-03, grad_scale: 4.0 2023-05-02 12:14:35,299 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=275408.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 12:14:40,638 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7400, 3.9713, 4.0317, 2.8963, 3.5559, 4.0816, 3.7163, 2.6864], device='cuda:4'), covar=tensor([0.0491, 0.0213, 0.0070, 0.0398, 0.0150, 0.0130, 0.0126, 0.0424], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0090, 0.0091, 0.0135, 0.0101, 0.0114, 0.0098, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 12:15:28,604 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=275446.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:15:33,204 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-05-02 12:15:37,714 INFO [train.py:904] (4/8) Epoch 28, batch 1400, loss[loss=0.1778, simple_loss=0.2472, pruned_loss=0.05416, over 16890.00 frames. ], tot_loss[loss=0.1581, simple_loss=0.244, pruned_loss=0.03604, over 3332531.05 frames. ], batch size: 83, lr: 2.42e-03, grad_scale: 4.0 2023-05-02 12:16:06,675 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.472e+02 2.102e+02 2.484e+02 3.057e+02 6.401e+02, threshold=4.968e+02, percent-clipped=1.0 2023-05-02 12:16:07,066 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4946, 5.8908, 5.6456, 5.7195, 5.2373, 5.3003, 5.2853, 6.0121], device='cuda:4'), covar=tensor([0.1484, 0.1003, 0.1036, 0.0940, 0.1060, 0.0814, 0.1440, 0.0997], device='cuda:4'), in_proj_covar=tensor([0.0722, 0.0879, 0.0715, 0.0676, 0.0554, 0.0553, 0.0738, 0.0685], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:16:46,821 INFO [train.py:904] (4/8) Epoch 28, batch 1450, loss[loss=0.1585, simple_loss=0.2338, pruned_loss=0.04164, over 16496.00 frames. ], tot_loss[loss=0.1578, simple_loss=0.2434, pruned_loss=0.03612, over 3317560.20 frames. ], batch size: 146, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:17:03,294 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=275514.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:17:53,999 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=275550.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:17:57,211 INFO [train.py:904] (4/8) Epoch 28, batch 1500, loss[loss=0.1537, simple_loss=0.235, pruned_loss=0.03615, over 16838.00 frames. ], tot_loss[loss=0.1582, simple_loss=0.2439, pruned_loss=0.03626, over 3316471.90 frames. ], batch size: 102, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:18:26,334 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.442e+02 2.124e+02 2.433e+02 3.002e+02 9.704e+02, threshold=4.866e+02, percent-clipped=1.0 2023-05-02 12:19:01,053 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=275598.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:19:07,364 INFO [train.py:904] (4/8) Epoch 28, batch 1550, loss[loss=0.1938, simple_loss=0.271, pruned_loss=0.05827, over 16889.00 frames. ], tot_loss[loss=0.1594, simple_loss=0.2446, pruned_loss=0.03708, over 3303011.61 frames. ], batch size: 109, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:19:15,942 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-02 12:19:16,813 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8638, 2.8716, 2.5410, 2.7995, 3.1548, 2.9538, 3.4250, 3.3614], device='cuda:4'), covar=tensor([0.0185, 0.0475, 0.0569, 0.0492, 0.0340, 0.0436, 0.0299, 0.0324], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0250, 0.0239, 0.0239, 0.0252, 0.0249, 0.0249, 0.0248], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:20:18,406 INFO [train.py:904] (4/8) Epoch 28, batch 1600, loss[loss=0.1681, simple_loss=0.2589, pruned_loss=0.03859, over 16708.00 frames. ], tot_loss[loss=0.1596, simple_loss=0.2455, pruned_loss=0.03682, over 3309728.44 frames. ], batch size: 62, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:20:33,171 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=275663.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:20:48,272 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.650e+02 2.228e+02 2.653e+02 3.158e+02 7.607e+02, threshold=5.306e+02, percent-clipped=1.0 2023-05-02 12:21:28,984 INFO [train.py:904] (4/8) Epoch 28, batch 1650, loss[loss=0.1488, simple_loss=0.2325, pruned_loss=0.03259, over 16815.00 frames. ], tot_loss[loss=0.1614, simple_loss=0.2474, pruned_loss=0.03772, over 3302717.08 frames. ], batch size: 102, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:21:29,316 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=275703.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 12:21:59,581 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0702, 2.2305, 2.3450, 3.7302, 2.2202, 2.5501, 2.3234, 2.3711], device='cuda:4'), covar=tensor([0.1699, 0.3712, 0.3066, 0.0768, 0.3972, 0.2556, 0.3742, 0.3127], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0474, 0.0388, 0.0338, 0.0447, 0.0543, 0.0446, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:22:21,191 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=275741.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:22:37,918 INFO [train.py:904] (4/8) Epoch 28, batch 1700, loss[loss=0.2031, simple_loss=0.2902, pruned_loss=0.05802, over 12147.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2494, pruned_loss=0.03839, over 3311488.76 frames. ], batch size: 248, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:23:08,709 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.133e+02 2.503e+02 2.990e+02 4.977e+02, threshold=5.005e+02, percent-clipped=0.0 2023-05-02 12:23:26,763 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.83 vs. limit=5.0 2023-05-02 12:23:49,162 INFO [train.py:904] (4/8) Epoch 28, batch 1750, loss[loss=0.1393, simple_loss=0.2275, pruned_loss=0.02557, over 17182.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2499, pruned_loss=0.03857, over 3319560.75 frames. ], batch size: 46, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:24:05,553 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=275814.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:24:20,408 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8895, 4.4164, 4.4396, 3.1399, 3.6545, 4.4330, 4.0247, 2.6948], device='cuda:4'), covar=tensor([0.0541, 0.0077, 0.0052, 0.0395, 0.0165, 0.0091, 0.0085, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0092, 0.0092, 0.0137, 0.0103, 0.0116, 0.0100, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 12:24:29,070 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8559, 5.2038, 4.9521, 4.9936, 4.7497, 4.6762, 4.6601, 5.2908], device='cuda:4'), covar=tensor([0.1341, 0.0949, 0.1125, 0.0885, 0.0881, 0.1154, 0.1319, 0.0976], device='cuda:4'), in_proj_covar=tensor([0.0728, 0.0887, 0.0721, 0.0681, 0.0559, 0.0558, 0.0746, 0.0691], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:24:50,688 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6820, 3.8039, 2.4021, 4.2859, 3.0080, 4.2020, 2.4883, 3.0767], device='cuda:4'), covar=tensor([0.0335, 0.0401, 0.1732, 0.0385, 0.0884, 0.0718, 0.1606, 0.0860], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0183, 0.0199, 0.0175, 0.0181, 0.0223, 0.0207, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 12:24:59,130 INFO [train.py:904] (4/8) Epoch 28, batch 1800, loss[loss=0.1558, simple_loss=0.2557, pruned_loss=0.02793, over 17139.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2507, pruned_loss=0.03865, over 3313840.88 frames. ], batch size: 48, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:25:08,984 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=275860.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:25:11,868 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=275862.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:25:19,206 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8814, 4.6682, 4.8653, 5.1199, 5.2864, 4.7205, 5.2712, 5.3034], device='cuda:4'), covar=tensor([0.2157, 0.1504, 0.2211, 0.0967, 0.0724, 0.1142, 0.0801, 0.0784], device='cuda:4'), in_proj_covar=tensor([0.0691, 0.0842, 0.0980, 0.0856, 0.0652, 0.0683, 0.0718, 0.0831], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:25:30,538 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.248e+02 2.316e+02 2.618e+02 3.157e+02 1.143e+03, threshold=5.235e+02, percent-clipped=6.0 2023-05-02 12:25:59,926 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5389, 4.4390, 4.4688, 4.1374, 4.2136, 4.4800, 4.2508, 4.2427], device='cuda:4'), covar=tensor([0.0625, 0.0936, 0.0328, 0.0334, 0.0830, 0.0620, 0.0575, 0.0675], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0478, 0.0372, 0.0374, 0.0370, 0.0429, 0.0255, 0.0446], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 12:26:08,320 INFO [train.py:904] (4/8) Epoch 28, batch 1850, loss[loss=0.1589, simple_loss=0.2603, pruned_loss=0.02875, over 17272.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2512, pruned_loss=0.03827, over 3325520.79 frames. ], batch size: 52, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:26:33,765 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=275921.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:27:16,666 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=275952.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:27:17,449 INFO [train.py:904] (4/8) Epoch 28, batch 1900, loss[loss=0.1635, simple_loss=0.2641, pruned_loss=0.03145, over 17035.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2501, pruned_loss=0.03734, over 3326508.60 frames. ], batch size: 50, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:27:32,225 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=275963.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:27:47,688 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.245e+02 2.028e+02 2.382e+02 2.719e+02 5.605e+02, threshold=4.765e+02, percent-clipped=1.0 2023-05-02 12:28:17,406 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 12:28:30,501 INFO [train.py:904] (4/8) Epoch 28, batch 1950, loss[loss=0.163, simple_loss=0.2489, pruned_loss=0.03854, over 16651.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2506, pruned_loss=0.03733, over 3321040.71 frames. ], batch size: 75, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:28:30,927 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=276003.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 12:28:40,998 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=276011.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:28:43,405 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=276013.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:28:56,203 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5415, 3.4904, 2.7531, 2.1916, 2.2672, 2.3730, 3.6019, 3.0835], device='cuda:4'), covar=tensor([0.2797, 0.0656, 0.1738, 0.3223, 0.2740, 0.2171, 0.0575, 0.1619], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0276, 0.0313, 0.0327, 0.0305, 0.0278, 0.0305, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 12:29:24,685 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=276041.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:29:36,486 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9945, 5.2204, 5.3664, 5.1446, 5.2083, 5.7944, 5.2440, 4.9806], device='cuda:4'), covar=tensor([0.1285, 0.2035, 0.2781, 0.2322, 0.2590, 0.0982, 0.1687, 0.2476], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0648, 0.0712, 0.0530, 0.0699, 0.0740, 0.0555, 0.0700], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 12:29:37,514 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=276051.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 12:29:40,564 INFO [train.py:904] (4/8) Epoch 28, batch 2000, loss[loss=0.1905, simple_loss=0.2712, pruned_loss=0.0549, over 12234.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2511, pruned_loss=0.03776, over 3305225.81 frames. ], batch size: 246, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:29:57,047 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7559, 4.7246, 4.6413, 4.1148, 4.7256, 1.8599, 4.4716, 4.3519], device='cuda:4'), covar=tensor([0.0190, 0.0164, 0.0208, 0.0371, 0.0127, 0.3101, 0.0191, 0.0260], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0177, 0.0214, 0.0187, 0.0190, 0.0220, 0.0203, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:30:11,364 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.574e+02 2.104e+02 2.511e+02 2.919e+02 7.567e+02, threshold=5.023e+02, percent-clipped=3.0 2023-05-02 12:30:31,277 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=276089.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:30:32,502 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9991, 5.0542, 5.4475, 5.4157, 5.4450, 5.1311, 5.0485, 4.9009], device='cuda:4'), covar=tensor([0.0349, 0.0543, 0.0389, 0.0448, 0.0453, 0.0384, 0.0956, 0.0473], device='cuda:4'), in_proj_covar=tensor([0.0442, 0.0496, 0.0480, 0.0442, 0.0528, 0.0506, 0.0581, 0.0405], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 12:30:39,210 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.61 vs. limit=5.0 2023-05-02 12:30:50,228 INFO [train.py:904] (4/8) Epoch 28, batch 2050, loss[loss=0.1395, simple_loss=0.2321, pruned_loss=0.02343, over 17228.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2505, pruned_loss=0.03775, over 3314468.87 frames. ], batch size: 44, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:32:00,357 INFO [train.py:904] (4/8) Epoch 28, batch 2100, loss[loss=0.1815, simple_loss=0.2599, pruned_loss=0.05153, over 16363.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2509, pruned_loss=0.03785, over 3323334.66 frames. ], batch size: 146, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:32:30,698 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.546e+02 2.091e+02 2.490e+02 2.967e+02 8.308e+02, threshold=4.979e+02, percent-clipped=2.0 2023-05-02 12:32:40,916 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-02 12:33:09,363 INFO [train.py:904] (4/8) Epoch 28, batch 2150, loss[loss=0.2178, simple_loss=0.2916, pruned_loss=0.07198, over 12092.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.252, pruned_loss=0.03839, over 3327651.76 frames. ], batch size: 246, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:33:27,926 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=276216.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:34:18,058 INFO [train.py:904] (4/8) Epoch 28, batch 2200, loss[loss=0.1536, simple_loss=0.2533, pruned_loss=0.02693, over 17148.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2525, pruned_loss=0.03861, over 3330518.25 frames. ], batch size: 48, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:34:39,355 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0747, 5.3986, 5.1735, 5.2141, 4.9314, 4.8807, 4.8691, 5.5365], device='cuda:4'), covar=tensor([0.1279, 0.0916, 0.1133, 0.0935, 0.0881, 0.1070, 0.1257, 0.0874], device='cuda:4'), in_proj_covar=tensor([0.0729, 0.0890, 0.0722, 0.0684, 0.0559, 0.0558, 0.0746, 0.0692], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:34:50,518 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.368e+02 2.258e+02 2.667e+02 3.359e+02 8.724e+02, threshold=5.333e+02, percent-clipped=6.0 2023-05-02 12:35:10,550 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0074, 4.9764, 4.8766, 4.4714, 4.5808, 4.9307, 4.7971, 4.6305], device='cuda:4'), covar=tensor([0.0661, 0.0736, 0.0328, 0.0390, 0.0982, 0.0572, 0.0459, 0.0765], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0483, 0.0377, 0.0379, 0.0373, 0.0434, 0.0258, 0.0452], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 12:35:27,983 INFO [train.py:904] (4/8) Epoch 28, batch 2250, loss[loss=0.1645, simple_loss=0.262, pruned_loss=0.03345, over 17101.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2539, pruned_loss=0.03915, over 3321778.34 frames. ], batch size: 53, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:35:28,710 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.43 vs. limit=5.0 2023-05-02 12:35:35,329 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=276308.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:35:38,137 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-02 12:35:48,048 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=276317.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:36:37,085 INFO [train.py:904] (4/8) Epoch 28, batch 2300, loss[loss=0.1748, simple_loss=0.271, pruned_loss=0.03929, over 17015.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2546, pruned_loss=0.03951, over 3317350.18 frames. ], batch size: 55, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:37:02,831 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8862, 4.5174, 3.1455, 2.4400, 2.7497, 2.7489, 4.8379, 3.5768], device='cuda:4'), covar=tensor([0.3070, 0.0480, 0.1903, 0.3222, 0.3203, 0.2259, 0.0334, 0.1681], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0278, 0.0315, 0.0329, 0.0307, 0.0280, 0.0308, 0.0356], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 12:37:08,703 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.645e+02 2.043e+02 2.409e+02 2.878e+02 4.638e+02, threshold=4.818e+02, percent-clipped=0.0 2023-05-02 12:37:12,907 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=276378.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:37:37,110 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3482, 4.1771, 4.4195, 4.5784, 4.6761, 4.2461, 4.5140, 4.6500], device='cuda:4'), covar=tensor([0.1964, 0.1452, 0.1625, 0.0852, 0.0720, 0.1276, 0.2867, 0.1146], device='cuda:4'), in_proj_covar=tensor([0.0702, 0.0853, 0.0990, 0.0869, 0.0660, 0.0692, 0.0727, 0.0842], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:37:46,502 INFO [train.py:904] (4/8) Epoch 28, batch 2350, loss[loss=0.186, simple_loss=0.2642, pruned_loss=0.05389, over 16774.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.255, pruned_loss=0.03995, over 3312039.46 frames. ], batch size: 124, lr: 2.41e-03, grad_scale: 4.0 2023-05-02 12:38:54,360 INFO [train.py:904] (4/8) Epoch 28, batch 2400, loss[loss=0.1659, simple_loss=0.2608, pruned_loss=0.0355, over 17133.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2565, pruned_loss=0.04021, over 3303007.40 frames. ], batch size: 47, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:39:26,374 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.588e+02 2.076e+02 2.394e+02 2.729e+02 1.080e+03, threshold=4.789e+02, percent-clipped=2.0 2023-05-02 12:40:03,651 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8610, 3.5982, 4.0085, 2.1037, 4.1769, 4.2569, 3.2759, 3.3260], device='cuda:4'), covar=tensor([0.0785, 0.0300, 0.0274, 0.1269, 0.0106, 0.0249, 0.0465, 0.0412], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0110, 0.0102, 0.0139, 0.0086, 0.0131, 0.0130, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 12:40:04,310 INFO [train.py:904] (4/8) Epoch 28, batch 2450, loss[loss=0.1778, simple_loss=0.2768, pruned_loss=0.03943, over 16541.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2564, pruned_loss=0.03953, over 3310301.26 frames. ], batch size: 68, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:40:23,625 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=276516.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:41:13,985 INFO [train.py:904] (4/8) Epoch 28, batch 2500, loss[loss=0.1585, simple_loss=0.2619, pruned_loss=0.02758, over 17144.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.256, pruned_loss=0.0395, over 3305462.54 frames. ], batch size: 49, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:41:30,323 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=276564.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:41:33,195 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 12:41:45,567 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.411e+02 2.004e+02 2.344e+02 2.885e+02 4.375e+02, threshold=4.688e+02, percent-clipped=0.0 2023-05-02 12:42:24,130 INFO [train.py:904] (4/8) Epoch 28, batch 2550, loss[loss=0.1543, simple_loss=0.2478, pruned_loss=0.03041, over 17217.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2556, pruned_loss=0.03912, over 3314208.45 frames. ], batch size: 46, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:42:28,650 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3159, 5.5787, 5.3285, 5.4022, 5.0956, 5.0314, 5.1012, 5.7151], device='cuda:4'), covar=tensor([0.1236, 0.0897, 0.1159, 0.0911, 0.0824, 0.0879, 0.1254, 0.0888], device='cuda:4'), in_proj_covar=tensor([0.0728, 0.0890, 0.0723, 0.0684, 0.0559, 0.0557, 0.0745, 0.0691], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:42:31,779 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=276608.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:42:36,042 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=276611.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:42:48,044 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8480, 2.0499, 2.4705, 2.8645, 2.7611, 3.3755, 2.3642, 3.3656], device='cuda:4'), covar=tensor([0.0307, 0.0592, 0.0431, 0.0394, 0.0401, 0.0232, 0.0546, 0.0208], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0201, 0.0189, 0.0196, 0.0211, 0.0169, 0.0205, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 12:43:33,179 INFO [train.py:904] (4/8) Epoch 28, batch 2600, loss[loss=0.1581, simple_loss=0.2454, pruned_loss=0.0354, over 16779.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2555, pruned_loss=0.03886, over 3319756.56 frames. ], batch size: 83, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:43:37,975 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=276656.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:43:55,771 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=276668.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:44:01,041 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=276672.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:44:02,601 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=276673.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:44:05,671 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.592e+02 2.267e+02 2.519e+02 3.133e+02 7.416e+02, threshold=5.037e+02, percent-clipped=2.0 2023-05-02 12:44:43,680 INFO [train.py:904] (4/8) Epoch 28, batch 2650, loss[loss=0.1698, simple_loss=0.2553, pruned_loss=0.04214, over 16869.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2564, pruned_loss=0.03885, over 3313045.24 frames. ], batch size: 96, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:44:52,534 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8971, 2.8067, 2.4768, 2.7853, 3.1439, 2.9591, 3.4686, 3.4135], device='cuda:4'), covar=tensor([0.0182, 0.0496, 0.0609, 0.0499, 0.0345, 0.0436, 0.0292, 0.0326], device='cuda:4'), in_proj_covar=tensor([0.0239, 0.0252, 0.0240, 0.0241, 0.0254, 0.0251, 0.0251, 0.0250], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:45:00,590 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0383, 4.7863, 5.0807, 5.2650, 5.4745, 4.7361, 5.4157, 5.4507], device='cuda:4'), covar=tensor([0.2035, 0.1447, 0.1803, 0.0866, 0.0585, 0.1108, 0.0714, 0.0715], device='cuda:4'), in_proj_covar=tensor([0.0700, 0.0853, 0.0989, 0.0868, 0.0660, 0.0693, 0.0728, 0.0840], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:45:19,136 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=276729.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:45:50,738 INFO [train.py:904] (4/8) Epoch 28, batch 2700, loss[loss=0.1595, simple_loss=0.2528, pruned_loss=0.03311, over 16806.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2562, pruned_loss=0.03826, over 3322332.55 frames. ], batch size: 39, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:46:23,524 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.276e+02 2.127e+02 2.476e+02 3.244e+02 1.015e+03, threshold=4.951e+02, percent-clipped=5.0 2023-05-02 12:46:38,505 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-02 12:46:42,584 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=276790.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:47:00,576 INFO [train.py:904] (4/8) Epoch 28, batch 2750, loss[loss=0.1824, simple_loss=0.2639, pruned_loss=0.0505, over 16883.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.256, pruned_loss=0.03788, over 3327148.95 frames. ], batch size: 96, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:48:07,281 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=276851.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:48:09,125 INFO [train.py:904] (4/8) Epoch 28, batch 2800, loss[loss=0.1797, simple_loss=0.2694, pruned_loss=0.04504, over 16867.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.256, pruned_loss=0.03783, over 3325822.84 frames. ], batch size: 96, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:48:23,751 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6442, 5.9810, 5.7521, 5.8235, 5.3763, 5.4553, 5.3615, 6.1297], device='cuda:4'), covar=tensor([0.1440, 0.1012, 0.1046, 0.0903, 0.0936, 0.0644, 0.1215, 0.0956], device='cuda:4'), in_proj_covar=tensor([0.0729, 0.0891, 0.0726, 0.0686, 0.0561, 0.0557, 0.0746, 0.0693], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:48:24,980 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=276863.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:48:41,264 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.121e+02 2.382e+02 2.775e+02 4.197e+02, threshold=4.764e+02, percent-clipped=0.0 2023-05-02 12:49:20,035 INFO [train.py:904] (4/8) Epoch 28, batch 2850, loss[loss=0.1419, simple_loss=0.2353, pruned_loss=0.0243, over 17176.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.255, pruned_loss=0.03764, over 3319088.66 frames. ], batch size: 46, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:49:50,204 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=276924.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:49:57,695 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=276931.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:50:27,301 INFO [train.py:904] (4/8) Epoch 28, batch 2900, loss[loss=0.1633, simple_loss=0.2538, pruned_loss=0.03633, over 16520.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2534, pruned_loss=0.03768, over 3326600.74 frames. ], batch size: 68, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:50:46,673 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=276967.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:50:54,970 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=276973.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:50:58,047 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.403e+02 2.124e+02 2.391e+02 2.943e+02 6.733e+02, threshold=4.783e+02, percent-clipped=5.0 2023-05-02 12:51:21,988 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=276992.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 12:51:36,310 INFO [train.py:904] (4/8) Epoch 28, batch 2950, loss[loss=0.1863, simple_loss=0.2592, pruned_loss=0.05675, over 16727.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2535, pruned_loss=0.03863, over 3330077.69 frames. ], batch size: 134, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:51:48,556 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=277011.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:52:01,178 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277021.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:52:05,335 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277024.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:52:45,412 INFO [train.py:904] (4/8) Epoch 28, batch 3000, loss[loss=0.1491, simple_loss=0.2424, pruned_loss=0.02789, over 16827.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2538, pruned_loss=0.0388, over 3321313.08 frames. ], batch size: 42, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:52:45,413 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 12:52:54,789 INFO [train.py:938] (4/8) Epoch 28, validation: loss=0.1335, simple_loss=0.2385, pruned_loss=0.01427, over 944034.00 frames. 2023-05-02 12:52:54,790 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 12:53:06,904 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8867, 4.0980, 3.1377, 4.7403, 3.3407, 4.6296, 2.9092, 3.4572], device='cuda:4'), covar=tensor([0.0339, 0.0398, 0.1245, 0.0332, 0.0740, 0.0609, 0.1419, 0.0732], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0184, 0.0198, 0.0177, 0.0181, 0.0224, 0.0207, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 12:53:21,136 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9080, 2.1543, 2.5064, 2.8535, 2.8222, 3.0198, 2.1275, 3.1499], device='cuda:4'), covar=tensor([0.0256, 0.0538, 0.0418, 0.0356, 0.0353, 0.0313, 0.0617, 0.0193], device='cuda:4'), in_proj_covar=tensor([0.0203, 0.0202, 0.0191, 0.0197, 0.0212, 0.0170, 0.0207, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 12:53:21,145 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=277072.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:53:25,661 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.470e+02 2.148e+02 2.585e+02 3.132e+02 1.139e+03, threshold=5.170e+02, percent-clipped=2.0 2023-05-02 12:54:02,140 INFO [train.py:904] (4/8) Epoch 28, batch 3050, loss[loss=0.1741, simple_loss=0.2683, pruned_loss=0.03991, over 17097.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2539, pruned_loss=0.03892, over 3318373.25 frames. ], batch size: 53, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:54:11,365 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5138, 2.4810, 2.4739, 4.3029, 2.4400, 2.9251, 2.5166, 2.5993], device='cuda:4'), covar=tensor([0.1426, 0.3878, 0.3316, 0.0597, 0.4249, 0.2575, 0.3686, 0.3824], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0478, 0.0390, 0.0341, 0.0448, 0.0548, 0.0450, 0.0559], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 12:55:01,328 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277146.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:55:11,783 INFO [train.py:904] (4/8) Epoch 28, batch 3100, loss[loss=0.1363, simple_loss=0.2263, pruned_loss=0.02317, over 16764.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2533, pruned_loss=0.03857, over 3318222.06 frames. ], batch size: 39, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:55:43,285 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.340e+02 2.064e+02 2.462e+02 2.812e+02 4.738e+02, threshold=4.925e+02, percent-clipped=0.0 2023-05-02 12:55:45,353 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=277177.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:56:21,091 INFO [train.py:904] (4/8) Epoch 28, batch 3150, loss[loss=0.1555, simple_loss=0.2539, pruned_loss=0.02856, over 17120.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2527, pruned_loss=0.03845, over 3313917.18 frames. ], batch size: 49, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:56:26,337 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-05-02 12:56:44,546 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277219.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:56:44,668 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=277219.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:57:10,203 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=277238.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:57:30,727 INFO [train.py:904] (4/8) Epoch 28, batch 3200, loss[loss=0.1382, simple_loss=0.2314, pruned_loss=0.02252, over 17089.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2508, pruned_loss=0.03764, over 3319535.87 frames. ], batch size: 47, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:57:50,755 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277267.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:58:01,813 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.445e+02 1.963e+02 2.375e+02 2.950e+02 8.294e+02, threshold=4.751e+02, percent-clipped=4.0 2023-05-02 12:58:07,094 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=277280.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:58:17,913 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277287.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 12:58:29,227 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6309, 3.8101, 3.9458, 2.7335, 3.5870, 4.0288, 3.6568, 2.4830], device='cuda:4'), covar=tensor([0.0538, 0.0282, 0.0078, 0.0425, 0.0133, 0.0113, 0.0130, 0.0493], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0092, 0.0093, 0.0137, 0.0104, 0.0117, 0.0100, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 12:58:39,077 INFO [train.py:904] (4/8) Epoch 28, batch 3250, loss[loss=0.1621, simple_loss=0.25, pruned_loss=0.03708, over 16477.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2518, pruned_loss=0.03877, over 3304080.85 frames. ], batch size: 68, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 12:58:42,418 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=277305.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:58:55,899 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277315.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:59:09,076 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277324.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 12:59:29,930 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6132, 4.6803, 4.8219, 4.6395, 4.6876, 5.2429, 4.7147, 4.4308], device='cuda:4'), covar=tensor([0.1610, 0.1951, 0.2613, 0.2358, 0.2749, 0.1148, 0.1889, 0.2814], device='cuda:4'), in_proj_covar=tensor([0.0436, 0.0654, 0.0718, 0.0532, 0.0708, 0.0741, 0.0557, 0.0708], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 12:59:47,702 INFO [train.py:904] (4/8) Epoch 28, batch 3300, loss[loss=0.2099, simple_loss=0.287, pruned_loss=0.06645, over 16293.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2538, pruned_loss=0.04008, over 3307372.98 frames. ], batch size: 165, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:00:05,308 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=277366.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:00:06,238 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277367.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:00:10,412 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1988, 4.9292, 5.1962, 5.3810, 5.6075, 4.9095, 5.5287, 5.5807], device='cuda:4'), covar=tensor([0.2109, 0.1464, 0.1890, 0.0852, 0.0553, 0.0842, 0.0619, 0.0703], device='cuda:4'), in_proj_covar=tensor([0.0709, 0.0862, 0.1002, 0.0879, 0.0667, 0.0701, 0.0734, 0.0847], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 13:00:12,570 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277372.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:00:18,088 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.541e+02 2.125e+02 2.555e+02 3.019e+02 5.327e+02, threshold=5.111e+02, percent-clipped=1.0 2023-05-02 13:00:45,971 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-02 13:00:55,902 INFO [train.py:904] (4/8) Epoch 28, batch 3350, loss[loss=0.1794, simple_loss=0.2714, pruned_loss=0.04372, over 12374.00 frames. ], tot_loss[loss=0.167, simple_loss=0.254, pruned_loss=0.04003, over 3300181.31 frames. ], batch size: 246, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:01:23,724 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-02 13:01:54,972 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277446.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:02:03,950 INFO [train.py:904] (4/8) Epoch 28, batch 3400, loss[loss=0.1575, simple_loss=0.2501, pruned_loss=0.03243, over 17222.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2528, pruned_loss=0.03913, over 3307733.60 frames. ], batch size: 45, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:02:34,599 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.382e+02 2.063e+02 2.423e+02 2.740e+02 5.592e+02, threshold=4.846e+02, percent-clipped=2.0 2023-05-02 13:03:01,407 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277494.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:03:13,659 INFO [train.py:904] (4/8) Epoch 28, batch 3450, loss[loss=0.1583, simple_loss=0.237, pruned_loss=0.0398, over 16919.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2523, pruned_loss=0.03868, over 3310181.39 frames. ], batch size: 90, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:03:35,902 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277519.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:03:55,458 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277533.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:04:23,507 INFO [train.py:904] (4/8) Epoch 28, batch 3500, loss[loss=0.1462, simple_loss=0.2281, pruned_loss=0.03212, over 16445.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2508, pruned_loss=0.03832, over 3309977.36 frames. ], batch size: 146, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:04:25,814 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=277554.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:04:42,186 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277567.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:04:47,062 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=277570.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:04:54,525 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277575.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:04:55,301 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.480e+02 2.163e+02 2.511e+02 2.754e+02 5.260e+02, threshold=5.022e+02, percent-clipped=2.0 2023-05-02 13:05:10,708 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277587.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 13:05:32,630 INFO [train.py:904] (4/8) Epoch 28, batch 3550, loss[loss=0.1439, simple_loss=0.2435, pruned_loss=0.02213, over 17250.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2498, pruned_loss=0.0378, over 3306445.92 frames. ], batch size: 52, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:05:35,722 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-05-02 13:05:49,386 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=277615.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:06:11,018 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=277631.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:06:16,855 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277635.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:06:35,262 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0562, 3.0709, 2.0776, 3.2518, 2.5015, 3.3062, 2.2121, 2.6225], device='cuda:4'), covar=tensor([0.0338, 0.0450, 0.1548, 0.0365, 0.0800, 0.0646, 0.1448, 0.0793], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0186, 0.0199, 0.0179, 0.0183, 0.0227, 0.0207, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 13:06:42,325 INFO [train.py:904] (4/8) Epoch 28, batch 3600, loss[loss=0.1664, simple_loss=0.2621, pruned_loss=0.03537, over 17212.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2492, pruned_loss=0.03786, over 3294640.68 frames. ], batch size: 45, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:06:53,439 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277661.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:07:01,958 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277667.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:07:14,462 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.038e+02 2.278e+02 2.834e+02 5.503e+02, threshold=4.555e+02, percent-clipped=2.0 2023-05-02 13:07:37,014 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=277691.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:07:53,647 INFO [train.py:904] (4/8) Epoch 28, batch 3650, loss[loss=0.1389, simple_loss=0.2136, pruned_loss=0.03205, over 16746.00 frames. ], tot_loss[loss=0.1619, simple_loss=0.2476, pruned_loss=0.03814, over 3289254.32 frames. ], batch size: 83, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:08:12,452 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277715.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:08:56,173 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1538, 3.4817, 3.5412, 3.5226, 3.5362, 3.4135, 3.2332, 3.4438], device='cuda:4'), covar=tensor([0.0714, 0.0782, 0.0603, 0.0588, 0.0757, 0.0666, 0.1293, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0449, 0.0507, 0.0488, 0.0450, 0.0536, 0.0514, 0.0593, 0.0413], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 13:09:06,647 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=277752.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 13:09:07,229 INFO [train.py:904] (4/8) Epoch 28, batch 3700, loss[loss=0.1809, simple_loss=0.2589, pruned_loss=0.05148, over 11755.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2463, pruned_loss=0.03938, over 3263863.87 frames. ], batch size: 246, lr: 2.41e-03, grad_scale: 8.0 2023-05-02 13:09:41,445 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.544e+02 2.185e+02 2.658e+02 3.107e+02 6.146e+02, threshold=5.316e+02, percent-clipped=3.0 2023-05-02 13:10:22,496 INFO [train.py:904] (4/8) Epoch 28, batch 3750, loss[loss=0.1714, simple_loss=0.2524, pruned_loss=0.04521, over 16477.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2469, pruned_loss=0.04077, over 3245569.15 frames. ], batch size: 75, lr: 2.40e-03, grad_scale: 4.0 2023-05-02 13:10:53,627 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-05-02 13:11:07,009 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277833.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:11:35,738 INFO [train.py:904] (4/8) Epoch 28, batch 3800, loss[loss=0.1863, simple_loss=0.2808, pruned_loss=0.04592, over 17039.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2486, pruned_loss=0.04193, over 3256956.28 frames. ], batch size: 55, lr: 2.40e-03, grad_scale: 4.0 2023-05-02 13:12:08,804 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277875.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:12:11,082 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.552e+02 2.049e+02 2.321e+02 2.888e+02 4.993e+02, threshold=4.641e+02, percent-clipped=0.0 2023-05-02 13:12:17,662 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277881.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:12:48,152 INFO [train.py:904] (4/8) Epoch 28, batch 3850, loss[loss=0.1656, simple_loss=0.2394, pruned_loss=0.0459, over 16901.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.249, pruned_loss=0.04286, over 3262925.56 frames. ], batch size: 109, lr: 2.40e-03, grad_scale: 4.0 2023-05-02 13:12:57,558 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9397, 5.1633, 5.3953, 5.1370, 5.2337, 5.7866, 5.3387, 5.0222], device='cuda:4'), covar=tensor([0.1328, 0.2005, 0.2306, 0.2220, 0.2850, 0.1097, 0.1714, 0.2577], device='cuda:4'), in_proj_covar=tensor([0.0437, 0.0655, 0.0721, 0.0534, 0.0709, 0.0743, 0.0560, 0.0708], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:12:59,659 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277910.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:13:00,006 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-05-02 13:13:12,430 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7941, 1.8657, 1.6495, 1.5435, 2.0090, 1.6887, 1.6568, 2.0047], device='cuda:4'), covar=tensor([0.0231, 0.0369, 0.0500, 0.0439, 0.0254, 0.0338, 0.0172, 0.0244], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0253, 0.0240, 0.0240, 0.0253, 0.0251, 0.0252, 0.0251], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 13:13:18,278 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=277923.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:13:23,015 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=277926.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:14:00,817 INFO [train.py:904] (4/8) Epoch 28, batch 3900, loss[loss=0.1876, simple_loss=0.272, pruned_loss=0.0516, over 12225.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2494, pruned_loss=0.04354, over 3251523.95 frames. ], batch size: 246, lr: 2.40e-03, grad_scale: 4.0 2023-05-02 13:14:08,083 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6003, 2.8510, 2.6761, 4.5825, 2.6320, 3.1084, 2.8173, 2.9253], device='cuda:4'), covar=tensor([0.1272, 0.3015, 0.2698, 0.0424, 0.3534, 0.2112, 0.3086, 0.2749], device='cuda:4'), in_proj_covar=tensor([0.0426, 0.0480, 0.0390, 0.0342, 0.0448, 0.0550, 0.0451, 0.0561], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 13:14:13,109 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=277961.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:14:27,559 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=277971.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:14:36,141 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.618e+02 2.147e+02 2.438e+02 2.765e+02 5.958e+02, threshold=4.876e+02, percent-clipped=2.0 2023-05-02 13:15:16,801 INFO [train.py:904] (4/8) Epoch 28, batch 3950, loss[loss=0.1493, simple_loss=0.2244, pruned_loss=0.03711, over 16862.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2484, pruned_loss=0.04373, over 3244963.43 frames. ], batch size: 109, lr: 2.40e-03, grad_scale: 4.0 2023-05-02 13:15:25,690 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=278009.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:15:58,205 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=278032.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:16:15,594 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 13:16:19,770 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=278047.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 13:16:26,864 INFO [train.py:904] (4/8) Epoch 28, batch 4000, loss[loss=0.1782, simple_loss=0.2638, pruned_loss=0.04636, over 17039.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2485, pruned_loss=0.04414, over 3257317.81 frames. ], batch size: 53, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:17:01,537 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.521e+02 1.968e+02 2.308e+02 2.954e+02 7.046e+02, threshold=4.616e+02, percent-clipped=2.0 2023-05-02 13:17:38,520 INFO [train.py:904] (4/8) Epoch 28, batch 4050, loss[loss=0.146, simple_loss=0.2397, pruned_loss=0.02609, over 16875.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.249, pruned_loss=0.04335, over 3266539.06 frames. ], batch size: 96, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:18:46,000 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.59 vs. limit=5.0 2023-05-02 13:18:52,735 INFO [train.py:904] (4/8) Epoch 28, batch 4100, loss[loss=0.1604, simple_loss=0.2547, pruned_loss=0.03308, over 16321.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2509, pruned_loss=0.04293, over 3262730.23 frames. ], batch size: 35, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:19:06,553 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7214, 4.8117, 5.0029, 4.7596, 4.8596, 5.4103, 4.9003, 4.5716], device='cuda:4'), covar=tensor([0.1163, 0.1930, 0.2005, 0.1988, 0.2319, 0.0880, 0.1550, 0.2408], device='cuda:4'), in_proj_covar=tensor([0.0433, 0.0649, 0.0714, 0.0530, 0.0702, 0.0735, 0.0555, 0.0704], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:19:24,719 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3576, 2.7995, 3.0521, 1.9675, 2.7393, 2.0741, 3.0001, 3.0418], device='cuda:4'), covar=tensor([0.0269, 0.0897, 0.0584, 0.2191, 0.0902, 0.1053, 0.0622, 0.1037], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0174, 0.0171, 0.0158, 0.0150, 0.0133, 0.0148, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 13:19:29,075 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.475e+02 1.827e+02 2.078e+02 2.388e+02 3.753e+02, threshold=4.156e+02, percent-clipped=0.0 2023-05-02 13:20:09,078 INFO [train.py:904] (4/8) Epoch 28, batch 4150, loss[loss=0.2037, simple_loss=0.2935, pruned_loss=0.05695, over 16808.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2579, pruned_loss=0.04469, over 3244555.35 frames. ], batch size: 124, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:20:19,309 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=278210.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:20:42,374 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=278226.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:20:43,938 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-02 13:21:22,142 INFO [train.py:904] (4/8) Epoch 28, batch 4200, loss[loss=0.1771, simple_loss=0.2801, pruned_loss=0.03702, over 16776.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2652, pruned_loss=0.04614, over 3230191.13 frames. ], batch size: 83, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:21:30,672 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=278258.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:21:47,292 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-05-02 13:21:51,747 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0228, 2.3871, 1.9850, 2.1724, 2.7231, 2.2766, 2.6387, 2.8696], device='cuda:4'), covar=tensor([0.0239, 0.0513, 0.0677, 0.0568, 0.0364, 0.0567, 0.0228, 0.0348], device='cuda:4'), in_proj_covar=tensor([0.0238, 0.0250, 0.0237, 0.0238, 0.0251, 0.0248, 0.0249, 0.0249], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 13:21:54,357 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=278274.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:21:57,400 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0279, 4.1407, 3.9024, 3.6443, 3.6708, 4.0613, 3.7367, 3.8092], device='cuda:4'), covar=tensor([0.0626, 0.0581, 0.0316, 0.0296, 0.0753, 0.0422, 0.1140, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0485, 0.0377, 0.0379, 0.0375, 0.0435, 0.0257, 0.0451], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:21:58,171 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.712e+02 2.281e+02 2.720e+02 3.199e+02 7.364e+02, threshold=5.441e+02, percent-clipped=7.0 2023-05-02 13:22:01,786 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5456, 4.6276, 4.4224, 4.1377, 4.1416, 4.5800, 4.2870, 4.2354], device='cuda:4'), covar=tensor([0.0649, 0.0650, 0.0345, 0.0333, 0.0938, 0.0440, 0.0597, 0.0663], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0484, 0.0377, 0.0379, 0.0374, 0.0435, 0.0257, 0.0451], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:22:23,934 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-02 13:22:26,017 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.01 vs. limit=5.0 2023-05-02 13:22:37,391 INFO [train.py:904] (4/8) Epoch 28, batch 4250, loss[loss=0.178, simple_loss=0.2833, pruned_loss=0.0364, over 16869.00 frames. ], tot_loss[loss=0.1794, simple_loss=0.268, pruned_loss=0.04541, over 3218084.76 frames. ], batch size: 102, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:22:38,380 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0642, 5.3762, 5.5658, 5.2686, 5.3333, 5.8795, 5.3740, 5.1496], device='cuda:4'), covar=tensor([0.0872, 0.1599, 0.1734, 0.1892, 0.2444, 0.0792, 0.1399, 0.2162], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0645, 0.0709, 0.0528, 0.0698, 0.0732, 0.0551, 0.0699], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:23:07,788 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=278323.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:23:13,431 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=278327.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:23:43,990 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=278347.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 13:23:52,555 INFO [train.py:904] (4/8) Epoch 28, batch 4300, loss[loss=0.1898, simple_loss=0.2834, pruned_loss=0.04807, over 16468.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2696, pruned_loss=0.0449, over 3209174.58 frames. ], batch size: 75, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:24:25,560 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=278374.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:24:29,318 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.378e+02 2.019e+02 2.401e+02 2.918e+02 4.618e+02, threshold=4.803e+02, percent-clipped=0.0 2023-05-02 13:24:40,760 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=278384.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:24:56,294 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=278395.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:25:07,418 INFO [train.py:904] (4/8) Epoch 28, batch 4350, loss[loss=0.2027, simple_loss=0.2852, pruned_loss=0.06006, over 11942.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2732, pruned_loss=0.04612, over 3212129.16 frames. ], batch size: 248, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:25:57,139 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=278435.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:26:22,393 INFO [train.py:904] (4/8) Epoch 28, batch 4400, loss[loss=0.1743, simple_loss=0.2706, pruned_loss=0.039, over 16317.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2751, pruned_loss=0.04708, over 3219787.30 frames. ], batch size: 165, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:26:58,177 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.247e+02 2.621e+02 3.049e+02 4.749e+02, threshold=5.243e+02, percent-clipped=0.0 2023-05-02 13:27:10,401 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.57 vs. limit=2.0 2023-05-02 13:27:35,979 INFO [train.py:904] (4/8) Epoch 28, batch 4450, loss[loss=0.2079, simple_loss=0.2957, pruned_loss=0.06003, over 17031.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.279, pruned_loss=0.0491, over 3213919.06 frames. ], batch size: 55, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:27:47,277 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=278511.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:28:03,382 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2206, 4.3169, 4.5517, 4.5147, 4.5296, 4.2692, 4.2725, 4.1798], device='cuda:4'), covar=tensor([0.0304, 0.0491, 0.0360, 0.0373, 0.0387, 0.0389, 0.0824, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0439, 0.0496, 0.0476, 0.0441, 0.0525, 0.0502, 0.0579, 0.0406], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 13:28:26,725 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9386, 5.0098, 4.8214, 4.5036, 4.5512, 4.9208, 4.6257, 4.6231], device='cuda:4'), covar=tensor([0.0448, 0.0256, 0.0202, 0.0232, 0.0651, 0.0268, 0.0390, 0.0505], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0475, 0.0371, 0.0374, 0.0368, 0.0427, 0.0253, 0.0444], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 13:28:48,441 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3704, 3.2165, 2.6352, 2.2020, 2.1839, 2.2109, 3.4123, 2.9148], device='cuda:4'), covar=tensor([0.3215, 0.0746, 0.1952, 0.2757, 0.2713, 0.2455, 0.0569, 0.1420], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0278, 0.0315, 0.0329, 0.0308, 0.0279, 0.0307, 0.0356], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:28:49,072 INFO [train.py:904] (4/8) Epoch 28, batch 4500, loss[loss=0.2013, simple_loss=0.2872, pruned_loss=0.05765, over 15488.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2792, pruned_loss=0.04965, over 3221459.59 frames. ], batch size: 191, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:29:17,478 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=278572.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:29:23,508 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.361e+02 1.834e+02 2.148e+02 2.381e+02 3.720e+02, threshold=4.296e+02, percent-clipped=0.0 2023-05-02 13:29:50,832 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3460, 5.9046, 6.1163, 5.6522, 5.8125, 6.3829, 5.8706, 5.6109], device='cuda:4'), covar=tensor([0.0772, 0.1504, 0.2080, 0.1929, 0.2424, 0.0774, 0.1281, 0.2039], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0639, 0.0703, 0.0523, 0.0693, 0.0727, 0.0546, 0.0694], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:30:01,295 INFO [train.py:904] (4/8) Epoch 28, batch 4550, loss[loss=0.2028, simple_loss=0.29, pruned_loss=0.05782, over 16800.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2799, pruned_loss=0.05045, over 3216694.30 frames. ], batch size: 89, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:30:36,956 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=278627.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:31:12,168 INFO [train.py:904] (4/8) Epoch 28, batch 4600, loss[loss=0.1993, simple_loss=0.2928, pruned_loss=0.0529, over 16842.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.281, pruned_loss=0.05096, over 3223536.77 frames. ], batch size: 39, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:31:43,902 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=278675.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:31:46,039 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.292e+02 1.855e+02 2.019e+02 2.344e+02 4.106e+02, threshold=4.037e+02, percent-clipped=0.0 2023-05-02 13:31:49,267 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=278679.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:32:07,348 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1294, 4.0208, 4.1798, 4.3118, 4.3995, 3.9983, 4.3498, 4.4553], device='cuda:4'), covar=tensor([0.1546, 0.1055, 0.1279, 0.0630, 0.0481, 0.1284, 0.0822, 0.0585], device='cuda:4'), in_proj_covar=tensor([0.0683, 0.0834, 0.0967, 0.0849, 0.0645, 0.0675, 0.0707, 0.0818], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 13:32:22,021 INFO [train.py:904] (4/8) Epoch 28, batch 4650, loss[loss=0.1892, simple_loss=0.2694, pruned_loss=0.05449, over 17210.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.28, pruned_loss=0.05089, over 3230702.76 frames. ], batch size: 44, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:33:00,158 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.65 vs. limit=5.0 2023-05-02 13:33:00,868 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=278730.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:33:33,716 INFO [train.py:904] (4/8) Epoch 28, batch 4700, loss[loss=0.1721, simple_loss=0.2607, pruned_loss=0.04177, over 16504.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2771, pruned_loss=0.04973, over 3233557.45 frames. ], batch size: 68, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:33:48,030 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-05-02 13:33:55,585 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4237, 2.9674, 2.5665, 2.2137, 2.2350, 2.1799, 3.0257, 2.7640], device='cuda:4'), covar=tensor([0.3051, 0.0802, 0.1936, 0.2923, 0.2731, 0.2464, 0.0652, 0.1553], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0276, 0.0313, 0.0327, 0.0307, 0.0277, 0.0304, 0.0354], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:34:07,033 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.385e+02 1.813e+02 1.994e+02 2.322e+02 3.636e+02, threshold=3.989e+02, percent-clipped=0.0 2023-05-02 13:34:45,676 INFO [train.py:904] (4/8) Epoch 28, batch 4750, loss[loss=0.155, simple_loss=0.2457, pruned_loss=0.03218, over 16553.00 frames. ], tot_loss[loss=0.1845, simple_loss=0.2729, pruned_loss=0.04801, over 3227635.52 frames. ], batch size: 62, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:35:22,673 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8341, 4.0331, 2.9437, 2.4861, 2.8470, 2.6520, 4.6121, 3.5275], device='cuda:4'), covar=tensor([0.2958, 0.0708, 0.2054, 0.2669, 0.2545, 0.2039, 0.0465, 0.1435], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0276, 0.0312, 0.0327, 0.0306, 0.0276, 0.0304, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 13:35:57,545 INFO [train.py:904] (4/8) Epoch 28, batch 4800, loss[loss=0.1912, simple_loss=0.2842, pruned_loss=0.04909, over 16487.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2696, pruned_loss=0.04596, over 3223130.63 frames. ], batch size: 68, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:36:18,860 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=278867.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:36:32,582 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.390e+02 1.887e+02 2.183e+02 2.468e+02 4.688e+02, threshold=4.367e+02, percent-clipped=1.0 2023-05-02 13:37:13,014 INFO [train.py:904] (4/8) Epoch 28, batch 4850, loss[loss=0.1907, simple_loss=0.2834, pruned_loss=0.04896, over 16702.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2694, pruned_loss=0.04496, over 3206376.45 frames. ], batch size: 76, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:37:40,549 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0359, 4.1614, 3.9366, 3.6667, 3.6421, 4.0564, 3.7229, 3.8517], device='cuda:4'), covar=tensor([0.0589, 0.0449, 0.0296, 0.0281, 0.0687, 0.0436, 0.1124, 0.0515], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0471, 0.0368, 0.0369, 0.0364, 0.0421, 0.0250, 0.0438], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 13:38:13,411 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-02 13:38:26,575 INFO [train.py:904] (4/8) Epoch 28, batch 4900, loss[loss=0.1899, simple_loss=0.2832, pruned_loss=0.04826, over 16669.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2682, pruned_loss=0.04343, over 3198159.68 frames. ], batch size: 134, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:38:30,126 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7831, 2.7349, 2.6799, 5.1424, 3.8626, 4.2721, 1.7691, 3.1476], device='cuda:4'), covar=tensor([0.1348, 0.0889, 0.1291, 0.0129, 0.0327, 0.0420, 0.1613, 0.0869], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0181, 0.0201, 0.0204, 0.0207, 0.0218, 0.0210, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 13:39:00,724 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.404e+02 1.895e+02 2.133e+02 2.360e+02 3.975e+02, threshold=4.266e+02, percent-clipped=0.0 2023-05-02 13:39:04,635 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=278979.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:39:39,801 INFO [train.py:904] (4/8) Epoch 28, batch 4950, loss[loss=0.1603, simple_loss=0.2612, pruned_loss=0.02971, over 16715.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2682, pruned_loss=0.04281, over 3194444.62 frames. ], batch size: 89, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:40:14,413 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=279027.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:40:19,456 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=279030.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:40:51,497 INFO [train.py:904] (4/8) Epoch 28, batch 5000, loss[loss=0.1883, simple_loss=0.2752, pruned_loss=0.05067, over 17111.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2699, pruned_loss=0.04285, over 3203918.63 frames. ], batch size: 49, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:41:10,499 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-05-02 13:41:26,568 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.600e+02 2.098e+02 2.415e+02 2.843e+02 4.357e+02, threshold=4.831e+02, percent-clipped=1.0 2023-05-02 13:41:27,922 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=279078.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:41:30,766 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7724, 3.0881, 3.2918, 1.9276, 2.8337, 2.1609, 3.3056, 3.3014], device='cuda:4'), covar=tensor([0.0260, 0.0863, 0.0668, 0.2135, 0.0909, 0.1000, 0.0623, 0.0870], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0171, 0.0170, 0.0156, 0.0148, 0.0132, 0.0145, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 13:42:04,192 INFO [train.py:904] (4/8) Epoch 28, batch 5050, loss[loss=0.1877, simple_loss=0.2783, pruned_loss=0.04851, over 16696.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2705, pruned_loss=0.04283, over 3208215.74 frames. ], batch size: 57, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:42:45,717 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=279131.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:43:17,182 INFO [train.py:904] (4/8) Epoch 28, batch 5100, loss[loss=0.1948, simple_loss=0.2772, pruned_loss=0.05616, over 16891.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2685, pruned_loss=0.04238, over 3216378.29 frames. ], batch size: 116, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:43:20,417 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-05-02 13:43:37,603 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=279167.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:43:52,643 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.331e+02 1.882e+02 2.136e+02 2.468e+02 5.149e+02, threshold=4.272e+02, percent-clipped=1.0 2023-05-02 13:44:15,270 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=279192.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:44:16,488 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6394, 2.4859, 2.5664, 4.1716, 2.5710, 3.9188, 1.5703, 2.9371], device='cuda:4'), covar=tensor([0.1493, 0.0890, 0.1213, 0.0133, 0.0144, 0.0385, 0.1743, 0.0834], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0181, 0.0202, 0.0204, 0.0208, 0.0218, 0.0210, 0.0200], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 13:44:30,687 INFO [train.py:904] (4/8) Epoch 28, batch 5150, loss[loss=0.1635, simple_loss=0.2637, pruned_loss=0.03165, over 16912.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2683, pruned_loss=0.04197, over 3189034.36 frames. ], batch size: 96, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:44:48,994 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=279215.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:45:42,233 INFO [train.py:904] (4/8) Epoch 28, batch 5200, loss[loss=0.1665, simple_loss=0.2578, pruned_loss=0.03758, over 16208.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2668, pruned_loss=0.04117, over 3193827.29 frames. ], batch size: 165, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:46:00,485 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3253, 4.4105, 4.2380, 3.9670, 3.9344, 4.3320, 3.9988, 4.1099], device='cuda:4'), covar=tensor([0.0591, 0.0584, 0.0310, 0.0317, 0.0924, 0.0583, 0.0825, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0469, 0.0365, 0.0367, 0.0363, 0.0421, 0.0249, 0.0434], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 13:46:12,931 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-02 13:46:17,346 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.295e+02 1.964e+02 2.228e+02 2.622e+02 8.480e+02, threshold=4.456e+02, percent-clipped=3.0 2023-05-02 13:46:53,564 INFO [train.py:904] (4/8) Epoch 28, batch 5250, loss[loss=0.1766, simple_loss=0.2736, pruned_loss=0.03983, over 16369.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2641, pruned_loss=0.04072, over 3207776.92 frames. ], batch size: 146, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:48:06,861 INFO [train.py:904] (4/8) Epoch 28, batch 5300, loss[loss=0.173, simple_loss=0.2693, pruned_loss=0.03838, over 16312.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2609, pruned_loss=0.03958, over 3216021.72 frames. ], batch size: 165, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:48:41,218 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.441e+02 1.963e+02 2.333e+02 2.759e+02 6.747e+02, threshold=4.665e+02, percent-clipped=2.0 2023-05-02 13:49:20,977 INFO [train.py:904] (4/8) Epoch 28, batch 5350, loss[loss=0.1639, simple_loss=0.2619, pruned_loss=0.033, over 15443.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.26, pruned_loss=0.03915, over 3217722.58 frames. ], batch size: 190, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:50:32,504 INFO [train.py:904] (4/8) Epoch 28, batch 5400, loss[loss=0.1682, simple_loss=0.2675, pruned_loss=0.03445, over 16376.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2629, pruned_loss=0.03989, over 3211612.37 frames. ], batch size: 146, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:51:08,321 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.450e+02 1.928e+02 2.246e+02 2.506e+02 7.644e+02, threshold=4.493e+02, percent-clipped=1.0 2023-05-02 13:51:24,274 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=279487.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:51:50,344 INFO [train.py:904] (4/8) Epoch 28, batch 5450, loss[loss=0.2082, simple_loss=0.3043, pruned_loss=0.05604, over 16504.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2663, pruned_loss=0.04138, over 3208458.37 frames. ], batch size: 68, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:52:31,270 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=279529.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:52:52,410 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=279542.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:53:09,374 INFO [train.py:904] (4/8) Epoch 28, batch 5500, loss[loss=0.2524, simple_loss=0.3232, pruned_loss=0.09081, over 12071.00 frames. ], tot_loss[loss=0.182, simple_loss=0.2729, pruned_loss=0.04556, over 3164440.68 frames. ], batch size: 248, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:53:10,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2098, 3.3830, 3.6460, 2.0786, 3.0987, 2.3808, 3.5582, 3.7491], device='cuda:4'), covar=tensor([0.0265, 0.0868, 0.0576, 0.2260, 0.0872, 0.1040, 0.0632, 0.0933], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0172, 0.0171, 0.0158, 0.0148, 0.0133, 0.0146, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 13:53:47,383 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.670e+02 2.730e+02 3.219e+02 3.961e+02 9.894e+02, threshold=6.439e+02, percent-clipped=12.0 2023-05-02 13:54:00,162 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0398, 3.3522, 3.5929, 2.0731, 3.0746, 2.0774, 3.5407, 3.6811], device='cuda:4'), covar=tensor([0.0243, 0.0779, 0.0542, 0.2227, 0.0814, 0.1173, 0.0531, 0.0815], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0172, 0.0171, 0.0158, 0.0148, 0.0133, 0.0146, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 13:54:09,339 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=279590.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:54:28,765 INFO [train.py:904] (4/8) Epoch 28, batch 5550, loss[loss=0.1908, simple_loss=0.2828, pruned_loss=0.04944, over 16498.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2806, pruned_loss=0.05102, over 3124715.75 frames. ], batch size: 75, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:54:29,329 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=279603.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:54:32,734 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-05-02 13:55:02,965 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.55 vs. limit=5.0 2023-05-02 13:55:47,869 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0731, 2.3707, 2.5203, 1.9460, 2.6449, 2.7441, 2.4100, 2.3910], device='cuda:4'), covar=tensor([0.0676, 0.0257, 0.0240, 0.0869, 0.0140, 0.0286, 0.0439, 0.0426], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0110, 0.0101, 0.0137, 0.0086, 0.0129, 0.0128, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 13:55:48,648 INFO [train.py:904] (4/8) Epoch 28, batch 5600, loss[loss=0.2567, simple_loss=0.3247, pruned_loss=0.09436, over 11072.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2857, pruned_loss=0.05538, over 3081397.64 frames. ], batch size: 246, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:56:28,890 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.204e+02 3.162e+02 3.859e+02 4.792e+02 1.536e+03, threshold=7.717e+02, percent-clipped=11.0 2023-05-02 13:57:08,601 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.29 vs. limit=5.0 2023-05-02 13:57:11,933 INFO [train.py:904] (4/8) Epoch 28, batch 5650, loss[loss=0.1778, simple_loss=0.267, pruned_loss=0.04433, over 17106.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.2895, pruned_loss=0.05814, over 3089447.81 frames. ], batch size: 49, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:58:29,304 INFO [train.py:904] (4/8) Epoch 28, batch 5700, loss[loss=0.2327, simple_loss=0.2999, pruned_loss=0.08272, over 11393.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2919, pruned_loss=0.06058, over 3068304.05 frames. ], batch size: 247, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 13:59:05,453 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.265e+02 3.251e+02 3.778e+02 4.612e+02 7.364e+02, threshold=7.556e+02, percent-clipped=0.0 2023-05-02 13:59:22,696 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=279787.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 13:59:46,949 INFO [train.py:904] (4/8) Epoch 28, batch 5750, loss[loss=0.2035, simple_loss=0.2943, pruned_loss=0.05635, over 16724.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2943, pruned_loss=0.06233, over 3031941.11 frames. ], batch size: 134, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 14:00:39,833 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=279835.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:01:07,306 INFO [train.py:904] (4/8) Epoch 28, batch 5800, loss[loss=0.168, simple_loss=0.2589, pruned_loss=0.03855, over 16696.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2942, pruned_loss=0.06167, over 3024267.65 frames. ], batch size: 57, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 14:01:46,119 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.097e+02 2.839e+02 3.367e+02 4.183e+02 8.215e+02, threshold=6.734e+02, percent-clipped=2.0 2023-05-02 14:01:57,466 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=279885.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:02:16,872 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=279898.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:02:19,140 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3294, 3.2511, 3.6583, 1.9169, 3.7913, 3.7799, 2.9727, 2.8086], device='cuda:4'), covar=tensor([0.0894, 0.0303, 0.0200, 0.1173, 0.0085, 0.0187, 0.0428, 0.0495], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0110, 0.0102, 0.0138, 0.0086, 0.0130, 0.0129, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 14:02:25,207 INFO [train.py:904] (4/8) Epoch 28, batch 5850, loss[loss=0.1927, simple_loss=0.2801, pruned_loss=0.05269, over 17139.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2918, pruned_loss=0.05973, over 3039056.51 frames. ], batch size: 47, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 14:02:53,241 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4153, 4.6987, 4.4895, 4.5221, 4.2533, 4.2135, 4.2324, 4.7421], device='cuda:4'), covar=tensor([0.1218, 0.0917, 0.1030, 0.0961, 0.0778, 0.1480, 0.1093, 0.0871], device='cuda:4'), in_proj_covar=tensor([0.0715, 0.0870, 0.0709, 0.0669, 0.0548, 0.0543, 0.0724, 0.0675], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:02:57,174 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-05-02 14:03:42,763 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0744, 5.4078, 5.1698, 5.1854, 4.9240, 4.8861, 4.7834, 5.5135], device='cuda:4'), covar=tensor([0.1312, 0.0906, 0.0992, 0.0970, 0.0806, 0.0878, 0.1208, 0.0789], device='cuda:4'), in_proj_covar=tensor([0.0713, 0.0869, 0.0708, 0.0668, 0.0546, 0.0542, 0.0721, 0.0674], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:03:44,211 INFO [train.py:904] (4/8) Epoch 28, batch 5900, loss[loss=0.1745, simple_loss=0.2749, pruned_loss=0.03708, over 16736.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2904, pruned_loss=0.05869, over 3049913.70 frames. ], batch size: 89, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 14:04:26,171 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.117e+02 2.561e+02 3.002e+02 3.588e+02 6.021e+02, threshold=6.005e+02, percent-clipped=0.0 2023-05-02 14:05:07,641 INFO [train.py:904] (4/8) Epoch 28, batch 5950, loss[loss=0.1948, simple_loss=0.287, pruned_loss=0.0513, over 17227.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.2911, pruned_loss=0.05734, over 3065136.59 frames. ], batch size: 52, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 14:05:15,064 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1916, 3.3291, 3.5795, 2.0592, 3.0241, 2.3190, 3.6378, 3.6744], device='cuda:4'), covar=tensor([0.0255, 0.0910, 0.0645, 0.2295, 0.0879, 0.1076, 0.0623, 0.0924], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0172, 0.0171, 0.0158, 0.0149, 0.0133, 0.0147, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 14:06:20,885 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=280051.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:06:23,829 INFO [train.py:904] (4/8) Epoch 28, batch 6000, loss[loss=0.1822, simple_loss=0.2646, pruned_loss=0.04991, over 16498.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2907, pruned_loss=0.05748, over 3061993.90 frames. ], batch size: 75, lr: 2.40e-03, grad_scale: 8.0 2023-05-02 14:06:23,829 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 14:06:34,276 INFO [train.py:938] (4/8) Epoch 28, validation: loss=0.148, simple_loss=0.2602, pruned_loss=0.0179, over 944034.00 frames. 2023-05-02 14:06:34,277 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 14:07:11,167 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.966e+02 2.746e+02 3.426e+02 4.098e+02 7.990e+02, threshold=6.852e+02, percent-clipped=5.0 2023-05-02 14:07:18,455 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8630, 4.8334, 4.6720, 3.9233, 4.7782, 1.7998, 4.5285, 4.3168], device='cuda:4'), covar=tensor([0.0109, 0.0105, 0.0205, 0.0402, 0.0104, 0.3018, 0.0155, 0.0309], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0175, 0.0213, 0.0187, 0.0189, 0.0218, 0.0201, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:07:25,958 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=280088.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:07:51,266 INFO [train.py:904] (4/8) Epoch 28, batch 6050, loss[loss=0.2198, simple_loss=0.2969, pruned_loss=0.07129, over 11429.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2897, pruned_loss=0.05712, over 3068643.16 frames. ], batch size: 248, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:08:06,518 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=280112.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:09:02,523 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=280149.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:09:08,866 INFO [train.py:904] (4/8) Epoch 28, batch 6100, loss[loss=0.1921, simple_loss=0.2795, pruned_loss=0.05237, over 16861.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2892, pruned_loss=0.05623, over 3081532.36 frames. ], batch size: 116, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:09:30,658 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6613, 2.4986, 2.2437, 3.6491, 2.4763, 3.7361, 1.4919, 2.7932], device='cuda:4'), covar=tensor([0.1445, 0.0884, 0.1421, 0.0196, 0.0206, 0.0437, 0.1829, 0.0839], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0180, 0.0201, 0.0204, 0.0208, 0.0219, 0.0211, 0.0200], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 14:09:36,232 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.81 vs. limit=5.0 2023-05-02 14:09:51,397 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.727e+02 2.432e+02 2.900e+02 3.731e+02 6.785e+02, threshold=5.800e+02, percent-clipped=0.0 2023-05-02 14:10:01,927 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=280185.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:10:22,523 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=280198.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:10:29,421 INFO [train.py:904] (4/8) Epoch 28, batch 6150, loss[loss=0.1819, simple_loss=0.2689, pruned_loss=0.04743, over 16428.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2872, pruned_loss=0.05554, over 3080320.43 frames. ], batch size: 68, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:10:30,567 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-05-02 14:11:16,582 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=280233.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:11:36,302 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=280246.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:11:45,495 INFO [train.py:904] (4/8) Epoch 28, batch 6200, loss[loss=0.1837, simple_loss=0.2753, pruned_loss=0.04601, over 16661.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.285, pruned_loss=0.05474, over 3108815.64 frames. ], batch size: 89, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:12:24,024 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.005e+02 2.646e+02 3.232e+02 3.925e+02 8.814e+02, threshold=6.464e+02, percent-clipped=6.0 2023-05-02 14:12:56,625 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3289, 5.6531, 5.3874, 5.4247, 5.1605, 5.0948, 5.0161, 5.7564], device='cuda:4'), covar=tensor([0.1313, 0.0950, 0.1138, 0.0915, 0.0820, 0.0789, 0.1326, 0.0918], device='cuda:4'), in_proj_covar=tensor([0.0720, 0.0877, 0.0717, 0.0674, 0.0551, 0.0548, 0.0728, 0.0679], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:13:00,726 INFO [train.py:904] (4/8) Epoch 28, batch 6250, loss[loss=0.2136, simple_loss=0.2855, pruned_loss=0.07091, over 11569.00 frames. ], tot_loss[loss=0.1956, simple_loss=0.2836, pruned_loss=0.05376, over 3107461.54 frames. ], batch size: 246, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:13:36,587 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=280327.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:14:14,430 INFO [train.py:904] (4/8) Epoch 28, batch 6300, loss[loss=0.1949, simple_loss=0.278, pruned_loss=0.05594, over 11895.00 frames. ], tot_loss[loss=0.1945, simple_loss=0.2832, pruned_loss=0.05293, over 3114636.39 frames. ], batch size: 248, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:14:19,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5373, 3.6101, 3.3630, 3.0332, 3.2044, 3.5140, 3.3265, 3.2872], device='cuda:4'), covar=tensor([0.0627, 0.0723, 0.0291, 0.0307, 0.0530, 0.0524, 0.1405, 0.0505], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0471, 0.0365, 0.0367, 0.0362, 0.0421, 0.0249, 0.0435], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:14:53,563 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.897e+02 2.662e+02 3.040e+02 3.867e+02 7.586e+02, threshold=6.080e+02, percent-clipped=1.0 2023-05-02 14:15:02,542 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-05-02 14:15:09,902 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=280388.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:15:31,097 INFO [train.py:904] (4/8) Epoch 28, batch 6350, loss[loss=0.1963, simple_loss=0.2792, pruned_loss=0.05674, over 15301.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.284, pruned_loss=0.05412, over 3112309.83 frames. ], batch size: 190, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:15:36,783 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=280407.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:16:31,362 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=280444.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:16:43,549 INFO [train.py:904] (4/8) Epoch 28, batch 6400, loss[loss=0.2003, simple_loss=0.2867, pruned_loss=0.05699, over 16748.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2856, pruned_loss=0.05646, over 3066870.40 frames. ], batch size: 76, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:17:19,316 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.815e+02 2.958e+02 3.396e+02 3.952e+02 7.468e+02, threshold=6.793e+02, percent-clipped=6.0 2023-05-02 14:17:56,100 INFO [train.py:904] (4/8) Epoch 28, batch 6450, loss[loss=0.221, simple_loss=0.2909, pruned_loss=0.07559, over 11904.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2861, pruned_loss=0.05608, over 3068631.03 frames. ], batch size: 248, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:18:18,360 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=280518.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:18:41,532 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=280533.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:19:02,770 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=280547.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:19:12,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5252, 3.1385, 3.5465, 1.8702, 3.6503, 3.7006, 2.9407, 2.8393], device='cuda:4'), covar=tensor([0.0766, 0.0311, 0.0208, 0.1271, 0.0110, 0.0224, 0.0464, 0.0500], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0103, 0.0140, 0.0087, 0.0131, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 14:19:13,553 INFO [train.py:904] (4/8) Epoch 28, batch 6500, loss[loss=0.1775, simple_loss=0.2707, pruned_loss=0.04216, over 16732.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2844, pruned_loss=0.05543, over 3084974.82 frames. ], batch size: 134, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:19:49,978 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.728e+02 2.616e+02 3.245e+02 4.274e+02 6.880e+02, threshold=6.490e+02, percent-clipped=1.0 2023-05-02 14:19:52,137 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=280579.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:20:15,938 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=280594.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:20:18,974 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0323, 2.2452, 2.2097, 3.7313, 2.0406, 2.5749, 2.2838, 2.3861], device='cuda:4'), covar=tensor([0.1617, 0.3730, 0.3203, 0.0637, 0.4320, 0.2693, 0.3943, 0.3339], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0473, 0.0384, 0.0334, 0.0444, 0.0541, 0.0444, 0.0552], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:20:28,521 INFO [train.py:904] (4/8) Epoch 28, batch 6550, loss[loss=0.1783, simple_loss=0.2837, pruned_loss=0.03649, over 16719.00 frames. ], tot_loss[loss=0.1981, simple_loss=0.2862, pruned_loss=0.05503, over 3093733.13 frames. ], batch size: 134, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:20:37,429 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=280608.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:20:45,260 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=280614.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:21:09,560 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8453, 3.8807, 3.9288, 3.7434, 3.8920, 4.2776, 3.9469, 3.6298], device='cuda:4'), covar=tensor([0.2132, 0.2161, 0.2701, 0.2342, 0.2470, 0.1607, 0.1595, 0.2439], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0642, 0.0708, 0.0518, 0.0696, 0.0730, 0.0548, 0.0698], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 14:21:44,266 INFO [train.py:904] (4/8) Epoch 28, batch 6600, loss[loss=0.2447, simple_loss=0.3132, pruned_loss=0.08809, over 11458.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2881, pruned_loss=0.05583, over 3054042.17 frames. ], batch size: 246, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:22:18,038 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=280675.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:22:21,696 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.864e+02 2.836e+02 3.315e+02 4.097e+02 1.046e+03, threshold=6.629e+02, percent-clipped=1.0 2023-05-02 14:22:28,914 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=280683.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:22:59,892 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=280702.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:23:01,183 INFO [train.py:904] (4/8) Epoch 28, batch 6650, loss[loss=0.1919, simple_loss=0.2788, pruned_loss=0.0525, over 16900.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2877, pruned_loss=0.05612, over 3074351.93 frames. ], batch size: 109, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:23:07,284 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=280707.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:23:16,704 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2271, 5.2216, 5.0829, 4.2950, 5.1758, 1.7221, 4.8852, 4.7231], device='cuda:4'), covar=tensor([0.0088, 0.0093, 0.0198, 0.0429, 0.0086, 0.3054, 0.0126, 0.0267], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0174, 0.0213, 0.0185, 0.0189, 0.0217, 0.0200, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:24:02,741 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=280744.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:24:08,711 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-02 14:24:15,397 INFO [train.py:904] (4/8) Epoch 28, batch 6700, loss[loss=0.188, simple_loss=0.2816, pruned_loss=0.04719, over 16848.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2867, pruned_loss=0.05638, over 3082530.39 frames. ], batch size: 102, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:24:18,764 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=280755.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:24:30,606 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=280763.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:24:38,684 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7235, 4.9860, 5.1458, 4.8955, 5.0094, 5.5335, 5.0057, 4.7373], device='cuda:4'), covar=tensor([0.1196, 0.1921, 0.2327, 0.1915, 0.2234, 0.0994, 0.1626, 0.2491], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0637, 0.0702, 0.0514, 0.0690, 0.0725, 0.0543, 0.0693], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 14:24:51,655 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.083e+02 2.827e+02 3.469e+02 4.301e+02 6.613e+02, threshold=6.939e+02, percent-clipped=0.0 2023-05-02 14:25:06,282 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9796, 3.2411, 3.1746, 2.0055, 2.9632, 3.2116, 2.9930, 1.9416], device='cuda:4'), covar=tensor([0.0649, 0.0074, 0.0094, 0.0527, 0.0146, 0.0140, 0.0130, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0090, 0.0091, 0.0136, 0.0103, 0.0116, 0.0099, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 14:25:13,163 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=280792.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:25:24,596 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3226, 3.8732, 3.7768, 2.4179, 3.3880, 3.8329, 3.4365, 2.0798], device='cuda:4'), covar=tensor([0.0579, 0.0059, 0.0075, 0.0491, 0.0139, 0.0127, 0.0122, 0.0544], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0090, 0.0091, 0.0136, 0.0103, 0.0116, 0.0099, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 14:25:29,255 INFO [train.py:904] (4/8) Epoch 28, batch 6750, loss[loss=0.2368, simple_loss=0.3131, pruned_loss=0.08025, over 12198.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2855, pruned_loss=0.05649, over 3084253.61 frames. ], batch size: 248, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:25:32,672 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6761, 2.3387, 2.2156, 3.4559, 2.3178, 3.6007, 1.3681, 2.6033], device='cuda:4'), covar=tensor([0.1451, 0.0907, 0.1477, 0.0217, 0.0191, 0.0431, 0.1953, 0.0974], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0180, 0.0201, 0.0203, 0.0208, 0.0219, 0.0210, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 14:26:41,173 INFO [train.py:904] (4/8) Epoch 28, batch 6800, loss[loss=0.2579, simple_loss=0.3348, pruned_loss=0.09047, over 11617.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2869, pruned_loss=0.05664, over 3088478.56 frames. ], batch size: 246, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:27:13,224 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=280874.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:27:18,675 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.178e+02 2.733e+02 3.177e+02 3.942e+02 7.279e+02, threshold=6.355e+02, percent-clipped=1.0 2023-05-02 14:27:36,559 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=280889.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:27:55,353 INFO [train.py:904] (4/8) Epoch 28, batch 6850, loss[loss=0.1976, simple_loss=0.3066, pruned_loss=0.04427, over 16408.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2883, pruned_loss=0.05732, over 3084814.14 frames. ], batch size: 35, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:27:56,253 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=280903.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:29:06,784 INFO [train.py:904] (4/8) Epoch 28, batch 6900, loss[loss=0.1981, simple_loss=0.2912, pruned_loss=0.05249, over 16453.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2902, pruned_loss=0.05652, over 3096224.08 frames. ], batch size: 146, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:29:26,310 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6828, 4.9121, 5.0272, 4.8741, 4.9047, 5.4263, 4.9019, 4.6702], device='cuda:4'), covar=tensor([0.1131, 0.1884, 0.2338, 0.1815, 0.2300, 0.0911, 0.1644, 0.2385], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0639, 0.0704, 0.0515, 0.0693, 0.0727, 0.0545, 0.0694], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 14:29:34,341 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=280970.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:29:36,867 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2111, 4.1864, 4.1169, 3.3255, 4.1653, 1.6353, 3.9552, 3.7104], device='cuda:4'), covar=tensor([0.0146, 0.0137, 0.0220, 0.0393, 0.0123, 0.3118, 0.0167, 0.0359], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0175, 0.0213, 0.0186, 0.0189, 0.0218, 0.0201, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:29:45,359 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.637e+02 2.748e+02 3.135e+02 3.737e+02 6.416e+02, threshold=6.270e+02, percent-clipped=1.0 2023-05-02 14:29:53,929 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=280983.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:30:06,782 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5123, 3.5047, 3.4932, 2.7403, 3.3974, 2.1199, 3.2022, 2.8750], device='cuda:4'), covar=tensor([0.0184, 0.0168, 0.0202, 0.0253, 0.0118, 0.2385, 0.0160, 0.0302], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0175, 0.0213, 0.0186, 0.0189, 0.0218, 0.0201, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:30:23,411 INFO [train.py:904] (4/8) Epoch 28, batch 6950, loss[loss=0.1929, simple_loss=0.278, pruned_loss=0.05386, over 17059.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2917, pruned_loss=0.05822, over 3082137.89 frames. ], batch size: 53, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:31:03,953 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=281031.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:31:18,121 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 14:31:23,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1995, 4.2427, 4.5470, 4.4971, 4.5266, 4.2640, 4.2492, 4.2359], device='cuda:4'), covar=tensor([0.0400, 0.0680, 0.0479, 0.0476, 0.0508, 0.0510, 0.0962, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0489, 0.0474, 0.0439, 0.0521, 0.0500, 0.0576, 0.0401], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 14:31:30,262 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=281049.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 14:31:30,430 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3166, 2.4308, 2.3203, 4.1671, 2.2771, 2.8434, 2.4227, 2.5787], device='cuda:4'), covar=tensor([0.1423, 0.3632, 0.3261, 0.0521, 0.4201, 0.2436, 0.3994, 0.3134], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0472, 0.0384, 0.0334, 0.0445, 0.0541, 0.0444, 0.0552], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:31:35,802 INFO [train.py:904] (4/8) Epoch 28, batch 7000, loss[loss=0.2097, simple_loss=0.3051, pruned_loss=0.05718, over 16727.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.292, pruned_loss=0.0579, over 3084123.17 frames. ], batch size: 124, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:31:44,080 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=281058.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:32:10,952 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4310, 3.3804, 3.4505, 3.5386, 3.5722, 3.3253, 3.5567, 3.6199], device='cuda:4'), covar=tensor([0.1328, 0.1005, 0.1047, 0.0677, 0.0807, 0.2317, 0.1201, 0.0896], device='cuda:4'), in_proj_covar=tensor([0.0665, 0.0811, 0.0939, 0.0826, 0.0632, 0.0660, 0.0693, 0.0799], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:32:12,984 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.569e+02 2.882e+02 3.440e+02 4.031e+02 6.341e+02, threshold=6.881e+02, percent-clipped=1.0 2023-05-02 14:32:26,861 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3498, 3.0008, 3.4183, 1.7865, 3.5477, 3.5905, 2.8720, 2.6604], device='cuda:4'), covar=tensor([0.0827, 0.0343, 0.0238, 0.1282, 0.0098, 0.0205, 0.0471, 0.0518], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0104, 0.0140, 0.0087, 0.0132, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 14:32:50,411 INFO [train.py:904] (4/8) Epoch 28, batch 7050, loss[loss=0.1858, simple_loss=0.2868, pruned_loss=0.04238, over 17009.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.292, pruned_loss=0.05742, over 3096101.07 frames. ], batch size: 55, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:33:01,817 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=281110.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 14:34:04,176 INFO [train.py:904] (4/8) Epoch 28, batch 7100, loss[loss=0.1817, simple_loss=0.2749, pruned_loss=0.04425, over 17134.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2904, pruned_loss=0.05699, over 3093248.51 frames. ], batch size: 48, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:34:29,331 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=281169.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:34:37,929 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=281174.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:34:43,392 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.809e+02 2.877e+02 3.420e+02 4.394e+02 9.887e+02, threshold=6.841e+02, percent-clipped=1.0 2023-05-02 14:34:58,406 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5845, 4.4748, 4.2967, 2.8559, 3.7523, 4.4037, 3.7853, 2.4657], device='cuda:4'), covar=tensor([0.0623, 0.0051, 0.0064, 0.0458, 0.0123, 0.0131, 0.0113, 0.0478], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0090, 0.0091, 0.0135, 0.0102, 0.0116, 0.0098, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 14:34:58,424 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8046, 3.5503, 4.2271, 2.0648, 4.4223, 4.3986, 3.1739, 3.2313], device='cuda:4'), covar=tensor([0.0884, 0.0338, 0.0199, 0.1268, 0.0078, 0.0180, 0.0477, 0.0476], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0104, 0.0140, 0.0087, 0.0132, 0.0131, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 14:34:59,550 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=281189.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:35:20,726 INFO [train.py:904] (4/8) Epoch 28, batch 7150, loss[loss=0.1899, simple_loss=0.2779, pruned_loss=0.05093, over 16454.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2883, pruned_loss=0.05691, over 3086268.24 frames. ], batch size: 75, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:35:21,035 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=281203.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:35:26,746 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4018, 3.3943, 3.4440, 3.5101, 3.5411, 3.3227, 3.5358, 3.5810], device='cuda:4'), covar=tensor([0.1339, 0.0950, 0.0988, 0.0637, 0.0672, 0.2130, 0.0977, 0.0937], device='cuda:4'), in_proj_covar=tensor([0.0665, 0.0811, 0.0939, 0.0825, 0.0631, 0.0658, 0.0692, 0.0799], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:35:47,782 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=281222.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:36:00,727 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=281230.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 14:36:09,136 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=281237.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:36:29,668 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=281251.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:36:31,489 INFO [train.py:904] (4/8) Epoch 28, batch 7200, loss[loss=0.2055, simple_loss=0.2853, pruned_loss=0.06288, over 11521.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2861, pruned_loss=0.05558, over 3061410.64 frames. ], batch size: 246, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:36:34,355 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-05-02 14:36:56,525 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=281270.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:37:07,756 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.681e+02 2.572e+02 3.176e+02 3.806e+02 6.231e+02, threshold=6.351e+02, percent-clipped=0.0 2023-05-02 14:37:46,525 INFO [train.py:904] (4/8) Epoch 28, batch 7250, loss[loss=0.223, simple_loss=0.2937, pruned_loss=0.07617, over 11628.00 frames. ], tot_loss[loss=0.1963, simple_loss=0.2837, pruned_loss=0.05446, over 3063659.70 frames. ], batch size: 250, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:38:09,209 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=281318.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:38:59,569 INFO [train.py:904] (4/8) Epoch 28, batch 7300, loss[loss=0.1891, simple_loss=0.2829, pruned_loss=0.04768, over 17114.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2831, pruned_loss=0.05433, over 3055274.40 frames. ], batch size: 47, lr: 2.39e-03, grad_scale: 2.0 2023-05-02 14:39:05,104 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9199, 2.1342, 2.1253, 3.3495, 2.0780, 2.4416, 2.2499, 2.2750], device='cuda:4'), covar=tensor([0.1640, 0.3580, 0.3314, 0.0809, 0.4561, 0.2611, 0.3674, 0.3753], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0473, 0.0385, 0.0336, 0.0446, 0.0543, 0.0447, 0.0554], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:39:08,078 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=281358.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:39:39,517 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.942e+02 3.043e+02 3.727e+02 4.632e+02 1.394e+03, threshold=7.454e+02, percent-clipped=7.0 2023-05-02 14:40:05,001 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-05-02 14:40:13,645 INFO [train.py:904] (4/8) Epoch 28, batch 7350, loss[loss=0.2385, simple_loss=0.3065, pruned_loss=0.08531, over 11058.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.285, pruned_loss=0.05605, over 3033362.71 frames. ], batch size: 248, lr: 2.39e-03, grad_scale: 2.0 2023-05-02 14:40:16,308 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=281405.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 14:40:17,485 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=281406.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:40:50,064 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=281426.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:41:03,904 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.80 vs. limit=5.0 2023-05-02 14:41:28,056 INFO [train.py:904] (4/8) Epoch 28, batch 7400, loss[loss=0.201, simple_loss=0.2932, pruned_loss=0.05437, over 16915.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2862, pruned_loss=0.0564, over 3035411.66 frames. ], batch size: 90, lr: 2.39e-03, grad_scale: 2.0 2023-05-02 14:41:33,780 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=281456.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:41:36,499 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3281, 3.4732, 3.6093, 3.5850, 3.5967, 3.4419, 3.4648, 3.4930], device='cuda:4'), covar=tensor([0.0407, 0.0678, 0.0473, 0.0444, 0.0546, 0.0570, 0.0823, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0483, 0.0468, 0.0433, 0.0514, 0.0494, 0.0569, 0.0397], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 14:42:08,140 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.796e+02 2.701e+02 3.025e+02 3.387e+02 6.725e+02, threshold=6.049e+02, percent-clipped=0.0 2023-05-02 14:42:18,914 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=281487.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:42:42,678 INFO [train.py:904] (4/8) Epoch 28, batch 7450, loss[loss=0.1935, simple_loss=0.2794, pruned_loss=0.05378, over 16679.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2867, pruned_loss=0.05681, over 3048764.56 frames. ], batch size: 57, lr: 2.39e-03, grad_scale: 2.0 2023-05-02 14:43:06,541 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=281517.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:43:17,742 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=281525.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 14:43:58,013 INFO [train.py:904] (4/8) Epoch 28, batch 7500, loss[loss=0.2428, simple_loss=0.3125, pruned_loss=0.08657, over 11427.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2869, pruned_loss=0.05604, over 3066327.61 frames. ], batch size: 246, lr: 2.39e-03, grad_scale: 2.0 2023-05-02 14:44:36,766 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.668e+02 2.813e+02 3.506e+02 4.192e+02 1.275e+03, threshold=7.011e+02, percent-clipped=6.0 2023-05-02 14:44:42,917 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3617, 5.3810, 5.2440, 4.7980, 4.8817, 5.2768, 5.1616, 4.9459], device='cuda:4'), covar=tensor([0.0595, 0.0472, 0.0279, 0.0309, 0.0966, 0.0487, 0.0319, 0.0685], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0465, 0.0359, 0.0362, 0.0357, 0.0414, 0.0248, 0.0429], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:45:11,676 INFO [train.py:904] (4/8) Epoch 28, batch 7550, loss[loss=0.2146, simple_loss=0.3024, pruned_loss=0.06341, over 16233.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2865, pruned_loss=0.05677, over 3053283.79 frames. ], batch size: 165, lr: 2.39e-03, grad_scale: 2.0 2023-05-02 14:46:25,956 INFO [train.py:904] (4/8) Epoch 28, batch 7600, loss[loss=0.2138, simple_loss=0.2829, pruned_loss=0.07231, over 11196.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2859, pruned_loss=0.05747, over 3035605.25 frames. ], batch size: 246, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 14:47:04,823 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.856e+02 2.608e+02 3.017e+02 3.545e+02 6.203e+02, threshold=6.033e+02, percent-clipped=0.0 2023-05-02 14:47:40,202 INFO [train.py:904] (4/8) Epoch 28, batch 7650, loss[loss=0.1952, simple_loss=0.2814, pruned_loss=0.05449, over 16684.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2855, pruned_loss=0.05676, over 3063105.04 frames. ], batch size: 62, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 14:47:43,650 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=281705.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 14:48:51,546 INFO [train.py:904] (4/8) Epoch 28, batch 7700, loss[loss=0.1866, simple_loss=0.2798, pruned_loss=0.04676, over 16927.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2855, pruned_loss=0.05713, over 3069134.85 frames. ], batch size: 90, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 14:48:52,552 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=281753.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 14:49:31,077 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.035e+02 3.039e+02 3.444e+02 4.107e+02 7.486e+02, threshold=6.887e+02, percent-clipped=3.0 2023-05-02 14:49:35,834 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=281782.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:50:07,526 INFO [train.py:904] (4/8) Epoch 28, batch 7750, loss[loss=0.2026, simple_loss=0.2901, pruned_loss=0.05755, over 16345.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2853, pruned_loss=0.05641, over 3091966.53 frames. ], batch size: 146, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 14:50:21,322 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=281812.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:50:23,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5061, 5.8122, 5.5800, 5.6471, 5.2886, 5.2241, 5.1797, 5.9440], device='cuda:4'), covar=tensor([0.1264, 0.0906, 0.0954, 0.0886, 0.0884, 0.0735, 0.1283, 0.0813], device='cuda:4'), in_proj_covar=tensor([0.0716, 0.0870, 0.0712, 0.0670, 0.0547, 0.0551, 0.0724, 0.0675], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 14:50:26,955 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2967, 4.0912, 4.0493, 2.6605, 3.6863, 4.1129, 3.6051, 2.3856], device='cuda:4'), covar=tensor([0.0615, 0.0062, 0.0062, 0.0452, 0.0121, 0.0129, 0.0120, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0089, 0.0090, 0.0134, 0.0101, 0.0115, 0.0098, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 14:50:40,375 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=281825.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:51:21,111 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=281852.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:51:21,935 INFO [train.py:904] (4/8) Epoch 28, batch 7800, loss[loss=0.2369, simple_loss=0.3039, pruned_loss=0.085, over 11405.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2868, pruned_loss=0.05761, over 3072433.65 frames. ], batch size: 246, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 14:51:52,697 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=281873.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:52:03,257 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.973e+02 2.715e+02 3.377e+02 4.149e+02 7.003e+02, threshold=6.754e+02, percent-clipped=1.0 2023-05-02 14:52:05,966 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.58 vs. limit=2.0 2023-05-02 14:52:37,334 INFO [train.py:904] (4/8) Epoch 28, batch 7850, loss[loss=0.1995, simple_loss=0.2929, pruned_loss=0.05305, over 16736.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2872, pruned_loss=0.05674, over 3086347.62 frames. ], batch size: 124, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 14:52:47,403 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=281910.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 14:52:51,799 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=281913.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:53:31,666 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 14:53:38,086 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=281945.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:53:48,936 INFO [train.py:904] (4/8) Epoch 28, batch 7900, loss[loss=0.1838, simple_loss=0.2798, pruned_loss=0.04393, over 16689.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2865, pruned_loss=0.05599, over 3103712.14 frames. ], batch size: 83, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 14:54:16,302 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=281971.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 14:54:29,909 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.540e+02 2.510e+02 2.970e+02 3.651e+02 5.631e+02, threshold=5.940e+02, percent-clipped=0.0 2023-05-02 14:55:09,783 INFO [train.py:904] (4/8) Epoch 28, batch 7950, loss[loss=0.2143, simple_loss=0.3033, pruned_loss=0.06272, over 16362.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2867, pruned_loss=0.0564, over 3105628.99 frames. ], batch size: 35, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 14:55:14,752 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=282006.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:55:40,917 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 14:56:27,143 INFO [train.py:904] (4/8) Epoch 28, batch 8000, loss[loss=0.2291, simple_loss=0.2984, pruned_loss=0.07991, over 11434.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2873, pruned_loss=0.05719, over 3086369.99 frames. ], batch size: 248, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:57:07,771 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.895e+02 2.806e+02 3.354e+02 4.135e+02 1.069e+03, threshold=6.709e+02, percent-clipped=7.0 2023-05-02 14:57:11,293 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=282082.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:57:22,018 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-05-02 14:57:33,338 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=282096.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:57:42,452 INFO [train.py:904] (4/8) Epoch 28, batch 8050, loss[loss=0.2075, simple_loss=0.3091, pruned_loss=0.05299, over 16821.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2874, pruned_loss=0.05685, over 3099187.08 frames. ], batch size: 102, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:57:49,469 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4018, 3.2807, 2.6078, 2.1190, 2.2710, 2.2519, 3.5421, 3.0841], device='cuda:4'), covar=tensor([0.3379, 0.0874, 0.2096, 0.3320, 0.2855, 0.2436, 0.0586, 0.1450], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0273, 0.0313, 0.0328, 0.0305, 0.0278, 0.0305, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 14:57:56,089 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=282112.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:58:22,177 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=282130.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:58:57,914 INFO [train.py:904] (4/8) Epoch 28, batch 8100, loss[loss=0.1985, simple_loss=0.285, pruned_loss=0.05598, over 15375.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.287, pruned_loss=0.05603, over 3111952.50 frames. ], batch size: 191, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 14:59:03,733 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=282157.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:59:09,073 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=282160.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 14:59:38,410 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.025e+02 2.655e+02 3.116e+02 4.195e+02 8.434e+02, threshold=6.231e+02, percent-clipped=4.0 2023-05-02 15:00:13,683 INFO [train.py:904] (4/8) Epoch 28, batch 8150, loss[loss=0.1755, simple_loss=0.2756, pruned_loss=0.03774, over 16206.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2844, pruned_loss=0.05518, over 3110859.50 frames. ], batch size: 35, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 15:00:21,775 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=282208.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:01:29,255 INFO [train.py:904] (4/8) Epoch 28, batch 8200, loss[loss=0.1704, simple_loss=0.2641, pruned_loss=0.03835, over 16704.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2814, pruned_loss=0.05428, over 3118983.20 frames. ], batch size: 83, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 15:01:50,082 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=282266.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 15:01:54,577 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0456, 3.3085, 3.7780, 2.0694, 3.1797, 2.1091, 3.5526, 3.5004], device='cuda:4'), covar=tensor([0.0266, 0.0918, 0.0489, 0.2326, 0.0800, 0.1182, 0.0621, 0.0920], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0171, 0.0158, 0.0148, 0.0133, 0.0147, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 15:02:13,182 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.791e+02 2.646e+02 3.224e+02 3.830e+02 8.155e+02, threshold=6.449e+02, percent-clipped=2.0 2023-05-02 15:02:46,941 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=282301.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:02:50,325 INFO [train.py:904] (4/8) Epoch 28, batch 8250, loss[loss=0.1637, simple_loss=0.2658, pruned_loss=0.03083, over 15245.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2805, pruned_loss=0.05237, over 3090483.13 frames. ], batch size: 190, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 15:03:42,225 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4103, 4.5465, 4.7053, 4.4298, 4.5480, 5.0501, 4.5126, 4.2188], device='cuda:4'), covar=tensor([0.1421, 0.1882, 0.2154, 0.1977, 0.2440, 0.0909, 0.1515, 0.2406], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0633, 0.0698, 0.0512, 0.0687, 0.0723, 0.0544, 0.0689], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:04:07,365 INFO [train.py:904] (4/8) Epoch 28, batch 8300, loss[loss=0.1744, simple_loss=0.2799, pruned_loss=0.03448, over 16931.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2785, pruned_loss=0.04984, over 3072222.56 frames. ], batch size: 96, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 15:04:50,784 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.600e+02 2.126e+02 2.473e+02 2.954e+02 5.391e+02, threshold=4.946e+02, percent-clipped=0.0 2023-05-02 15:05:26,438 INFO [train.py:904] (4/8) Epoch 28, batch 8350, loss[loss=0.1832, simple_loss=0.2834, pruned_loss=0.04153, over 16711.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.278, pruned_loss=0.04804, over 3065624.42 frames. ], batch size: 134, lr: 2.39e-03, grad_scale: 4.0 2023-05-02 15:06:30,493 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7928, 1.4051, 1.7586, 1.6962, 1.8960, 1.8686, 1.7422, 1.7570], device='cuda:4'), covar=tensor([0.0281, 0.0446, 0.0235, 0.0321, 0.0302, 0.0236, 0.0457, 0.0164], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0196, 0.0184, 0.0189, 0.0206, 0.0163, 0.0200, 0.0163], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:06:43,064 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=282452.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:06:43,824 INFO [train.py:904] (4/8) Epoch 28, batch 8400, loss[loss=0.1534, simple_loss=0.2545, pruned_loss=0.02619, over 16715.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2754, pruned_loss=0.04592, over 3061541.70 frames. ], batch size: 89, lr: 2.39e-03, grad_scale: 8.0 2023-05-02 15:06:59,915 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5010, 3.7053, 3.8039, 2.6774, 3.3779, 3.7776, 3.5886, 2.2367], device='cuda:4'), covar=tensor([0.0502, 0.0080, 0.0059, 0.0396, 0.0143, 0.0113, 0.0096, 0.0534], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0088, 0.0090, 0.0133, 0.0101, 0.0114, 0.0097, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 15:07:26,987 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.421e+02 2.221e+02 2.637e+02 3.332e+02 6.864e+02, threshold=5.273e+02, percent-clipped=5.0 2023-05-02 15:07:40,679 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=282489.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:08:02,090 INFO [train.py:904] (4/8) Epoch 28, batch 8450, loss[loss=0.1671, simple_loss=0.2638, pruned_loss=0.03518, over 16238.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2737, pruned_loss=0.04435, over 3059275.01 frames. ], batch size: 165, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:08:11,392 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=282508.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:08:25,926 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7457, 4.0597, 3.0906, 2.3629, 2.4188, 2.6739, 4.3233, 3.4622], device='cuda:4'), covar=tensor([0.2922, 0.0516, 0.1836, 0.3166, 0.3249, 0.2187, 0.0358, 0.1341], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0270, 0.0308, 0.0324, 0.0301, 0.0274, 0.0301, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:09:17,297 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=282550.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:09:21,852 INFO [train.py:904] (4/8) Epoch 28, batch 8500, loss[loss=0.1785, simple_loss=0.2687, pruned_loss=0.04416, over 16860.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2695, pruned_loss=0.0421, over 3044065.76 frames. ], batch size: 116, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:09:27,449 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=282556.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:09:43,309 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=282566.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 15:10:07,072 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.338e+02 2.128e+02 2.662e+02 3.202e+02 5.740e+02, threshold=5.324e+02, percent-clipped=2.0 2023-05-02 15:10:42,399 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=282601.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:10:45,537 INFO [train.py:904] (4/8) Epoch 28, batch 8550, loss[loss=0.1705, simple_loss=0.2745, pruned_loss=0.03326, over 16700.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2675, pruned_loss=0.04101, over 3037858.17 frames. ], batch size: 89, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:11:05,484 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=282614.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 15:11:30,915 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1147, 4.1241, 4.4494, 4.4366, 4.4125, 4.1921, 4.1574, 4.1502], device='cuda:4'), covar=tensor([0.0394, 0.0667, 0.0478, 0.0447, 0.0521, 0.0483, 0.0956, 0.0569], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0485, 0.0469, 0.0433, 0.0515, 0.0493, 0.0569, 0.0398], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 15:11:42,719 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6005, 3.7530, 2.8249, 2.1940, 2.2523, 2.4502, 3.9814, 3.2652], device='cuda:4'), covar=tensor([0.3122, 0.0569, 0.1997, 0.3231, 0.3100, 0.2278, 0.0429, 0.1395], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0269, 0.0308, 0.0323, 0.0300, 0.0274, 0.0300, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:12:14,466 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=282649.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:12:21,560 INFO [train.py:904] (4/8) Epoch 28, batch 8600, loss[loss=0.1663, simple_loss=0.2618, pruned_loss=0.03533, over 16500.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2679, pruned_loss=0.04001, over 3040429.02 frames. ], batch size: 75, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:12:32,779 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8193, 3.7226, 3.8868, 3.9723, 4.0927, 3.6984, 4.0369, 4.1035], device='cuda:4'), covar=tensor([0.1708, 0.1198, 0.1325, 0.0750, 0.0575, 0.1863, 0.0790, 0.0775], device='cuda:4'), in_proj_covar=tensor([0.0654, 0.0802, 0.0925, 0.0815, 0.0622, 0.0649, 0.0681, 0.0789], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:12:32,802 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=282658.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:13:06,344 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6002, 3.8422, 2.8182, 2.1659, 2.3080, 2.4880, 4.0909, 3.2847], device='cuda:4'), covar=tensor([0.3221, 0.0596, 0.2080, 0.3296, 0.3098, 0.2358, 0.0409, 0.1430], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0269, 0.0309, 0.0324, 0.0300, 0.0275, 0.0301, 0.0347], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:13:12,556 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=282677.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:13:17,326 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.534e+02 2.220e+02 2.463e+02 3.072e+02 6.042e+02, threshold=4.927e+02, percent-clipped=1.0 2023-05-02 15:13:45,606 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.01 vs. limit=5.0 2023-05-02 15:13:50,830 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=282698.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 15:13:58,473 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.96 vs. limit=5.0 2023-05-02 15:13:58,782 INFO [train.py:904] (4/8) Epoch 28, batch 8650, loss[loss=0.174, simple_loss=0.2641, pruned_loss=0.04193, over 11940.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2664, pruned_loss=0.03902, over 3031644.36 frames. ], batch size: 248, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:14:16,201 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-05-02 15:14:36,887 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=282719.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:15:14,335 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=282738.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:15:40,252 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=282752.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:15:40,925 INFO [train.py:904] (4/8) Epoch 28, batch 8700, loss[loss=0.1689, simple_loss=0.2619, pruned_loss=0.03792, over 16723.00 frames. ], tot_loss[loss=0.17, simple_loss=0.264, pruned_loss=0.03805, over 3055571.75 frames. ], batch size: 134, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:15:53,194 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=282759.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 15:16:23,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1620, 2.5524, 2.6571, 1.9723, 2.8000, 2.8454, 2.5846, 2.5281], device='cuda:4'), covar=tensor([0.0658, 0.0274, 0.0252, 0.1003, 0.0124, 0.0271, 0.0446, 0.0446], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0108, 0.0099, 0.0135, 0.0083, 0.0127, 0.0126, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 15:16:29,664 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.456e+02 2.234e+02 2.610e+02 3.054e+02 5.139e+02, threshold=5.220e+02, percent-clipped=1.0 2023-05-02 15:17:06,889 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=282800.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:17:12,021 INFO [train.py:904] (4/8) Epoch 28, batch 8750, loss[loss=0.1581, simple_loss=0.2612, pruned_loss=0.02749, over 16848.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2636, pruned_loss=0.03738, over 3050148.80 frames. ], batch size: 90, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:17:18,895 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=282805.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:18:12,024 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8140, 3.1071, 3.5037, 2.0738, 2.9649, 2.1974, 3.2697, 3.2237], device='cuda:4'), covar=tensor([0.0282, 0.0968, 0.0467, 0.2144, 0.0790, 0.0970, 0.0714, 0.0989], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0167, 0.0168, 0.0155, 0.0145, 0.0130, 0.0144, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 15:18:46,445 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=282845.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:19:02,828 INFO [train.py:904] (4/8) Epoch 28, batch 8800, loss[loss=0.1816, simple_loss=0.2729, pruned_loss=0.04518, over 12370.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2622, pruned_loss=0.03638, over 3058221.51 frames. ], batch size: 248, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:19:29,127 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=282866.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:20:00,725 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.524e+02 2.123e+02 2.465e+02 3.098e+02 6.856e+02, threshold=4.929e+02, percent-clipped=4.0 2023-05-02 15:20:47,900 INFO [train.py:904] (4/8) Epoch 28, batch 8850, loss[loss=0.1551, simple_loss=0.246, pruned_loss=0.03208, over 12343.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2645, pruned_loss=0.03583, over 3056744.31 frames. ], batch size: 248, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:21:53,238 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1299, 5.1031, 4.7802, 4.1149, 4.9449, 1.8179, 4.6750, 4.5877], device='cuda:4'), covar=tensor([0.0089, 0.0086, 0.0251, 0.0383, 0.0093, 0.2975, 0.0146, 0.0270], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0169, 0.0207, 0.0180, 0.0183, 0.0213, 0.0196, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:22:35,071 INFO [train.py:904] (4/8) Epoch 28, batch 8900, loss[loss=0.163, simple_loss=0.2659, pruned_loss=0.03004, over 16251.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2651, pruned_loss=0.03494, over 3074130.80 frames. ], batch size: 165, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:23:38,028 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.086e+02 2.432e+02 2.867e+02 4.868e+02, threshold=4.864e+02, percent-clipped=0.0 2023-05-02 15:24:06,730 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8662, 3.9135, 4.0018, 3.8103, 3.9496, 4.3603, 4.0023, 3.7285], device='cuda:4'), covar=tensor([0.2077, 0.2182, 0.2340, 0.2385, 0.2702, 0.1579, 0.1623, 0.2616], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0616, 0.0680, 0.0498, 0.0669, 0.0707, 0.0531, 0.0671], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:24:09,663 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4143, 4.2695, 4.4226, 4.5814, 4.7543, 4.3634, 4.7547, 4.7516], device='cuda:4'), covar=tensor([0.1936, 0.1207, 0.1694, 0.0819, 0.0594, 0.0974, 0.0553, 0.0802], device='cuda:4'), in_proj_covar=tensor([0.0642, 0.0787, 0.0908, 0.0804, 0.0611, 0.0638, 0.0671, 0.0776], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:24:36,948 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5258, 3.7861, 2.7507, 2.1833, 2.2442, 2.3294, 3.9474, 3.1377], device='cuda:4'), covar=tensor([0.3317, 0.0536, 0.2024, 0.3200, 0.3036, 0.2435, 0.0420, 0.1457], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0269, 0.0308, 0.0323, 0.0299, 0.0274, 0.0300, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:24:39,871 INFO [train.py:904] (4/8) Epoch 28, batch 8950, loss[loss=0.1613, simple_loss=0.2565, pruned_loss=0.03304, over 16159.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2648, pruned_loss=0.03523, over 3091594.61 frames. ], batch size: 35, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:25:04,559 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=283014.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:25:12,649 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-05-02 15:25:45,368 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=283033.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:26:27,356 INFO [train.py:904] (4/8) Epoch 28, batch 9000, loss[loss=0.1501, simple_loss=0.2365, pruned_loss=0.03191, over 12268.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2619, pruned_loss=0.03416, over 3092097.60 frames. ], batch size: 248, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:26:27,357 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 15:26:38,046 INFO [train.py:938] (4/8) Epoch 28, validation: loss=0.1436, simple_loss=0.2472, pruned_loss=0.02006, over 944034.00 frames. 2023-05-02 15:26:38,047 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 15:26:41,524 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=283054.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 15:26:54,817 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3978, 4.6708, 4.5270, 4.5112, 4.2453, 4.2377, 4.2203, 4.7376], device='cuda:4'), covar=tensor([0.1140, 0.0957, 0.0922, 0.0889, 0.0825, 0.1506, 0.1155, 0.0864], device='cuda:4'), in_proj_covar=tensor([0.0701, 0.0848, 0.0694, 0.0655, 0.0535, 0.0538, 0.0705, 0.0660], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:27:36,920 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.442e+02 2.008e+02 2.381e+02 2.804e+02 4.985e+02, threshold=4.761e+02, percent-clipped=1.0 2023-05-02 15:28:21,106 INFO [train.py:904] (4/8) Epoch 28, batch 9050, loss[loss=0.1361, simple_loss=0.2282, pruned_loss=0.02202, over 16856.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2615, pruned_loss=0.03449, over 3077243.19 frames. ], batch size: 96, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:28:24,599 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6436, 3.7111, 3.4948, 3.1663, 3.3347, 3.6000, 3.3931, 3.4372], device='cuda:4'), covar=tensor([0.0552, 0.0628, 0.0278, 0.0245, 0.0480, 0.0484, 0.1200, 0.0480], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0455, 0.0354, 0.0353, 0.0349, 0.0408, 0.0244, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:28:48,323 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=283115.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:28:50,487 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.69 vs. limit=5.0 2023-05-02 15:29:01,977 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.52 vs. limit=5.0 2023-05-02 15:29:04,094 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6265, 3.7054, 3.4805, 3.1415, 3.2807, 3.5686, 3.3524, 3.4115], device='cuda:4'), covar=tensor([0.0578, 0.0732, 0.0323, 0.0286, 0.0485, 0.0526, 0.1546, 0.0496], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0456, 0.0354, 0.0354, 0.0349, 0.0408, 0.0244, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:29:36,467 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1683, 1.5351, 1.9512, 2.1145, 2.2047, 2.3812, 1.8415, 2.2859], device='cuda:4'), covar=tensor([0.0268, 0.0607, 0.0357, 0.0373, 0.0384, 0.0246, 0.0568, 0.0168], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0195, 0.0184, 0.0188, 0.0205, 0.0162, 0.0199, 0.0162], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:29:44,620 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.80 vs. limit=2.0 2023-05-02 15:29:46,371 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=283144.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:29:48,421 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=283145.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:30:04,160 INFO [train.py:904] (4/8) Epoch 28, batch 9100, loss[loss=0.1676, simple_loss=0.2623, pruned_loss=0.03648, over 16624.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2603, pruned_loss=0.03465, over 3050839.98 frames. ], batch size: 134, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:30:20,757 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=283161.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:30:58,218 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=283176.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:31:08,548 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.518e+02 2.173e+02 2.551e+02 2.852e+02 5.102e+02, threshold=5.103e+02, percent-clipped=2.0 2023-05-02 15:31:38,531 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=283193.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:31:54,945 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8636, 3.9189, 4.1480, 4.1421, 4.1325, 3.9532, 3.9469, 3.9383], device='cuda:4'), covar=tensor([0.0351, 0.0656, 0.0443, 0.0426, 0.0468, 0.0445, 0.0757, 0.0504], device='cuda:4'), in_proj_covar=tensor([0.0424, 0.0475, 0.0463, 0.0426, 0.0508, 0.0485, 0.0558, 0.0393], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 15:32:01,009 INFO [train.py:904] (4/8) Epoch 28, batch 9150, loss[loss=0.1697, simple_loss=0.2616, pruned_loss=0.03885, over 16346.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2606, pruned_loss=0.03414, over 3049060.05 frames. ], batch size: 146, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:32:06,356 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=283205.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:32:48,907 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-05-02 15:33:44,279 INFO [train.py:904] (4/8) Epoch 28, batch 9200, loss[loss=0.1466, simple_loss=0.244, pruned_loss=0.02456, over 12060.00 frames. ], tot_loss[loss=0.1617, simple_loss=0.2566, pruned_loss=0.03339, over 3049718.09 frames. ], batch size: 250, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:33:48,317 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 15:34:34,270 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.493e+02 2.118e+02 2.676e+02 3.077e+02 8.618e+02, threshold=5.351e+02, percent-clipped=1.0 2023-05-02 15:34:39,381 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=283283.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:35:20,274 INFO [train.py:904] (4/8) Epoch 28, batch 9250, loss[loss=0.1524, simple_loss=0.2512, pruned_loss=0.02679, over 15387.00 frames. ], tot_loss[loss=0.1616, simple_loss=0.2563, pruned_loss=0.03346, over 3050811.96 frames. ], batch size: 191, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:35:26,319 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8127, 3.8088, 3.9295, 3.6584, 3.9306, 4.3092, 3.9424, 3.6323], device='cuda:4'), covar=tensor([0.2136, 0.2576, 0.2749, 0.2569, 0.2941, 0.1736, 0.1751, 0.2664], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0611, 0.0675, 0.0495, 0.0664, 0.0703, 0.0526, 0.0665], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:35:44,169 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=283314.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:36:25,614 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-02 15:36:28,134 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=283333.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:36:36,690 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-02 15:36:44,840 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4779, 4.7308, 4.5581, 4.5695, 4.2642, 4.2839, 4.2650, 4.7617], device='cuda:4'), covar=tensor([0.0937, 0.0815, 0.0883, 0.0844, 0.0744, 0.1449, 0.1008, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0696, 0.0842, 0.0690, 0.0651, 0.0532, 0.0535, 0.0700, 0.0655], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:36:55,350 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=283344.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:37:14,581 INFO [train.py:904] (4/8) Epoch 28, batch 9300, loss[loss=0.1715, simple_loss=0.2588, pruned_loss=0.04212, over 15318.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2552, pruned_loss=0.03321, over 3068761.13 frames. ], batch size: 190, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:37:16,891 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=283354.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 15:37:34,850 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=283362.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:37:35,102 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4342, 3.0911, 2.7491, 2.2738, 2.1658, 2.3350, 3.0258, 2.8351], device='cuda:4'), covar=tensor([0.2726, 0.0670, 0.1712, 0.2933, 0.2933, 0.2366, 0.0470, 0.1537], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0268, 0.0307, 0.0321, 0.0297, 0.0272, 0.0298, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:37:35,127 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9183, 2.1448, 2.4140, 3.2000, 2.1724, 2.3372, 2.3306, 2.2656], device='cuda:4'), covar=tensor([0.1395, 0.3881, 0.3022, 0.0751, 0.4598, 0.2762, 0.3664, 0.3961], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0464, 0.0380, 0.0326, 0.0438, 0.0530, 0.0438, 0.0542], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:38:16,509 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.266e+02 1.947e+02 2.341e+02 2.771e+02 4.646e+02, threshold=4.683e+02, percent-clipped=0.0 2023-05-02 15:38:17,947 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=283381.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:38:58,082 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=283402.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 15:38:59,593 INFO [train.py:904] (4/8) Epoch 28, batch 9350, loss[loss=0.2011, simple_loss=0.288, pruned_loss=0.05713, over 16332.00 frames. ], tot_loss[loss=0.1619, simple_loss=0.256, pruned_loss=0.03386, over 3057881.24 frames. ], batch size: 146, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:40:41,025 INFO [train.py:904] (4/8) Epoch 28, batch 9400, loss[loss=0.1553, simple_loss=0.2536, pruned_loss=0.02843, over 15297.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2562, pruned_loss=0.03396, over 3049311.43 frames. ], batch size: 192, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:41:00,039 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=283461.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:41:19,507 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=283471.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:41:39,543 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.482e+02 2.116e+02 2.396e+02 2.984e+02 4.135e+02, threshold=4.791e+02, percent-clipped=0.0 2023-05-02 15:42:20,564 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=283500.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:42:24,700 INFO [train.py:904] (4/8) Epoch 28, batch 9450, loss[loss=0.1559, simple_loss=0.2503, pruned_loss=0.0307, over 16945.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2582, pruned_loss=0.03404, over 3065420.19 frames. ], batch size: 109, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:42:36,844 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=283509.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:43:04,121 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7246, 4.6750, 4.4804, 3.8373, 4.5898, 1.6772, 4.3398, 4.2019], device='cuda:4'), covar=tensor([0.0094, 0.0114, 0.0194, 0.0321, 0.0096, 0.2971, 0.0124, 0.0270], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0169, 0.0206, 0.0178, 0.0183, 0.0212, 0.0195, 0.0173], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:43:41,157 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2029, 4.0538, 4.2428, 4.3631, 4.4845, 4.0757, 4.4627, 4.5204], device='cuda:4'), covar=tensor([0.1737, 0.1159, 0.1426, 0.0720, 0.0522, 0.1230, 0.0683, 0.0748], device='cuda:4'), in_proj_covar=tensor([0.0641, 0.0783, 0.0904, 0.0799, 0.0608, 0.0633, 0.0668, 0.0772], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:43:48,948 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=283544.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 15:44:06,789 INFO [train.py:904] (4/8) Epoch 28, batch 9500, loss[loss=0.1461, simple_loss=0.2382, pruned_loss=0.02701, over 17141.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2579, pruned_loss=0.03403, over 3056968.10 frames. ], batch size: 47, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:44:32,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7652, 3.0945, 3.4031, 2.0422, 2.9065, 2.1659, 3.2928, 3.2534], device='cuda:4'), covar=tensor([0.0277, 0.0940, 0.0529, 0.2204, 0.0817, 0.1113, 0.0620, 0.0991], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0164, 0.0165, 0.0154, 0.0143, 0.0129, 0.0141, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 15:45:03,522 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 15:45:03,874 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.031e+02 2.419e+02 2.973e+02 5.266e+02, threshold=4.838e+02, percent-clipped=2.0 2023-05-02 15:45:10,594 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4156, 4.5305, 4.6654, 4.4395, 4.5626, 5.0536, 4.5873, 4.2512], device='cuda:4'), covar=tensor([0.1501, 0.2211, 0.2488, 0.2083, 0.2473, 0.0973, 0.1639, 0.2598], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0612, 0.0675, 0.0497, 0.0665, 0.0702, 0.0526, 0.0665], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:45:39,803 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-05-02 15:45:53,231 INFO [train.py:904] (4/8) Epoch 28, batch 9550, loss[loss=0.1602, simple_loss=0.2515, pruned_loss=0.03444, over 12474.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2576, pruned_loss=0.03395, over 3065340.31 frames. ], batch size: 246, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:45:59,240 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=283605.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 15:47:09,793 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=283639.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:47:34,670 INFO [train.py:904] (4/8) Epoch 28, batch 9600, loss[loss=0.1771, simple_loss=0.2718, pruned_loss=0.04117, over 16977.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2583, pruned_loss=0.03435, over 3076495.17 frames. ], batch size: 109, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:48:29,440 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 2.192e+02 2.621e+02 3.097e+02 5.622e+02, threshold=5.243e+02, percent-clipped=5.0 2023-05-02 15:48:58,332 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3987, 3.0100, 2.7333, 2.2817, 2.2230, 2.3527, 3.0723, 2.8439], device='cuda:4'), covar=tensor([0.2703, 0.0694, 0.1731, 0.2967, 0.2594, 0.2270, 0.0481, 0.1523], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0266, 0.0305, 0.0319, 0.0295, 0.0271, 0.0296, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 15:49:23,007 INFO [train.py:904] (4/8) Epoch 28, batch 9650, loss[loss=0.1398, simple_loss=0.2395, pruned_loss=0.02011, over 16471.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.259, pruned_loss=0.03409, over 3055544.85 frames. ], batch size: 68, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:51:10,207 INFO [train.py:904] (4/8) Epoch 28, batch 9700, loss[loss=0.1684, simple_loss=0.2598, pruned_loss=0.03855, over 16192.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2579, pruned_loss=0.0335, over 3061812.55 frames. ], batch size: 165, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:51:46,257 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 15:51:47,244 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=283771.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:52:08,986 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.467e+02 2.180e+02 2.423e+02 2.886e+02 5.203e+02, threshold=4.846e+02, percent-clipped=0.0 2023-05-02 15:52:48,128 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=283800.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:52:52,808 INFO [train.py:904] (4/8) Epoch 28, batch 9750, loss[loss=0.1639, simple_loss=0.2473, pruned_loss=0.04029, over 12208.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2575, pruned_loss=0.03428, over 3042633.62 frames. ], batch size: 247, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:53:24,762 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=283819.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:54:13,272 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=283842.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:54:24,070 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=283848.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:54:31,904 INFO [train.py:904] (4/8) Epoch 28, batch 9800, loss[loss=0.1513, simple_loss=0.2569, pruned_loss=0.0228, over 16792.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2576, pruned_loss=0.03343, over 3051855.85 frames. ], batch size: 83, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:54:41,146 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4257, 3.4756, 3.6992, 3.6779, 3.6921, 3.4945, 3.5478, 3.5926], device='cuda:4'), covar=tensor([0.0436, 0.0939, 0.0554, 0.0528, 0.0578, 0.0629, 0.0887, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0423, 0.0475, 0.0462, 0.0423, 0.0507, 0.0483, 0.0556, 0.0391], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 15:55:23,050 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.531e+02 2.061e+02 2.416e+02 3.000e+02 5.773e+02, threshold=4.831e+02, percent-clipped=3.0 2023-05-02 15:55:37,849 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4398, 3.3545, 3.3823, 3.5468, 3.5489, 3.3153, 3.5753, 3.6112], device='cuda:4'), covar=tensor([0.1342, 0.1144, 0.1353, 0.0853, 0.0848, 0.2558, 0.1054, 0.0962], device='cuda:4'), in_proj_covar=tensor([0.0640, 0.0784, 0.0903, 0.0799, 0.0607, 0.0633, 0.0667, 0.0773], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:56:11,008 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=283900.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 15:56:15,638 INFO [train.py:904] (4/8) Epoch 28, batch 9850, loss[loss=0.1638, simple_loss=0.2618, pruned_loss=0.03289, over 15477.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2586, pruned_loss=0.03321, over 3052593.90 frames. ], batch size: 191, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:56:17,432 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=283903.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:57:36,905 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=283939.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:57:51,611 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1606, 1.5711, 1.9293, 2.1358, 2.2347, 2.3579, 1.7957, 2.2919], device='cuda:4'), covar=tensor([0.0287, 0.0630, 0.0342, 0.0385, 0.0399, 0.0296, 0.0608, 0.0185], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0194, 0.0181, 0.0185, 0.0203, 0.0160, 0.0198, 0.0160], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 15:58:06,899 INFO [train.py:904] (4/8) Epoch 28, batch 9900, loss[loss=0.1679, simple_loss=0.2689, pruned_loss=0.03338, over 16252.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2593, pruned_loss=0.03289, over 3056988.57 frames. ], batch size: 165, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 15:58:38,566 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=283966.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 15:59:13,294 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.328e+02 2.134e+02 2.412e+02 2.887e+02 8.278e+02, threshold=4.823e+02, percent-clipped=3.0 2023-05-02 15:59:30,051 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=283987.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:00:07,766 INFO [train.py:904] (4/8) Epoch 28, batch 9950, loss[loss=0.172, simple_loss=0.2738, pruned_loss=0.0351, over 15346.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2612, pruned_loss=0.03329, over 3054891.69 frames. ], batch size: 190, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 16:00:44,602 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6928, 3.8614, 2.3327, 4.2164, 2.9520, 4.1356, 2.4308, 3.1198], device='cuda:4'), covar=tensor([0.0302, 0.0318, 0.1687, 0.0327, 0.0770, 0.0540, 0.1728, 0.0718], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0176, 0.0193, 0.0166, 0.0176, 0.0213, 0.0201, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 16:01:08,227 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=284027.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:02:08,461 INFO [train.py:904] (4/8) Epoch 28, batch 10000, loss[loss=0.1532, simple_loss=0.257, pruned_loss=0.02469, over 15487.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2596, pruned_loss=0.03277, over 3079626.58 frames. ], batch size: 192, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 16:02:52,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9197, 3.7070, 3.6525, 4.0338, 4.1007, 3.8305, 4.0680, 4.1349], device='cuda:4'), covar=tensor([0.1854, 0.1725, 0.2959, 0.1429, 0.1079, 0.2289, 0.1499, 0.1519], device='cuda:4'), in_proj_covar=tensor([0.0637, 0.0780, 0.0898, 0.0795, 0.0604, 0.0630, 0.0665, 0.0770], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:03:03,885 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.562e+02 2.120e+02 2.490e+02 3.242e+02 7.022e+02, threshold=4.980e+02, percent-clipped=2.0 2023-05-02 16:03:50,387 INFO [train.py:904] (4/8) Epoch 28, batch 10050, loss[loss=0.1566, simple_loss=0.2591, pruned_loss=0.02699, over 15256.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2599, pruned_loss=0.03285, over 3087310.02 frames. ], batch size: 190, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 16:05:23,769 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6131, 3.6552, 3.4931, 3.1772, 3.3655, 3.6013, 3.3366, 3.3874], device='cuda:4'), covar=tensor([0.0682, 0.0942, 0.0349, 0.0339, 0.0560, 0.0940, 0.1449, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0446, 0.0348, 0.0347, 0.0341, 0.0401, 0.0239, 0.0413], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:05:24,422 INFO [train.py:904] (4/8) Epoch 28, batch 10100, loss[loss=0.1578, simple_loss=0.2426, pruned_loss=0.0365, over 12535.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2602, pruned_loss=0.03301, over 3071488.92 frames. ], batch size: 247, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 16:06:20,723 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.509e+02 2.222e+02 2.734e+02 3.195e+02 6.018e+02, threshold=5.468e+02, percent-clipped=8.0 2023-05-02 16:06:39,386 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=284198.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:06:42,000 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=284200.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 16:06:44,464 INFO [train.py:904] (4/8) Epoch 28, batch 10150, loss[loss=0.1612, simple_loss=0.2459, pruned_loss=0.03824, over 12113.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2596, pruned_loss=0.03355, over 3057429.38 frames. ], batch size: 248, lr: 2.38e-03, grad_scale: 8.0 2023-05-02 16:07:10,315 INFO [train.py:904] (4/8) Epoch 29, batch 0, loss[loss=0.1882, simple_loss=0.2859, pruned_loss=0.04521, over 17040.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2859, pruned_loss=0.04521, over 17040.00 frames. ], batch size: 50, lr: 2.34e-03, grad_scale: 8.0 2023-05-02 16:07:10,315 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 16:07:17,743 INFO [train.py:938] (4/8) Epoch 29, validation: loss=0.1427, simple_loss=0.246, pruned_loss=0.0197, over 944034.00 frames. 2023-05-02 16:07:17,743 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 16:07:53,512 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-02 16:08:07,725 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6004, 3.8347, 4.0055, 2.8514, 3.5638, 4.0631, 3.6900, 2.2484], device='cuda:4'), covar=tensor([0.0578, 0.0378, 0.0078, 0.0428, 0.0165, 0.0128, 0.0134, 0.0588], device='cuda:4'), in_proj_covar=tensor([0.0135, 0.0087, 0.0088, 0.0132, 0.0100, 0.0111, 0.0095, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 16:08:10,068 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5650, 3.5525, 4.1085, 2.4179, 3.3418, 2.6397, 3.9293, 3.8763], device='cuda:4'), covar=tensor([0.0285, 0.1143, 0.0549, 0.2207, 0.0893, 0.1125, 0.0676, 0.1383], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0163, 0.0165, 0.0153, 0.0143, 0.0129, 0.0141, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 16:08:18,234 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=284248.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 16:08:26,888 INFO [train.py:904] (4/8) Epoch 29, batch 50, loss[loss=0.1837, simple_loss=0.2792, pruned_loss=0.0441, over 17012.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2649, pruned_loss=0.04434, over 749546.47 frames. ], batch size: 55, lr: 2.34e-03, grad_scale: 2.0 2023-05-02 16:08:28,878 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 16:09:08,277 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.709e+02 2.422e+02 3.076e+02 3.842e+02 2.178e+03, threshold=6.152e+02, percent-clipped=5.0 2023-05-02 16:09:37,377 INFO [train.py:904] (4/8) Epoch 29, batch 100, loss[loss=0.1695, simple_loss=0.2683, pruned_loss=0.03536, over 17182.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2597, pruned_loss=0.04146, over 1316967.52 frames. ], batch size: 46, lr: 2.34e-03, grad_scale: 2.0 2023-05-02 16:10:02,631 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=284322.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:10:46,097 INFO [train.py:904] (4/8) Epoch 29, batch 150, loss[loss=0.1948, simple_loss=0.2709, pruned_loss=0.05937, over 16714.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2592, pruned_loss=0.04087, over 1765073.34 frames. ], batch size: 134, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:11:25,625 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.248e+02 2.614e+02 3.025e+02 1.081e+03, threshold=5.228e+02, percent-clipped=2.0 2023-05-02 16:11:49,871 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-05-02 16:11:53,383 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0126, 4.7768, 4.9601, 5.1984, 5.3864, 4.8157, 5.3498, 5.4041], device='cuda:4'), covar=tensor([0.2246, 0.1705, 0.2424, 0.1126, 0.0803, 0.0893, 0.0757, 0.0877], device='cuda:4'), in_proj_covar=tensor([0.0652, 0.0797, 0.0918, 0.0811, 0.0617, 0.0642, 0.0681, 0.0787], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:11:55,140 INFO [train.py:904] (4/8) Epoch 29, batch 200, loss[loss=0.1756, simple_loss=0.271, pruned_loss=0.04015, over 17063.00 frames. ], tot_loss[loss=0.17, simple_loss=0.259, pruned_loss=0.04056, over 2110217.09 frames. ], batch size: 55, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:12:51,470 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=284443.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:13:04,663 INFO [train.py:904] (4/8) Epoch 29, batch 250, loss[loss=0.1469, simple_loss=0.2381, pruned_loss=0.02788, over 17240.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2571, pruned_loss=0.03989, over 2385648.81 frames. ], batch size: 43, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:13:44,133 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.52 vs. limit=5.0 2023-05-02 16:13:47,576 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.177e+02 2.500e+02 3.084e+02 6.085e+02, threshold=5.001e+02, percent-clipped=1.0 2023-05-02 16:14:08,303 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=284498.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:14:16,992 INFO [train.py:904] (4/8) Epoch 29, batch 300, loss[loss=0.1352, simple_loss=0.2287, pruned_loss=0.0208, over 17231.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2549, pruned_loss=0.03949, over 2588738.40 frames. ], batch size: 44, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:14:17,383 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=284504.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:15:14,272 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=284546.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:15:23,224 INFO [train.py:904] (4/8) Epoch 29, batch 350, loss[loss=0.1644, simple_loss=0.2499, pruned_loss=0.0394, over 16728.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2521, pruned_loss=0.03862, over 2748224.23 frames. ], batch size: 134, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:16:02,935 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.536e+02 2.083e+02 2.374e+02 2.884e+02 4.611e+02, threshold=4.748e+02, percent-clipped=0.0 2023-05-02 16:16:31,992 INFO [train.py:904] (4/8) Epoch 29, batch 400, loss[loss=0.1638, simple_loss=0.2422, pruned_loss=0.04268, over 16865.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.25, pruned_loss=0.03792, over 2885123.28 frames. ], batch size: 83, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:16:57,124 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=284622.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:17:38,901 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8078, 2.7341, 2.1811, 2.3320, 3.0655, 2.8254, 3.4139, 3.3804], device='cuda:4'), covar=tensor([0.0197, 0.0603, 0.0764, 0.0693, 0.0403, 0.0501, 0.0328, 0.0340], device='cuda:4'), in_proj_covar=tensor([0.0231, 0.0246, 0.0236, 0.0236, 0.0246, 0.0245, 0.0240, 0.0243], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:17:41,206 INFO [train.py:904] (4/8) Epoch 29, batch 450, loss[loss=0.1738, simple_loss=0.2593, pruned_loss=0.0442, over 16428.00 frames. ], tot_loss[loss=0.1617, simple_loss=0.2489, pruned_loss=0.03728, over 2985792.17 frames. ], batch size: 68, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:17:44,845 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9239, 1.4661, 1.7494, 1.8019, 1.8847, 2.0009, 1.7432, 1.8793], device='cuda:4'), covar=tensor([0.0266, 0.0496, 0.0260, 0.0354, 0.0324, 0.0245, 0.0506, 0.0200], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0199, 0.0187, 0.0192, 0.0209, 0.0166, 0.0203, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:18:02,994 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=284670.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:18:18,521 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.454e+02 1.978e+02 2.312e+02 2.858e+02 4.933e+02, threshold=4.623e+02, percent-clipped=2.0 2023-05-02 16:18:47,400 INFO [train.py:904] (4/8) Epoch 29, batch 500, loss[loss=0.1572, simple_loss=0.2386, pruned_loss=0.03793, over 16839.00 frames. ], tot_loss[loss=0.1612, simple_loss=0.2483, pruned_loss=0.03701, over 3060665.31 frames. ], batch size: 96, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:19:56,201 INFO [train.py:904] (4/8) Epoch 29, batch 550, loss[loss=0.1831, simple_loss=0.2613, pruned_loss=0.0525, over 16772.00 frames. ], tot_loss[loss=0.1607, simple_loss=0.2473, pruned_loss=0.03701, over 3118836.69 frames. ], batch size: 124, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:20:33,818 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9688, 2.0825, 2.2946, 3.5405, 2.1516, 2.2955, 2.2323, 2.2241], device='cuda:4'), covar=tensor([0.1687, 0.3937, 0.3276, 0.0839, 0.4325, 0.2880, 0.3878, 0.3638], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0473, 0.0388, 0.0334, 0.0446, 0.0540, 0.0446, 0.0553], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:20:35,714 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.399e+02 1.886e+02 2.233e+02 2.634e+02 5.354e+02, threshold=4.465e+02, percent-clipped=1.0 2023-05-02 16:20:57,907 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=284799.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:21:04,186 INFO [train.py:904] (4/8) Epoch 29, batch 600, loss[loss=0.1765, simple_loss=0.2595, pruned_loss=0.04677, over 17079.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2466, pruned_loss=0.03751, over 3154055.71 frames. ], batch size: 53, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:21:22,014 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=284816.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:22:12,346 INFO [train.py:904] (4/8) Epoch 29, batch 650, loss[loss=0.1712, simple_loss=0.2483, pruned_loss=0.047, over 16941.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2451, pruned_loss=0.03732, over 3188613.46 frames. ], batch size: 109, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:22:17,375 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8545, 4.2706, 3.0303, 2.3741, 2.6681, 2.6382, 4.6142, 3.4104], device='cuda:4'), covar=tensor([0.2894, 0.0573, 0.1941, 0.3057, 0.2972, 0.2191, 0.0358, 0.1581], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0273, 0.0314, 0.0327, 0.0303, 0.0279, 0.0303, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 16:22:37,685 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4221, 4.4446, 4.7903, 4.7656, 4.7872, 4.4950, 4.4752, 4.4127], device='cuda:4'), covar=tensor([0.0466, 0.0752, 0.0437, 0.0435, 0.0501, 0.0532, 0.0949, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0437, 0.0492, 0.0475, 0.0438, 0.0523, 0.0499, 0.0571, 0.0401], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 16:22:46,701 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=284877.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:22:53,843 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.402e+02 2.111e+02 2.518e+02 3.065e+02 7.660e+02, threshold=5.037e+02, percent-clipped=4.0 2023-05-02 16:23:22,615 INFO [train.py:904] (4/8) Epoch 29, batch 700, loss[loss=0.1596, simple_loss=0.2398, pruned_loss=0.03967, over 16696.00 frames. ], tot_loss[loss=0.1591, simple_loss=0.2448, pruned_loss=0.03672, over 3225236.26 frames. ], batch size: 89, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:24:04,503 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2593, 2.3511, 2.3276, 3.9631, 2.2709, 2.6823, 2.3960, 2.5213], device='cuda:4'), covar=tensor([0.1566, 0.3718, 0.3376, 0.0701, 0.4331, 0.2741, 0.3800, 0.3693], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0474, 0.0389, 0.0336, 0.0447, 0.0542, 0.0448, 0.0555], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:24:12,247 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=284939.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:24:30,116 INFO [train.py:904] (4/8) Epoch 29, batch 750, loss[loss=0.136, simple_loss=0.2243, pruned_loss=0.02387, over 16854.00 frames. ], tot_loss[loss=0.1595, simple_loss=0.2457, pruned_loss=0.03667, over 3251794.72 frames. ], batch size: 42, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:25:13,140 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.375e+02 2.091e+02 2.322e+02 2.629e+02 4.322e+02, threshold=4.644e+02, percent-clipped=0.0 2023-05-02 16:25:30,051 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=284996.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:25:37,941 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=285000.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:25:42,931 INFO [train.py:904] (4/8) Epoch 29, batch 800, loss[loss=0.1685, simple_loss=0.2664, pruned_loss=0.03526, over 17129.00 frames. ], tot_loss[loss=0.1604, simple_loss=0.2464, pruned_loss=0.03721, over 3272859.52 frames. ], batch size: 49, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:26:08,943 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.6071, 6.0143, 5.7524, 5.8383, 5.3961, 5.5218, 5.3775, 6.1311], device='cuda:4'), covar=tensor([0.1469, 0.1005, 0.1035, 0.0865, 0.0974, 0.0602, 0.1373, 0.0957], device='cuda:4'), in_proj_covar=tensor([0.0720, 0.0869, 0.0712, 0.0673, 0.0549, 0.0549, 0.0730, 0.0681], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:26:34,282 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=285042.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 16:26:49,075 INFO [train.py:904] (4/8) Epoch 29, batch 850, loss[loss=0.1484, simple_loss=0.2339, pruned_loss=0.03141, over 16837.00 frames. ], tot_loss[loss=0.1596, simple_loss=0.2459, pruned_loss=0.03671, over 3285620.57 frames. ], batch size: 96, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:26:53,006 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=285057.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:26:54,210 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9602, 2.6638, 2.7404, 4.3425, 3.5809, 4.1575, 1.6439, 3.1081], device='cuda:4'), covar=tensor([0.1403, 0.0766, 0.1158, 0.0168, 0.0177, 0.0372, 0.1710, 0.0834], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0180, 0.0200, 0.0203, 0.0204, 0.0217, 0.0210, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 16:27:00,069 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7235, 2.5217, 2.1264, 2.2987, 2.8634, 2.6265, 2.7570, 2.9228], device='cuda:4'), covar=tensor([0.0287, 0.0462, 0.0562, 0.0463, 0.0261, 0.0335, 0.0240, 0.0313], device='cuda:4'), in_proj_covar=tensor([0.0232, 0.0246, 0.0236, 0.0236, 0.0247, 0.0246, 0.0241, 0.0245], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:27:17,995 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-02 16:27:31,496 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.555e+02 2.142e+02 2.515e+02 3.015e+02 5.883e+02, threshold=5.030e+02, percent-clipped=4.0 2023-05-02 16:27:49,517 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=285099.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:27:54,229 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=285103.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 16:27:55,639 INFO [train.py:904] (4/8) Epoch 29, batch 900, loss[loss=0.1529, simple_loss=0.2471, pruned_loss=0.02928, over 17142.00 frames. ], tot_loss[loss=0.1587, simple_loss=0.2453, pruned_loss=0.03605, over 3305184.93 frames. ], batch size: 47, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:28:56,110 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=285147.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:29:05,599 INFO [train.py:904] (4/8) Epoch 29, batch 950, loss[loss=0.1414, simple_loss=0.2219, pruned_loss=0.03043, over 16772.00 frames. ], tot_loss[loss=0.1586, simple_loss=0.2451, pruned_loss=0.036, over 3312092.05 frames. ], batch size: 83, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:29:30,490 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=285172.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:29:47,163 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.394e+02 2.122e+02 2.524e+02 3.029e+02 7.170e+02, threshold=5.047e+02, percent-clipped=3.0 2023-05-02 16:30:14,195 INFO [train.py:904] (4/8) Epoch 29, batch 1000, loss[loss=0.1587, simple_loss=0.2548, pruned_loss=0.03132, over 17097.00 frames. ], tot_loss[loss=0.1579, simple_loss=0.2443, pruned_loss=0.03576, over 3300713.21 frames. ], batch size: 49, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:30:35,929 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5142, 3.4433, 4.0658, 2.2304, 3.2882, 2.5722, 3.9163, 3.7853], device='cuda:4'), covar=tensor([0.0266, 0.1090, 0.0499, 0.2194, 0.0864, 0.1050, 0.0582, 0.1109], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0171, 0.0158, 0.0148, 0.0133, 0.0146, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 16:31:24,214 INFO [train.py:904] (4/8) Epoch 29, batch 1050, loss[loss=0.1444, simple_loss=0.2303, pruned_loss=0.02924, over 16847.00 frames. ], tot_loss[loss=0.1576, simple_loss=0.2435, pruned_loss=0.03582, over 3301364.71 frames. ], batch size: 39, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:32:05,304 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.383e+02 1.938e+02 2.240e+02 2.554e+02 8.528e+02, threshold=4.479e+02, percent-clipped=1.0 2023-05-02 16:32:19,794 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=285295.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:32:31,241 INFO [train.py:904] (4/8) Epoch 29, batch 1100, loss[loss=0.1406, simple_loss=0.2207, pruned_loss=0.03026, over 16462.00 frames. ], tot_loss[loss=0.1574, simple_loss=0.2439, pruned_loss=0.03552, over 3301148.80 frames. ], batch size: 75, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:33:20,394 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-05-02 16:33:37,332 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=285352.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:33:40,187 INFO [train.py:904] (4/8) Epoch 29, batch 1150, loss[loss=0.1352, simple_loss=0.221, pruned_loss=0.02473, over 15972.00 frames. ], tot_loss[loss=0.1563, simple_loss=0.2425, pruned_loss=0.03509, over 3303581.81 frames. ], batch size: 35, lr: 2.33e-03, grad_scale: 2.0 2023-05-02 16:34:21,637 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7103, 4.0915, 2.9435, 2.3858, 2.6633, 2.6345, 4.3491, 3.4488], device='cuda:4'), covar=tensor([0.3076, 0.0620, 0.2095, 0.3064, 0.2861, 0.2254, 0.0441, 0.1549], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0274, 0.0314, 0.0328, 0.0305, 0.0280, 0.0305, 0.0354], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 16:34:22,232 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.496e+02 2.113e+02 2.491e+02 2.895e+02 1.258e+03, threshold=4.983e+02, percent-clipped=2.0 2023-05-02 16:34:40,261 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=285398.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 16:34:47,458 INFO [train.py:904] (4/8) Epoch 29, batch 1200, loss[loss=0.1615, simple_loss=0.2353, pruned_loss=0.04386, over 16231.00 frames. ], tot_loss[loss=0.1558, simple_loss=0.2418, pruned_loss=0.03493, over 3305193.23 frames. ], batch size: 165, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:35:22,825 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 16:35:31,714 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-02 16:35:37,968 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6385, 2.6438, 2.3642, 2.3913, 2.9369, 2.6676, 3.2521, 3.1355], device='cuda:4'), covar=tensor([0.0198, 0.0494, 0.0549, 0.0548, 0.0327, 0.0481, 0.0265, 0.0321], device='cuda:4'), in_proj_covar=tensor([0.0234, 0.0247, 0.0237, 0.0237, 0.0248, 0.0247, 0.0243, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:35:56,648 INFO [train.py:904] (4/8) Epoch 29, batch 1250, loss[loss=0.1431, simple_loss=0.2265, pruned_loss=0.02984, over 16770.00 frames. ], tot_loss[loss=0.1568, simple_loss=0.2427, pruned_loss=0.03541, over 3316226.80 frames. ], batch size: 102, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:36:21,858 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=285472.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:36:38,699 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.468e+02 2.107e+02 2.455e+02 3.038e+02 4.730e+02, threshold=4.911e+02, percent-clipped=0.0 2023-05-02 16:37:05,018 INFO [train.py:904] (4/8) Epoch 29, batch 1300, loss[loss=0.1651, simple_loss=0.2561, pruned_loss=0.03702, over 17049.00 frames. ], tot_loss[loss=0.1566, simple_loss=0.2426, pruned_loss=0.03534, over 3314066.24 frames. ], batch size: 55, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:37:27,839 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=285520.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:37:36,050 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.79 vs. limit=5.0 2023-05-02 16:38:13,879 INFO [train.py:904] (4/8) Epoch 29, batch 1350, loss[loss=0.1514, simple_loss=0.2467, pruned_loss=0.02805, over 17204.00 frames. ], tot_loss[loss=0.157, simple_loss=0.2433, pruned_loss=0.03538, over 3304425.69 frames. ], batch size: 46, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:38:21,625 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5557, 3.1284, 3.4837, 2.0494, 3.5538, 3.5730, 3.0382, 2.7907], device='cuda:4'), covar=tensor([0.0724, 0.0298, 0.0205, 0.1102, 0.0131, 0.0237, 0.0452, 0.0457], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0103, 0.0141, 0.0087, 0.0133, 0.0131, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 16:38:58,360 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.516e+02 2.007e+02 2.357e+02 2.705e+02 4.553e+02, threshold=4.714e+02, percent-clipped=0.0 2023-05-02 16:39:04,697 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6770, 2.7089, 2.4915, 2.5866, 2.9616, 2.8177, 3.2294, 3.2503], device='cuda:4'), covar=tensor([0.0206, 0.0510, 0.0562, 0.0530, 0.0350, 0.0481, 0.0349, 0.0302], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0248, 0.0237, 0.0238, 0.0248, 0.0248, 0.0244, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:39:11,812 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=285595.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:39:24,823 INFO [train.py:904] (4/8) Epoch 29, batch 1400, loss[loss=0.1688, simple_loss=0.2422, pruned_loss=0.04768, over 16711.00 frames. ], tot_loss[loss=0.1573, simple_loss=0.2434, pruned_loss=0.03555, over 3300729.46 frames. ], batch size: 124, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:39:49,649 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=285622.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:40:19,786 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=285643.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:40:31,657 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=285652.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:40:33,553 INFO [train.py:904] (4/8) Epoch 29, batch 1450, loss[loss=0.169, simple_loss=0.2628, pruned_loss=0.03755, over 17271.00 frames. ], tot_loss[loss=0.1577, simple_loss=0.2434, pruned_loss=0.03602, over 3299789.48 frames. ], batch size: 52, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:40:52,165 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-05-02 16:41:13,944 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=285683.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:41:16,389 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.632e+02 2.247e+02 2.721e+02 3.292e+02 6.741e+02, threshold=5.441e+02, percent-clipped=4.0 2023-05-02 16:41:34,412 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=285698.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 16:41:37,506 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=285700.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:41:43,180 INFO [train.py:904] (4/8) Epoch 29, batch 1500, loss[loss=0.1739, simple_loss=0.253, pruned_loss=0.04742, over 12020.00 frames. ], tot_loss[loss=0.1584, simple_loss=0.2439, pruned_loss=0.0364, over 3300774.53 frames. ], batch size: 247, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:41:45,287 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4938, 4.5640, 4.8783, 4.8660, 4.9031, 4.5858, 4.5812, 4.4482], device='cuda:4'), covar=tensor([0.0469, 0.0743, 0.0452, 0.0452, 0.0532, 0.0495, 0.0920, 0.0766], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0505, 0.0487, 0.0448, 0.0534, 0.0511, 0.0586, 0.0410], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 16:42:19,606 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4426, 5.4369, 5.1304, 4.6207, 5.2228, 2.2266, 4.9960, 4.9850], device='cuda:4'), covar=tensor([0.0087, 0.0084, 0.0250, 0.0421, 0.0110, 0.2658, 0.0149, 0.0252], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0176, 0.0215, 0.0185, 0.0192, 0.0220, 0.0202, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:42:40,635 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=285746.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 16:42:51,734 INFO [train.py:904] (4/8) Epoch 29, batch 1550, loss[loss=0.1695, simple_loss=0.2508, pruned_loss=0.04409, over 16738.00 frames. ], tot_loss[loss=0.16, simple_loss=0.2451, pruned_loss=0.0374, over 3315938.96 frames. ], batch size: 134, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:43:34,663 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.318e+02 2.209e+02 2.639e+02 3.034e+02 6.721e+02, threshold=5.278e+02, percent-clipped=2.0 2023-05-02 16:44:00,607 INFO [train.py:904] (4/8) Epoch 29, batch 1600, loss[loss=0.1509, simple_loss=0.24, pruned_loss=0.03091, over 17213.00 frames. ], tot_loss[loss=0.1612, simple_loss=0.2468, pruned_loss=0.03785, over 3310361.97 frames. ], batch size: 44, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:44:04,313 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7256, 3.8514, 2.1319, 4.4039, 2.9324, 4.3016, 2.1132, 3.0187], device='cuda:4'), covar=tensor([0.0364, 0.0386, 0.2061, 0.0381, 0.0917, 0.0447, 0.2181, 0.0882], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0183, 0.0199, 0.0177, 0.0182, 0.0223, 0.0208, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 16:44:33,766 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6523, 2.7239, 2.7482, 4.4802, 2.6279, 3.0469, 2.6800, 2.8010], device='cuda:4'), covar=tensor([0.1328, 0.3365, 0.2999, 0.0574, 0.3830, 0.2548, 0.3659, 0.3425], device='cuda:4'), in_proj_covar=tensor([0.0424, 0.0478, 0.0391, 0.0339, 0.0449, 0.0548, 0.0451, 0.0562], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:45:03,080 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4136, 3.5898, 3.9753, 2.2553, 3.1671, 2.6424, 3.8013, 3.8008], device='cuda:4'), covar=tensor([0.0286, 0.0977, 0.0493, 0.2115, 0.0853, 0.0981, 0.0608, 0.1124], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0171, 0.0171, 0.0158, 0.0149, 0.0134, 0.0146, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 16:45:05,265 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-05-02 16:45:09,516 INFO [train.py:904] (4/8) Epoch 29, batch 1650, loss[loss=0.161, simple_loss=0.242, pruned_loss=0.04005, over 16699.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2488, pruned_loss=0.03868, over 3318591.92 frames. ], batch size: 124, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:45:50,989 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.518e+02 2.202e+02 2.447e+02 3.087e+02 4.596e+02, threshold=4.895e+02, percent-clipped=0.0 2023-05-02 16:46:16,869 INFO [train.py:904] (4/8) Epoch 29, batch 1700, loss[loss=0.1513, simple_loss=0.2407, pruned_loss=0.03092, over 16760.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2498, pruned_loss=0.03914, over 3315204.31 frames. ], batch size: 39, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:47:18,516 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7587, 2.3838, 1.9196, 2.2004, 2.7422, 2.5486, 2.7126, 2.8741], device='cuda:4'), covar=tensor([0.0247, 0.0483, 0.0633, 0.0493, 0.0288, 0.0400, 0.0235, 0.0306], device='cuda:4'), in_proj_covar=tensor([0.0238, 0.0251, 0.0240, 0.0240, 0.0251, 0.0250, 0.0247, 0.0250], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:47:24,308 INFO [train.py:904] (4/8) Epoch 29, batch 1750, loss[loss=0.1817, simple_loss=0.261, pruned_loss=0.05121, over 16455.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2512, pruned_loss=0.03943, over 3301580.21 frames. ], batch size: 146, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:47:49,279 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2359, 5.2218, 5.0680, 4.4531, 4.7122, 5.1357, 5.1153, 4.6993], device='cuda:4'), covar=tensor([0.0664, 0.0539, 0.0403, 0.0494, 0.1246, 0.0568, 0.0367, 0.0985], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0485, 0.0377, 0.0380, 0.0374, 0.0436, 0.0260, 0.0451], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 16:47:58,305 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=285978.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:48:07,053 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.447e+02 2.051e+02 2.355e+02 2.951e+02 5.834e+02, threshold=4.710e+02, percent-clipped=3.0 2023-05-02 16:48:36,742 INFO [train.py:904] (4/8) Epoch 29, batch 1800, loss[loss=0.1667, simple_loss=0.2619, pruned_loss=0.03578, over 16667.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2518, pruned_loss=0.0388, over 3308411.23 frames. ], batch size: 57, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:48:55,266 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0407, 3.9937, 3.9007, 3.1890, 3.9467, 1.7381, 3.7225, 3.3130], device='cuda:4'), covar=tensor([0.0173, 0.0164, 0.0245, 0.0333, 0.0119, 0.3227, 0.0160, 0.0361], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0177, 0.0215, 0.0186, 0.0192, 0.0220, 0.0203, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 16:49:03,075 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9011, 4.8172, 4.7704, 4.3635, 4.4970, 4.8315, 4.6369, 4.5206], device='cuda:4'), covar=tensor([0.0669, 0.0847, 0.0352, 0.0400, 0.0950, 0.0528, 0.0489, 0.0736], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0486, 0.0378, 0.0380, 0.0375, 0.0437, 0.0260, 0.0452], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 16:49:18,348 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-05-02 16:49:43,955 INFO [train.py:904] (4/8) Epoch 29, batch 1850, loss[loss=0.1578, simple_loss=0.2496, pruned_loss=0.03302, over 17127.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2527, pruned_loss=0.03862, over 3314468.94 frames. ], batch size: 48, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:49:44,504 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8350, 4.0633, 2.8239, 4.6983, 3.3049, 4.5599, 2.9517, 3.3906], device='cuda:4'), covar=tensor([0.0383, 0.0429, 0.1609, 0.0317, 0.0836, 0.0574, 0.1480, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0184, 0.0200, 0.0178, 0.0182, 0.0224, 0.0208, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 16:50:28,094 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 2.081e+02 2.345e+02 2.768e+02 6.474e+02, threshold=4.689e+02, percent-clipped=2.0 2023-05-02 16:50:32,584 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0202, 5.0965, 5.5231, 5.4938, 5.5066, 5.1439, 5.1057, 4.9362], device='cuda:4'), covar=tensor([0.0378, 0.0571, 0.0375, 0.0407, 0.0520, 0.0451, 0.0967, 0.0477], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0506, 0.0489, 0.0450, 0.0537, 0.0513, 0.0589, 0.0412], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 16:50:53,098 INFO [train.py:904] (4/8) Epoch 29, batch 1900, loss[loss=0.161, simple_loss=0.2603, pruned_loss=0.03082, over 17095.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2516, pruned_loss=0.03798, over 3313699.86 frames. ], batch size: 53, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:51:20,263 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=286123.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 16:51:26,351 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8942, 2.0531, 2.5300, 2.8417, 2.8439, 3.4143, 2.4557, 3.4150], device='cuda:4'), covar=tensor([0.0337, 0.0694, 0.0446, 0.0470, 0.0414, 0.0257, 0.0573, 0.0214], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0201, 0.0189, 0.0195, 0.0212, 0.0169, 0.0205, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 16:51:29,329 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=286129.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:52:04,245 INFO [train.py:904] (4/8) Epoch 29, batch 1950, loss[loss=0.1703, simple_loss=0.2565, pruned_loss=0.04208, over 16870.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2523, pruned_loss=0.03805, over 3299622.06 frames. ], batch size: 83, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:52:46,911 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=286184.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 16:52:48,709 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.778e+02 2.194e+02 2.537e+02 3.082e+02 2.053e+03, threshold=5.073e+02, percent-clipped=1.0 2023-05-02 16:52:55,331 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=286190.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:53:10,064 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5099, 3.6281, 3.9197, 2.7101, 3.5332, 3.9594, 3.6054, 2.3415], device='cuda:4'), covar=tensor([0.0559, 0.0261, 0.0072, 0.0451, 0.0144, 0.0122, 0.0126, 0.0541], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0091, 0.0092, 0.0137, 0.0104, 0.0117, 0.0099, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 16:53:13,092 INFO [train.py:904] (4/8) Epoch 29, batch 2000, loss[loss=0.1582, simple_loss=0.2356, pruned_loss=0.04044, over 16700.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2512, pruned_loss=0.03783, over 3295763.04 frames. ], batch size: 89, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:54:21,795 INFO [train.py:904] (4/8) Epoch 29, batch 2050, loss[loss=0.1466, simple_loss=0.2455, pruned_loss=0.02386, over 17110.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2513, pruned_loss=0.03728, over 3305511.63 frames. ], batch size: 47, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:54:54,805 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=286278.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:55:04,677 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.605e+02 2.125e+02 2.436e+02 2.931e+02 6.185e+02, threshold=4.871e+02, percent-clipped=3.0 2023-05-02 16:55:18,577 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=286295.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:55:30,407 INFO [train.py:904] (4/8) Epoch 29, batch 2100, loss[loss=0.1931, simple_loss=0.2737, pruned_loss=0.0562, over 16849.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2519, pruned_loss=0.03782, over 3306161.08 frames. ], batch size: 116, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:56:00,346 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=286326.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:56:26,228 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=286344.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:56:40,455 INFO [train.py:904] (4/8) Epoch 29, batch 2150, loss[loss=0.1354, simple_loss=0.2243, pruned_loss=0.02331, over 17022.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2521, pruned_loss=0.03759, over 3317204.89 frames. ], batch size: 41, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 16:56:44,674 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=286356.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:57:00,749 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-05-02 16:57:24,981 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.220e+02 2.758e+02 3.117e+02 6.077e+02, threshold=5.516e+02, percent-clipped=2.0 2023-05-02 16:57:50,675 INFO [train.py:904] (4/8) Epoch 29, batch 2200, loss[loss=0.1858, simple_loss=0.2667, pruned_loss=0.05244, over 15485.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2524, pruned_loss=0.03785, over 3326149.95 frames. ], batch size: 190, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:57:52,907 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=286405.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:58:53,616 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-02 16:58:59,516 INFO [train.py:904] (4/8) Epoch 29, batch 2250, loss[loss=0.1868, simple_loss=0.2603, pruned_loss=0.05671, over 16935.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2524, pruned_loss=0.03854, over 3324200.77 frames. ], batch size: 109, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 16:59:33,869 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=286479.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 16:59:43,679 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=286485.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 16:59:45,653 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.691e+02 2.269e+02 2.585e+02 3.162e+02 6.397e+02, threshold=5.169e+02, percent-clipped=1.0 2023-05-02 17:00:08,670 INFO [train.py:904] (4/8) Epoch 29, batch 2300, loss[loss=0.1774, simple_loss=0.257, pruned_loss=0.04892, over 16719.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2533, pruned_loss=0.03906, over 3322774.50 frames. ], batch size: 134, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 17:00:32,366 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1467, 5.2112, 5.6575, 5.6217, 5.6351, 5.2691, 5.2498, 5.0560], device='cuda:4'), covar=tensor([0.0389, 0.0597, 0.0329, 0.0402, 0.0494, 0.0375, 0.0956, 0.0478], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0506, 0.0490, 0.0450, 0.0537, 0.0513, 0.0589, 0.0413], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 17:01:17,608 INFO [train.py:904] (4/8) Epoch 29, batch 2350, loss[loss=0.1493, simple_loss=0.2398, pruned_loss=0.0294, over 16801.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2539, pruned_loss=0.03965, over 3327386.91 frames. ], batch size: 39, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 17:02:03,068 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.493e+02 2.083e+02 2.366e+02 2.716e+02 6.133e+02, threshold=4.732e+02, percent-clipped=1.0 2023-05-02 17:02:04,899 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=286588.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:02:17,360 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-05-02 17:02:27,446 INFO [train.py:904] (4/8) Epoch 29, batch 2400, loss[loss=0.1548, simple_loss=0.2463, pruned_loss=0.03168, over 17197.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2541, pruned_loss=0.03923, over 3337799.60 frames. ], batch size: 44, lr: 2.33e-03, grad_scale: 8.0 2023-05-02 17:03:30,983 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=286649.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:03:34,191 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=286651.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:03:37,416 INFO [train.py:904] (4/8) Epoch 29, batch 2450, loss[loss=0.1392, simple_loss=0.2229, pruned_loss=0.02775, over 16799.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.255, pruned_loss=0.03917, over 3330701.76 frames. ], batch size: 102, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 17:03:59,928 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=286671.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:04:06,474 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-02 17:04:23,750 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.201e+02 2.590e+02 3.137e+02 8.541e+02, threshold=5.179e+02, percent-clipped=2.0 2023-05-02 17:04:41,082 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=286700.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:04:46,152 INFO [train.py:904] (4/8) Epoch 29, batch 2500, loss[loss=0.2048, simple_loss=0.2897, pruned_loss=0.05995, over 12363.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2542, pruned_loss=0.03871, over 3326352.65 frames. ], batch size: 248, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 17:05:24,967 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=286732.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:05:55,671 INFO [train.py:904] (4/8) Epoch 29, batch 2550, loss[loss=0.1683, simple_loss=0.2527, pruned_loss=0.04194, over 16091.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2541, pruned_loss=0.03868, over 3315801.52 frames. ], batch size: 35, lr: 2.33e-03, grad_scale: 4.0 2023-05-02 17:06:16,818 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1527, 2.4280, 2.5674, 1.9335, 2.6894, 2.7342, 2.4370, 2.3305], device='cuda:4'), covar=tensor([0.0706, 0.0299, 0.0261, 0.0956, 0.0152, 0.0309, 0.0476, 0.0447], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0103, 0.0141, 0.0088, 0.0133, 0.0132, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 17:06:18,992 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=286770.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:06:23,257 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-05-02 17:06:33,020 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=286779.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 17:06:40,933 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=286785.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:06:44,659 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.642e+02 2.042e+02 2.418e+02 2.906e+02 6.403e+02, threshold=4.836e+02, percent-clipped=2.0 2023-05-02 17:06:51,132 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-05-02 17:07:07,733 INFO [train.py:904] (4/8) Epoch 29, batch 2600, loss[loss=0.1522, simple_loss=0.2319, pruned_loss=0.0363, over 16836.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2532, pruned_loss=0.038, over 3320975.20 frames. ], batch size: 96, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:07:24,213 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3180, 2.5989, 2.1587, 2.2829, 2.9068, 2.6440, 2.9792, 3.0661], device='cuda:4'), covar=tensor([0.0225, 0.0497, 0.0654, 0.0588, 0.0337, 0.0433, 0.0328, 0.0300], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0252, 0.0240, 0.0241, 0.0253, 0.0251, 0.0249, 0.0251], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:07:39,059 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=286827.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 17:07:45,092 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=286831.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:07:47,330 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=286833.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:08:15,898 INFO [train.py:904] (4/8) Epoch 29, batch 2650, loss[loss=0.1732, simple_loss=0.2539, pruned_loss=0.04623, over 16865.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2547, pruned_loss=0.03829, over 3322014.21 frames. ], batch size: 96, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:09:00,291 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.410e+02 2.057e+02 2.433e+02 2.859e+02 4.818e+02, threshold=4.867e+02, percent-clipped=0.0 2023-05-02 17:09:22,331 INFO [train.py:904] (4/8) Epoch 29, batch 2700, loss[loss=0.1561, simple_loss=0.2527, pruned_loss=0.02982, over 17186.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2543, pruned_loss=0.03756, over 3320987.75 frames. ], batch size: 46, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:09:30,179 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.08 vs. limit=2.0 2023-05-02 17:09:46,354 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1837, 3.9662, 4.3931, 2.3249, 4.6035, 4.6606, 3.3591, 3.5564], device='cuda:4'), covar=tensor([0.0696, 0.0264, 0.0295, 0.1139, 0.0095, 0.0199, 0.0464, 0.0422], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0112, 0.0103, 0.0140, 0.0088, 0.0133, 0.0131, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 17:10:17,438 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=286944.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:10:26,951 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=286951.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:10:30,032 INFO [train.py:904] (4/8) Epoch 29, batch 2750, loss[loss=0.1656, simple_loss=0.2548, pruned_loss=0.03816, over 16447.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2552, pruned_loss=0.03732, over 3324368.18 frames. ], batch size: 75, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:11:09,652 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0046, 4.1838, 4.4030, 2.4853, 4.7003, 4.7952, 3.3674, 3.4342], device='cuda:4'), covar=tensor([0.1018, 0.0206, 0.0281, 0.1200, 0.0103, 0.0217, 0.0472, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0112, 0.0103, 0.0140, 0.0088, 0.0133, 0.0131, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 17:11:17,993 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 1.930e+02 2.341e+02 2.772e+02 4.818e+02, threshold=4.681e+02, percent-clipped=0.0 2023-05-02 17:11:32,916 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=286999.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:11:35,415 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=287000.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:11:40,740 INFO [train.py:904] (4/8) Epoch 29, batch 2800, loss[loss=0.1478, simple_loss=0.2447, pruned_loss=0.02545, over 17122.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2553, pruned_loss=0.03722, over 3319050.54 frames. ], batch size: 47, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:12:02,151 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7805, 4.2679, 3.0735, 2.3764, 2.6635, 2.7261, 4.6320, 3.6361], device='cuda:4'), covar=tensor([0.3057, 0.0582, 0.1851, 0.3179, 0.2954, 0.2169, 0.0396, 0.1409], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0278, 0.0318, 0.0332, 0.0309, 0.0282, 0.0309, 0.0359], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 17:12:12,974 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=287027.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:12:42,704 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=287048.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:12:50,053 INFO [train.py:904] (4/8) Epoch 29, batch 2850, loss[loss=0.1654, simple_loss=0.2449, pruned_loss=0.04297, over 16902.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2551, pruned_loss=0.03755, over 3318094.67 frames. ], batch size: 116, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:13:21,718 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4908, 4.4994, 4.6958, 4.4798, 4.5438, 5.1339, 4.6213, 4.3379], device='cuda:4'), covar=tensor([0.1679, 0.2253, 0.2561, 0.2142, 0.2585, 0.1144, 0.1760, 0.2476], device='cuda:4'), in_proj_covar=tensor([0.0442, 0.0656, 0.0731, 0.0535, 0.0715, 0.0751, 0.0564, 0.0710], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 17:13:39,988 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.623e+02 2.226e+02 2.685e+02 3.371e+02 7.491e+02, threshold=5.370e+02, percent-clipped=5.0 2023-05-02 17:13:50,507 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3170, 5.6287, 5.3732, 5.4625, 5.1635, 5.0451, 5.0887, 5.7367], device='cuda:4'), covar=tensor([0.1266, 0.0881, 0.1140, 0.0902, 0.0792, 0.0850, 0.1280, 0.0867], device='cuda:4'), in_proj_covar=tensor([0.0732, 0.0886, 0.0726, 0.0687, 0.0561, 0.0558, 0.0743, 0.0696], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:14:00,573 INFO [train.py:904] (4/8) Epoch 29, batch 2900, loss[loss=0.1756, simple_loss=0.2693, pruned_loss=0.04096, over 16639.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.253, pruned_loss=0.03761, over 3326501.21 frames. ], batch size: 57, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:14:31,519 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=287126.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:14:52,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0240, 4.5632, 4.5382, 3.3177, 3.6924, 4.5069, 4.0397, 2.7165], device='cuda:4'), covar=tensor([0.0512, 0.0083, 0.0055, 0.0365, 0.0171, 0.0100, 0.0091, 0.0494], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0091, 0.0093, 0.0137, 0.0105, 0.0118, 0.0100, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 17:14:53,680 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9816, 5.0620, 5.4952, 5.4911, 5.5041, 5.1517, 5.1237, 4.9567], device='cuda:4'), covar=tensor([0.0419, 0.0579, 0.0371, 0.0393, 0.0510, 0.0423, 0.1005, 0.0452], device='cuda:4'), in_proj_covar=tensor([0.0451, 0.0511, 0.0494, 0.0452, 0.0540, 0.0516, 0.0595, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 17:15:10,984 INFO [train.py:904] (4/8) Epoch 29, batch 2950, loss[loss=0.1729, simple_loss=0.2542, pruned_loss=0.04583, over 16264.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2529, pruned_loss=0.03824, over 3324485.12 frames. ], batch size: 165, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:15:59,393 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.489e+02 2.092e+02 2.416e+02 3.171e+02 5.496e+02, threshold=4.831e+02, percent-clipped=1.0 2023-05-02 17:16:20,171 INFO [train.py:904] (4/8) Epoch 29, batch 3000, loss[loss=0.1816, simple_loss=0.2698, pruned_loss=0.04673, over 16461.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2526, pruned_loss=0.03885, over 3325702.86 frames. ], batch size: 75, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:16:20,172 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 17:16:28,746 INFO [train.py:938] (4/8) Epoch 29, validation: loss=0.1336, simple_loss=0.2385, pruned_loss=0.01438, over 944034.00 frames. 2023-05-02 17:16:28,746 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 17:16:35,463 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1967, 4.0293, 4.2655, 4.3926, 4.4566, 4.0773, 4.2827, 4.4795], device='cuda:4'), covar=tensor([0.1754, 0.1379, 0.1332, 0.0683, 0.0647, 0.1402, 0.2200, 0.0776], device='cuda:4'), in_proj_covar=tensor([0.0709, 0.0870, 0.1003, 0.0883, 0.0673, 0.0700, 0.0737, 0.0855], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:17:12,452 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9652, 4.9493, 4.7473, 4.2035, 4.8945, 1.9679, 4.6046, 4.3725], device='cuda:4'), covar=tensor([0.0170, 0.0117, 0.0235, 0.0359, 0.0119, 0.2823, 0.0138, 0.0271], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0179, 0.0217, 0.0188, 0.0195, 0.0221, 0.0205, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:17:12,708 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-02 17:17:21,712 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.69 vs. limit=5.0 2023-05-02 17:17:25,956 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=287244.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:17:30,437 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7725, 2.7829, 2.6327, 5.0524, 4.0058, 4.2972, 1.5933, 3.1540], device='cuda:4'), covar=tensor([0.1450, 0.0842, 0.1339, 0.0161, 0.0202, 0.0402, 0.1749, 0.0815], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0182, 0.0201, 0.0208, 0.0207, 0.0221, 0.0212, 0.0200], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 17:17:40,002 INFO [train.py:904] (4/8) Epoch 29, batch 3050, loss[loss=0.1387, simple_loss=0.2235, pruned_loss=0.02695, over 16805.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2528, pruned_loss=0.03874, over 3334524.60 frames. ], batch size: 39, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:17:43,539 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4012, 2.9845, 2.6729, 2.2929, 2.2535, 2.3305, 3.0090, 2.8348], device='cuda:4'), covar=tensor([0.2573, 0.0712, 0.1678, 0.2547, 0.2380, 0.2202, 0.0527, 0.1444], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0276, 0.0315, 0.0330, 0.0308, 0.0281, 0.0307, 0.0357], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 17:17:47,881 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4606, 3.4486, 3.4955, 3.5607, 3.5945, 3.3189, 3.5372, 3.6680], device='cuda:4'), covar=tensor([0.1250, 0.0976, 0.0964, 0.0567, 0.0650, 0.2470, 0.1412, 0.0775], device='cuda:4'), in_proj_covar=tensor([0.0709, 0.0870, 0.1002, 0.0882, 0.0672, 0.0699, 0.0736, 0.0854], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:18:29,494 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.465e+02 2.177e+02 2.556e+02 3.004e+02 5.914e+02, threshold=5.112e+02, percent-clipped=3.0 2023-05-02 17:18:34,330 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=287292.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:18:49,717 INFO [train.py:904] (4/8) Epoch 29, batch 3100, loss[loss=0.1463, simple_loss=0.2268, pruned_loss=0.03285, over 16822.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2526, pruned_loss=0.03875, over 3335606.55 frames. ], batch size: 102, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:19:22,313 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=287327.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:20:00,323 INFO [train.py:904] (4/8) Epoch 29, batch 3150, loss[loss=0.145, simple_loss=0.2274, pruned_loss=0.03131, over 16738.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2513, pruned_loss=0.03836, over 3343871.00 frames. ], batch size: 39, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:20:30,291 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=287375.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:20:49,321 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.513e+02 2.202e+02 2.547e+02 3.057e+02 5.331e+02, threshold=5.094e+02, percent-clipped=1.0 2023-05-02 17:21:10,100 INFO [train.py:904] (4/8) Epoch 29, batch 3200, loss[loss=0.1665, simple_loss=0.2638, pruned_loss=0.03456, over 16759.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2501, pruned_loss=0.03766, over 3336375.90 frames. ], batch size: 62, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:21:20,060 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=287411.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:21:40,277 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=287426.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:22:18,321 INFO [train.py:904] (4/8) Epoch 29, batch 3250, loss[loss=0.2003, simple_loss=0.2664, pruned_loss=0.06713, over 16905.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2496, pruned_loss=0.03753, over 3342488.75 frames. ], batch size: 109, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:22:43,498 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=287472.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:22:45,675 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=287474.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:23:05,921 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.595e+02 2.139e+02 2.558e+02 3.030e+02 4.692e+02, threshold=5.116e+02, percent-clipped=0.0 2023-05-02 17:23:26,857 INFO [train.py:904] (4/8) Epoch 29, batch 3300, loss[loss=0.1486, simple_loss=0.2459, pruned_loss=0.02568, over 17226.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2505, pruned_loss=0.03781, over 3338346.80 frames. ], batch size: 44, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:23:33,169 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8945, 4.1060, 2.7456, 4.7555, 3.2830, 4.6631, 2.8345, 3.5382], device='cuda:4'), covar=tensor([0.0375, 0.0411, 0.1566, 0.0329, 0.0832, 0.0521, 0.1521, 0.0699], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0185, 0.0200, 0.0180, 0.0183, 0.0227, 0.0208, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 17:23:43,483 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-05-02 17:24:03,625 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3370, 5.2886, 5.0529, 4.4904, 5.1294, 2.0008, 4.8328, 4.9617], device='cuda:4'), covar=tensor([0.0117, 0.0113, 0.0261, 0.0501, 0.0126, 0.2883, 0.0181, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0180, 0.0219, 0.0190, 0.0196, 0.0222, 0.0207, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:24:11,395 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=287536.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:24:34,314 INFO [train.py:904] (4/8) Epoch 29, batch 3350, loss[loss=0.1564, simple_loss=0.2499, pruned_loss=0.03149, over 17192.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2517, pruned_loss=0.03822, over 3333645.55 frames. ], batch size: 45, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:24:39,340 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2190, 5.7971, 5.9445, 5.5603, 5.7171, 6.2703, 5.7693, 5.4508], device='cuda:4'), covar=tensor([0.0948, 0.2046, 0.2702, 0.2232, 0.2724, 0.1000, 0.1511, 0.2555], device='cuda:4'), in_proj_covar=tensor([0.0440, 0.0653, 0.0727, 0.0533, 0.0714, 0.0744, 0.0561, 0.0707], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 17:24:46,596 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=287562.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:25:14,647 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=287583.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:25:21,673 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.413e+02 2.115e+02 2.472e+02 3.004e+02 5.841e+02, threshold=4.943e+02, percent-clipped=1.0 2023-05-02 17:25:33,276 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=287597.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:25:43,316 INFO [train.py:904] (4/8) Epoch 29, batch 3400, loss[loss=0.1433, simple_loss=0.2406, pruned_loss=0.02304, over 17198.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.252, pruned_loss=0.03818, over 3333337.13 frames. ], batch size: 44, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:26:10,340 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=287623.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:26:40,153 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=287644.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:26:50,138 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.14 vs. limit=5.0 2023-05-02 17:26:53,540 INFO [train.py:904] (4/8) Epoch 29, batch 3450, loss[loss=0.1625, simple_loss=0.2476, pruned_loss=0.03874, over 16811.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2505, pruned_loss=0.03733, over 3325819.46 frames. ], batch size: 83, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:27:43,291 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.405e+02 1.986e+02 2.391e+02 2.898e+02 4.066e+02, threshold=4.783e+02, percent-clipped=0.0 2023-05-02 17:28:04,222 INFO [train.py:904] (4/8) Epoch 29, batch 3500, loss[loss=0.1737, simple_loss=0.2516, pruned_loss=0.04792, over 16743.00 frames. ], tot_loss[loss=0.1613, simple_loss=0.2493, pruned_loss=0.03663, over 3333904.13 frames. ], batch size: 134, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:28:18,077 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-05-02 17:28:26,818 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5949, 5.5416, 5.4536, 4.9593, 5.0823, 5.5030, 5.3628, 5.1171], device='cuda:4'), covar=tensor([0.0569, 0.0564, 0.0295, 0.0365, 0.1059, 0.0465, 0.0297, 0.0758], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0498, 0.0386, 0.0388, 0.0382, 0.0446, 0.0265, 0.0461], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 17:28:26,847 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=287720.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:29:14,560 INFO [train.py:904] (4/8) Epoch 29, batch 3550, loss[loss=0.1696, simple_loss=0.2516, pruned_loss=0.04384, over 16227.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.2491, pruned_loss=0.03634, over 3336370.40 frames. ], batch size: 165, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:29:33,117 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=287767.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:29:42,450 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=287773.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:29:53,933 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=287781.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 17:30:07,185 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.331e+02 2.118e+02 2.537e+02 3.010e+02 5.172e+02, threshold=5.074e+02, percent-clipped=1.0 2023-05-02 17:30:26,881 INFO [train.py:904] (4/8) Epoch 29, batch 3600, loss[loss=0.1628, simple_loss=0.2441, pruned_loss=0.04074, over 16748.00 frames. ], tot_loss[loss=0.1604, simple_loss=0.2481, pruned_loss=0.03636, over 3331642.89 frames. ], batch size: 134, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:31:11,578 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=287834.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:31:40,142 INFO [train.py:904] (4/8) Epoch 29, batch 3650, loss[loss=0.1584, simple_loss=0.2371, pruned_loss=0.0399, over 16817.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2477, pruned_loss=0.03696, over 3325031.83 frames. ], batch size: 102, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:32:35,499 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.746e+02 2.232e+02 2.562e+02 3.214e+02 7.635e+02, threshold=5.123e+02, percent-clipped=1.0 2023-05-02 17:32:38,207 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=287892.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:32:47,426 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7959, 3.8474, 2.5028, 4.3352, 3.1664, 4.2138, 2.7108, 3.1768], device='cuda:4'), covar=tensor([0.0342, 0.0425, 0.1636, 0.0346, 0.0780, 0.0703, 0.1399, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0186, 0.0201, 0.0181, 0.0184, 0.0229, 0.0209, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005, 0.0004, 0.0004], device='cuda:4') 2023-05-02 17:32:57,080 INFO [train.py:904] (4/8) Epoch 29, batch 3700, loss[loss=0.1735, simple_loss=0.2485, pruned_loss=0.04921, over 10992.00 frames. ], tot_loss[loss=0.1616, simple_loss=0.2461, pruned_loss=0.03858, over 3300836.21 frames. ], batch size: 247, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:33:18,786 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=287918.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:33:45,983 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8900, 4.8443, 4.7868, 4.4311, 4.4696, 4.8133, 4.6215, 4.5972], device='cuda:4'), covar=tensor([0.0634, 0.0696, 0.0302, 0.0337, 0.0934, 0.0553, 0.0451, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0497, 0.0385, 0.0388, 0.0381, 0.0445, 0.0264, 0.0460], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 17:33:48,746 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=287939.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:33:58,079 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.06 vs. limit=5.0 2023-05-02 17:34:09,993 INFO [train.py:904] (4/8) Epoch 29, batch 3750, loss[loss=0.1547, simple_loss=0.2332, pruned_loss=0.03811, over 16507.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2479, pruned_loss=0.04016, over 3291586.26 frames. ], batch size: 75, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:34:50,090 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4947, 3.5979, 3.4273, 1.9558, 2.9250, 2.1385, 3.7222, 3.9972], device='cuda:4'), covar=tensor([0.0209, 0.0844, 0.0814, 0.2552, 0.1106, 0.1271, 0.0624, 0.0844], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0173, 0.0171, 0.0158, 0.0149, 0.0134, 0.0147, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 17:35:05,541 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.573e+02 2.186e+02 2.475e+02 3.056e+02 8.739e+02, threshold=4.949e+02, percent-clipped=4.0 2023-05-02 17:35:29,219 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=288003.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:35:30,025 INFO [train.py:904] (4/8) Epoch 29, batch 3800, loss[loss=0.1843, simple_loss=0.2659, pruned_loss=0.05139, over 15709.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2483, pruned_loss=0.04074, over 3293447.35 frames. ], batch size: 191, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:35:37,633 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0342, 5.1026, 5.3896, 5.3534, 5.4328, 5.0624, 5.0227, 4.8232], device='cuda:4'), covar=tensor([0.0320, 0.0562, 0.0358, 0.0409, 0.0474, 0.0427, 0.1063, 0.0532], device='cuda:4'), in_proj_covar=tensor([0.0454, 0.0515, 0.0496, 0.0456, 0.0545, 0.0523, 0.0601, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 17:36:40,374 INFO [train.py:904] (4/8) Epoch 29, batch 3850, loss[loss=0.1702, simple_loss=0.2433, pruned_loss=0.04857, over 16803.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2488, pruned_loss=0.04161, over 3296785.78 frames. ], batch size: 124, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:36:55,450 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=288064.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:36:59,628 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288067.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:37:12,443 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288076.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 17:37:15,283 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7951, 2.8938, 2.3078, 2.7639, 3.1760, 2.8593, 3.3123, 3.4607], device='cuda:4'), covar=tensor([0.0105, 0.0492, 0.0640, 0.0477, 0.0310, 0.0412, 0.0221, 0.0266], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0250, 0.0239, 0.0240, 0.0252, 0.0250, 0.0248, 0.0250], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:37:30,812 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.513e+02 2.131e+02 2.477e+02 2.774e+02 5.510e+02, threshold=4.954e+02, percent-clipped=1.0 2023-05-02 17:37:50,142 INFO [train.py:904] (4/8) Epoch 29, batch 3900, loss[loss=0.1693, simple_loss=0.2429, pruned_loss=0.04782, over 16741.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2484, pruned_loss=0.04203, over 3280506.09 frames. ], batch size: 124, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:37:53,591 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=288106.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 17:38:06,534 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288115.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:38:27,296 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288129.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:39:00,615 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=288152.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 17:39:03,102 INFO [train.py:904] (4/8) Epoch 29, batch 3950, loss[loss=0.1583, simple_loss=0.2364, pruned_loss=0.04003, over 16757.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2474, pruned_loss=0.0424, over 3287265.94 frames. ], batch size: 89, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:39:22,287 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=288167.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 17:39:55,368 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.587e+02 2.270e+02 2.585e+02 3.079e+02 6.589e+02, threshold=5.170e+02, percent-clipped=2.0 2023-05-02 17:39:58,691 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288192.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:40:16,237 INFO [train.py:904] (4/8) Epoch 29, batch 4000, loss[loss=0.1789, simple_loss=0.2634, pruned_loss=0.04715, over 16426.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2477, pruned_loss=0.04306, over 3286930.00 frames. ], batch size: 75, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:40:30,278 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=288213.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 17:40:37,355 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288218.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:41:07,361 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288239.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:41:08,414 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288240.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:41:29,239 INFO [train.py:904] (4/8) Epoch 29, batch 4050, loss[loss=0.1994, simple_loss=0.2774, pruned_loss=0.0607, over 12057.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.249, pruned_loss=0.0427, over 3273932.15 frames. ], batch size: 248, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:41:47,212 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288266.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:42:17,219 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288287.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:42:21,065 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.369e+02 1.858e+02 2.084e+02 2.415e+02 5.689e+02, threshold=4.167e+02, percent-clipped=1.0 2023-05-02 17:42:32,376 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.5445, 5.8422, 5.5841, 5.6685, 5.3208, 5.2212, 5.2726, 6.0086], device='cuda:4'), covar=tensor([0.1225, 0.0792, 0.0937, 0.0854, 0.0760, 0.0694, 0.1115, 0.0768], device='cuda:4'), in_proj_covar=tensor([0.0741, 0.0891, 0.0730, 0.0693, 0.0568, 0.0565, 0.0750, 0.0701], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:42:42,062 INFO [train.py:904] (4/8) Epoch 29, batch 4100, loss[loss=0.1873, simple_loss=0.2885, pruned_loss=0.04298, over 16693.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2509, pruned_loss=0.04242, over 3266066.63 frames. ], batch size: 89, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:43:17,896 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=288327.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:43:45,824 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=288345.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:43:59,219 INFO [train.py:904] (4/8) Epoch 29, batch 4150, loss[loss=0.1807, simple_loss=0.273, pruned_loss=0.04419, over 16806.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2577, pruned_loss=0.04438, over 3252724.16 frames. ], batch size: 83, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:44:07,903 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288359.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:44:33,294 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288376.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:44:51,051 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=288388.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:44:52,908 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.550e+02 2.295e+02 2.546e+02 3.002e+02 5.285e+02, threshold=5.091e+02, percent-clipped=5.0 2023-05-02 17:45:02,600 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=288396.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:45:13,932 INFO [train.py:904] (4/8) Epoch 29, batch 4200, loss[loss=0.1999, simple_loss=0.2911, pruned_loss=0.05437, over 15218.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2645, pruned_loss=0.04579, over 3228255.34 frames. ], batch size: 190, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:45:18,724 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=288406.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:45:31,010 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6032, 2.5229, 2.0104, 2.6819, 2.1189, 2.7313, 2.1851, 2.3529], device='cuda:4'), covar=tensor([0.0323, 0.0348, 0.1164, 0.0243, 0.0606, 0.0403, 0.1130, 0.0611], device='cuda:4'), in_proj_covar=tensor([0.0180, 0.0184, 0.0198, 0.0179, 0.0183, 0.0226, 0.0207, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 17:45:31,226 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.31 vs. limit=5.0 2023-05-02 17:45:45,766 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288424.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:45:53,528 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288429.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:46:29,545 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1577, 2.4146, 2.0473, 2.1290, 2.8040, 2.3673, 2.6315, 2.9679], device='cuda:4'), covar=tensor([0.0210, 0.0527, 0.0671, 0.0652, 0.0331, 0.0505, 0.0220, 0.0320], device='cuda:4'), in_proj_covar=tensor([0.0239, 0.0249, 0.0238, 0.0239, 0.0250, 0.0248, 0.0247, 0.0249], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:46:30,280 INFO [train.py:904] (4/8) Epoch 29, batch 4250, loss[loss=0.1597, simple_loss=0.2627, pruned_loss=0.0283, over 16768.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2681, pruned_loss=0.0458, over 3211229.43 frames. ], batch size: 89, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:46:36,714 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=288457.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:46:43,212 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288462.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 17:47:05,781 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288477.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:47:22,494 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5278, 2.2377, 1.9481, 2.0475, 2.5618, 2.2219, 2.2817, 2.6916], device='cuda:4'), covar=tensor([0.0270, 0.0482, 0.0572, 0.0534, 0.0281, 0.0422, 0.0231, 0.0297], device='cuda:4'), in_proj_covar=tensor([0.0238, 0.0249, 0.0237, 0.0238, 0.0249, 0.0248, 0.0246, 0.0248], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:47:24,428 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.453e+02 2.222e+02 2.492e+02 3.034e+02 5.623e+02, threshold=4.984e+02, percent-clipped=1.0 2023-05-02 17:47:45,986 INFO [train.py:904] (4/8) Epoch 29, batch 4300, loss[loss=0.1797, simple_loss=0.2756, pruned_loss=0.04192, over 16482.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2697, pruned_loss=0.04545, over 3208582.20 frames. ], batch size: 68, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:47:51,123 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288508.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 17:48:41,891 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1542, 3.8075, 3.6687, 2.3460, 3.4012, 3.7849, 3.4459, 2.2078], device='cuda:4'), covar=tensor([0.0597, 0.0049, 0.0081, 0.0489, 0.0124, 0.0110, 0.0121, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0091, 0.0093, 0.0137, 0.0104, 0.0117, 0.0100, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 17:48:59,342 INFO [train.py:904] (4/8) Epoch 29, batch 4350, loss[loss=0.184, simple_loss=0.27, pruned_loss=0.049, over 17047.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2727, pruned_loss=0.04672, over 3202199.77 frames. ], batch size: 53, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:49:09,473 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9829, 4.9888, 4.7087, 4.1262, 4.9407, 1.7463, 4.6802, 4.3272], device='cuda:4'), covar=tensor([0.0071, 0.0056, 0.0188, 0.0364, 0.0072, 0.3178, 0.0103, 0.0288], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0181, 0.0221, 0.0192, 0.0197, 0.0224, 0.0208, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:49:53,411 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 2.193e+02 2.447e+02 2.913e+02 5.150e+02, threshold=4.894e+02, percent-clipped=1.0 2023-05-02 17:50:13,840 INFO [train.py:904] (4/8) Epoch 29, batch 4400, loss[loss=0.1907, simple_loss=0.2855, pruned_loss=0.04798, over 16691.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2755, pruned_loss=0.04806, over 3207050.77 frames. ], batch size: 89, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:50:22,229 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=288609.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:51:04,821 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6530, 2.7756, 2.4001, 2.5093, 3.0926, 2.6461, 3.1733, 3.2830], device='cuda:4'), covar=tensor([0.0102, 0.0369, 0.0492, 0.0443, 0.0242, 0.0390, 0.0227, 0.0223], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0247, 0.0236, 0.0236, 0.0248, 0.0246, 0.0244, 0.0246], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:51:12,393 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.78 vs. limit=5.0 2023-05-02 17:51:26,299 INFO [train.py:904] (4/8) Epoch 29, batch 4450, loss[loss=0.2001, simple_loss=0.2811, pruned_loss=0.05951, over 16708.00 frames. ], tot_loss[loss=0.1894, simple_loss=0.2793, pruned_loss=0.04977, over 3192574.44 frames. ], batch size: 57, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:51:34,247 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288659.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:51:49,577 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=288670.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 17:51:59,993 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0756, 3.3429, 3.3807, 5.3350, 4.2457, 4.5394, 2.1061, 3.5412], device='cuda:4'), covar=tensor([0.1197, 0.0674, 0.0862, 0.0097, 0.0291, 0.0306, 0.1440, 0.0655], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0180, 0.0200, 0.0205, 0.0206, 0.0218, 0.0209, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 17:52:08,508 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288683.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:52:18,785 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.398e+02 1.892e+02 2.239e+02 2.604e+02 5.228e+02, threshold=4.478e+02, percent-clipped=1.0 2023-05-02 17:52:34,494 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288701.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:52:37,594 INFO [train.py:904] (4/8) Epoch 29, batch 4500, loss[loss=0.2012, simple_loss=0.2934, pruned_loss=0.05446, over 16867.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.28, pruned_loss=0.05076, over 3187894.28 frames. ], batch size: 116, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:52:42,934 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288707.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:53:04,282 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8073, 5.0494, 4.8780, 4.8615, 4.6063, 4.5515, 4.5605, 5.1448], device='cuda:4'), covar=tensor([0.1174, 0.0836, 0.0912, 0.0887, 0.0817, 0.1177, 0.1084, 0.0797], device='cuda:4'), in_proj_covar=tensor([0.0728, 0.0874, 0.0715, 0.0681, 0.0557, 0.0554, 0.0733, 0.0688], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:53:18,699 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1346, 2.4614, 2.4032, 3.9440, 2.2533, 2.7397, 2.4622, 2.5297], device='cuda:4'), covar=tensor([0.1563, 0.3170, 0.3039, 0.0637, 0.4259, 0.2323, 0.3134, 0.3354], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0480, 0.0389, 0.0340, 0.0449, 0.0552, 0.0452, 0.0562], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 17:53:22,731 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5323, 3.5390, 2.6973, 2.3328, 2.4152, 2.3452, 3.8793, 3.2498], device='cuda:4'), covar=tensor([0.2937, 0.0723, 0.1955, 0.2548, 0.2670, 0.2337, 0.0485, 0.1304], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0276, 0.0314, 0.0329, 0.0309, 0.0280, 0.0307, 0.0355], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 17:53:49,128 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288752.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:53:51,862 INFO [train.py:904] (4/8) Epoch 29, batch 4550, loss[loss=0.1984, simple_loss=0.2842, pruned_loss=0.0563, over 16578.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2804, pruned_loss=0.05177, over 3192996.99 frames. ], batch size: 62, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:54:03,019 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288762.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 17:54:44,107 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 1.747e+02 2.008e+02 2.421e+02 6.234e+02, threshold=4.017e+02, percent-clipped=2.0 2023-05-02 17:54:53,658 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=288797.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:55:04,694 INFO [train.py:904] (4/8) Epoch 29, batch 4600, loss[loss=0.1766, simple_loss=0.2723, pruned_loss=0.04046, over 16908.00 frames. ], tot_loss[loss=0.192, simple_loss=0.281, pruned_loss=0.05151, over 3199230.31 frames. ], batch size: 96, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:55:11,559 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288808.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 17:55:14,423 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288810.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 17:56:18,390 INFO [train.py:904] (4/8) Epoch 29, batch 4650, loss[loss=0.2274, simple_loss=0.3064, pruned_loss=0.07421, over 11693.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2805, pruned_loss=0.05171, over 3212219.83 frames. ], batch size: 246, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 17:56:21,004 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=288856.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 17:56:24,823 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=288858.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:56:28,386 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2769, 4.3823, 4.5936, 4.5709, 4.6156, 4.3481, 4.3540, 4.1795], device='cuda:4'), covar=tensor([0.0301, 0.0444, 0.0357, 0.0360, 0.0398, 0.0351, 0.0811, 0.0553], device='cuda:4'), in_proj_covar=tensor([0.0437, 0.0495, 0.0478, 0.0437, 0.0525, 0.0502, 0.0581, 0.0406], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 17:57:00,653 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9343, 5.3074, 5.5204, 5.2287, 5.3226, 5.8675, 5.3729, 5.0639], device='cuda:4'), covar=tensor([0.1031, 0.1795, 0.2000, 0.1871, 0.2191, 0.0892, 0.1349, 0.2163], device='cuda:4'), in_proj_covar=tensor([0.0431, 0.0639, 0.0705, 0.0518, 0.0695, 0.0729, 0.0548, 0.0690], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 17:57:10,668 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.194e+02 1.769e+02 2.025e+02 2.417e+02 4.191e+02, threshold=4.050e+02, percent-clipped=1.0 2023-05-02 17:57:29,736 INFO [train.py:904] (4/8) Epoch 29, batch 4700, loss[loss=0.1744, simple_loss=0.2481, pruned_loss=0.05037, over 11396.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2777, pruned_loss=0.05073, over 3197121.79 frames. ], batch size: 247, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:58:32,766 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-05-02 17:58:41,594 INFO [train.py:904] (4/8) Epoch 29, batch 4750, loss[loss=0.2129, simple_loss=0.2859, pruned_loss=0.06993, over 12274.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2745, pruned_loss=0.04912, over 3191876.82 frames. ], batch size: 248, lr: 2.32e-03, grad_scale: 4.0 2023-05-02 17:58:57,734 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=288965.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 17:59:23,551 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=288983.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:59:35,273 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.313e+02 1.780e+02 1.987e+02 2.347e+02 4.131e+02, threshold=3.973e+02, percent-clipped=1.0 2023-05-02 17:59:50,950 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=289001.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 17:59:54,497 INFO [train.py:904] (4/8) Epoch 29, batch 4800, loss[loss=0.167, simple_loss=0.2685, pruned_loss=0.03277, over 16865.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2707, pruned_loss=0.04701, over 3195812.24 frames. ], batch size: 96, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 18:00:37,056 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=289031.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:00:52,585 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5713, 3.6680, 2.3448, 4.2934, 2.9387, 4.1566, 2.4422, 2.9879], device='cuda:4'), covar=tensor([0.0339, 0.0385, 0.1716, 0.0169, 0.0840, 0.0673, 0.1615, 0.0881], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0183, 0.0199, 0.0177, 0.0182, 0.0225, 0.0207, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:01:00,476 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=289046.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 18:01:05,208 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=289049.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:01:08,923 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=289052.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:01:12,124 INFO [train.py:904] (4/8) Epoch 29, batch 4850, loss[loss=0.1969, simple_loss=0.282, pruned_loss=0.05589, over 11867.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2712, pruned_loss=0.04595, over 3187465.40 frames. ], batch size: 248, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 18:01:41,284 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-02 18:02:08,678 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.294e+02 1.829e+02 2.187e+02 2.631e+02 3.612e+02, threshold=4.373e+02, percent-clipped=0.0 2023-05-02 18:02:22,995 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=289100.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:02:28,897 INFO [train.py:904] (4/8) Epoch 29, batch 4900, loss[loss=0.1839, simple_loss=0.2776, pruned_loss=0.04509, over 15363.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2699, pruned_loss=0.04431, over 3171977.17 frames. ], batch size: 190, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 18:02:33,584 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=289107.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 18:03:16,513 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=289136.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:03:36,366 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1834, 2.4858, 2.5052, 4.0625, 2.2794, 2.8676, 2.4730, 2.6380], device='cuda:4'), covar=tensor([0.1619, 0.3561, 0.3069, 0.0575, 0.4011, 0.2447, 0.3634, 0.3170], device='cuda:4'), in_proj_covar=tensor([0.0424, 0.0480, 0.0388, 0.0339, 0.0447, 0.0550, 0.0451, 0.0561], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:03:42,152 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=289153.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:03:43,041 INFO [train.py:904] (4/8) Epoch 29, batch 4950, loss[loss=0.1892, simple_loss=0.2755, pruned_loss=0.0515, over 12155.00 frames. ], tot_loss[loss=0.178, simple_loss=0.269, pruned_loss=0.04345, over 3171438.83 frames. ], batch size: 247, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 18:04:38,081 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.471e+02 2.040e+02 2.385e+02 2.853e+02 5.286e+02, threshold=4.771e+02, percent-clipped=2.0 2023-05-02 18:04:46,892 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=289197.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:04:56,967 INFO [train.py:904] (4/8) Epoch 29, batch 5000, loss[loss=0.1735, simple_loss=0.2737, pruned_loss=0.03664, over 16793.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2703, pruned_loss=0.04297, over 3189673.65 frames. ], batch size: 124, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 18:05:47,657 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5634, 4.6483, 4.4223, 4.1065, 4.0955, 4.5256, 4.2883, 4.2646], device='cuda:4'), covar=tensor([0.0600, 0.0461, 0.0342, 0.0321, 0.0979, 0.0543, 0.0533, 0.0562], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0475, 0.0368, 0.0370, 0.0364, 0.0425, 0.0253, 0.0440], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:06:00,787 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2829, 2.2655, 2.8290, 3.1896, 3.1857, 3.8231, 2.3531, 3.7782], device='cuda:4'), covar=tensor([0.0244, 0.0542, 0.0376, 0.0343, 0.0312, 0.0158, 0.0621, 0.0129], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0201, 0.0189, 0.0196, 0.0213, 0.0169, 0.0206, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 18:06:09,972 INFO [train.py:904] (4/8) Epoch 29, batch 5050, loss[loss=0.1683, simple_loss=0.2624, pruned_loss=0.03712, over 16483.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2708, pruned_loss=0.04285, over 3189088.46 frames. ], batch size: 75, lr: 2.32e-03, grad_scale: 8.0 2023-05-02 18:06:26,929 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=289265.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:07:03,090 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.434e+02 1.941e+02 2.202e+02 2.540e+02 3.761e+02, threshold=4.404e+02, percent-clipped=0.0 2023-05-02 18:07:22,109 INFO [train.py:904] (4/8) Epoch 29, batch 5100, loss[loss=0.1694, simple_loss=0.2627, pruned_loss=0.03803, over 16397.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2688, pruned_loss=0.04218, over 3195126.84 frames. ], batch size: 146, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:07:34,856 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=289313.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:08:06,089 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.4651, 2.5538, 2.4894, 4.1491, 2.9274, 3.9415, 1.4437, 3.0072], device='cuda:4'), covar=tensor([0.1517, 0.0836, 0.1222, 0.0128, 0.0163, 0.0321, 0.1756, 0.0769], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0180, 0.0199, 0.0203, 0.0205, 0.0217, 0.0209, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:08:36,361 INFO [train.py:904] (4/8) Epoch 29, batch 5150, loss[loss=0.1589, simple_loss=0.2622, pruned_loss=0.02786, over 16404.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2686, pruned_loss=0.04125, over 3187670.20 frames. ], batch size: 146, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:08:57,722 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-05-02 18:09:04,554 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1605, 2.4384, 2.4151, 3.9476, 2.2409, 2.7690, 2.4447, 2.5621], device='cuda:4'), covar=tensor([0.1500, 0.3448, 0.2933, 0.0575, 0.3939, 0.2512, 0.3583, 0.3131], device='cuda:4'), in_proj_covar=tensor([0.0424, 0.0480, 0.0389, 0.0340, 0.0448, 0.0551, 0.0452, 0.0561], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:09:29,038 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.215e+02 1.894e+02 2.237e+02 2.654e+02 4.017e+02, threshold=4.473e+02, percent-clipped=0.0 2023-05-02 18:09:44,630 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1910, 4.2537, 4.5580, 4.5065, 4.5065, 4.2768, 4.2204, 4.2205], device='cuda:4'), covar=tensor([0.0369, 0.0561, 0.0354, 0.0417, 0.0471, 0.0401, 0.0885, 0.0515], device='cuda:4'), in_proj_covar=tensor([0.0437, 0.0494, 0.0479, 0.0438, 0.0525, 0.0502, 0.0582, 0.0406], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 18:09:45,672 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=289402.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 18:09:47,652 INFO [train.py:904] (4/8) Epoch 29, batch 5200, loss[loss=0.1625, simple_loss=0.2534, pruned_loss=0.03581, over 16522.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2671, pruned_loss=0.04082, over 3191273.44 frames. ], batch size: 68, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:10:20,796 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0112, 2.8401, 2.0609, 2.4133, 3.2529, 2.8513, 3.5577, 3.5777], device='cuda:4'), covar=tensor([0.0119, 0.0590, 0.0892, 0.0676, 0.0326, 0.0522, 0.0321, 0.0284], device='cuda:4'), in_proj_covar=tensor([0.0235, 0.0247, 0.0236, 0.0237, 0.0248, 0.0247, 0.0243, 0.0248], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:10:27,495 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-02 18:10:59,663 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=289453.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:11:00,510 INFO [train.py:904] (4/8) Epoch 29, batch 5250, loss[loss=0.1706, simple_loss=0.2661, pruned_loss=0.03752, over 16427.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2649, pruned_loss=0.04069, over 3200887.34 frames. ], batch size: 146, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:11:55,497 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-05-02 18:11:55,945 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.356e+02 1.894e+02 2.157e+02 2.543e+02 4.571e+02, threshold=4.315e+02, percent-clipped=1.0 2023-05-02 18:11:58,122 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=289492.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:12:10,717 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=289501.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:12:14,338 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=289503.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:12:15,673 INFO [train.py:904] (4/8) Epoch 29, batch 5300, loss[loss=0.1539, simple_loss=0.2481, pruned_loss=0.02981, over 16307.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.261, pruned_loss=0.03934, over 3214972.40 frames. ], batch size: 165, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:12:31,676 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4727, 3.7147, 3.9136, 1.9761, 4.2541, 4.2193, 3.0757, 3.0160], device='cuda:4'), covar=tensor([0.1134, 0.0211, 0.0229, 0.1349, 0.0072, 0.0134, 0.0447, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0102, 0.0139, 0.0087, 0.0131, 0.0130, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:12:43,416 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.86 vs. limit=5.0 2023-05-02 18:13:26,912 INFO [train.py:904] (4/8) Epoch 29, batch 5350, loss[loss=0.1961, simple_loss=0.2772, pruned_loss=0.05744, over 17026.00 frames. ], tot_loss[loss=0.169, simple_loss=0.26, pruned_loss=0.03896, over 3225313.26 frames. ], batch size: 53, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:13:42,032 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=289564.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:14:21,403 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.205e+02 1.853e+02 2.169e+02 2.585e+02 3.843e+02, threshold=4.338e+02, percent-clipped=0.0 2023-05-02 18:14:41,092 INFO [train.py:904] (4/8) Epoch 29, batch 5400, loss[loss=0.1703, simple_loss=0.2569, pruned_loss=0.04189, over 16252.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2622, pruned_loss=0.03976, over 3198913.69 frames. ], batch size: 35, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:15:57,431 INFO [train.py:904] (4/8) Epoch 29, batch 5450, loss[loss=0.2179, simple_loss=0.3071, pruned_loss=0.06434, over 15249.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2646, pruned_loss=0.04078, over 3208368.60 frames. ], batch size: 190, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:16:38,894 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1436, 3.3719, 3.5901, 2.0593, 3.1887, 2.4247, 3.5473, 3.7306], device='cuda:4'), covar=tensor([0.0245, 0.0821, 0.0588, 0.2264, 0.0804, 0.0985, 0.0613, 0.0930], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0172, 0.0172, 0.0158, 0.0149, 0.0133, 0.0147, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 18:16:53,634 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.432e+02 2.272e+02 2.915e+02 3.734e+02 7.781e+02, threshold=5.831e+02, percent-clipped=14.0 2023-05-02 18:17:12,138 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=289702.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 18:17:14,644 INFO [train.py:904] (4/8) Epoch 29, batch 5500, loss[loss=0.2665, simple_loss=0.3336, pruned_loss=0.09976, over 11640.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2711, pruned_loss=0.04452, over 3165932.48 frames. ], batch size: 247, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:17:20,481 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-05-02 18:17:51,880 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3820, 3.3018, 2.6865, 2.2238, 2.3089, 2.2992, 3.4245, 3.1062], device='cuda:4'), covar=tensor([0.2970, 0.0635, 0.1743, 0.2662, 0.2510, 0.2261, 0.0512, 0.1249], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0274, 0.0312, 0.0327, 0.0305, 0.0277, 0.0305, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 18:18:02,410 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3680, 3.1364, 3.5192, 1.8884, 3.6524, 3.6297, 2.8285, 2.8017], device='cuda:4'), covar=tensor([0.0863, 0.0334, 0.0205, 0.1187, 0.0105, 0.0194, 0.0479, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0102, 0.0138, 0.0087, 0.0132, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:18:26,454 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=289750.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 18:18:31,706 INFO [train.py:904] (4/8) Epoch 29, batch 5550, loss[loss=0.1792, simple_loss=0.2778, pruned_loss=0.04032, over 16869.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2777, pruned_loss=0.04856, over 3157382.89 frames. ], batch size: 102, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:19:30,745 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.874e+02 2.953e+02 3.616e+02 4.343e+02 9.960e+02, threshold=7.231e+02, percent-clipped=12.0 2023-05-02 18:19:34,042 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=289792.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:19:53,194 INFO [train.py:904] (4/8) Epoch 29, batch 5600, loss[loss=0.2004, simple_loss=0.2919, pruned_loss=0.05442, over 16546.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.2817, pruned_loss=0.0516, over 3138910.00 frames. ], batch size: 62, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:20:32,146 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7302, 4.7659, 4.5547, 3.7975, 4.6782, 1.6849, 4.4637, 4.3319], device='cuda:4'), covar=tensor([0.0141, 0.0128, 0.0237, 0.0456, 0.0119, 0.3231, 0.0149, 0.0293], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0179, 0.0219, 0.0190, 0.0195, 0.0222, 0.0206, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:20:54,481 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=289840.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:21:16,881 INFO [train.py:904] (4/8) Epoch 29, batch 5650, loss[loss=0.2087, simple_loss=0.3057, pruned_loss=0.05582, over 17208.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2864, pruned_loss=0.05558, over 3109795.22 frames. ], batch size: 44, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:21:18,067 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7366, 3.4602, 3.9675, 1.9422, 4.1494, 4.1152, 3.0466, 3.1211], device='cuda:4'), covar=tensor([0.0786, 0.0302, 0.0212, 0.1341, 0.0078, 0.0199, 0.0444, 0.0501], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0103, 0.0138, 0.0087, 0.0131, 0.0129, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:21:25,654 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=289859.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:21:32,050 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7863, 2.7602, 2.4349, 4.8740, 3.3210, 3.9256, 1.7774, 2.7530], device='cuda:4'), covar=tensor([0.1507, 0.0955, 0.1552, 0.0201, 0.0436, 0.0543, 0.1856, 0.1115], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0181, 0.0200, 0.0204, 0.0206, 0.0217, 0.0210, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:21:35,823 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.6238, 2.3995, 2.3245, 3.7070, 2.6206, 3.7730, 1.4844, 2.6421], device='cuda:4'), covar=tensor([0.1420, 0.0911, 0.1343, 0.0196, 0.0215, 0.0415, 0.1790, 0.0945], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0181, 0.0200, 0.0204, 0.0206, 0.0217, 0.0210, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:22:15,728 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.920e+02 3.088e+02 3.696e+02 4.617e+02 9.164e+02, threshold=7.392e+02, percent-clipped=2.0 2023-05-02 18:22:35,917 INFO [train.py:904] (4/8) Epoch 29, batch 5700, loss[loss=0.2527, simple_loss=0.3223, pruned_loss=0.09157, over 11469.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.288, pruned_loss=0.05768, over 3097943.48 frames. ], batch size: 248, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:23:01,348 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=289920.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 18:23:55,705 INFO [train.py:904] (4/8) Epoch 29, batch 5750, loss[loss=0.2407, simple_loss=0.3003, pruned_loss=0.09059, over 11475.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2908, pruned_loss=0.05944, over 3065209.95 frames. ], batch size: 248, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:24:34,826 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7583, 3.6029, 4.0533, 2.1098, 4.2812, 4.2511, 3.1774, 3.2315], device='cuda:4'), covar=tensor([0.0816, 0.0299, 0.0225, 0.1253, 0.0086, 0.0181, 0.0430, 0.0480], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0103, 0.0138, 0.0087, 0.0132, 0.0129, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:24:40,826 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=289981.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 18:24:56,848 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.808e+02 2.660e+02 3.372e+02 4.165e+02 9.302e+02, threshold=6.745e+02, percent-clipped=2.0 2023-05-02 18:25:08,688 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=289998.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:25:20,988 INFO [train.py:904] (4/8) Epoch 29, batch 5800, loss[loss=0.2063, simple_loss=0.2933, pruned_loss=0.05963, over 15384.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.29, pruned_loss=0.05764, over 3083460.07 frames. ], batch size: 190, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:25:28,669 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1634, 4.0487, 4.2369, 4.3556, 4.4823, 4.0654, 4.4145, 4.5178], device='cuda:4'), covar=tensor([0.2021, 0.1308, 0.1573, 0.0795, 0.0669, 0.1405, 0.0958, 0.0745], device='cuda:4'), in_proj_covar=tensor([0.0685, 0.0834, 0.0963, 0.0849, 0.0646, 0.0669, 0.0706, 0.0821], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:25:58,180 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3854, 4.5396, 4.7133, 4.4916, 4.5555, 5.0515, 4.5939, 4.3013], device='cuda:4'), covar=tensor([0.1673, 0.1803, 0.2032, 0.1932, 0.2263, 0.0977, 0.1636, 0.2498], device='cuda:4'), in_proj_covar=tensor([0.0432, 0.0639, 0.0706, 0.0519, 0.0694, 0.0728, 0.0549, 0.0693], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 18:26:05,036 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-05-02 18:26:31,304 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4928, 3.4612, 3.4451, 2.7061, 3.3255, 2.1525, 3.1540, 2.7922], device='cuda:4'), covar=tensor([0.0202, 0.0171, 0.0198, 0.0243, 0.0125, 0.2456, 0.0161, 0.0298], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0178, 0.0217, 0.0188, 0.0193, 0.0220, 0.0204, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:26:31,356 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5890, 2.4945, 1.9228, 2.6226, 2.1265, 2.7270, 2.1366, 2.3151], device='cuda:4'), covar=tensor([0.0345, 0.0354, 0.1372, 0.0319, 0.0671, 0.0470, 0.1326, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0177, 0.0182, 0.0197, 0.0174, 0.0181, 0.0222, 0.0205, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:26:34,258 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=290050.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:26:39,145 INFO [train.py:904] (4/8) Epoch 29, batch 5850, loss[loss=0.1857, simple_loss=0.2834, pruned_loss=0.04399, over 16289.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2881, pruned_loss=0.05597, over 3080410.90 frames. ], batch size: 165, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:26:48,293 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=290059.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:27:26,389 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.79 vs. limit=5.0 2023-05-02 18:27:39,837 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.636e+02 2.628e+02 3.224e+02 3.685e+02 6.132e+02, threshold=6.447e+02, percent-clipped=0.0 2023-05-02 18:28:01,356 INFO [train.py:904] (4/8) Epoch 29, batch 5900, loss[loss=0.1734, simple_loss=0.2747, pruned_loss=0.03609, over 16907.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2878, pruned_loss=0.05609, over 3076189.95 frames. ], batch size: 96, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:28:15,154 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=290111.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:29:05,008 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3384, 3.6734, 3.7912, 2.3353, 3.2766, 2.4893, 3.7713, 3.9721], device='cuda:4'), covar=tensor([0.0217, 0.0713, 0.0585, 0.1912, 0.0764, 0.0973, 0.0511, 0.0824], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0171, 0.0156, 0.0148, 0.0133, 0.0145, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 18:29:16,662 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8073, 2.7289, 2.8606, 2.2099, 2.7024, 2.1379, 2.7385, 2.9212], device='cuda:4'), covar=tensor([0.0247, 0.0881, 0.0493, 0.1821, 0.0813, 0.0956, 0.0543, 0.0752], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0171, 0.0156, 0.0148, 0.0133, 0.0145, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 18:29:20,949 INFO [train.py:904] (4/8) Epoch 29, batch 5950, loss[loss=0.1998, simple_loss=0.2867, pruned_loss=0.05645, over 16690.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.289, pruned_loss=0.05567, over 3050941.72 frames. ], batch size: 124, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:29:30,138 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=290159.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:30:18,753 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.883e+02 2.594e+02 3.044e+02 3.971e+02 8.061e+02, threshold=6.088e+02, percent-clipped=1.0 2023-05-02 18:30:32,092 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.95 vs. limit=5.0 2023-05-02 18:30:40,225 INFO [train.py:904] (4/8) Epoch 29, batch 6000, loss[loss=0.1908, simple_loss=0.2784, pruned_loss=0.05159, over 16443.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.288, pruned_loss=0.05475, over 3074229.07 frames. ], batch size: 146, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:30:40,226 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 18:30:50,250 INFO [train.py:938] (4/8) Epoch 29, validation: loss=0.1475, simple_loss=0.2594, pruned_loss=0.01778, over 944034.00 frames. 2023-05-02 18:30:50,251 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 18:30:55,860 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=290207.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:32:11,220 INFO [train.py:904] (4/8) Epoch 29, batch 6050, loss[loss=0.1994, simple_loss=0.2929, pruned_loss=0.05294, over 16369.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2865, pruned_loss=0.05417, over 3089420.23 frames. ], batch size: 35, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:32:13,344 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3126, 2.4786, 2.4985, 3.9610, 2.3470, 2.8398, 2.5165, 2.5811], device='cuda:4'), covar=tensor([0.1398, 0.3457, 0.2858, 0.0583, 0.3860, 0.2359, 0.3545, 0.3128], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0475, 0.0386, 0.0337, 0.0446, 0.0546, 0.0448, 0.0557], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:32:15,113 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7857, 2.7367, 2.8704, 2.1596, 2.6875, 2.1545, 2.7034, 2.9117], device='cuda:4'), covar=tensor([0.0258, 0.0833, 0.0463, 0.1856, 0.0802, 0.0912, 0.0580, 0.0756], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0172, 0.0171, 0.0157, 0.0148, 0.0133, 0.0146, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 18:32:42,690 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=290276.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 18:33:04,446 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.589e+02 2.468e+02 2.905e+02 3.476e+02 8.072e+02, threshold=5.810e+02, percent-clipped=1.0 2023-05-02 18:33:29,598 INFO [train.py:904] (4/8) Epoch 29, batch 6100, loss[loss=0.2176, simple_loss=0.2984, pruned_loss=0.06841, over 16546.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2856, pruned_loss=0.05368, over 3085044.43 frames. ], batch size: 57, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:33:53,073 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=290318.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:34:32,673 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3125, 2.8098, 2.9925, 2.0021, 2.6824, 1.9930, 2.9912, 3.0559], device='cuda:4'), covar=tensor([0.0288, 0.0953, 0.0644, 0.2163, 0.0983, 0.1131, 0.0687, 0.0960], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0172, 0.0172, 0.0157, 0.0149, 0.0133, 0.0146, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 18:34:37,570 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4960, 3.4463, 3.4262, 2.6355, 3.2621, 2.1670, 3.1152, 2.7377], device='cuda:4'), covar=tensor([0.0176, 0.0156, 0.0200, 0.0230, 0.0123, 0.2429, 0.0156, 0.0313], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0179, 0.0217, 0.0189, 0.0194, 0.0221, 0.0205, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:34:46,605 INFO [train.py:904] (4/8) Epoch 29, batch 6150, loss[loss=0.1576, simple_loss=0.245, pruned_loss=0.03511, over 17044.00 frames. ], tot_loss[loss=0.196, simple_loss=0.2849, pruned_loss=0.05359, over 3090310.06 frames. ], batch size: 55, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:34:47,112 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=290354.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:35:27,382 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=290379.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:35:30,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2000, 2.8016, 3.0769, 1.8325, 3.2300, 3.2495, 2.7153, 2.5420], device='cuda:4'), covar=tensor([0.0892, 0.0358, 0.0245, 0.1246, 0.0122, 0.0267, 0.0485, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0103, 0.0139, 0.0088, 0.0133, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:35:44,979 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.708e+02 2.649e+02 3.165e+02 3.863e+02 6.871e+02, threshold=6.329e+02, percent-clipped=1.0 2023-05-02 18:35:50,326 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8939, 4.6762, 4.9003, 5.0983, 5.2900, 4.7671, 5.2887, 5.2940], device='cuda:4'), covar=tensor([0.2177, 0.1445, 0.1938, 0.0815, 0.0644, 0.0957, 0.0726, 0.0757], device='cuda:4'), in_proj_covar=tensor([0.0687, 0.0836, 0.0964, 0.0850, 0.0648, 0.0669, 0.0708, 0.0825], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:35:50,595 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-02 18:36:04,348 INFO [train.py:904] (4/8) Epoch 29, batch 6200, loss[loss=0.2132, simple_loss=0.2817, pruned_loss=0.07236, over 11644.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2834, pruned_loss=0.0536, over 3082495.83 frames. ], batch size: 248, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:36:08,330 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=290406.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:36:16,580 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4769, 3.2729, 3.6956, 1.9259, 3.8599, 3.8187, 2.9523, 2.8938], device='cuda:4'), covar=tensor([0.0826, 0.0327, 0.0219, 0.1217, 0.0083, 0.0220, 0.0464, 0.0524], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0111, 0.0103, 0.0139, 0.0088, 0.0133, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 18:37:21,098 INFO [train.py:904] (4/8) Epoch 29, batch 6250, loss[loss=0.1647, simple_loss=0.2563, pruned_loss=0.03659, over 17119.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2829, pruned_loss=0.05316, over 3095490.83 frames. ], batch size: 49, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:38:15,680 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.611e+02 2.626e+02 3.170e+02 3.673e+02 5.508e+02, threshold=6.341e+02, percent-clipped=0.0 2023-05-02 18:38:34,375 INFO [train.py:904] (4/8) Epoch 29, batch 6300, loss[loss=0.1735, simple_loss=0.265, pruned_loss=0.04094, over 17110.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.2827, pruned_loss=0.05261, over 3096802.63 frames. ], batch size: 49, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:39:52,799 INFO [train.py:904] (4/8) Epoch 29, batch 6350, loss[loss=0.1839, simple_loss=0.2766, pruned_loss=0.04561, over 16826.00 frames. ], tot_loss[loss=0.1952, simple_loss=0.283, pruned_loss=0.05366, over 3086729.33 frames. ], batch size: 96, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:40:28,180 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=290576.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 18:40:46,403 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0767, 4.0555, 3.9865, 3.0974, 3.9810, 1.8323, 3.8177, 3.4462], device='cuda:4'), covar=tensor([0.0148, 0.0121, 0.0214, 0.0308, 0.0104, 0.3114, 0.0153, 0.0348], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0179, 0.0218, 0.0189, 0.0194, 0.0222, 0.0205, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:40:50,035 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.932e+02 2.799e+02 3.423e+02 4.274e+02 8.913e+02, threshold=6.847e+02, percent-clipped=6.0 2023-05-02 18:41:03,796 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7705, 5.0714, 5.2383, 5.0165, 5.0751, 5.6130, 5.1180, 4.8922], device='cuda:4'), covar=tensor([0.1210, 0.2073, 0.2706, 0.2007, 0.2466, 0.1121, 0.1893, 0.2689], device='cuda:4'), in_proj_covar=tensor([0.0433, 0.0644, 0.0714, 0.0520, 0.0698, 0.0732, 0.0553, 0.0695], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 18:41:09,296 INFO [train.py:904] (4/8) Epoch 29, batch 6400, loss[loss=0.1843, simple_loss=0.2754, pruned_loss=0.04657, over 15275.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.2835, pruned_loss=0.05469, over 3085612.83 frames. ], batch size: 191, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:41:39,257 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=290624.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 18:42:23,291 INFO [train.py:904] (4/8) Epoch 29, batch 6450, loss[loss=0.169, simple_loss=0.2619, pruned_loss=0.03805, over 16810.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2841, pruned_loss=0.05451, over 3095610.23 frames. ], batch size: 83, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:42:23,804 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=290654.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:42:34,223 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 18:42:38,744 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0010, 3.9151, 4.0716, 4.1785, 4.2568, 3.8833, 4.2259, 4.3173], device='cuda:4'), covar=tensor([0.1739, 0.1177, 0.1280, 0.0684, 0.0627, 0.1487, 0.0918, 0.0677], device='cuda:4'), in_proj_covar=tensor([0.0687, 0.0836, 0.0964, 0.0847, 0.0645, 0.0667, 0.0708, 0.0823], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:42:43,402 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-02 18:42:53,394 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=290674.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:43:19,913 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.946e+02 2.663e+02 2.912e+02 3.666e+02 7.754e+02, threshold=5.824e+02, percent-clipped=2.0 2023-05-02 18:43:37,903 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=290702.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:43:40,038 INFO [train.py:904] (4/8) Epoch 29, batch 6500, loss[loss=0.1995, simple_loss=0.2807, pruned_loss=0.05914, over 16566.00 frames. ], tot_loss[loss=0.1956, simple_loss=0.2828, pruned_loss=0.05423, over 3100214.79 frames. ], batch size: 62, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:43:44,349 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=290706.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:43:53,178 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5102, 3.4879, 2.5677, 2.1399, 2.3831, 2.2127, 3.6278, 3.0985], device='cuda:4'), covar=tensor([0.3196, 0.0664, 0.2151, 0.2960, 0.2853, 0.2534, 0.0508, 0.1534], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0275, 0.0313, 0.0328, 0.0306, 0.0278, 0.0305, 0.0351], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 18:44:23,894 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-02 18:44:51,376 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.61 vs. limit=5.0 2023-05-02 18:44:59,498 INFO [train.py:904] (4/8) Epoch 29, batch 6550, loss[loss=0.176, simple_loss=0.2847, pruned_loss=0.03363, over 16823.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2853, pruned_loss=0.05446, over 3115007.17 frames. ], batch size: 102, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:44:59,827 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=290754.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:45:01,565 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 18:45:58,154 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.498e+02 2.774e+02 3.264e+02 3.867e+02 7.263e+02, threshold=6.527e+02, percent-clipped=1.0 2023-05-02 18:46:19,466 INFO [train.py:904] (4/8) Epoch 29, batch 6600, loss[loss=0.2397, simple_loss=0.3113, pruned_loss=0.08403, over 11869.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2874, pruned_loss=0.05469, over 3118980.94 frames. ], batch size: 248, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:47:38,810 INFO [train.py:904] (4/8) Epoch 29, batch 6650, loss[loss=0.1797, simple_loss=0.2703, pruned_loss=0.04449, over 16715.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2875, pruned_loss=0.05559, over 3118515.90 frames. ], batch size: 76, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:47:44,095 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7073, 3.7653, 3.8754, 3.6678, 3.8782, 4.1936, 3.8775, 3.6116], device='cuda:4'), covar=tensor([0.2624, 0.2434, 0.2671, 0.2568, 0.2496, 0.1913, 0.1794, 0.2781], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0643, 0.0714, 0.0521, 0.0698, 0.0733, 0.0554, 0.0698], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 18:48:35,966 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.884e+02 2.668e+02 3.227e+02 3.787e+02 6.260e+02, threshold=6.453e+02, percent-clipped=0.0 2023-05-02 18:48:55,746 INFO [train.py:904] (4/8) Epoch 29, batch 6700, loss[loss=0.2503, simple_loss=0.3102, pruned_loss=0.09522, over 11722.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2866, pruned_loss=0.05636, over 3097381.83 frames. ], batch size: 248, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:50:07,299 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4119, 5.7294, 5.4214, 5.5400, 5.1844, 5.2152, 5.0864, 5.8559], device='cuda:4'), covar=tensor([0.1336, 0.0801, 0.0977, 0.0796, 0.0835, 0.0666, 0.1246, 0.0768], device='cuda:4'), in_proj_covar=tensor([0.0721, 0.0864, 0.0713, 0.0673, 0.0551, 0.0553, 0.0726, 0.0680], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:50:12,969 INFO [train.py:904] (4/8) Epoch 29, batch 6750, loss[loss=0.1915, simple_loss=0.274, pruned_loss=0.05447, over 16877.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.286, pruned_loss=0.05672, over 3100720.93 frames. ], batch size: 116, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:50:40,238 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4153, 3.3062, 2.6888, 2.1949, 2.2308, 2.2865, 3.4620, 3.0002], device='cuda:4'), covar=tensor([0.3084, 0.0692, 0.1931, 0.3049, 0.2784, 0.2358, 0.0515, 0.1467], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0276, 0.0315, 0.0330, 0.0308, 0.0279, 0.0307, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 18:50:41,460 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6447, 2.7923, 2.4275, 2.5798, 3.1244, 2.7813, 3.2147, 3.3134], device='cuda:4'), covar=tensor([0.0135, 0.0439, 0.0516, 0.0438, 0.0282, 0.0403, 0.0255, 0.0259], device='cuda:4'), in_proj_covar=tensor([0.0231, 0.0244, 0.0234, 0.0235, 0.0245, 0.0243, 0.0239, 0.0243], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:50:43,264 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=290974.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:51:10,928 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.985e+02 2.623e+02 3.095e+02 3.882e+02 8.048e+02, threshold=6.190e+02, percent-clipped=2.0 2023-05-02 18:51:28,424 INFO [train.py:904] (4/8) Epoch 29, batch 6800, loss[loss=0.2243, simple_loss=0.2983, pruned_loss=0.07514, over 11437.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2875, pruned_loss=0.05735, over 3104519.95 frames. ], batch size: 246, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:51:57,164 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=291022.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:52:18,498 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9450, 2.7743, 2.9051, 2.1991, 2.7261, 2.2012, 2.7643, 3.0037], device='cuda:4'), covar=tensor([0.0314, 0.0862, 0.0512, 0.1850, 0.0832, 0.0967, 0.0709, 0.0790], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0173, 0.0172, 0.0158, 0.0149, 0.0134, 0.0146, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 18:52:19,620 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9109, 2.7661, 2.9304, 2.1301, 2.7475, 2.1962, 2.7899, 2.9663], device='cuda:4'), covar=tensor([0.0262, 0.0811, 0.0491, 0.1907, 0.0807, 0.0970, 0.0566, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0173, 0.0172, 0.0158, 0.0149, 0.0134, 0.0146, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 18:52:39,332 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8401, 4.8122, 4.6897, 3.7279, 4.7224, 1.7050, 4.4411, 4.2537], device='cuda:4'), covar=tensor([0.0174, 0.0147, 0.0242, 0.0505, 0.0150, 0.3256, 0.0275, 0.0332], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0178, 0.0217, 0.0188, 0.0193, 0.0221, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 18:52:46,127 INFO [train.py:904] (4/8) Epoch 29, batch 6850, loss[loss=0.1733, simple_loss=0.2743, pruned_loss=0.03614, over 16828.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.288, pruned_loss=0.05715, over 3117361.36 frames. ], batch size: 42, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:53:09,384 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=291069.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:53:42,653 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.921e+02 2.621e+02 3.195e+02 3.750e+02 8.276e+02, threshold=6.389e+02, percent-clipped=2.0 2023-05-02 18:54:02,736 INFO [train.py:904] (4/8) Epoch 29, batch 6900, loss[loss=0.2512, simple_loss=0.3173, pruned_loss=0.09251, over 11630.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2907, pruned_loss=0.05706, over 3113325.26 frames. ], batch size: 248, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:54:44,567 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=291130.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 18:55:20,068 INFO [train.py:904] (4/8) Epoch 29, batch 6950, loss[loss=0.2019, simple_loss=0.2904, pruned_loss=0.05668, over 16866.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.2929, pruned_loss=0.05938, over 3080950.77 frames. ], batch size: 109, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:55:51,123 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9026, 2.7000, 2.6040, 1.8905, 2.5314, 2.6848, 2.5847, 1.8815], device='cuda:4'), covar=tensor([0.0525, 0.0112, 0.0115, 0.0443, 0.0160, 0.0157, 0.0151, 0.0460], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0091, 0.0092, 0.0137, 0.0104, 0.0117, 0.0100, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 18:56:18,587 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.816e+02 3.069e+02 3.631e+02 4.410e+02 8.633e+02, threshold=7.261e+02, percent-clipped=4.0 2023-05-02 18:56:36,625 INFO [train.py:904] (4/8) Epoch 29, batch 7000, loss[loss=0.2047, simple_loss=0.299, pruned_loss=0.05525, over 15421.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2922, pruned_loss=0.05822, over 3079037.81 frames. ], batch size: 190, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:57:45,291 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.82 vs. limit=5.0 2023-05-02 18:57:50,282 INFO [train.py:904] (4/8) Epoch 29, batch 7050, loss[loss=0.2155, simple_loss=0.2997, pruned_loss=0.06571, over 15320.00 frames. ], tot_loss[loss=0.2035, simple_loss=0.2924, pruned_loss=0.05732, over 3102905.16 frames. ], batch size: 190, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:58:49,441 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.848e+02 2.753e+02 3.384e+02 4.060e+02 1.222e+03, threshold=6.768e+02, percent-clipped=4.0 2023-05-02 18:59:07,325 INFO [train.py:904] (4/8) Epoch 29, batch 7100, loss[loss=0.1837, simple_loss=0.2741, pruned_loss=0.04664, over 16936.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.2908, pruned_loss=0.05671, over 3109503.25 frames. ], batch size: 109, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 18:59:19,272 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3384, 3.3202, 2.5137, 2.2022, 2.2257, 2.1540, 3.3161, 2.8981], device='cuda:4'), covar=tensor([0.3443, 0.0857, 0.2343, 0.3043, 0.2824, 0.2690, 0.0728, 0.1610], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0276, 0.0315, 0.0329, 0.0307, 0.0279, 0.0306, 0.0353], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 19:00:25,886 INFO [train.py:904] (4/8) Epoch 29, batch 7150, loss[loss=0.1815, simple_loss=0.2731, pruned_loss=0.04494, over 16489.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2887, pruned_loss=0.05618, over 3117259.28 frames. ], batch size: 68, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 19:01:23,564 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.870e+02 2.798e+02 3.396e+02 4.128e+02 9.099e+02, threshold=6.792e+02, percent-clipped=1.0 2023-05-02 19:01:41,834 INFO [train.py:904] (4/8) Epoch 29, batch 7200, loss[loss=0.1979, simple_loss=0.2987, pruned_loss=0.04853, over 16675.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.286, pruned_loss=0.05454, over 3113193.05 frames. ], batch size: 134, lr: 2.31e-03, grad_scale: 8.0 2023-05-02 19:01:45,981 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5393, 3.5932, 3.3757, 3.0355, 3.2165, 3.4747, 3.3449, 3.3166], device='cuda:4'), covar=tensor([0.0538, 0.0640, 0.0277, 0.0251, 0.0480, 0.0491, 0.1137, 0.0435], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0470, 0.0363, 0.0363, 0.0360, 0.0418, 0.0249, 0.0435], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:02:14,489 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=291425.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:02:40,555 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8418, 4.9301, 5.2166, 5.1868, 5.2409, 4.9228, 4.8946, 4.7313], device='cuda:4'), covar=tensor([0.0323, 0.0519, 0.0372, 0.0394, 0.0465, 0.0341, 0.0952, 0.0454], device='cuda:4'), in_proj_covar=tensor([0.0438, 0.0498, 0.0478, 0.0440, 0.0524, 0.0505, 0.0583, 0.0405], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 19:02:40,703 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9761, 2.7609, 2.6557, 4.8050, 3.5394, 4.1339, 1.6049, 3.1174], device='cuda:4'), covar=tensor([0.1292, 0.0850, 0.1258, 0.0156, 0.0336, 0.0424, 0.1708, 0.0814], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0182, 0.0203, 0.0206, 0.0209, 0.0220, 0.0213, 0.0202], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 19:03:00,075 INFO [train.py:904] (4/8) Epoch 29, batch 7250, loss[loss=0.18, simple_loss=0.2628, pruned_loss=0.04865, over 11549.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2834, pruned_loss=0.05336, over 3097981.89 frames. ], batch size: 246, lr: 2.31e-03, grad_scale: 4.0 2023-05-02 19:03:06,881 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0429, 3.1974, 3.3832, 2.1431, 3.0202, 2.2667, 3.4737, 3.4974], device='cuda:4'), covar=tensor([0.0285, 0.0948, 0.0639, 0.2148, 0.0890, 0.1095, 0.0669, 0.1118], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0173, 0.0172, 0.0158, 0.0149, 0.0134, 0.0147, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 19:03:58,886 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.627e+02 2.585e+02 3.000e+02 3.629e+02 8.129e+02, threshold=6.000e+02, percent-clipped=2.0 2023-05-02 19:04:16,115 INFO [train.py:904] (4/8) Epoch 29, batch 7300, loss[loss=0.235, simple_loss=0.3004, pruned_loss=0.08483, over 11672.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2823, pruned_loss=0.05286, over 3101147.40 frames. ], batch size: 246, lr: 2.31e-03, grad_scale: 4.0 2023-05-02 19:05:20,423 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0485, 4.1361, 3.9386, 3.6509, 3.7287, 4.0569, 3.6704, 3.8614], device='cuda:4'), covar=tensor([0.0602, 0.0581, 0.0282, 0.0271, 0.0650, 0.0458, 0.1167, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0468, 0.0361, 0.0361, 0.0358, 0.0416, 0.0248, 0.0433], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:05:32,025 INFO [train.py:904] (4/8) Epoch 29, batch 7350, loss[loss=0.2452, simple_loss=0.3087, pruned_loss=0.09088, over 10831.00 frames. ], tot_loss[loss=0.1952, simple_loss=0.2833, pruned_loss=0.05357, over 3077025.97 frames. ], batch size: 246, lr: 2.31e-03, grad_scale: 4.0 2023-05-02 19:06:32,138 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.793e+02 2.867e+02 3.369e+02 3.998e+02 9.840e+02, threshold=6.738e+02, percent-clipped=8.0 2023-05-02 19:06:49,563 INFO [train.py:904] (4/8) Epoch 29, batch 7400, loss[loss=0.229, simple_loss=0.3034, pruned_loss=0.07732, over 11383.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2847, pruned_loss=0.0546, over 3072819.00 frames. ], batch size: 248, lr: 2.31e-03, grad_scale: 4.0 2023-05-02 19:08:07,280 INFO [train.py:904] (4/8) Epoch 29, batch 7450, loss[loss=0.2063, simple_loss=0.2842, pruned_loss=0.06417, over 11204.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2849, pruned_loss=0.05493, over 3091375.99 frames. ], batch size: 248, lr: 2.31e-03, grad_scale: 4.0 2023-05-02 19:09:05,200 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5082, 4.0779, 4.0736, 2.6404, 3.6510, 4.1498, 3.6792, 2.3351], device='cuda:4'), covar=tensor([0.0575, 0.0063, 0.0064, 0.0455, 0.0117, 0.0105, 0.0104, 0.0483], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0091, 0.0092, 0.0136, 0.0103, 0.0117, 0.0099, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 19:09:10,902 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.991e+02 2.789e+02 3.286e+02 3.823e+02 6.627e+02, threshold=6.571e+02, percent-clipped=0.0 2023-05-02 19:09:28,108 INFO [train.py:904] (4/8) Epoch 29, batch 7500, loss[loss=0.1913, simple_loss=0.2799, pruned_loss=0.0513, over 16719.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2849, pruned_loss=0.05457, over 3078241.80 frames. ], batch size: 124, lr: 2.31e-03, grad_scale: 4.0 2023-05-02 19:10:01,612 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=291725.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:10:45,759 INFO [train.py:904] (4/8) Epoch 29, batch 7550, loss[loss=0.1775, simple_loss=0.2625, pruned_loss=0.0462, over 17066.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.284, pruned_loss=0.05459, over 3085513.18 frames. ], batch size: 55, lr: 2.31e-03, grad_scale: 4.0 2023-05-02 19:11:15,343 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=291773.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:11:19,706 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=291776.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:11:26,752 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0039, 5.0605, 4.8593, 4.4671, 4.5514, 4.9403, 4.7915, 4.6474], device='cuda:4'), covar=tensor([0.0626, 0.0454, 0.0334, 0.0358, 0.0943, 0.0478, 0.0425, 0.0707], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0468, 0.0361, 0.0362, 0.0357, 0.0417, 0.0248, 0.0434], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:11:44,715 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.882e+02 2.637e+02 3.162e+02 3.774e+02 1.302e+03, threshold=6.323e+02, percent-clipped=2.0 2023-05-02 19:12:01,406 INFO [train.py:904] (4/8) Epoch 29, batch 7600, loss[loss=0.1992, simple_loss=0.2863, pruned_loss=0.05601, over 16958.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2834, pruned_loss=0.05488, over 3075029.65 frames. ], batch size: 109, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:12:19,363 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6708, 1.8705, 2.2571, 2.5954, 2.6178, 2.9046, 1.9747, 2.9082], device='cuda:4'), covar=tensor([0.0252, 0.0610, 0.0386, 0.0377, 0.0370, 0.0244, 0.0697, 0.0183], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0197, 0.0186, 0.0191, 0.0209, 0.0166, 0.0204, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:12:52,640 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=291837.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:13:18,969 INFO [train.py:904] (4/8) Epoch 29, batch 7650, loss[loss=0.1941, simple_loss=0.2853, pruned_loss=0.05147, over 16738.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2838, pruned_loss=0.05516, over 3081614.05 frames. ], batch size: 124, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:14:09,163 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-05-02 19:14:20,910 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.976e+02 2.915e+02 3.313e+02 4.275e+02 8.487e+02, threshold=6.627e+02, percent-clipped=5.0 2023-05-02 19:14:36,034 INFO [train.py:904] (4/8) Epoch 29, batch 7700, loss[loss=0.1828, simple_loss=0.2721, pruned_loss=0.04677, over 16720.00 frames. ], tot_loss[loss=0.198, simple_loss=0.284, pruned_loss=0.05598, over 3060581.41 frames. ], batch size: 124, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:14:43,780 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4581, 3.6555, 3.9428, 1.9966, 4.1511, 4.2849, 3.0362, 2.8996], device='cuda:4'), covar=tensor([0.1173, 0.0241, 0.0247, 0.1357, 0.0090, 0.0153, 0.0486, 0.0632], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0104, 0.0139, 0.0088, 0.0132, 0.0130, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 19:15:53,493 INFO [train.py:904] (4/8) Epoch 29, batch 7750, loss[loss=0.222, simple_loss=0.3085, pruned_loss=0.06771, over 16432.00 frames. ], tot_loss[loss=0.1981, simple_loss=0.2847, pruned_loss=0.0558, over 3066875.99 frames. ], batch size: 146, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:16:55,756 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.177e+02 2.797e+02 3.314e+02 3.963e+02 7.468e+02, threshold=6.627e+02, percent-clipped=1.0 2023-05-02 19:17:14,568 INFO [train.py:904] (4/8) Epoch 29, batch 7800, loss[loss=0.2066, simple_loss=0.2925, pruned_loss=0.06037, over 16927.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2858, pruned_loss=0.05659, over 3063479.56 frames. ], batch size: 109, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:18:30,015 INFO [train.py:904] (4/8) Epoch 29, batch 7850, loss[loss=0.1759, simple_loss=0.2737, pruned_loss=0.03902, over 16736.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2868, pruned_loss=0.05675, over 3060056.28 frames. ], batch size: 89, lr: 2.30e-03, grad_scale: 2.0 2023-05-02 19:18:30,478 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=292054.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:19:30,094 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.904e+02 2.705e+02 3.148e+02 3.782e+02 7.710e+02, threshold=6.297e+02, percent-clipped=1.0 2023-05-02 19:19:43,447 INFO [train.py:904] (4/8) Epoch 29, batch 7900, loss[loss=0.1777, simple_loss=0.2691, pruned_loss=0.04318, over 16914.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2861, pruned_loss=0.05619, over 3070345.41 frames. ], batch size: 83, lr: 2.30e-03, grad_scale: 2.0 2023-05-02 19:19:59,953 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=292115.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:20:07,749 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1221, 2.5129, 2.0548, 2.2482, 2.8049, 2.4802, 2.7382, 2.9572], device='cuda:4'), covar=tensor([0.0212, 0.0446, 0.0633, 0.0527, 0.0319, 0.0402, 0.0300, 0.0310], device='cuda:4'), in_proj_covar=tensor([0.0230, 0.0243, 0.0233, 0.0233, 0.0244, 0.0241, 0.0239, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:20:08,004 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-02 19:20:19,975 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7942, 4.8662, 4.6706, 4.3469, 4.4000, 4.7723, 4.6148, 4.4866], device='cuda:4'), covar=tensor([0.0641, 0.0576, 0.0324, 0.0321, 0.0844, 0.0541, 0.0399, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0470, 0.0362, 0.0363, 0.0359, 0.0418, 0.0250, 0.0435], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:20:26,271 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=292132.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:20:26,381 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2034, 4.2871, 4.1010, 3.8420, 3.8785, 4.2144, 3.9086, 3.9937], device='cuda:4'), covar=tensor([0.0668, 0.0730, 0.0337, 0.0323, 0.0727, 0.0586, 0.0853, 0.0642], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0470, 0.0362, 0.0363, 0.0359, 0.0418, 0.0250, 0.0435], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:21:01,246 INFO [train.py:904] (4/8) Epoch 29, batch 7950, loss[loss=0.2008, simple_loss=0.2835, pruned_loss=0.05901, over 16784.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2862, pruned_loss=0.05651, over 3075793.78 frames. ], batch size: 124, lr: 2.30e-03, grad_scale: 2.0 2023-05-02 19:21:07,672 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-02 19:21:10,850 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-02 19:22:03,657 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.684e+02 2.577e+02 3.154e+02 3.832e+02 7.252e+02, threshold=6.308e+02, percent-clipped=1.0 2023-05-02 19:22:18,472 INFO [train.py:904] (4/8) Epoch 29, batch 8000, loss[loss=0.2636, simple_loss=0.3269, pruned_loss=0.1002, over 11560.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2869, pruned_loss=0.05648, over 3100584.25 frames. ], batch size: 246, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:22:21,316 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3026, 4.3763, 4.6690, 4.6296, 4.6641, 4.3893, 4.3816, 4.3476], device='cuda:4'), covar=tensor([0.0336, 0.0629, 0.0409, 0.0439, 0.0464, 0.0435, 0.0930, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0440, 0.0499, 0.0479, 0.0443, 0.0527, 0.0506, 0.0584, 0.0406], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 19:23:31,105 INFO [train.py:904] (4/8) Epoch 29, batch 8050, loss[loss=0.1739, simple_loss=0.2735, pruned_loss=0.03715, over 16776.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2871, pruned_loss=0.05634, over 3102495.07 frames. ], batch size: 102, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:24:32,478 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.919e+02 2.613e+02 3.145e+02 3.848e+02 6.977e+02, threshold=6.289e+02, percent-clipped=2.0 2023-05-02 19:24:46,378 INFO [train.py:904] (4/8) Epoch 29, batch 8100, loss[loss=0.198, simple_loss=0.2914, pruned_loss=0.05229, over 16685.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2868, pruned_loss=0.05606, over 3084111.99 frames. ], batch size: 124, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:25:02,664 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-05-02 19:25:03,324 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7365, 3.8075, 3.9224, 3.7132, 3.8949, 4.2247, 3.9041, 3.6039], device='cuda:4'), covar=tensor([0.2346, 0.2205, 0.2519, 0.2293, 0.2447, 0.1917, 0.1715, 0.2650], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0639, 0.0708, 0.0518, 0.0691, 0.0730, 0.0549, 0.0693], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 19:25:03,486 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7220, 1.8002, 1.5690, 1.4193, 1.9024, 1.5620, 1.6196, 1.8942], device='cuda:4'), covar=tensor([0.0245, 0.0386, 0.0529, 0.0455, 0.0283, 0.0363, 0.0220, 0.0293], device='cuda:4'), in_proj_covar=tensor([0.0230, 0.0243, 0.0233, 0.0233, 0.0244, 0.0241, 0.0239, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:25:09,317 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7525, 2.9961, 3.2510, 2.0335, 2.8924, 2.1085, 3.2949, 3.3709], device='cuda:4'), covar=tensor([0.0305, 0.0948, 0.0660, 0.2214, 0.0893, 0.1111, 0.0664, 0.0954], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0171, 0.0171, 0.0157, 0.0148, 0.0134, 0.0145, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 19:26:01,419 INFO [train.py:904] (4/8) Epoch 29, batch 8150, loss[loss=0.1839, simple_loss=0.2723, pruned_loss=0.04782, over 16808.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2847, pruned_loss=0.05534, over 3094366.47 frames. ], batch size: 39, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:27:01,238 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.703e+02 2.736e+02 3.221e+02 3.909e+02 8.294e+02, threshold=6.443e+02, percent-clipped=4.0 2023-05-02 19:27:04,901 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-05-02 19:27:15,046 INFO [train.py:904] (4/8) Epoch 29, batch 8200, loss[loss=0.2031, simple_loss=0.277, pruned_loss=0.06454, over 11344.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2821, pruned_loss=0.05474, over 3091463.01 frames. ], batch size: 248, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:27:25,887 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=292410.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:28:00,577 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=292432.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:28:34,797 INFO [train.py:904] (4/8) Epoch 29, batch 8250, loss[loss=0.1745, simple_loss=0.2705, pruned_loss=0.03929, over 16673.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2815, pruned_loss=0.05219, over 3076077.29 frames. ], batch size: 62, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:28:37,928 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0752, 2.1406, 2.6368, 3.0031, 2.8040, 3.3955, 2.3281, 3.3959], device='cuda:4'), covar=tensor([0.0224, 0.0564, 0.0388, 0.0319, 0.0392, 0.0222, 0.0619, 0.0227], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0196, 0.0185, 0.0191, 0.0207, 0.0166, 0.0203, 0.0167], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:29:17,990 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=292480.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:29:40,597 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4295, 4.5296, 4.6990, 4.4812, 4.5253, 5.0521, 4.5708, 4.2424], device='cuda:4'), covar=tensor([0.1420, 0.1957, 0.2286, 0.1844, 0.2531, 0.1006, 0.1594, 0.2534], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0635, 0.0703, 0.0513, 0.0686, 0.0726, 0.0547, 0.0689], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 19:29:41,434 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.407e+02 2.184e+02 2.675e+02 3.477e+02 6.200e+02, threshold=5.351e+02, percent-clipped=0.0 2023-05-02 19:29:55,807 INFO [train.py:904] (4/8) Epoch 29, batch 8300, loss[loss=0.1677, simple_loss=0.2718, pruned_loss=0.03177, over 16787.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2785, pruned_loss=0.04871, over 3094470.34 frames. ], batch size: 102, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:31:15,559 INFO [train.py:904] (4/8) Epoch 29, batch 8350, loss[loss=0.1761, simple_loss=0.2764, pruned_loss=0.03789, over 16659.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2776, pruned_loss=0.04702, over 3081401.15 frames. ], batch size: 134, lr: 2.30e-03, grad_scale: 4.0 2023-05-02 19:32:20,504 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.032e+02 2.659e+02 3.191e+02 5.260e+02, threshold=5.318e+02, percent-clipped=0.0 2023-05-02 19:32:36,080 INFO [train.py:904] (4/8) Epoch 29, batch 8400, loss[loss=0.1742, simple_loss=0.2655, pruned_loss=0.04142, over 16840.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2752, pruned_loss=0.0455, over 3056590.85 frames. ], batch size: 116, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:33:56,324 INFO [train.py:904] (4/8) Epoch 29, batch 8450, loss[loss=0.1792, simple_loss=0.2858, pruned_loss=0.03631, over 16394.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2737, pruned_loss=0.044, over 3057074.61 frames. ], batch size: 146, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:34:05,445 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8528, 3.7378, 3.9182, 4.0159, 4.1078, 3.7026, 4.0611, 4.1430], device='cuda:4'), covar=tensor([0.1758, 0.1222, 0.1315, 0.0710, 0.0615, 0.1863, 0.0772, 0.0899], device='cuda:4'), in_proj_covar=tensor([0.0670, 0.0817, 0.0941, 0.0832, 0.0635, 0.0654, 0.0695, 0.0812], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:35:03,480 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.472e+02 2.196e+02 2.552e+02 2.926e+02 6.173e+02, threshold=5.103e+02, percent-clipped=1.0 2023-05-02 19:35:19,248 INFO [train.py:904] (4/8) Epoch 29, batch 8500, loss[loss=0.1577, simple_loss=0.2561, pruned_loss=0.02967, over 16764.00 frames. ], tot_loss[loss=0.177, simple_loss=0.27, pruned_loss=0.042, over 3040378.99 frames. ], batch size: 102, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:35:28,844 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=292710.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:35:42,658 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4055, 3.5030, 3.6591, 3.6501, 3.6531, 3.5040, 3.5256, 3.5538], device='cuda:4'), covar=tensor([0.0529, 0.0965, 0.0669, 0.0597, 0.0614, 0.0743, 0.0946, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0436, 0.0496, 0.0476, 0.0440, 0.0521, 0.0502, 0.0579, 0.0403], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 19:35:50,930 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9221, 2.7368, 2.5949, 2.0050, 2.4561, 2.7426, 2.6110, 1.8967], device='cuda:4'), covar=tensor([0.0421, 0.0105, 0.0099, 0.0327, 0.0180, 0.0125, 0.0120, 0.0466], device='cuda:4'), in_proj_covar=tensor([0.0136, 0.0089, 0.0090, 0.0133, 0.0101, 0.0114, 0.0097, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 19:35:50,999 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1196, 2.3645, 2.4729, 2.9938, 1.8235, 3.2433, 1.9318, 2.8502], device='cuda:4'), covar=tensor([0.1117, 0.0671, 0.0956, 0.0177, 0.0095, 0.0351, 0.1416, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0182, 0.0201, 0.0204, 0.0207, 0.0218, 0.0211, 0.0200], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 19:36:08,531 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5520, 4.7145, 4.8451, 4.6058, 4.6875, 5.1917, 4.7047, 4.4038], device='cuda:4'), covar=tensor([0.1329, 0.1822, 0.2174, 0.2101, 0.2298, 0.0990, 0.1626, 0.2597], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0627, 0.0695, 0.0508, 0.0679, 0.0717, 0.0541, 0.0681], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 19:36:42,994 INFO [train.py:904] (4/8) Epoch 29, batch 8550, loss[loss=0.2052, simple_loss=0.3027, pruned_loss=0.05388, over 16835.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.268, pruned_loss=0.04122, over 3017348.31 frames. ], batch size: 116, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:36:52,467 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=292758.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:38:04,312 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.654e+02 2.162e+02 2.570e+02 3.257e+02 5.024e+02, threshold=5.139e+02, percent-clipped=0.0 2023-05-02 19:38:21,503 INFO [train.py:904] (4/8) Epoch 29, batch 8600, loss[loss=0.1661, simple_loss=0.2645, pruned_loss=0.03388, over 15412.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2682, pruned_loss=0.04002, over 3030958.24 frames. ], batch size: 191, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:38:59,574 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-02 19:39:03,803 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7327, 2.3417, 2.3695, 3.3820, 1.6207, 3.6021, 1.6058, 2.7676], device='cuda:4'), covar=tensor([0.1512, 0.0930, 0.1259, 0.0171, 0.0097, 0.0369, 0.1825, 0.0810], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0182, 0.0201, 0.0203, 0.0206, 0.0218, 0.0211, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 19:39:21,427 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9682, 2.6928, 3.0139, 2.1675, 2.7465, 2.1994, 2.7239, 2.8624], device='cuda:4'), covar=tensor([0.0270, 0.0939, 0.0450, 0.1898, 0.0780, 0.0978, 0.0636, 0.0808], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0168, 0.0168, 0.0155, 0.0146, 0.0132, 0.0144, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 19:39:41,634 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9791, 3.8511, 4.0685, 4.1664, 4.2709, 3.8558, 4.2240, 4.2987], device='cuda:4'), covar=tensor([0.1742, 0.1135, 0.1323, 0.0721, 0.0566, 0.1647, 0.0738, 0.0715], device='cuda:4'), in_proj_covar=tensor([0.0667, 0.0815, 0.0937, 0.0831, 0.0633, 0.0653, 0.0692, 0.0810], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:39:53,063 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9564, 2.6158, 2.9918, 2.0629, 2.7340, 2.1399, 2.6476, 2.7577], device='cuda:4'), covar=tensor([0.0315, 0.1151, 0.0472, 0.2112, 0.0818, 0.1040, 0.0660, 0.0967], device='cuda:4'), in_proj_covar=tensor([0.0158, 0.0168, 0.0168, 0.0155, 0.0146, 0.0132, 0.0144, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 19:40:01,866 INFO [train.py:904] (4/8) Epoch 29, batch 8650, loss[loss=0.1508, simple_loss=0.2481, pruned_loss=0.0268, over 12133.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2658, pruned_loss=0.03855, over 3022109.71 frames. ], batch size: 246, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:40:33,394 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7471, 4.5488, 4.8242, 4.9638, 5.1034, 4.5989, 5.1174, 5.0988], device='cuda:4'), covar=tensor([0.1879, 0.1230, 0.1586, 0.0707, 0.0562, 0.1060, 0.0568, 0.0745], device='cuda:4'), in_proj_covar=tensor([0.0666, 0.0813, 0.0936, 0.0829, 0.0632, 0.0652, 0.0691, 0.0809], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:41:31,353 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.553e+02 2.024e+02 2.393e+02 2.870e+02 4.192e+02, threshold=4.785e+02, percent-clipped=0.0 2023-05-02 19:41:48,487 INFO [train.py:904] (4/8) Epoch 29, batch 8700, loss[loss=0.1464, simple_loss=0.2443, pruned_loss=0.02419, over 16781.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2635, pruned_loss=0.03749, over 3033302.57 frames. ], batch size: 83, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:42:08,135 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.81 vs. limit=5.0 2023-05-02 19:43:26,842 INFO [train.py:904] (4/8) Epoch 29, batch 8750, loss[loss=0.1919, simple_loss=0.2923, pruned_loss=0.04576, over 16676.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2628, pruned_loss=0.03706, over 3023707.10 frames. ], batch size: 89, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:44:01,878 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=292967.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:45:00,946 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.419e+02 2.165e+02 2.592e+02 3.174e+02 6.684e+02, threshold=5.184e+02, percent-clipped=4.0 2023-05-02 19:45:20,130 INFO [train.py:904] (4/8) Epoch 29, batch 8800, loss[loss=0.1791, simple_loss=0.2707, pruned_loss=0.04369, over 16609.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2611, pruned_loss=0.03589, over 3042112.10 frames. ], batch size: 62, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:46:00,091 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2993, 3.4109, 2.0628, 3.8284, 2.5517, 3.7470, 2.0851, 2.6475], device='cuda:4'), covar=tensor([0.0435, 0.0439, 0.1830, 0.0268, 0.1003, 0.0572, 0.1880, 0.1018], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0176, 0.0191, 0.0167, 0.0176, 0.0214, 0.0200, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 19:46:12,245 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=293028.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:47:05,839 INFO [train.py:904] (4/8) Epoch 29, batch 8850, loss[loss=0.1398, simple_loss=0.2367, pruned_loss=0.0215, over 12148.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2642, pruned_loss=0.03518, over 3050728.18 frames. ], batch size: 247, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:47:12,803 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5576, 3.6670, 3.4101, 3.0643, 3.1786, 3.5415, 3.3162, 3.3174], device='cuda:4'), covar=tensor([0.0585, 0.0612, 0.0356, 0.0325, 0.0547, 0.0432, 0.1767, 0.0529], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0460, 0.0354, 0.0355, 0.0350, 0.0408, 0.0245, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:48:05,792 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=293081.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:48:31,037 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8193, 1.4381, 1.7109, 1.7471, 1.9065, 1.9193, 1.7589, 1.8667], device='cuda:4'), covar=tensor([0.0296, 0.0516, 0.0261, 0.0382, 0.0357, 0.0282, 0.0515, 0.0197], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0195, 0.0184, 0.0188, 0.0206, 0.0163, 0.0201, 0.0165], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:48:35,604 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.537e+02 2.279e+02 2.601e+02 3.166e+02 4.771e+02, threshold=5.201e+02, percent-clipped=0.0 2023-05-02 19:48:55,720 INFO [train.py:904] (4/8) Epoch 29, batch 8900, loss[loss=0.1764, simple_loss=0.2753, pruned_loss=0.03881, over 16426.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2645, pruned_loss=0.03502, over 3035749.94 frames. ], batch size: 146, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:49:23,372 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=293117.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:50:06,881 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-02 19:50:34,629 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=293142.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:51:01,960 INFO [train.py:904] (4/8) Epoch 29, batch 8950, loss[loss=0.1488, simple_loss=0.2427, pruned_loss=0.02745, over 15351.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2641, pruned_loss=0.03538, over 3053712.11 frames. ], batch size: 192, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:51:52,446 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=293178.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:52:32,341 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.460e+02 2.109e+02 2.425e+02 3.053e+02 5.475e+02, threshold=4.851e+02, percent-clipped=2.0 2023-05-02 19:52:53,145 INFO [train.py:904] (4/8) Epoch 29, batch 9000, loss[loss=0.1418, simple_loss=0.2325, pruned_loss=0.0256, over 12266.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2609, pruned_loss=0.03418, over 3048131.09 frames. ], batch size: 248, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:52:53,145 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 19:53:02,751 INFO [train.py:938] (4/8) Epoch 29, validation: loss=0.1431, simple_loss=0.2465, pruned_loss=0.01987, over 944034.00 frames. 2023-05-02 19:53:02,752 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 19:54:21,861 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1053, 5.4614, 5.2641, 5.2852, 4.9493, 4.9515, 4.8831, 5.5479], device='cuda:4'), covar=tensor([0.1289, 0.0846, 0.0867, 0.0796, 0.0773, 0.0833, 0.1218, 0.0853], device='cuda:4'), in_proj_covar=tensor([0.0705, 0.0850, 0.0697, 0.0659, 0.0541, 0.0543, 0.0711, 0.0668], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 19:54:33,792 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4892, 3.5316, 3.7016, 3.6908, 3.7005, 3.5447, 3.5684, 3.6000], device='cuda:4'), covar=tensor([0.0369, 0.0872, 0.0507, 0.0447, 0.0485, 0.0588, 0.0668, 0.0460], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0487, 0.0470, 0.0433, 0.0514, 0.0494, 0.0568, 0.0397], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 19:54:46,359 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.58 vs. limit=5.0 2023-05-02 19:54:47,962 INFO [train.py:904] (4/8) Epoch 29, batch 9050, loss[loss=0.1525, simple_loss=0.2473, pruned_loss=0.02885, over 16923.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2608, pruned_loss=0.03409, over 3056958.21 frames. ], batch size: 102, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:56:15,025 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.590e+02 2.113e+02 2.508e+02 2.955e+02 4.534e+02, threshold=5.015e+02, percent-clipped=0.0 2023-05-02 19:56:34,919 INFO [train.py:904] (4/8) Epoch 29, batch 9100, loss[loss=0.162, simple_loss=0.2664, pruned_loss=0.02879, over 16724.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.26, pruned_loss=0.03445, over 3064407.72 frames. ], batch size: 83, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:57:17,599 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=293323.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:58:34,110 INFO [train.py:904] (4/8) Epoch 29, batch 9150, loss[loss=0.1432, simple_loss=0.2447, pruned_loss=0.0208, over 16865.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2608, pruned_loss=0.03434, over 3070438.59 frames. ], batch size: 96, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 19:58:51,990 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=293361.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 19:58:56,702 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=293364.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:00:04,472 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.264e+02 2.108e+02 2.614e+02 3.370e+02 9.014e+02, threshold=5.228e+02, percent-clipped=3.0 2023-05-02 20:00:20,590 INFO [train.py:904] (4/8) Epoch 29, batch 9200, loss[loss=0.1507, simple_loss=0.243, pruned_loss=0.02919, over 17133.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2571, pruned_loss=0.03363, over 3079467.31 frames. ], batch size: 47, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:00:55,065 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=293422.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:01:00,757 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=293425.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:01:23,544 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=293437.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:01:58,342 INFO [train.py:904] (4/8) Epoch 29, batch 9250, loss[loss=0.1571, simple_loss=0.2577, pruned_loss=0.02826, over 16872.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2571, pruned_loss=0.03381, over 3071975.57 frames. ], batch size: 102, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:02:33,895 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2579, 4.4243, 4.1641, 3.8769, 3.8129, 4.3330, 4.0265, 3.9777], device='cuda:4'), covar=tensor([0.0721, 0.0668, 0.0412, 0.0400, 0.0931, 0.0534, 0.0736, 0.0705], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0458, 0.0354, 0.0354, 0.0349, 0.0407, 0.0244, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 20:02:39,017 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=293473.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:03:28,146 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.02 vs. limit=5.0 2023-05-02 20:03:28,486 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.628e+02 2.175e+02 2.514e+02 3.115e+02 5.451e+02, threshold=5.029e+02, percent-clipped=2.0 2023-05-02 20:03:49,264 INFO [train.py:904] (4/8) Epoch 29, batch 9300, loss[loss=0.1363, simple_loss=0.2307, pruned_loss=0.02095, over 15104.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.2556, pruned_loss=0.03309, over 3076886.79 frames. ], batch size: 190, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:04:37,725 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0185, 2.2810, 2.3881, 3.0966, 1.7900, 3.2818, 1.7904, 2.8312], device='cuda:4'), covar=tensor([0.1253, 0.0742, 0.1118, 0.0182, 0.0092, 0.0396, 0.1646, 0.0724], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0179, 0.0198, 0.0200, 0.0202, 0.0214, 0.0209, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:05:33,313 INFO [train.py:904] (4/8) Epoch 29, batch 9350, loss[loss=0.1632, simple_loss=0.2579, pruned_loss=0.03425, over 16350.00 frames. ], tot_loss[loss=0.1605, simple_loss=0.2553, pruned_loss=0.03285, over 3079627.64 frames. ], batch size: 146, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:06:53,156 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-05-02 20:06:56,975 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.575e+02 1.973e+02 2.358e+02 2.873e+02 5.255e+02, threshold=4.716e+02, percent-clipped=1.0 2023-05-02 20:07:15,889 INFO [train.py:904] (4/8) Epoch 29, batch 9400, loss[loss=0.1462, simple_loss=0.2378, pruned_loss=0.0273, over 12332.00 frames. ], tot_loss[loss=0.16, simple_loss=0.2547, pruned_loss=0.03263, over 3054330.47 frames. ], batch size: 246, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:07:53,879 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=293623.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:08:22,146 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-02 20:08:55,053 INFO [train.py:904] (4/8) Epoch 29, batch 9450, loss[loss=0.1675, simple_loss=0.2613, pruned_loss=0.03684, over 16902.00 frames. ], tot_loss[loss=0.161, simple_loss=0.2562, pruned_loss=0.03285, over 3060382.35 frames. ], batch size: 109, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:09:29,416 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=293671.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:10:18,287 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.478e+02 2.101e+02 2.359e+02 2.998e+02 5.430e+02, threshold=4.718e+02, percent-clipped=2.0 2023-05-02 20:10:34,114 INFO [train.py:904] (4/8) Epoch 29, batch 9500, loss[loss=0.1473, simple_loss=0.2481, pruned_loss=0.02326, over 16742.00 frames. ], tot_loss[loss=0.1604, simple_loss=0.2557, pruned_loss=0.03256, over 3064404.92 frames. ], batch size: 83, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:11:02,456 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=293717.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:11:08,991 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=293720.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:11:30,819 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-05-02 20:11:38,121 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3273, 2.9511, 3.0814, 1.8061, 3.2598, 3.3413, 2.8155, 2.7156], device='cuda:4'), covar=tensor([0.0739, 0.0303, 0.0249, 0.1256, 0.0112, 0.0195, 0.0450, 0.0464], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0108, 0.0099, 0.0135, 0.0084, 0.0127, 0.0126, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 20:11:41,932 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=293737.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:11:59,888 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4591, 4.5326, 4.3236, 3.9633, 4.0691, 4.4324, 4.1431, 4.1556], device='cuda:4'), covar=tensor([0.0608, 0.0644, 0.0360, 0.0376, 0.0822, 0.0637, 0.0625, 0.0658], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0455, 0.0353, 0.0353, 0.0348, 0.0405, 0.0243, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 20:12:19,483 INFO [train.py:904] (4/8) Epoch 29, batch 9550, loss[loss=0.1651, simple_loss=0.2677, pruned_loss=0.03123, over 16736.00 frames. ], tot_loss[loss=0.1603, simple_loss=0.2552, pruned_loss=0.03267, over 3067141.81 frames. ], batch size: 134, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:12:53,434 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1499, 2.5393, 2.5915, 1.9491, 2.7635, 2.8276, 2.5262, 2.5046], device='cuda:4'), covar=tensor([0.0668, 0.0298, 0.0260, 0.1036, 0.0141, 0.0250, 0.0483, 0.0460], device='cuda:4'), in_proj_covar=tensor([0.0145, 0.0109, 0.0099, 0.0136, 0.0085, 0.0127, 0.0127, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 20:12:54,818 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3325, 3.3750, 1.9700, 3.7308, 2.4946, 3.6695, 2.3148, 2.8230], device='cuda:4'), covar=tensor([0.0369, 0.0468, 0.1905, 0.0307, 0.1030, 0.0723, 0.1627, 0.0843], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0175, 0.0191, 0.0167, 0.0176, 0.0213, 0.0201, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:12:59,207 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=293773.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:13:25,491 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=293785.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:13:25,601 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4614, 3.3855, 3.5082, 3.5636, 3.6060, 3.3388, 3.5846, 3.6756], device='cuda:4'), covar=tensor([0.1302, 0.0940, 0.0981, 0.0629, 0.0629, 0.2259, 0.0925, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0655, 0.0799, 0.0917, 0.0817, 0.0621, 0.0640, 0.0681, 0.0791], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 20:13:28,657 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.01 vs. limit=5.0 2023-05-02 20:13:43,637 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.327e+02 2.273e+02 2.766e+02 3.522e+02 6.419e+02, threshold=5.532e+02, percent-clipped=5.0 2023-05-02 20:13:59,134 INFO [train.py:904] (4/8) Epoch 29, batch 9600, loss[loss=0.1797, simple_loss=0.2765, pruned_loss=0.04146, over 16203.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2571, pruned_loss=0.03374, over 3057784.16 frames. ], batch size: 165, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:14:32,482 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=293821.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:15:45,740 INFO [train.py:904] (4/8) Epoch 29, batch 9650, loss[loss=0.1605, simple_loss=0.2468, pruned_loss=0.03708, over 11936.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2585, pruned_loss=0.03382, over 3040827.46 frames. ], batch size: 250, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:16:38,057 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0169, 2.2098, 2.3106, 2.9744, 1.5569, 3.2370, 1.9269, 2.6872], device='cuda:4'), covar=tensor([0.1248, 0.0876, 0.1199, 0.0226, 0.0109, 0.0560, 0.1496, 0.0804], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0179, 0.0198, 0.0199, 0.0201, 0.0214, 0.0209, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:17:16,474 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.571e+02 2.271e+02 2.644e+02 3.266e+02 5.877e+02, threshold=5.288e+02, percent-clipped=1.0 2023-05-02 20:17:35,788 INFO [train.py:904] (4/8) Epoch 29, batch 9700, loss[loss=0.1494, simple_loss=0.2529, pruned_loss=0.023, over 16865.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2578, pruned_loss=0.03372, over 3025575.54 frames. ], batch size: 102, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:18:08,950 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6018, 3.0400, 3.3249, 1.9891, 2.8609, 2.2158, 3.1708, 3.2805], device='cuda:4'), covar=tensor([0.0324, 0.0878, 0.0543, 0.2315, 0.0857, 0.1075, 0.0694, 0.0942], device='cuda:4'), in_proj_covar=tensor([0.0157, 0.0166, 0.0167, 0.0154, 0.0145, 0.0130, 0.0142, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:19:17,209 INFO [train.py:904] (4/8) Epoch 29, batch 9750, loss[loss=0.15, simple_loss=0.2358, pruned_loss=0.03205, over 12665.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2566, pruned_loss=0.03382, over 3033085.22 frames. ], batch size: 248, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:19:21,623 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.09 vs. limit=2.0 2023-05-02 20:20:17,748 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9815, 2.3043, 2.3522, 3.1007, 1.7941, 3.2411, 1.8045, 2.7660], device='cuda:4'), covar=tensor([0.1274, 0.0721, 0.1124, 0.0175, 0.0078, 0.0345, 0.1591, 0.0715], device='cuda:4'), in_proj_covar=tensor([0.0171, 0.0178, 0.0197, 0.0198, 0.0201, 0.0213, 0.0208, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:20:24,559 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=293987.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:20:38,044 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.578e+02 2.282e+02 2.658e+02 3.259e+02 1.207e+03, threshold=5.315e+02, percent-clipped=3.0 2023-05-02 20:20:56,835 INFO [train.py:904] (4/8) Epoch 29, batch 9800, loss[loss=0.1706, simple_loss=0.2765, pruned_loss=0.03229, over 15440.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2579, pruned_loss=0.03351, over 3044777.94 frames. ], batch size: 191, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:21:22,935 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=294017.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:21:27,409 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=294020.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:21:40,147 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5807, 3.5851, 2.7285, 2.2346, 2.2680, 2.4151, 3.8548, 3.1286], device='cuda:4'), covar=tensor([0.3091, 0.0664, 0.2014, 0.3295, 0.3168, 0.2315, 0.0420, 0.1555], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0268, 0.0307, 0.0322, 0.0297, 0.0273, 0.0298, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 20:21:57,321 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1542, 3.2608, 3.2106, 2.2635, 2.9672, 3.3037, 3.1305, 2.0415], device='cuda:4'), covar=tensor([0.0527, 0.0063, 0.0084, 0.0427, 0.0141, 0.0094, 0.0101, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0088, 0.0090, 0.0133, 0.0101, 0.0114, 0.0097, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-02 20:22:29,373 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=294048.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:22:39,768 INFO [train.py:904] (4/8) Epoch 29, batch 9850, loss[loss=0.1464, simple_loss=0.2503, pruned_loss=0.0213, over 16871.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2591, pruned_loss=0.03318, over 3056034.77 frames. ], batch size: 102, lr: 2.30e-03, grad_scale: 16.0 2023-05-02 20:23:02,394 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=294065.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:23:09,658 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=294068.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:23:27,196 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.73 vs. limit=2.0 2023-05-02 20:24:14,911 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.468e+02 2.054e+02 2.475e+02 2.961e+02 4.915e+02, threshold=4.950e+02, percent-clipped=0.0 2023-05-02 20:24:32,429 INFO [train.py:904] (4/8) Epoch 29, batch 9900, loss[loss=0.1687, simple_loss=0.2523, pruned_loss=0.04261, over 12539.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2593, pruned_loss=0.0332, over 3049326.04 frames. ], batch size: 248, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:25:37,989 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9669, 3.1351, 3.5902, 2.0630, 3.0004, 2.2264, 3.4675, 3.4217], device='cuda:4'), covar=tensor([0.0257, 0.0896, 0.0521, 0.2130, 0.0783, 0.1048, 0.0566, 0.0968], device='cuda:4'), in_proj_covar=tensor([0.0156, 0.0164, 0.0166, 0.0153, 0.0144, 0.0129, 0.0141, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:26:27,775 INFO [train.py:904] (4/8) Epoch 29, batch 9950, loss[loss=0.1504, simple_loss=0.2529, pruned_loss=0.02391, over 17268.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2619, pruned_loss=0.03334, over 3067346.80 frames. ], batch size: 52, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:27:06,740 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=294170.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:27:24,923 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-05-02 20:28:08,138 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.560e+02 2.053e+02 2.347e+02 2.830e+02 4.908e+02, threshold=4.695e+02, percent-clipped=0.0 2023-05-02 20:28:27,256 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-05-02 20:28:27,670 INFO [train.py:904] (4/8) Epoch 29, batch 10000, loss[loss=0.1506, simple_loss=0.2454, pruned_loss=0.02791, over 12528.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2602, pruned_loss=0.03268, over 3081718.26 frames. ], batch size: 246, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:29:21,500 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=294231.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:30:08,536 INFO [train.py:904] (4/8) Epoch 29, batch 10050, loss[loss=0.1651, simple_loss=0.2602, pruned_loss=0.03502, over 12224.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2601, pruned_loss=0.03302, over 3080585.73 frames. ], batch size: 250, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:31:27,888 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.622e+02 2.082e+02 2.498e+02 3.055e+02 5.920e+02, threshold=4.997e+02, percent-clipped=2.0 2023-05-02 20:31:35,460 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0124, 1.8838, 1.6821, 1.4899, 2.0018, 1.6144, 1.4813, 1.9671], device='cuda:4'), covar=tensor([0.0236, 0.0368, 0.0515, 0.0458, 0.0283, 0.0355, 0.0206, 0.0278], device='cuda:4'), in_proj_covar=tensor([0.0219, 0.0237, 0.0226, 0.0227, 0.0235, 0.0234, 0.0229, 0.0233], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 20:31:41,186 INFO [train.py:904] (4/8) Epoch 29, batch 10100, loss[loss=0.1529, simple_loss=0.242, pruned_loss=0.03192, over 15314.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2596, pruned_loss=0.03294, over 3071934.20 frames. ], batch size: 190, lr: 2.30e-03, grad_scale: 8.0 2023-05-02 20:32:42,407 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=294339.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:32:46,796 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=294343.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:33:20,748 INFO [train.py:904] (4/8) Epoch 30, batch 0, loss[loss=0.1542, simple_loss=0.2454, pruned_loss=0.03147, over 16997.00 frames. ], tot_loss[loss=0.1542, simple_loss=0.2454, pruned_loss=0.03147, over 16997.00 frames. ], batch size: 41, lr: 2.26e-03, grad_scale: 8.0 2023-05-02 20:33:20,748 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 20:33:28,210 INFO [train.py:938] (4/8) Epoch 30, validation: loss=0.1426, simple_loss=0.2458, pruned_loss=0.01974, over 944034.00 frames. 2023-05-02 20:33:28,211 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 20:34:10,648 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7695, 2.6588, 2.4264, 4.0594, 3.0742, 3.9471, 1.5567, 2.9736], device='cuda:4'), covar=tensor([0.1521, 0.0748, 0.1387, 0.0177, 0.0127, 0.0435, 0.1827, 0.0867], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0178, 0.0198, 0.0199, 0.0200, 0.0214, 0.0209, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:34:28,805 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.544e+02 2.426e+02 2.887e+02 3.382e+02 7.415e+02, threshold=5.774e+02, percent-clipped=4.0 2023-05-02 20:34:29,604 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=294400.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:34:36,207 INFO [train.py:904] (4/8) Epoch 30, batch 50, loss[loss=0.175, simple_loss=0.2657, pruned_loss=0.04217, over 16025.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2629, pruned_loss=0.04417, over 754537.13 frames. ], batch size: 35, lr: 2.26e-03, grad_scale: 1.0 2023-05-02 20:35:02,114 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=294423.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:35:11,661 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-02 20:35:43,884 INFO [train.py:904] (4/8) Epoch 30, batch 100, loss[loss=0.1838, simple_loss=0.2762, pruned_loss=0.04569, over 16697.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2617, pruned_loss=0.04408, over 1306513.03 frames. ], batch size: 62, lr: 2.26e-03, grad_scale: 1.0 2023-05-02 20:36:23,327 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0169, 4.3933, 4.4036, 3.1472, 3.7155, 4.3899, 3.9651, 2.6799], device='cuda:4'), covar=tensor([0.0492, 0.0095, 0.0055, 0.0400, 0.0160, 0.0102, 0.0105, 0.0485], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0090, 0.0091, 0.0135, 0.0103, 0.0116, 0.0098, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 20:36:24,485 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=294484.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:36:32,244 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=294488.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:36:46,207 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.508e+02 2.392e+02 2.778e+02 3.445e+02 1.255e+03, threshold=5.557e+02, percent-clipped=5.0 2023-05-02 20:36:51,463 INFO [train.py:904] (4/8) Epoch 30, batch 150, loss[loss=0.17, simple_loss=0.2544, pruned_loss=0.04276, over 16485.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2606, pruned_loss=0.04263, over 1750780.74 frames. ], batch size: 146, lr: 2.25e-03, grad_scale: 1.0 2023-05-02 20:37:20,775 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=294526.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:37:32,809 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-05-02 20:37:50,413 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6437, 3.6670, 2.2584, 3.9212, 2.9521, 3.8631, 2.3918, 3.0211], device='cuda:4'), covar=tensor([0.0314, 0.0447, 0.1695, 0.0430, 0.0777, 0.0763, 0.1535, 0.0804], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0179, 0.0194, 0.0170, 0.0178, 0.0217, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:37:53,416 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=294549.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:37:59,954 INFO [train.py:904] (4/8) Epoch 30, batch 200, loss[loss=0.1447, simple_loss=0.2442, pruned_loss=0.02262, over 16806.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2597, pruned_loss=0.04286, over 2110103.78 frames. ], batch size: 39, lr: 2.25e-03, grad_scale: 1.0 2023-05-02 20:38:23,409 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=294571.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:38:41,559 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7436, 3.7856, 2.4521, 4.1233, 3.0931, 4.0602, 2.5134, 3.1605], device='cuda:4'), covar=tensor([0.0337, 0.0475, 0.1574, 0.0444, 0.0863, 0.0757, 0.1567, 0.0736], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0179, 0.0194, 0.0171, 0.0179, 0.0217, 0.0204, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:39:01,687 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.576e+02 2.189e+02 2.574e+02 3.148e+02 2.009e+03, threshold=5.148e+02, percent-clipped=2.0 2023-05-02 20:39:06,655 INFO [train.py:904] (4/8) Epoch 30, batch 250, loss[loss=0.1675, simple_loss=0.2591, pruned_loss=0.03795, over 16788.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2588, pruned_loss=0.04257, over 2376853.95 frames. ], batch size: 62, lr: 2.25e-03, grad_scale: 1.0 2023-05-02 20:39:45,415 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=294632.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:40:01,624 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=294643.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:40:02,903 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6000, 4.7671, 4.9328, 4.7292, 4.7540, 5.3426, 4.8133, 4.4800], device='cuda:4'), covar=tensor([0.1703, 0.2275, 0.2478, 0.2342, 0.2856, 0.1113, 0.1931, 0.2741], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0634, 0.0707, 0.0513, 0.0686, 0.0726, 0.0544, 0.0681], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 20:40:11,766 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8483, 2.8887, 2.7849, 5.0809, 4.0902, 4.3940, 1.7554, 3.1887], device='cuda:4'), covar=tensor([0.1331, 0.0818, 0.1188, 0.0201, 0.0214, 0.0396, 0.1643, 0.0789], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0181, 0.0201, 0.0204, 0.0204, 0.0217, 0.0212, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:40:15,312 INFO [train.py:904] (4/8) Epoch 30, batch 300, loss[loss=0.1545, simple_loss=0.2445, pruned_loss=0.03231, over 16809.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2543, pruned_loss=0.04017, over 2588234.56 frames. ], batch size: 42, lr: 2.25e-03, grad_scale: 1.0 2023-05-02 20:41:05,683 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3470, 4.2211, 4.2621, 3.9717, 4.0731, 4.2663, 4.0716, 4.1152], device='cuda:4'), covar=tensor([0.0696, 0.1081, 0.0403, 0.0364, 0.0806, 0.0604, 0.0687, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0466, 0.0361, 0.0362, 0.0356, 0.0415, 0.0248, 0.0431], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 20:41:06,777 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=294691.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:41:12,839 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=294695.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:41:19,992 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.370e+02 1.963e+02 2.186e+02 2.662e+02 5.332e+02, threshold=4.373e+02, percent-clipped=1.0 2023-05-02 20:41:25,715 INFO [train.py:904] (4/8) Epoch 30, batch 350, loss[loss=0.1833, simple_loss=0.2605, pruned_loss=0.05306, over 16456.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2523, pruned_loss=0.03913, over 2753061.52 frames. ], batch size: 75, lr: 2.25e-03, grad_scale: 1.0 2023-05-02 20:42:33,021 INFO [train.py:904] (4/8) Epoch 30, batch 400, loss[loss=0.1542, simple_loss=0.2496, pruned_loss=0.02941, over 17125.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2505, pruned_loss=0.03875, over 2876710.22 frames. ], batch size: 48, lr: 2.25e-03, grad_scale: 2.0 2023-05-02 20:42:55,445 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6812, 3.7220, 4.3849, 2.5568, 3.4990, 2.8695, 4.0329, 3.9670], device='cuda:4'), covar=tensor([0.0259, 0.0943, 0.0439, 0.2007, 0.0768, 0.0937, 0.0553, 0.1176], device='cuda:4'), in_proj_covar=tensor([0.0160, 0.0168, 0.0169, 0.0157, 0.0147, 0.0132, 0.0144, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:43:06,960 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=294779.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:43:34,529 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.286e+02 2.158e+02 2.521e+02 2.985e+02 1.602e+03, threshold=5.041e+02, percent-clipped=2.0 2023-05-02 20:43:41,853 INFO [train.py:904] (4/8) Epoch 30, batch 450, loss[loss=0.1503, simple_loss=0.2409, pruned_loss=0.0298, over 17140.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2498, pruned_loss=0.03865, over 2975088.65 frames. ], batch size: 47, lr: 2.25e-03, grad_scale: 2.0 2023-05-02 20:43:50,102 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9608, 2.9134, 2.6479, 5.0402, 3.7854, 4.4280, 1.9124, 3.2862], device='cuda:4'), covar=tensor([0.1422, 0.0939, 0.1361, 0.0261, 0.0232, 0.0433, 0.1734, 0.0788], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0181, 0.0200, 0.0203, 0.0203, 0.0217, 0.0211, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:44:12,087 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=294826.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:44:26,745 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-02 20:44:31,305 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3194, 3.3515, 3.6331, 2.4619, 3.2890, 3.6977, 3.4231, 2.2387], device='cuda:4'), covar=tensor([0.0581, 0.0197, 0.0070, 0.0444, 0.0145, 0.0134, 0.0121, 0.0509], device='cuda:4'), in_proj_covar=tensor([0.0141, 0.0092, 0.0092, 0.0137, 0.0104, 0.0117, 0.0100, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 20:44:36,118 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=294844.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:44:49,592 INFO [train.py:904] (4/8) Epoch 30, batch 500, loss[loss=0.152, simple_loss=0.2428, pruned_loss=0.03064, over 17206.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.249, pruned_loss=0.0378, over 3053712.78 frames. ], batch size: 46, lr: 2.25e-03, grad_scale: 2.0 2023-05-02 20:45:17,739 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=294874.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:45:19,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1587, 2.4658, 2.6010, 1.9038, 2.7208, 2.7837, 2.4935, 2.3786], device='cuda:4'), covar=tensor([0.0776, 0.0278, 0.0320, 0.1054, 0.0191, 0.0291, 0.0519, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0147, 0.0110, 0.0101, 0.0138, 0.0086, 0.0130, 0.0129, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:45:52,584 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.499e+02 2.033e+02 2.399e+02 2.919e+02 5.551e+02, threshold=4.798e+02, percent-clipped=2.0 2023-05-02 20:45:57,299 INFO [train.py:904] (4/8) Epoch 30, batch 550, loss[loss=0.1578, simple_loss=0.2479, pruned_loss=0.03384, over 16397.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2487, pruned_loss=0.03777, over 3102298.93 frames. ], batch size: 68, lr: 2.25e-03, grad_scale: 2.0 2023-05-02 20:46:13,069 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0423, 2.2233, 2.6462, 3.0220, 2.8155, 3.5338, 2.5071, 3.5385], device='cuda:4'), covar=tensor([0.0325, 0.0637, 0.0436, 0.0405, 0.0430, 0.0230, 0.0609, 0.0207], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0200, 0.0188, 0.0193, 0.0212, 0.0168, 0.0206, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 20:46:28,989 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=294927.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:47:05,460 INFO [train.py:904] (4/8) Epoch 30, batch 600, loss[loss=0.168, simple_loss=0.2357, pruned_loss=0.05018, over 16912.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2489, pruned_loss=0.03807, over 3151128.19 frames. ], batch size: 109, lr: 2.25e-03, grad_scale: 2.0 2023-05-02 20:47:52,983 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=294989.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:48:01,881 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=294995.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:48:07,356 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.536e+02 2.157e+02 2.572e+02 3.014e+02 4.443e+02, threshold=5.143e+02, percent-clipped=0.0 2023-05-02 20:48:13,112 INFO [train.py:904] (4/8) Epoch 30, batch 650, loss[loss=0.1666, simple_loss=0.2587, pruned_loss=0.0373, over 17138.00 frames. ], tot_loss[loss=0.1614, simple_loss=0.2472, pruned_loss=0.03775, over 3180411.48 frames. ], batch size: 48, lr: 2.25e-03, grad_scale: 2.0 2023-05-02 20:49:04,907 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=295043.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:49:15,138 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295050.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 20:49:20,779 INFO [train.py:904] (4/8) Epoch 30, batch 700, loss[loss=0.1529, simple_loss=0.2352, pruned_loss=0.03534, over 16701.00 frames. ], tot_loss[loss=0.161, simple_loss=0.2471, pruned_loss=0.03748, over 3214008.17 frames. ], batch size: 83, lr: 2.25e-03, grad_scale: 2.0 2023-05-02 20:49:39,333 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295068.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:49:54,481 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=295079.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:50:03,933 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-05-02 20:50:23,317 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.300e+02 2.027e+02 2.322e+02 2.790e+02 6.383e+02, threshold=4.644e+02, percent-clipped=2.0 2023-05-02 20:50:27,886 INFO [train.py:904] (4/8) Epoch 30, batch 750, loss[loss=0.1363, simple_loss=0.2186, pruned_loss=0.02705, over 16819.00 frames. ], tot_loss[loss=0.1612, simple_loss=0.2473, pruned_loss=0.03756, over 3238529.12 frames. ], batch size: 39, lr: 2.25e-03, grad_scale: 2.0 2023-05-02 20:50:56,714 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5657, 3.5566, 3.7780, 2.6534, 3.4707, 3.8358, 3.5970, 2.1729], device='cuda:4'), covar=tensor([0.0585, 0.0298, 0.0082, 0.0473, 0.0144, 0.0142, 0.0123, 0.0620], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0092, 0.0093, 0.0137, 0.0104, 0.0117, 0.0100, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 20:51:00,102 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=295127.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:51:03,349 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295129.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:51:24,558 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=295144.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:51:38,201 INFO [train.py:904] (4/8) Epoch 30, batch 800, loss[loss=0.1525, simple_loss=0.2484, pruned_loss=0.0283, over 17214.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.2472, pruned_loss=0.03728, over 3258777.49 frames. ], batch size: 44, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 20:51:39,889 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9625, 2.2357, 2.6529, 2.9489, 2.8652, 3.4665, 2.5058, 3.4782], device='cuda:4'), covar=tensor([0.0327, 0.0574, 0.0389, 0.0358, 0.0404, 0.0247, 0.0563, 0.0219], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0199, 0.0187, 0.0193, 0.0212, 0.0168, 0.0205, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 20:51:49,486 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.40 vs. limit=5.0 2023-05-02 20:51:51,506 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7758, 3.9915, 2.8185, 4.6009, 3.3500, 4.5280, 2.9054, 3.4037], device='cuda:4'), covar=tensor([0.0422, 0.0476, 0.1524, 0.0402, 0.0792, 0.0583, 0.1442, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0183, 0.0197, 0.0175, 0.0181, 0.0222, 0.0207, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:52:27,496 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2819, 3.2923, 3.4802, 2.3871, 3.1981, 3.5618, 3.2790, 2.1254], device='cuda:4'), covar=tensor([0.0555, 0.0153, 0.0077, 0.0450, 0.0144, 0.0128, 0.0115, 0.0516], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0091, 0.0092, 0.0137, 0.0104, 0.0117, 0.0100, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 20:52:29,857 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=295192.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:52:42,510 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.377e+02 2.068e+02 2.393e+02 2.794e+02 7.464e+02, threshold=4.786e+02, percent-clipped=2.0 2023-05-02 20:52:44,676 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9255, 2.6710, 2.8799, 2.1269, 2.7031, 2.0995, 2.7449, 2.8628], device='cuda:4'), covar=tensor([0.0374, 0.0954, 0.0573, 0.2009, 0.0926, 0.0982, 0.0620, 0.0900], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0170, 0.0171, 0.0158, 0.0148, 0.0133, 0.0145, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 20:52:48,745 INFO [train.py:904] (4/8) Epoch 30, batch 850, loss[loss=0.1608, simple_loss=0.2596, pruned_loss=0.03097, over 16718.00 frames. ], tot_loss[loss=0.1606, simple_loss=0.2464, pruned_loss=0.03736, over 3270959.83 frames. ], batch size: 57, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 20:53:18,704 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295225.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:53:21,299 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=295227.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:53:56,732 INFO [train.py:904] (4/8) Epoch 30, batch 900, loss[loss=0.1655, simple_loss=0.2624, pruned_loss=0.0343, over 16714.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2461, pruned_loss=0.03683, over 3278742.78 frames. ], batch size: 57, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 20:54:27,486 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=295275.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:54:30,571 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.4047, 5.4091, 5.2987, 4.8046, 4.9387, 5.3646, 5.2395, 4.9479], device='cuda:4'), covar=tensor([0.0650, 0.0654, 0.0353, 0.0351, 0.1119, 0.0508, 0.0309, 0.0836], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0483, 0.0373, 0.0376, 0.0370, 0.0432, 0.0258, 0.0448], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 20:54:34,087 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295279.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:54:41,807 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8513, 4.3671, 4.4467, 3.2074, 3.6263, 4.3814, 3.9354, 2.4958], device='cuda:4'), covar=tensor([0.0537, 0.0088, 0.0052, 0.0378, 0.0165, 0.0109, 0.0097, 0.0535], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0091, 0.0092, 0.0136, 0.0104, 0.0117, 0.0099, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 20:54:43,661 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295286.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:55:02,819 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 2.063e+02 2.465e+02 2.923e+02 4.814e+02, threshold=4.929e+02, percent-clipped=1.0 2023-05-02 20:55:08,082 INFO [train.py:904] (4/8) Epoch 30, batch 950, loss[loss=0.155, simple_loss=0.2465, pruned_loss=0.0318, over 17242.00 frames. ], tot_loss[loss=0.1598, simple_loss=0.246, pruned_loss=0.03679, over 3279669.74 frames. ], batch size: 45, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 20:55:48,097 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6139, 3.4821, 4.1165, 2.2604, 3.3862, 2.5628, 4.0735, 3.7947], device='cuda:4'), covar=tensor([0.0268, 0.1134, 0.0482, 0.2201, 0.0848, 0.1007, 0.0599, 0.1280], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0171, 0.0158, 0.0149, 0.0133, 0.0146, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 20:55:49,186 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295333.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:55:59,251 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295340.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:56:06,364 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295345.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 20:56:17,470 INFO [train.py:904] (4/8) Epoch 30, batch 1000, loss[loss=0.1516, simple_loss=0.2532, pruned_loss=0.02496, over 17271.00 frames. ], tot_loss[loss=0.159, simple_loss=0.2446, pruned_loss=0.03672, over 3278703.69 frames. ], batch size: 52, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 20:56:17,863 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7576, 3.9654, 2.1852, 4.4398, 3.0611, 4.3735, 2.3032, 3.1970], device='cuda:4'), covar=tensor([0.0381, 0.0426, 0.2080, 0.0399, 0.0941, 0.0511, 0.2033, 0.0828], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0183, 0.0197, 0.0176, 0.0181, 0.0223, 0.0207, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 20:56:24,317 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295359.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:56:39,488 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-02 20:56:53,382 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-05-02 20:57:12,896 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295394.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:57:20,046 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.691e+02 2.023e+02 2.384e+02 2.929e+02 5.067e+02, threshold=4.769e+02, percent-clipped=1.0 2023-05-02 20:57:26,447 INFO [train.py:904] (4/8) Epoch 30, batch 1050, loss[loss=0.1549, simple_loss=0.2556, pruned_loss=0.02711, over 17137.00 frames. ], tot_loss[loss=0.1587, simple_loss=0.2446, pruned_loss=0.03637, over 3286332.33 frames. ], batch size: 49, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 20:57:48,350 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295420.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:57:54,256 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295424.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 20:58:36,215 INFO [train.py:904] (4/8) Epoch 30, batch 1100, loss[loss=0.1472, simple_loss=0.2491, pruned_loss=0.02263, over 17260.00 frames. ], tot_loss[loss=0.1582, simple_loss=0.2443, pruned_loss=0.03602, over 3296953.18 frames. ], batch size: 52, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 20:59:32,027 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4663, 3.5718, 3.7808, 2.6715, 3.4277, 3.8533, 3.5624, 2.2173], device='cuda:4'), covar=tensor([0.0552, 0.0209, 0.0070, 0.0415, 0.0140, 0.0124, 0.0118, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0091, 0.0092, 0.0136, 0.0104, 0.0117, 0.0099, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 20:59:33,173 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5259, 4.5736, 4.8565, 4.8747, 4.9054, 4.6196, 4.5834, 4.4763], device='cuda:4'), covar=tensor([0.0424, 0.0783, 0.0535, 0.0452, 0.0560, 0.0568, 0.0947, 0.0683], device='cuda:4'), in_proj_covar=tensor([0.0446, 0.0509, 0.0488, 0.0450, 0.0535, 0.0517, 0.0592, 0.0415], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 20:59:38,542 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.377e+02 2.061e+02 2.413e+02 2.897e+02 1.352e+03, threshold=4.827e+02, percent-clipped=9.0 2023-05-02 20:59:43,304 INFO [train.py:904] (4/8) Epoch 30, batch 1150, loss[loss=0.1499, simple_loss=0.236, pruned_loss=0.03195, over 11983.00 frames. ], tot_loss[loss=0.1572, simple_loss=0.2437, pruned_loss=0.03536, over 3292840.67 frames. ], batch size: 248, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:00:30,551 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8989, 4.1281, 2.7187, 4.6826, 3.3893, 4.6708, 2.7226, 3.4230], device='cuda:4'), covar=tensor([0.0354, 0.0413, 0.1581, 0.0431, 0.0766, 0.0500, 0.1581, 0.0775], device='cuda:4'), in_proj_covar=tensor([0.0178, 0.0182, 0.0197, 0.0175, 0.0181, 0.0222, 0.0207, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 21:00:41,682 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.95 vs. limit=5.0 2023-05-02 21:00:52,295 INFO [train.py:904] (4/8) Epoch 30, batch 1200, loss[loss=0.1369, simple_loss=0.2174, pruned_loss=0.02815, over 16958.00 frames. ], tot_loss[loss=0.1557, simple_loss=0.2425, pruned_loss=0.03448, over 3306370.30 frames. ], batch size: 41, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:00:54,969 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295556.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:00:57,994 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295558.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 21:01:29,883 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295581.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:01:55,306 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.513e+02 2.025e+02 2.431e+02 2.981e+02 5.369e+02, threshold=4.863e+02, percent-clipped=2.0 2023-05-02 21:02:00,804 INFO [train.py:904] (4/8) Epoch 30, batch 1250, loss[loss=0.1576, simple_loss=0.2432, pruned_loss=0.03598, over 12213.00 frames. ], tot_loss[loss=0.1556, simple_loss=0.2419, pruned_loss=0.03464, over 3307621.05 frames. ], batch size: 246, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:02:19,199 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295617.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 21:02:21,400 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295619.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 21:02:34,278 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295629.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 21:02:39,652 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8563, 2.8967, 2.4843, 2.7615, 3.1896, 2.9586, 3.5072, 3.4563], device='cuda:4'), covar=tensor([0.0199, 0.0547, 0.0731, 0.0574, 0.0385, 0.0485, 0.0353, 0.0317], device='cuda:4'), in_proj_covar=tensor([0.0239, 0.0252, 0.0239, 0.0240, 0.0250, 0.0249, 0.0247, 0.0250], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:02:43,120 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8239, 5.0732, 5.2757, 5.0685, 5.1011, 5.7079, 5.2058, 4.8744], device='cuda:4'), covar=tensor([0.1550, 0.2263, 0.3025, 0.2175, 0.2679, 0.1192, 0.1933, 0.2783], device='cuda:4'), in_proj_covar=tensor([0.0435, 0.0653, 0.0727, 0.0528, 0.0704, 0.0745, 0.0558, 0.0702], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:02:43,130 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295635.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:02:46,363 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7057, 2.4574, 2.0052, 2.2385, 2.7962, 2.5262, 2.6984, 2.8587], device='cuda:4'), covar=tensor([0.0291, 0.0520, 0.0635, 0.0516, 0.0294, 0.0402, 0.0262, 0.0334], device='cuda:4'), in_proj_covar=tensor([0.0239, 0.0252, 0.0239, 0.0240, 0.0250, 0.0249, 0.0247, 0.0250], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:02:58,028 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=295645.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:03:08,457 INFO [train.py:904] (4/8) Epoch 30, batch 1300, loss[loss=0.155, simple_loss=0.2338, pruned_loss=0.03812, over 16204.00 frames. ], tot_loss[loss=0.1563, simple_loss=0.2425, pruned_loss=0.03507, over 3317506.88 frames. ], batch size: 165, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:03:52,670 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295685.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:03:57,720 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295689.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:03:59,041 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295690.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 21:04:02,884 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=295693.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:04:11,929 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.484e+02 2.150e+02 2.443e+02 2.917e+02 6.484e+02, threshold=4.887e+02, percent-clipped=2.0 2023-05-02 21:04:18,058 INFO [train.py:904] (4/8) Epoch 30, batch 1350, loss[loss=0.161, simple_loss=0.2362, pruned_loss=0.04287, over 11574.00 frames. ], tot_loss[loss=0.1561, simple_loss=0.2424, pruned_loss=0.03493, over 3305944.69 frames. ], batch size: 246, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:04:29,326 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5420, 3.5308, 4.0866, 2.2841, 3.3863, 2.5846, 3.9838, 3.7162], device='cuda:4'), covar=tensor([0.0235, 0.1018, 0.0493, 0.2172, 0.0798, 0.1010, 0.0563, 0.1220], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0172, 0.0172, 0.0158, 0.0149, 0.0133, 0.0147, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 21:04:33,097 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295715.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:04:46,299 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=295724.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:05:17,254 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295746.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:05:26,448 INFO [train.py:904] (4/8) Epoch 30, batch 1400, loss[loss=0.1657, simple_loss=0.2453, pruned_loss=0.04308, over 16885.00 frames. ], tot_loss[loss=0.1562, simple_loss=0.2424, pruned_loss=0.03505, over 3312591.20 frames. ], batch size: 96, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:05:52,092 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=295772.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:06:07,151 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3768, 4.2573, 4.3031, 4.0122, 4.1199, 4.3364, 4.0378, 4.1491], device='cuda:4'), covar=tensor([0.0665, 0.0955, 0.0352, 0.0310, 0.0664, 0.0523, 0.0799, 0.0659], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0490, 0.0380, 0.0381, 0.0376, 0.0438, 0.0261, 0.0454], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:06:30,027 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.296e+02 2.078e+02 2.399e+02 2.795e+02 5.908e+02, threshold=4.799e+02, percent-clipped=2.0 2023-05-02 21:06:35,127 INFO [train.py:904] (4/8) Epoch 30, batch 1450, loss[loss=0.1344, simple_loss=0.2236, pruned_loss=0.02261, over 17243.00 frames. ], tot_loss[loss=0.1558, simple_loss=0.2418, pruned_loss=0.0349, over 3318081.43 frames. ], batch size: 45, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:06:57,461 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295820.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:07:44,323 INFO [train.py:904] (4/8) Epoch 30, batch 1500, loss[loss=0.1305, simple_loss=0.2184, pruned_loss=0.02126, over 15881.00 frames. ], tot_loss[loss=0.1564, simple_loss=0.2418, pruned_loss=0.03552, over 3314534.63 frames. ], batch size: 35, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:08:11,940 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=295873.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:08:21,945 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=295881.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:08:22,062 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295881.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:08:47,908 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.572e+02 2.156e+02 2.462e+02 3.031e+02 5.873e+02, threshold=4.924e+02, percent-clipped=1.0 2023-05-02 21:08:54,105 INFO [train.py:904] (4/8) Epoch 30, batch 1550, loss[loss=0.1672, simple_loss=0.2434, pruned_loss=0.04552, over 16469.00 frames. ], tot_loss[loss=0.1577, simple_loss=0.2437, pruned_loss=0.0359, over 3307833.79 frames. ], batch size: 146, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:09:05,571 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295912.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 21:09:08,344 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295914.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 21:09:29,567 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=295929.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:09:36,270 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=295934.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 21:09:37,303 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=295935.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:10:02,957 INFO [train.py:904] (4/8) Epoch 30, batch 1600, loss[loss=0.156, simple_loss=0.2384, pruned_loss=0.03683, over 16716.00 frames. ], tot_loss[loss=0.1593, simple_loss=0.2456, pruned_loss=0.03653, over 3302225.65 frames. ], batch size: 76, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:10:09,896 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5736, 4.4896, 4.4930, 4.1506, 4.2596, 4.5153, 4.2935, 4.3158], device='cuda:4'), covar=tensor([0.0625, 0.0745, 0.0342, 0.0336, 0.0776, 0.0519, 0.0567, 0.0645], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0490, 0.0379, 0.0380, 0.0376, 0.0437, 0.0261, 0.0453], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:10:14,698 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.94 vs. limit=5.0 2023-05-02 21:10:42,975 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=295983.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:10:45,977 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=295985.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 21:10:51,153 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=295989.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:11:05,441 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.556e+02 2.204e+02 2.497e+02 2.999e+02 5.396e+02, threshold=4.994e+02, percent-clipped=1.0 2023-05-02 21:11:14,068 INFO [train.py:904] (4/8) Epoch 30, batch 1650, loss[loss=0.1724, simple_loss=0.2749, pruned_loss=0.03491, over 17045.00 frames. ], tot_loss[loss=0.1605, simple_loss=0.247, pruned_loss=0.03703, over 3307761.25 frames. ], batch size: 53, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:11:30,060 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296015.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:11:33,119 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2206, 2.3138, 2.8959, 3.1581, 3.1467, 3.6626, 2.6300, 3.7273], device='cuda:4'), covar=tensor([0.0300, 0.0596, 0.0390, 0.0377, 0.0369, 0.0225, 0.0594, 0.0167], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0202, 0.0190, 0.0196, 0.0214, 0.0171, 0.0208, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 21:12:00,954 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296037.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:12:06,274 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296041.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:12:06,492 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8555, 2.6464, 2.5233, 4.3539, 3.2458, 4.0653, 1.7388, 2.9465], device='cuda:4'), covar=tensor([0.1518, 0.0942, 0.1377, 0.0224, 0.0190, 0.0454, 0.1804, 0.0923], device='cuda:4'), in_proj_covar=tensor([0.0176, 0.0184, 0.0203, 0.0209, 0.0208, 0.0221, 0.0214, 0.0202], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 21:12:16,872 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296049.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:12:24,440 INFO [train.py:904] (4/8) Epoch 30, batch 1700, loss[loss=0.1668, simple_loss=0.2488, pruned_loss=0.04238, over 16921.00 frames. ], tot_loss[loss=0.162, simple_loss=0.2494, pruned_loss=0.03734, over 3315429.17 frames. ], batch size: 90, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:12:36,067 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296063.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:12:44,496 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6962, 3.3287, 3.6253, 2.0569, 3.7187, 3.7621, 3.1438, 2.8574], device='cuda:4'), covar=tensor([0.0725, 0.0280, 0.0211, 0.1164, 0.0140, 0.0225, 0.0423, 0.0487], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0113, 0.0104, 0.0141, 0.0089, 0.0136, 0.0133, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 21:13:14,668 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0225, 3.1029, 2.7476, 2.9254, 3.3659, 2.9874, 3.6648, 3.5273], device='cuda:4'), covar=tensor([0.0162, 0.0419, 0.0523, 0.0457, 0.0273, 0.0457, 0.0218, 0.0288], device='cuda:4'), in_proj_covar=tensor([0.0241, 0.0253, 0.0240, 0.0242, 0.0252, 0.0250, 0.0249, 0.0252], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:13:26,992 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.214e+02 2.660e+02 3.287e+02 5.248e+02, threshold=5.320e+02, percent-clipped=2.0 2023-05-02 21:13:32,339 INFO [train.py:904] (4/8) Epoch 30, batch 1750, loss[loss=0.148, simple_loss=0.2348, pruned_loss=0.03059, over 16553.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2496, pruned_loss=0.03735, over 3316010.15 frames. ], batch size: 75, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:13:38,688 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6281, 3.6678, 2.7879, 2.2236, 2.3581, 2.4012, 3.7793, 3.1629], device='cuda:4'), covar=tensor([0.2855, 0.0593, 0.1867, 0.3151, 0.2996, 0.2250, 0.0544, 0.1611], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0277, 0.0316, 0.0330, 0.0308, 0.0282, 0.0306, 0.0356], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:13:40,751 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296110.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:14:23,003 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0322, 4.1697, 3.9780, 3.7544, 3.5719, 4.1881, 3.8672, 3.9056], device='cuda:4'), covar=tensor([0.1001, 0.1001, 0.0505, 0.0474, 0.1234, 0.0585, 0.1097, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0491, 0.0380, 0.0381, 0.0376, 0.0438, 0.0261, 0.0455], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:14:41,095 INFO [train.py:904] (4/8) Epoch 30, batch 1800, loss[loss=0.164, simple_loss=0.2625, pruned_loss=0.03274, over 17124.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2506, pruned_loss=0.03735, over 3315178.20 frames. ], batch size: 48, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:15:12,949 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296176.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:15:47,286 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 2.165e+02 2.480e+02 2.953e+02 5.408e+02, threshold=4.959e+02, percent-clipped=1.0 2023-05-02 21:15:51,605 INFO [train.py:904] (4/8) Epoch 30, batch 1850, loss[loss=0.1415, simple_loss=0.2365, pruned_loss=0.02329, over 17211.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2512, pruned_loss=0.03704, over 3317574.54 frames. ], batch size: 44, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:16:03,775 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296212.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 21:16:06,110 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296214.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 21:16:27,871 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296229.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 21:17:02,018 INFO [train.py:904] (4/8) Epoch 30, batch 1900, loss[loss=0.1779, simple_loss=0.2751, pruned_loss=0.04039, over 16735.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2504, pruned_loss=0.03694, over 3315395.40 frames. ], batch size: 57, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:17:09,976 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296260.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:17:13,526 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296262.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 21:17:45,578 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296285.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 21:17:50,106 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-05-02 21:18:06,423 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.514e+02 2.062e+02 2.316e+02 2.814e+02 5.874e+02, threshold=4.633e+02, percent-clipped=2.0 2023-05-02 21:18:10,597 INFO [train.py:904] (4/8) Epoch 30, batch 1950, loss[loss=0.1559, simple_loss=0.2606, pruned_loss=0.02555, over 17142.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2504, pruned_loss=0.03688, over 3314224.84 frames. ], batch size: 48, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:18:26,919 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296315.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:18:52,566 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296333.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 21:19:02,347 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296341.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:19:21,053 INFO [train.py:904] (4/8) Epoch 30, batch 2000, loss[loss=0.1451, simple_loss=0.2264, pruned_loss=0.0319, over 16809.00 frames. ], tot_loss[loss=0.1616, simple_loss=0.2498, pruned_loss=0.03673, over 3301228.60 frames. ], batch size: 102, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:19:32,070 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0933, 5.0160, 4.9572, 4.5040, 4.6594, 4.9914, 4.9006, 4.6640], device='cuda:4'), covar=tensor([0.0682, 0.0801, 0.0608, 0.0447, 0.1294, 0.0615, 0.0462, 0.0901], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0496, 0.0383, 0.0386, 0.0380, 0.0443, 0.0264, 0.0459], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:19:51,024 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296376.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:19:56,771 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3831, 4.4163, 4.7286, 4.7184, 4.7769, 4.4873, 4.4472, 4.3321], device='cuda:4'), covar=tensor([0.0468, 0.0833, 0.0602, 0.0588, 0.0592, 0.0527, 0.1033, 0.0813], device='cuda:4'), in_proj_covar=tensor([0.0450, 0.0512, 0.0489, 0.0453, 0.0537, 0.0519, 0.0596, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 21:20:09,533 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296389.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:20:25,431 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.330e+02 2.133e+02 2.447e+02 2.850e+02 4.858e+02, threshold=4.894e+02, percent-clipped=1.0 2023-05-02 21:20:30,258 INFO [train.py:904] (4/8) Epoch 30, batch 2050, loss[loss=0.1562, simple_loss=0.2548, pruned_loss=0.02885, over 17120.00 frames. ], tot_loss[loss=0.162, simple_loss=0.2501, pruned_loss=0.03696, over 3306532.15 frames. ], batch size: 48, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:20:32,264 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296405.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:20:34,243 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9057, 4.2004, 4.0456, 4.0847, 3.7321, 3.7995, 3.8547, 4.1996], device='cuda:4'), covar=tensor([0.1293, 0.1014, 0.1071, 0.0799, 0.0840, 0.1783, 0.1008, 0.1012], device='cuda:4'), in_proj_covar=tensor([0.0743, 0.0892, 0.0736, 0.0697, 0.0572, 0.0565, 0.0751, 0.0703], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:21:12,430 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296434.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:21:18,948 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2893, 3.5255, 3.8516, 2.2518, 3.1246, 2.4097, 3.6587, 3.7415], device='cuda:4'), covar=tensor([0.0322, 0.0977, 0.0540, 0.2121, 0.0873, 0.1051, 0.0649, 0.1084], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0173, 0.0172, 0.0159, 0.0149, 0.0134, 0.0147, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 21:21:39,722 INFO [train.py:904] (4/8) Epoch 30, batch 2100, loss[loss=0.1588, simple_loss=0.2556, pruned_loss=0.03099, over 17145.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2517, pruned_loss=0.03758, over 3303340.16 frames. ], batch size: 49, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:22:09,636 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296476.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:22:23,081 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7194, 3.9592, 2.7952, 2.2561, 2.6508, 2.3369, 4.0105, 3.4544], device='cuda:4'), covar=tensor([0.2945, 0.0519, 0.2007, 0.3196, 0.2647, 0.2222, 0.0586, 0.1485], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0280, 0.0318, 0.0333, 0.0310, 0.0284, 0.0309, 0.0359], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:22:33,120 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296492.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:22:36,781 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296495.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:22:39,004 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296497.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:22:46,671 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.431e+02 2.088e+02 2.487e+02 2.838e+02 5.578e+02, threshold=4.974e+02, percent-clipped=1.0 2023-05-02 21:22:48,891 INFO [train.py:904] (4/8) Epoch 30, batch 2150, loss[loss=0.1586, simple_loss=0.2448, pruned_loss=0.03621, over 15933.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2524, pruned_loss=0.03833, over 3301743.12 frames. ], batch size: 35, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:23:03,758 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296514.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:23:16,223 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296524.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:23:23,346 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296529.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 21:23:56,772 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296553.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 21:23:57,533 INFO [train.py:904] (4/8) Epoch 30, batch 2200, loss[loss=0.1828, simple_loss=0.276, pruned_loss=0.04479, over 17073.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2526, pruned_loss=0.03859, over 3292675.36 frames. ], batch size: 53, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:24:04,153 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296558.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:24:27,036 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296575.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:24:30,753 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296577.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:24:32,435 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 21:25:04,059 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.512e+02 2.151e+02 2.512e+02 3.134e+02 5.584e+02, threshold=5.024e+02, percent-clipped=2.0 2023-05-02 21:25:06,331 INFO [train.py:904] (4/8) Epoch 30, batch 2250, loss[loss=0.1806, simple_loss=0.2599, pruned_loss=0.05069, over 16900.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2532, pruned_loss=0.03857, over 3302158.61 frames. ], batch size: 109, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:25:20,115 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296613.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:25:20,151 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2902, 2.6345, 2.1536, 2.5053, 2.9512, 2.7120, 3.0174, 3.1230], device='cuda:4'), covar=tensor([0.0253, 0.0506, 0.0665, 0.0556, 0.0346, 0.0422, 0.0315, 0.0337], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0254, 0.0241, 0.0244, 0.0254, 0.0252, 0.0251, 0.0254], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:26:17,730 INFO [train.py:904] (4/8) Epoch 30, batch 2300, loss[loss=0.147, simple_loss=0.2373, pruned_loss=0.02841, over 16696.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2528, pruned_loss=0.03836, over 3312927.75 frames. ], batch size: 89, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:26:41,378 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296671.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:26:45,109 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296674.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:27:05,615 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296688.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:27:18,262 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9530, 5.2831, 5.4964, 5.1827, 5.2895, 5.8922, 5.3831, 5.0973], device='cuda:4'), covar=tensor([0.1322, 0.2143, 0.2569, 0.2170, 0.2683, 0.1039, 0.1738, 0.2357], device='cuda:4'), in_proj_covar=tensor([0.0437, 0.0652, 0.0728, 0.0530, 0.0708, 0.0746, 0.0559, 0.0704], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:27:24,077 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.359e+02 2.179e+02 2.509e+02 3.030e+02 4.955e+02, threshold=5.019e+02, percent-clipped=0.0 2023-05-02 21:27:27,333 INFO [train.py:904] (4/8) Epoch 30, batch 2350, loss[loss=0.1585, simple_loss=0.2467, pruned_loss=0.03521, over 17068.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2534, pruned_loss=0.03884, over 3310966.87 frames. ], batch size: 53, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:27:28,738 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296705.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:27:53,416 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1013, 2.2984, 2.4546, 3.8179, 2.2448, 2.5987, 2.3207, 2.4426], device='cuda:4'), covar=tensor([0.1676, 0.3852, 0.3112, 0.0719, 0.4069, 0.2624, 0.4118, 0.3148], device='cuda:4'), in_proj_covar=tensor([0.0431, 0.0484, 0.0395, 0.0346, 0.0452, 0.0555, 0.0458, 0.0568], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:28:30,579 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296749.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:28:36,423 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=296753.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:28:37,278 INFO [train.py:904] (4/8) Epoch 30, batch 2400, loss[loss=0.1473, simple_loss=0.2371, pruned_loss=0.02879, over 17014.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.254, pruned_loss=0.03841, over 3317273.58 frames. ], batch size: 41, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:29:20,242 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 21:29:27,116 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296790.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:29:43,969 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.575e+02 2.201e+02 2.707e+02 3.099e+02 6.549e+02, threshold=5.413e+02, percent-clipped=3.0 2023-05-02 21:29:46,349 INFO [train.py:904] (4/8) Epoch 30, batch 2450, loss[loss=0.1629, simple_loss=0.2644, pruned_loss=0.03074, over 17084.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2543, pruned_loss=0.03849, over 3319758.28 frames. ], batch size: 50, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:30:20,356 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9909, 2.1720, 2.7132, 3.0402, 2.8029, 3.5374, 2.4615, 3.5434], device='cuda:4'), covar=tensor([0.0327, 0.0666, 0.0424, 0.0404, 0.0456, 0.0217, 0.0637, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0206, 0.0205, 0.0194, 0.0199, 0.0218, 0.0174, 0.0210, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 21:30:48,316 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296848.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 21:30:55,421 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296853.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:30:56,382 INFO [train.py:904] (4/8) Epoch 30, batch 2500, loss[loss=0.2103, simple_loss=0.2929, pruned_loss=0.06383, over 12329.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2543, pruned_loss=0.03876, over 3320339.62 frames. ], batch size: 246, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:31:19,657 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296870.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:31:23,970 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296873.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:31:41,013 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.3457, 5.3337, 5.0301, 4.5299, 5.1711, 1.9381, 4.8514, 4.8440], device='cuda:4'), covar=tensor([0.0089, 0.0079, 0.0238, 0.0394, 0.0101, 0.3040, 0.0160, 0.0257], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0178, 0.0217, 0.0187, 0.0195, 0.0222, 0.0206, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:32:04,914 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.483e+02 2.203e+02 2.647e+02 3.147e+02 6.132e+02, threshold=5.294e+02, percent-clipped=2.0 2023-05-02 21:32:07,797 INFO [train.py:904] (4/8) Epoch 30, batch 2550, loss[loss=0.158, simple_loss=0.2517, pruned_loss=0.03211, over 17109.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2547, pruned_loss=0.03875, over 3318794.25 frames. ], batch size: 47, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:32:22,200 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9908, 4.7859, 5.0553, 5.2271, 5.4101, 4.7744, 5.4151, 5.4467], device='cuda:4'), covar=tensor([0.2200, 0.1433, 0.1842, 0.0847, 0.0607, 0.1016, 0.0606, 0.0651], device='cuda:4'), in_proj_covar=tensor([0.0718, 0.0870, 0.1006, 0.0890, 0.0674, 0.0701, 0.0739, 0.0861], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:32:49,084 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296934.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 21:32:54,104 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296938.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:32:58,379 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3968, 2.4303, 2.3542, 4.2076, 2.3546, 2.8037, 2.4900, 2.6317], device='cuda:4'), covar=tensor([0.1455, 0.3854, 0.3584, 0.0622, 0.4419, 0.2688, 0.4003, 0.3728], device='cuda:4'), in_proj_covar=tensor([0.0432, 0.0485, 0.0396, 0.0347, 0.0453, 0.0557, 0.0458, 0.0569], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:33:01,403 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0981, 3.1530, 3.2245, 2.2143, 3.0678, 3.3518, 3.0935, 1.7605], device='cuda:4'), covar=tensor([0.0599, 0.0155, 0.0116, 0.0497, 0.0178, 0.0155, 0.0157, 0.0707], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0091, 0.0093, 0.0136, 0.0104, 0.0117, 0.0100, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 21:33:15,497 INFO [train.py:904] (4/8) Epoch 30, batch 2600, loss[loss=0.166, simple_loss=0.2659, pruned_loss=0.03302, over 17062.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2547, pruned_loss=0.03865, over 3324126.33 frames. ], batch size: 53, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:33:36,722 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=296969.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:33:39,729 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=296971.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:34:18,428 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=296999.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:34:22,260 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.341e+02 2.007e+02 2.417e+02 2.925e+02 5.072e+02, threshold=4.833e+02, percent-clipped=0.0 2023-05-02 21:34:24,414 INFO [train.py:904] (4/8) Epoch 30, batch 2650, loss[loss=0.1683, simple_loss=0.2495, pruned_loss=0.04353, over 16210.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2549, pruned_loss=0.03784, over 3333737.01 frames. ], batch size: 165, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:34:36,854 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1180, 5.6251, 5.7565, 5.4304, 5.4937, 6.0951, 5.5084, 5.1857], device='cuda:4'), covar=tensor([0.1001, 0.1898, 0.2622, 0.2005, 0.2654, 0.0961, 0.1647, 0.2547], device='cuda:4'), in_proj_covar=tensor([0.0435, 0.0647, 0.0722, 0.0526, 0.0704, 0.0743, 0.0556, 0.0701], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:34:38,422 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-05-02 21:34:46,004 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297019.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:35:20,548 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=297044.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:35:34,415 INFO [train.py:904] (4/8) Epoch 30, batch 2700, loss[loss=0.1793, simple_loss=0.2707, pruned_loss=0.04394, over 16549.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2551, pruned_loss=0.03702, over 3342359.24 frames. ], batch size: 68, lr: 2.25e-03, grad_scale: 8.0 2023-05-02 21:35:35,397 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4185, 4.4437, 4.7391, 4.7266, 4.7801, 4.5047, 4.4910, 4.3856], device='cuda:4'), covar=tensor([0.0418, 0.0673, 0.0463, 0.0451, 0.0563, 0.0472, 0.0909, 0.0653], device='cuda:4'), in_proj_covar=tensor([0.0450, 0.0512, 0.0489, 0.0452, 0.0533, 0.0517, 0.0595, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 21:35:58,067 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-05-02 21:36:23,908 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297090.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:36:42,009 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.416e+02 2.120e+02 2.454e+02 2.893e+02 5.534e+02, threshold=4.909e+02, percent-clipped=1.0 2023-05-02 21:36:43,153 INFO [train.py:904] (4/8) Epoch 30, batch 2750, loss[loss=0.1803, simple_loss=0.2592, pruned_loss=0.05077, over 16854.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2554, pruned_loss=0.037, over 3340151.53 frames. ], batch size: 116, lr: 2.25e-03, grad_scale: 4.0 2023-05-02 21:37:30,682 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297138.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:37:44,368 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297148.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:37:50,759 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297153.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:37:51,587 INFO [train.py:904] (4/8) Epoch 30, batch 2800, loss[loss=0.1673, simple_loss=0.2685, pruned_loss=0.03305, over 17015.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2555, pruned_loss=0.03713, over 3343735.89 frames. ], batch size: 50, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:38:14,244 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297170.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:38:26,518 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-05-02 21:38:49,782 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297196.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:38:57,321 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297201.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:39:00,128 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.644e+02 2.070e+02 2.512e+02 3.097e+02 5.652e+02, threshold=5.024e+02, percent-clipped=1.0 2023-05-02 21:39:01,278 INFO [train.py:904] (4/8) Epoch 30, batch 2850, loss[loss=0.1491, simple_loss=0.2392, pruned_loss=0.02951, over 16848.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2537, pruned_loss=0.03686, over 3342851.50 frames. ], batch size: 42, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:39:15,373 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6786, 3.8623, 4.0558, 2.7366, 3.6398, 4.1136, 3.7394, 2.5448], device='cuda:4'), covar=tensor([0.0518, 0.0347, 0.0071, 0.0437, 0.0140, 0.0120, 0.0110, 0.0486], device='cuda:4'), in_proj_covar=tensor([0.0139, 0.0091, 0.0092, 0.0135, 0.0104, 0.0116, 0.0099, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 21:39:21,721 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297218.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:39:26,941 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=297222.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:39:37,163 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=297229.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 21:40:10,536 INFO [train.py:904] (4/8) Epoch 30, batch 2900, loss[loss=0.164, simple_loss=0.2417, pruned_loss=0.04319, over 16391.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2523, pruned_loss=0.03661, over 3344592.06 frames. ], batch size: 75, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:40:33,095 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297269.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:40:51,673 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=297283.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:41:08,380 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=297294.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:41:09,517 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1796, 5.6971, 5.8397, 5.4594, 5.6128, 6.1970, 5.5993, 5.2783], device='cuda:4'), covar=tensor([0.0967, 0.2179, 0.2809, 0.2291, 0.2632, 0.0973, 0.1611, 0.2571], device='cuda:4'), in_proj_covar=tensor([0.0440, 0.0657, 0.0732, 0.0532, 0.0712, 0.0750, 0.0563, 0.0710], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:41:22,344 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.582e+02 2.045e+02 2.480e+02 2.886e+02 5.342e+02, threshold=4.960e+02, percent-clipped=3.0 2023-05-02 21:41:22,359 INFO [train.py:904] (4/8) Epoch 30, batch 2950, loss[loss=0.1664, simple_loss=0.2493, pruned_loss=0.0418, over 16806.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2515, pruned_loss=0.03744, over 3342835.00 frames. ], batch size: 102, lr: 2.24e-03, grad_scale: 4.0 2023-05-02 21:41:41,509 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297317.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:42:18,965 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297344.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:42:33,297 INFO [train.py:904] (4/8) Epoch 30, batch 3000, loss[loss=0.2575, simple_loss=0.3217, pruned_loss=0.09664, over 11862.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2524, pruned_loss=0.03832, over 3334663.47 frames. ], batch size: 246, lr: 2.24e-03, grad_scale: 4.0 2023-05-02 21:42:33,297 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 21:42:42,092 INFO [train.py:938] (4/8) Epoch 30, validation: loss=0.1331, simple_loss=0.238, pruned_loss=0.01415, over 944034.00 frames. 2023-05-02 21:42:42,093 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 21:43:35,185 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297392.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:43:45,078 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9879, 2.2845, 2.3958, 2.9184, 2.1180, 3.2095, 1.8190, 2.7099], device='cuda:4'), covar=tensor([0.1170, 0.0711, 0.1141, 0.0207, 0.0141, 0.0352, 0.1566, 0.0769], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0185, 0.0204, 0.0210, 0.0209, 0.0222, 0.0214, 0.0202], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 21:43:51,743 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.593e+02 2.205e+02 2.625e+02 3.080e+02 5.870e+02, threshold=5.251e+02, percent-clipped=1.0 2023-05-02 21:43:51,758 INFO [train.py:904] (4/8) Epoch 30, batch 3050, loss[loss=0.1625, simple_loss=0.2401, pruned_loss=0.04246, over 16440.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2516, pruned_loss=0.03799, over 3337129.69 frames. ], batch size: 146, lr: 2.24e-03, grad_scale: 4.0 2023-05-02 21:45:01,424 INFO [train.py:904] (4/8) Epoch 30, batch 3100, loss[loss=0.1782, simple_loss=0.2564, pruned_loss=0.04999, over 15635.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2513, pruned_loss=0.03805, over 3334517.01 frames. ], batch size: 190, lr: 2.24e-03, grad_scale: 4.0 2023-05-02 21:45:55,817 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4138, 4.3778, 4.5689, 4.3227, 4.4142, 5.0109, 4.5170, 4.1570], device='cuda:4'), covar=tensor([0.1858, 0.2440, 0.2472, 0.2317, 0.2900, 0.1244, 0.1725, 0.2828], device='cuda:4'), in_proj_covar=tensor([0.0442, 0.0659, 0.0734, 0.0533, 0.0714, 0.0751, 0.0563, 0.0712], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:46:07,374 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.515e+02 2.135e+02 2.482e+02 3.007e+02 9.801e+02, threshold=4.963e+02, percent-clipped=2.0 2023-05-02 21:46:07,389 INFO [train.py:904] (4/8) Epoch 30, batch 3150, loss[loss=0.1457, simple_loss=0.2273, pruned_loss=0.03211, over 16993.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.251, pruned_loss=0.0376, over 3341277.95 frames. ], batch size: 41, lr: 2.24e-03, grad_scale: 4.0 2023-05-02 21:46:42,713 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297529.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 21:47:17,202 INFO [train.py:904] (4/8) Epoch 30, batch 3200, loss[loss=0.1686, simple_loss=0.2485, pruned_loss=0.04433, over 16835.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2512, pruned_loss=0.03786, over 3329114.52 frames. ], batch size: 116, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:47:38,039 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-05-02 21:47:49,841 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297577.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:47:51,694 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=297578.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:47:51,788 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9730, 4.9530, 4.8862, 4.4892, 4.5942, 4.9252, 4.8176, 4.6290], device='cuda:4'), covar=tensor([0.0728, 0.0738, 0.0354, 0.0367, 0.1006, 0.0525, 0.0511, 0.0826], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0501, 0.0387, 0.0390, 0.0383, 0.0448, 0.0267, 0.0465], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:48:13,651 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297594.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:48:27,238 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.469e+02 2.037e+02 2.359e+02 2.785e+02 6.437e+02, threshold=4.718e+02, percent-clipped=1.0 2023-05-02 21:48:27,253 INFO [train.py:904] (4/8) Epoch 30, batch 3250, loss[loss=0.1921, simple_loss=0.261, pruned_loss=0.06167, over 16891.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2513, pruned_loss=0.03781, over 3323594.98 frames. ], batch size: 109, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:49:20,230 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297642.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:49:36,042 INFO [train.py:904] (4/8) Epoch 30, batch 3300, loss[loss=0.1716, simple_loss=0.2683, pruned_loss=0.03744, over 17057.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2526, pruned_loss=0.03792, over 3328036.31 frames. ], batch size: 50, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:50:46,297 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.603e+02 2.242e+02 2.585e+02 3.281e+02 4.783e+02, threshold=5.170e+02, percent-clipped=1.0 2023-05-02 21:50:46,312 INFO [train.py:904] (4/8) Epoch 30, batch 3350, loss[loss=0.1329, simple_loss=0.2207, pruned_loss=0.02251, over 16768.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2525, pruned_loss=0.03805, over 3318295.84 frames. ], batch size: 39, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:50:52,737 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 21:50:58,018 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5329, 2.5141, 2.4680, 4.3394, 2.4562, 2.8766, 2.5785, 2.6949], device='cuda:4'), covar=tensor([0.1415, 0.3827, 0.3392, 0.0577, 0.4205, 0.2762, 0.3848, 0.3520], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0486, 0.0396, 0.0347, 0.0454, 0.0558, 0.0459, 0.0570], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:51:56,062 INFO [train.py:904] (4/8) Epoch 30, batch 3400, loss[loss=0.2084, simple_loss=0.2975, pruned_loss=0.0597, over 12008.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2527, pruned_loss=0.03791, over 3311152.17 frames. ], batch size: 247, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:52:57,874 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5361, 3.0811, 3.4604, 2.0201, 3.5592, 3.5584, 3.0251, 2.8157], device='cuda:4'), covar=tensor([0.0751, 0.0306, 0.0226, 0.1105, 0.0134, 0.0238, 0.0438, 0.0455], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0104, 0.0139, 0.0089, 0.0135, 0.0132, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 21:53:07,131 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.363e+02 1.981e+02 2.351e+02 2.739e+02 4.469e+02, threshold=4.702e+02, percent-clipped=0.0 2023-05-02 21:53:07,146 INFO [train.py:904] (4/8) Epoch 30, batch 3450, loss[loss=0.1671, simple_loss=0.262, pruned_loss=0.03614, over 17114.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2514, pruned_loss=0.03763, over 3309564.39 frames. ], batch size: 55, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:53:23,364 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2056, 2.2817, 2.4107, 3.8678, 2.3366, 2.6256, 2.3681, 2.4343], device='cuda:4'), covar=tensor([0.1654, 0.3897, 0.3353, 0.0818, 0.4119, 0.2698, 0.4274, 0.3224], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0487, 0.0396, 0.0348, 0.0454, 0.0559, 0.0459, 0.0571], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 21:54:09,063 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9597, 2.9057, 2.6711, 4.8162, 3.7845, 4.2534, 1.6675, 3.1302], device='cuda:4'), covar=tensor([0.1361, 0.0817, 0.1297, 0.0261, 0.0254, 0.0451, 0.1779, 0.0851], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0185, 0.0204, 0.0211, 0.0210, 0.0222, 0.0214, 0.0202], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 21:54:17,197 INFO [train.py:904] (4/8) Epoch 30, batch 3500, loss[loss=0.186, simple_loss=0.2726, pruned_loss=0.0497, over 16448.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2508, pruned_loss=0.03743, over 3300101.46 frames. ], batch size: 146, lr: 2.24e-03, grad_scale: 4.0 2023-05-02 21:54:27,834 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-05-02 21:54:36,644 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2248, 3.3060, 2.0949, 3.4854, 2.7009, 3.5054, 2.2418, 2.7206], device='cuda:4'), covar=tensor([0.0345, 0.0444, 0.1672, 0.0365, 0.0783, 0.0701, 0.1525, 0.0783], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0186, 0.0199, 0.0180, 0.0183, 0.0226, 0.0208, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 21:54:51,994 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=297878.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:55:09,836 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8891, 1.4392, 1.7698, 1.6947, 1.8429, 1.9826, 1.6559, 1.8779], device='cuda:4'), covar=tensor([0.0288, 0.0493, 0.0292, 0.0362, 0.0334, 0.0247, 0.0528, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0206, 0.0204, 0.0194, 0.0199, 0.0218, 0.0174, 0.0210, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 21:55:28,082 INFO [train.py:904] (4/8) Epoch 30, batch 3550, loss[loss=0.1483, simple_loss=0.2448, pruned_loss=0.02588, over 17113.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.251, pruned_loss=0.03769, over 3301657.52 frames. ], batch size: 49, lr: 2.24e-03, grad_scale: 4.0 2023-05-02 21:55:29,302 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.419e+02 1.998e+02 2.319e+02 2.785e+02 6.034e+02, threshold=4.637e+02, percent-clipped=1.0 2023-05-02 21:55:59,067 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=297926.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:56:38,331 INFO [train.py:904] (4/8) Epoch 30, batch 3600, loss[loss=0.1646, simple_loss=0.2607, pruned_loss=0.03422, over 16770.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2505, pruned_loss=0.03779, over 3294222.86 frames. ], batch size: 57, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:56:42,396 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2762, 4.2890, 4.5895, 4.5764, 4.6171, 4.3458, 4.3430, 4.2613], device='cuda:4'), covar=tensor([0.0394, 0.0736, 0.0444, 0.0422, 0.0533, 0.0468, 0.0821, 0.0697], device='cuda:4'), in_proj_covar=tensor([0.0458, 0.0523, 0.0498, 0.0459, 0.0544, 0.0525, 0.0608, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 21:56:42,437 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=297957.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:56:54,684 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7612, 1.8362, 1.7036, 1.5460, 1.9678, 1.6047, 1.6902, 1.9296], device='cuda:4'), covar=tensor([0.0245, 0.0327, 0.0431, 0.0385, 0.0223, 0.0319, 0.0183, 0.0227], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0252, 0.0240, 0.0242, 0.0252, 0.0250, 0.0251, 0.0253], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 21:57:53,330 INFO [train.py:904] (4/8) Epoch 30, batch 3650, loss[loss=0.1759, simple_loss=0.2554, pruned_loss=0.04824, over 11748.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2491, pruned_loss=0.03845, over 3278978.84 frames. ], batch size: 249, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:57:55,116 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.721e+02 2.298e+02 2.680e+02 3.365e+02 6.432e+02, threshold=5.360e+02, percent-clipped=4.0 2023-05-02 21:58:07,009 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.0032, 2.1266, 2.6231, 2.8661, 2.8589, 2.9510, 2.2188, 3.1579], device='cuda:4'), covar=tensor([0.0217, 0.0558, 0.0367, 0.0300, 0.0363, 0.0372, 0.0623, 0.0212], device='cuda:4'), in_proj_covar=tensor([0.0206, 0.0205, 0.0194, 0.0199, 0.0219, 0.0174, 0.0210, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 21:58:14,254 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=298018.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 21:59:06,223 INFO [train.py:904] (4/8) Epoch 30, batch 3700, loss[loss=0.1857, simple_loss=0.261, pruned_loss=0.0552, over 11389.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2473, pruned_loss=0.03964, over 3267609.37 frames. ], batch size: 248, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 21:59:24,682 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.79 vs. limit=5.0 2023-05-02 22:00:17,693 INFO [train.py:904] (4/8) Epoch 30, batch 3750, loss[loss=0.1721, simple_loss=0.2451, pruned_loss=0.04954, over 16922.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2476, pruned_loss=0.04087, over 3275446.35 frames. ], batch size: 109, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:00:19,704 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.710e+02 2.048e+02 2.384e+02 2.845e+02 4.773e+02, threshold=4.768e+02, percent-clipped=0.0 2023-05-02 22:00:26,414 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3292, 3.4760, 3.6783, 2.3385, 3.2531, 2.5088, 3.7949, 3.9306], device='cuda:4'), covar=tensor([0.0248, 0.0848, 0.0595, 0.2066, 0.0809, 0.0994, 0.0507, 0.0822], device='cuda:4'), in_proj_covar=tensor([0.0164, 0.0174, 0.0172, 0.0159, 0.0150, 0.0134, 0.0147, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 22:01:30,455 INFO [train.py:904] (4/8) Epoch 30, batch 3800, loss[loss=0.16, simple_loss=0.2451, pruned_loss=0.03747, over 16295.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2497, pruned_loss=0.04178, over 3271305.99 frames. ], batch size: 165, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:02:22,052 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5569, 3.1722, 3.6737, 1.9555, 3.7158, 3.7177, 3.1418, 2.8036], device='cuda:4'), covar=tensor([0.0805, 0.0334, 0.0206, 0.1206, 0.0137, 0.0238, 0.0418, 0.0486], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0113, 0.0105, 0.0141, 0.0089, 0.0136, 0.0133, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:02:43,875 INFO [train.py:904] (4/8) Epoch 30, batch 3850, loss[loss=0.1676, simple_loss=0.2383, pruned_loss=0.04846, over 16927.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2494, pruned_loss=0.04258, over 3282084.28 frames. ], batch size: 109, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:02:44,948 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.556e+02 2.341e+02 2.538e+02 2.759e+02 9.437e+02, threshold=5.075e+02, percent-clipped=1.0 2023-05-02 22:03:53,264 INFO [train.py:904] (4/8) Epoch 30, batch 3900, loss[loss=0.162, simple_loss=0.2546, pruned_loss=0.03471, over 17118.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2489, pruned_loss=0.04269, over 3275727.19 frames. ], batch size: 49, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:03:54,929 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7063, 3.8608, 2.9585, 2.2920, 2.4045, 2.4927, 3.9757, 3.3141], device='cuda:4'), covar=tensor([0.2933, 0.0575, 0.1841, 0.3351, 0.3076, 0.2193, 0.0564, 0.1603], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0279, 0.0317, 0.0333, 0.0312, 0.0283, 0.0309, 0.0359], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:04:59,547 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6191, 4.5434, 4.5326, 3.8584, 4.5466, 1.7259, 4.2582, 4.0518], device='cuda:4'), covar=tensor([0.0203, 0.0211, 0.0227, 0.0428, 0.0150, 0.3139, 0.0210, 0.0319], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0180, 0.0219, 0.0190, 0.0197, 0.0223, 0.0209, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 22:05:04,297 INFO [train.py:904] (4/8) Epoch 30, batch 3950, loss[loss=0.1561, simple_loss=0.231, pruned_loss=0.04062, over 16831.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2484, pruned_loss=0.04331, over 3280637.78 frames. ], batch size: 102, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:05:05,533 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.526e+02 2.189e+02 2.590e+02 3.423e+02 7.666e+02, threshold=5.180e+02, percent-clipped=4.0 2023-05-02 22:05:16,583 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-05-02 22:05:17,508 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=298313.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:06:02,132 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3638, 4.4075, 4.6767, 4.6483, 4.6908, 4.4187, 4.4293, 4.3200], device='cuda:4'), covar=tensor([0.0412, 0.0724, 0.0462, 0.0447, 0.0570, 0.0471, 0.0832, 0.0683], device='cuda:4'), in_proj_covar=tensor([0.0456, 0.0520, 0.0495, 0.0458, 0.0541, 0.0523, 0.0603, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 22:06:15,411 INFO [train.py:904] (4/8) Epoch 30, batch 4000, loss[loss=0.1628, simple_loss=0.2511, pruned_loss=0.03723, over 16800.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2481, pruned_loss=0.04337, over 3281407.98 frames. ], batch size: 39, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:06:28,589 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-05-02 22:07:25,649 INFO [train.py:904] (4/8) Epoch 30, batch 4050, loss[loss=0.1759, simple_loss=0.2542, pruned_loss=0.04877, over 16158.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.249, pruned_loss=0.04265, over 3283770.24 frames. ], batch size: 35, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:07:27,596 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.527e+02 1.991e+02 2.231e+02 2.617e+02 4.473e+02, threshold=4.462e+02, percent-clipped=0.0 2023-05-02 22:07:28,587 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7065, 4.8812, 5.1024, 5.0416, 5.1222, 4.8659, 4.6810, 4.6103], device='cuda:4'), covar=tensor([0.0478, 0.0683, 0.0458, 0.0595, 0.0640, 0.0545, 0.1223, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0455, 0.0519, 0.0493, 0.0457, 0.0540, 0.0522, 0.0601, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-02 22:07:36,171 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-05-02 22:07:37,015 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4511, 3.4265, 3.4846, 3.5615, 3.6129, 3.3477, 3.5599, 3.6793], device='cuda:4'), covar=tensor([0.1240, 0.0924, 0.1040, 0.0635, 0.0636, 0.2234, 0.1189, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0719, 0.0871, 0.1006, 0.0890, 0.0676, 0.0703, 0.0739, 0.0862], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 22:08:02,579 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7104, 2.3785, 1.9144, 2.1462, 2.6757, 2.3308, 2.3721, 2.7495], device='cuda:4'), covar=tensor([0.0221, 0.0470, 0.0638, 0.0527, 0.0285, 0.0393, 0.0224, 0.0291], device='cuda:4'), in_proj_covar=tensor([0.0241, 0.0250, 0.0240, 0.0241, 0.0251, 0.0249, 0.0250, 0.0252], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 22:08:22,526 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.42 vs. limit=5.0 2023-05-02 22:08:37,711 INFO [train.py:904] (4/8) Epoch 30, batch 4100, loss[loss=0.1989, simple_loss=0.2846, pruned_loss=0.05665, over 16686.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2504, pruned_loss=0.04209, over 3288794.49 frames. ], batch size: 124, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:08:58,543 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.63 vs. limit=5.0 2023-05-02 22:09:20,293 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=298482.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:09:53,961 INFO [train.py:904] (4/8) Epoch 30, batch 4150, loss[loss=0.1806, simple_loss=0.271, pruned_loss=0.04509, over 16368.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2569, pruned_loss=0.04424, over 3255363.39 frames. ], batch size: 35, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:09:56,054 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.391e+02 2.028e+02 2.326e+02 2.808e+02 4.631e+02, threshold=4.653e+02, percent-clipped=1.0 2023-05-02 22:10:35,422 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=298530.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:10:55,752 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=298543.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 22:11:12,252 INFO [train.py:904] (4/8) Epoch 30, batch 4200, loss[loss=0.1857, simple_loss=0.285, pruned_loss=0.04314, over 16384.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2634, pruned_loss=0.04558, over 3228106.81 frames. ], batch size: 68, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:12:08,221 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=298591.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:12:26,984 INFO [train.py:904] (4/8) Epoch 30, batch 4250, loss[loss=0.1563, simple_loss=0.2554, pruned_loss=0.02863, over 16337.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2669, pruned_loss=0.04542, over 3213401.37 frames. ], batch size: 165, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:12:28,280 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.166e+02 2.498e+02 2.846e+02 5.761e+02, threshold=4.997e+02, percent-clipped=3.0 2023-05-02 22:12:40,767 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=298613.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:13:04,708 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=298629.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 22:13:41,911 INFO [train.py:904] (4/8) Epoch 30, batch 4300, loss[loss=0.1902, simple_loss=0.285, pruned_loss=0.04767, over 16888.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2682, pruned_loss=0.04481, over 3188118.96 frames. ], batch size: 116, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:13:52,430 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=298661.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:14:14,614 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-05-02 22:14:35,022 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=298689.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:14:36,391 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=298690.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 22:14:56,148 INFO [train.py:904] (4/8) Epoch 30, batch 4350, loss[loss=0.198, simple_loss=0.2849, pruned_loss=0.05559, over 16485.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2714, pruned_loss=0.04612, over 3169909.63 frames. ], batch size: 68, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:14:57,381 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.491e+02 2.186e+02 2.538e+02 2.866e+02 4.288e+02, threshold=5.076e+02, percent-clipped=0.0 2023-05-02 22:15:08,159 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7258, 2.3322, 2.0023, 2.1294, 2.6076, 2.2485, 2.4491, 2.7279], device='cuda:4'), covar=tensor([0.0240, 0.0437, 0.0591, 0.0506, 0.0309, 0.0429, 0.0278, 0.0287], device='cuda:4'), in_proj_covar=tensor([0.0239, 0.0249, 0.0238, 0.0239, 0.0249, 0.0247, 0.0248, 0.0250], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 22:16:03,961 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=298750.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 22:16:09,613 INFO [train.py:904] (4/8) Epoch 30, batch 4400, loss[loss=0.1843, simple_loss=0.2821, pruned_loss=0.04322, over 16894.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2737, pruned_loss=0.04708, over 3183082.02 frames. ], batch size: 116, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:17:21,871 INFO [train.py:904] (4/8) Epoch 30, batch 4450, loss[loss=0.1947, simple_loss=0.2883, pruned_loss=0.05057, over 16964.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2776, pruned_loss=0.04878, over 3192615.47 frames. ], batch size: 55, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:17:23,559 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 2.049e+02 2.382e+02 3.028e+02 5.771e+02, threshold=4.764e+02, percent-clipped=2.0 2023-05-02 22:18:09,768 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=298838.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 22:18:32,504 INFO [train.py:904] (4/8) Epoch 30, batch 4500, loss[loss=0.1955, simple_loss=0.287, pruned_loss=0.05201, over 17237.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2781, pruned_loss=0.04933, over 3213299.63 frames. ], batch size: 44, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:19:00,964 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2991, 5.3828, 5.1832, 4.8143, 4.9039, 5.2622, 5.0458, 4.9561], device='cuda:4'), covar=tensor([0.0512, 0.0257, 0.0222, 0.0227, 0.0754, 0.0267, 0.0269, 0.0511], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0486, 0.0378, 0.0380, 0.0374, 0.0437, 0.0259, 0.0449], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:19:19,731 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=298886.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:19:22,907 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=5.06 vs. limit=5.0 2023-05-02 22:19:44,877 INFO [train.py:904] (4/8) Epoch 30, batch 4550, loss[loss=0.1936, simple_loss=0.2841, pruned_loss=0.05158, over 16403.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2794, pruned_loss=0.05022, over 3215497.81 frames. ], batch size: 68, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:19:46,101 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.328e+02 1.782e+02 2.097e+02 2.367e+02 4.731e+02, threshold=4.194e+02, percent-clipped=0.0 2023-05-02 22:20:57,286 INFO [train.py:904] (4/8) Epoch 30, batch 4600, loss[loss=0.2067, simple_loss=0.294, pruned_loss=0.05971, over 16523.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2804, pruned_loss=0.05054, over 3228646.50 frames. ], batch size: 68, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:21:33,908 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.0854, 5.1135, 4.8340, 4.1827, 5.0671, 2.0115, 4.7868, 4.3126], device='cuda:4'), covar=tensor([0.0060, 0.0057, 0.0147, 0.0328, 0.0055, 0.2852, 0.0086, 0.0288], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0179, 0.0218, 0.0190, 0.0196, 0.0222, 0.0207, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 22:21:42,882 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=298985.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 22:21:54,431 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4315, 3.4460, 2.0873, 4.0381, 2.7064, 3.9877, 2.3627, 2.8648], device='cuda:4'), covar=tensor([0.0356, 0.0443, 0.1910, 0.0160, 0.0889, 0.0475, 0.1602, 0.0894], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0184, 0.0197, 0.0176, 0.0182, 0.0223, 0.0205, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:22:09,580 INFO [train.py:904] (4/8) Epoch 30, batch 4650, loss[loss=0.174, simple_loss=0.2558, pruned_loss=0.04613, over 16495.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2795, pruned_loss=0.05055, over 3234715.38 frames. ], batch size: 35, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:22:10,878 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.373e+02 1.831e+02 2.100e+02 2.427e+02 4.549e+02, threshold=4.199e+02, percent-clipped=1.0 2023-05-02 22:23:09,774 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=299045.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 22:23:22,504 INFO [train.py:904] (4/8) Epoch 30, batch 4700, loss[loss=0.1618, simple_loss=0.2563, pruned_loss=0.03367, over 16735.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2761, pruned_loss=0.04908, over 3236866.02 frames. ], batch size: 124, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:23:37,410 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.2894, 3.2450, 1.6973, 3.5586, 2.3777, 3.5309, 1.8707, 2.5595], device='cuda:4'), covar=tensor([0.0315, 0.0463, 0.2222, 0.0213, 0.1019, 0.0593, 0.2155, 0.0973], device='cuda:4'), in_proj_covar=tensor([0.0179, 0.0184, 0.0197, 0.0176, 0.0182, 0.0224, 0.0206, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:24:36,507 INFO [train.py:904] (4/8) Epoch 30, batch 4750, loss[loss=0.1694, simple_loss=0.2716, pruned_loss=0.03358, over 16766.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2718, pruned_loss=0.04691, over 3248655.76 frames. ], batch size: 89, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:24:37,715 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.301e+02 1.904e+02 2.113e+02 2.472e+02 4.285e+02, threshold=4.226e+02, percent-clipped=1.0 2023-05-02 22:24:55,278 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 22:25:06,714 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8668, 2.7912, 2.7977, 4.6939, 3.2530, 4.0582, 1.7911, 2.9628], device='cuda:4'), covar=tensor([0.1302, 0.0788, 0.1181, 0.0164, 0.0233, 0.0365, 0.1586, 0.0865], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0183, 0.0202, 0.0208, 0.0208, 0.0219, 0.0212, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:25:08,348 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=299126.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:25:25,792 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=299138.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:25:49,338 INFO [train.py:904] (4/8) Epoch 30, batch 4800, loss[loss=0.1786, simple_loss=0.2618, pruned_loss=0.04774, over 16656.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2677, pruned_loss=0.04485, over 3244553.61 frames. ], batch size: 62, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:26:37,234 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=299186.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:26:37,282 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=299186.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:26:38,533 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=299187.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:26:44,995 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=299191.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:27:04,356 INFO [train.py:904] (4/8) Epoch 30, batch 4850, loss[loss=0.2348, simple_loss=0.3103, pruned_loss=0.07966, over 11891.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.269, pruned_loss=0.04423, over 3220817.19 frames. ], batch size: 247, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:27:06,457 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.292e+02 1.894e+02 2.171e+02 2.574e+02 8.129e+02, threshold=4.343e+02, percent-clipped=1.0 2023-05-02 22:27:36,209 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1727, 3.0468, 3.2893, 1.7754, 3.4477, 3.4462, 2.8403, 2.6398], device='cuda:4'), covar=tensor([0.0935, 0.0311, 0.0201, 0.1329, 0.0106, 0.0199, 0.0441, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0149, 0.0111, 0.0103, 0.0139, 0.0088, 0.0134, 0.0132, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:27:49,334 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=299234.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:28:03,000 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=299243.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:28:16,446 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=299252.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:28:18,999 INFO [train.py:904] (4/8) Epoch 30, batch 4900, loss[loss=0.1749, simple_loss=0.2625, pruned_loss=0.04364, over 11928.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2676, pruned_loss=0.04267, over 3206247.79 frames. ], batch size: 246, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:28:35,183 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.4688, 3.7356, 2.7254, 2.1490, 2.3877, 2.4684, 3.9615, 3.1134], device='cuda:4'), covar=tensor([0.3162, 0.0602, 0.1983, 0.3017, 0.2762, 0.2054, 0.0460, 0.1343], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0276, 0.0315, 0.0330, 0.0309, 0.0281, 0.0306, 0.0355], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:29:05,448 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-05-02 22:29:05,978 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=299285.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 22:29:33,788 INFO [train.py:904] (4/8) Epoch 30, batch 4950, loss[loss=0.1895, simple_loss=0.2836, pruned_loss=0.04771, over 15291.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2676, pruned_loss=0.04208, over 3191334.39 frames. ], batch size: 190, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:29:34,222 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=299304.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:29:34,846 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.269e+02 1.940e+02 2.221e+02 2.664e+02 5.987e+02, threshold=4.443e+02, percent-clipped=2.0 2023-05-02 22:30:02,784 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=299323.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:30:17,899 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=299333.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 22:30:32,477 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3506, 3.4847, 3.5441, 2.1775, 2.9258, 2.4160, 3.7354, 3.8380], device='cuda:4'), covar=tensor([0.0275, 0.0799, 0.0750, 0.2052, 0.0959, 0.0997, 0.0569, 0.0882], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0173, 0.0172, 0.0159, 0.0149, 0.0134, 0.0147, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 22:30:36,122 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=299345.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:30:47,899 INFO [train.py:904] (4/8) Epoch 30, batch 5000, loss[loss=0.1811, simple_loss=0.2797, pruned_loss=0.04129, over 16391.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2693, pruned_loss=0.04226, over 3191904.77 frames. ], batch size: 146, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:30:49,625 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8061, 4.6229, 4.8279, 4.9997, 5.1883, 4.6400, 5.2096, 5.1869], device='cuda:4'), covar=tensor([0.1859, 0.1414, 0.1697, 0.0793, 0.0523, 0.0981, 0.0612, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0688, 0.0835, 0.0965, 0.0854, 0.0646, 0.0673, 0.0706, 0.0822], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 22:31:33,566 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=299384.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:31:46,453 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=299393.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:32:01,705 INFO [train.py:904] (4/8) Epoch 30, batch 5050, loss[loss=0.1713, simple_loss=0.2667, pruned_loss=0.03793, over 15402.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2701, pruned_loss=0.04213, over 3202444.21 frames. ], batch size: 190, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:32:02,876 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.469e+02 2.023e+02 2.400e+02 2.818e+02 3.992e+02, threshold=4.801e+02, percent-clipped=0.0 2023-05-02 22:32:17,897 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5102, 3.7444, 2.8358, 2.2590, 2.4425, 2.5193, 3.9868, 3.2557], device='cuda:4'), covar=tensor([0.3279, 0.0650, 0.1988, 0.3089, 0.2669, 0.2138, 0.0511, 0.1433], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0278, 0.0316, 0.0331, 0.0310, 0.0282, 0.0307, 0.0357], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:33:09,038 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7387, 2.4982, 2.6088, 4.3782, 3.1359, 3.9836, 1.6648, 3.0096], device='cuda:4'), covar=tensor([0.1418, 0.0949, 0.1269, 0.0157, 0.0237, 0.0367, 0.1709, 0.0834], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0183, 0.0202, 0.0208, 0.0208, 0.0219, 0.0213, 0.0202], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:33:11,009 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.22 vs. limit=5.0 2023-05-02 22:33:13,773 INFO [train.py:904] (4/8) Epoch 30, batch 5100, loss[loss=0.171, simple_loss=0.2661, pruned_loss=0.03797, over 15428.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2681, pruned_loss=0.04155, over 3202987.66 frames. ], batch size: 191, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:33:58,015 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=299482.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:34:30,234 INFO [train.py:904] (4/8) Epoch 30, batch 5150, loss[loss=0.1769, simple_loss=0.2765, pruned_loss=0.03868, over 16587.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2683, pruned_loss=0.04111, over 3206175.31 frames. ], batch size: 57, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:34:31,338 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.423e+02 1.896e+02 2.234e+02 2.667e+02 3.487e+02, threshold=4.469e+02, percent-clipped=0.0 2023-05-02 22:35:19,722 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-05-02 22:35:33,156 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=299547.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:35:38,752 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8583, 3.1430, 3.4587, 2.0355, 2.9060, 2.3158, 3.3662, 3.4195], device='cuda:4'), covar=tensor([0.0262, 0.0823, 0.0585, 0.2094, 0.0883, 0.0990, 0.0579, 0.0771], device='cuda:4'), in_proj_covar=tensor([0.0163, 0.0173, 0.0173, 0.0159, 0.0150, 0.0134, 0.0147, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 22:35:43,064 INFO [train.py:904] (4/8) Epoch 30, batch 5200, loss[loss=0.1889, simple_loss=0.2725, pruned_loss=0.0527, over 16887.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2668, pruned_loss=0.04079, over 3206894.58 frames. ], batch size: 109, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:36:48,431 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=299599.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:36:55,738 INFO [train.py:904] (4/8) Epoch 30, batch 5250, loss[loss=0.1741, simple_loss=0.2587, pruned_loss=0.0448, over 12356.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2651, pruned_loss=0.0407, over 3200036.96 frames. ], batch size: 246, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:36:56,958 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.459e+02 2.002e+02 2.220e+02 2.747e+02 4.551e+02, threshold=4.439e+02, percent-clipped=1.0 2023-05-02 22:37:30,897 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-05-02 22:38:06,894 INFO [train.py:904] (4/8) Epoch 30, batch 5300, loss[loss=0.1549, simple_loss=0.2361, pruned_loss=0.03684, over 16486.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2624, pruned_loss=0.04026, over 3184670.25 frames. ], batch size: 68, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:38:42,816 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=299679.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:39:18,175 INFO [train.py:904] (4/8) Epoch 30, batch 5350, loss[loss=0.1635, simple_loss=0.2679, pruned_loss=0.02955, over 16862.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2606, pruned_loss=0.03956, over 3183108.98 frames. ], batch size: 102, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:39:19,343 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.376e+02 1.953e+02 2.222e+02 2.592e+02 5.137e+02, threshold=4.444e+02, percent-clipped=1.0 2023-05-02 22:39:26,150 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6436, 3.9514, 2.8891, 2.3140, 2.5902, 2.6616, 4.2408, 3.3631], device='cuda:4'), covar=tensor([0.3105, 0.0639, 0.1964, 0.2840, 0.2626, 0.2032, 0.0485, 0.1354], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0275, 0.0313, 0.0328, 0.0306, 0.0280, 0.0305, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:40:27,077 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5262, 4.3108, 4.2363, 2.7700, 3.7078, 4.2639, 3.6669, 2.3966], device='cuda:4'), covar=tensor([0.0588, 0.0050, 0.0053, 0.0441, 0.0115, 0.0101, 0.0129, 0.0486], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0092, 0.0093, 0.0136, 0.0104, 0.0117, 0.0101, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 22:40:29,705 INFO [train.py:904] (4/8) Epoch 30, batch 5400, loss[loss=0.1697, simple_loss=0.2634, pruned_loss=0.03794, over 16516.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2627, pruned_loss=0.03993, over 3197544.51 frames. ], batch size: 68, lr: 2.24e-03, grad_scale: 8.0 2023-05-02 22:40:51,533 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=299769.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:41:10,105 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=299782.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:41:34,621 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=299798.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 22:41:34,816 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-05-02 22:41:46,647 INFO [train.py:904] (4/8) Epoch 30, batch 5450, loss[loss=0.2253, simple_loss=0.3144, pruned_loss=0.06809, over 15337.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2659, pruned_loss=0.0412, over 3197764.39 frames. ], batch size: 190, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:41:47,795 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.535e+02 2.110e+02 2.402e+02 2.864e+02 4.092e+02, threshold=4.805e+02, percent-clipped=0.0 2023-05-02 22:42:25,882 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.2611, 5.5157, 5.2762, 5.3460, 5.1002, 5.0214, 4.9142, 5.6653], device='cuda:4'), covar=tensor([0.1333, 0.0831, 0.1097, 0.0931, 0.0774, 0.0824, 0.1319, 0.0782], device='cuda:4'), in_proj_covar=tensor([0.0721, 0.0872, 0.0715, 0.0677, 0.0558, 0.0551, 0.0730, 0.0686], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 22:42:27,030 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=299830.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:42:27,185 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=299830.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:42:52,015 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=299847.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:43:02,919 INFO [train.py:904] (4/8) Epoch 30, batch 5500, loss[loss=0.2363, simple_loss=0.3178, pruned_loss=0.0774, over 16466.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2722, pruned_loss=0.04484, over 3173802.20 frames. ], batch size: 35, lr: 2.23e-03, grad_scale: 16.0 2023-05-02 22:43:11,662 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=299859.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 22:44:09,163 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=299895.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:44:14,722 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=299899.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:44:19,309 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1863, 5.2722, 5.5620, 5.5033, 5.6077, 5.2369, 5.1653, 4.9375], device='cuda:4'), covar=tensor([0.0357, 0.0595, 0.0400, 0.0388, 0.0464, 0.0424, 0.1066, 0.0527], device='cuda:4'), in_proj_covar=tensor([0.0437, 0.0500, 0.0478, 0.0442, 0.0523, 0.0506, 0.0582, 0.0408], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 22:44:21,844 INFO [train.py:904] (4/8) Epoch 30, batch 5550, loss[loss=0.2285, simple_loss=0.3106, pruned_loss=0.07318, over 16414.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.279, pruned_loss=0.0494, over 3157078.85 frames. ], batch size: 146, lr: 2.23e-03, grad_scale: 16.0 2023-05-02 22:44:23,798 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.226e+02 2.904e+02 3.261e+02 4.114e+02 7.155e+02, threshold=6.523e+02, percent-clipped=13.0 2023-05-02 22:45:32,581 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=299947.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:45:43,005 INFO [train.py:904] (4/8) Epoch 30, batch 5600, loss[loss=0.1962, simple_loss=0.2872, pruned_loss=0.05259, over 16727.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2818, pruned_loss=0.05171, over 3148389.00 frames. ], batch size: 83, lr: 2.23e-03, grad_scale: 16.0 2023-05-02 22:45:47,960 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3163, 3.1390, 3.6056, 1.8270, 3.7755, 3.7987, 2.8836, 2.7709], device='cuda:4'), covar=tensor([0.0897, 0.0366, 0.0229, 0.1249, 0.0092, 0.0186, 0.0510, 0.0542], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0104, 0.0140, 0.0088, 0.0135, 0.0132, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:46:08,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4870, 4.5821, 4.7248, 4.4787, 4.6050, 5.0726, 4.5678, 4.3118], device='cuda:4'), covar=tensor([0.1384, 0.1941, 0.2410, 0.1930, 0.2135, 0.1005, 0.1814, 0.2467], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0636, 0.0708, 0.0516, 0.0688, 0.0726, 0.0546, 0.0687], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:46:27,283 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=299979.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:46:44,588 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2954, 3.5657, 3.5479, 2.1079, 3.0131, 2.4527, 3.6957, 4.0204], device='cuda:4'), covar=tensor([0.0330, 0.0888, 0.0723, 0.2289, 0.0987, 0.1101, 0.0691, 0.0911], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0171, 0.0157, 0.0148, 0.0133, 0.0145, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 22:46:53,349 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-05-02 22:47:11,368 INFO [train.py:904] (4/8) Epoch 30, batch 5650, loss[loss=0.1842, simple_loss=0.2748, pruned_loss=0.04678, over 16489.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2867, pruned_loss=0.05521, over 3133288.09 frames. ], batch size: 68, lr: 2.23e-03, grad_scale: 16.0 2023-05-02 22:47:13,271 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.022e+02 3.189e+02 3.868e+02 4.538e+02 6.994e+02, threshold=7.735e+02, percent-clipped=1.0 2023-05-02 22:47:48,225 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=300027.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:48:27,596 INFO [train.py:904] (4/8) Epoch 30, batch 5700, loss[loss=0.2046, simple_loss=0.2984, pruned_loss=0.05542, over 15313.00 frames. ], tot_loss[loss=0.2008, simple_loss=0.2878, pruned_loss=0.05693, over 3112255.61 frames. ], batch size: 191, lr: 2.23e-03, grad_scale: 16.0 2023-05-02 22:48:51,475 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5632, 3.6789, 2.8437, 2.3678, 2.5734, 2.5760, 4.0695, 3.3146], device='cuda:4'), covar=tensor([0.3277, 0.0766, 0.2013, 0.2907, 0.2689, 0.2202, 0.0541, 0.1424], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0277, 0.0316, 0.0331, 0.0308, 0.0281, 0.0306, 0.0354], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:48:57,335 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=300073.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:49:45,341 INFO [train.py:904] (4/8) Epoch 30, batch 5750, loss[loss=0.2037, simple_loss=0.2973, pruned_loss=0.05507, over 16873.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2917, pruned_loss=0.05923, over 3087853.63 frames. ], batch size: 116, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:49:49,226 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.972e+02 3.119e+02 3.894e+02 5.149e+02 1.112e+03, threshold=7.788e+02, percent-clipped=3.0 2023-05-02 22:50:18,678 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=300125.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:50:21,771 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-05-02 22:50:35,640 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=300134.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:51:06,732 INFO [train.py:904] (4/8) Epoch 30, batch 5800, loss[loss=0.2119, simple_loss=0.2852, pruned_loss=0.06925, over 11845.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2916, pruned_loss=0.05859, over 3062214.32 frames. ], batch size: 246, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:51:07,774 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=300154.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 22:52:25,619 INFO [train.py:904] (4/8) Epoch 30, batch 5850, loss[loss=0.1851, simple_loss=0.2836, pruned_loss=0.04331, over 16882.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2893, pruned_loss=0.05688, over 3071201.01 frames. ], batch size: 96, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:52:28,949 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.843e+02 2.638e+02 2.992e+02 3.617e+02 7.830e+02, threshold=5.985e+02, percent-clipped=1.0 2023-05-02 22:53:20,671 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7296, 2.7078, 2.6021, 3.8077, 2.9991, 3.8333, 1.4490, 3.0196], device='cuda:4'), covar=tensor([0.1409, 0.0758, 0.1239, 0.0180, 0.0210, 0.0340, 0.1817, 0.0762], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0182, 0.0202, 0.0207, 0.0208, 0.0218, 0.0212, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:53:36,292 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8972, 1.4173, 1.7629, 1.7091, 1.8574, 1.9650, 1.6870, 1.8690], device='cuda:4'), covar=tensor([0.0266, 0.0450, 0.0234, 0.0337, 0.0314, 0.0203, 0.0488, 0.0164], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0199, 0.0187, 0.0195, 0.0212, 0.0168, 0.0205, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 22:53:45,026 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0239, 4.0490, 4.3274, 4.2941, 4.3036, 4.0456, 4.0522, 4.0579], device='cuda:4'), covar=tensor([0.0381, 0.0650, 0.0450, 0.0403, 0.0488, 0.0498, 0.0950, 0.0591], device='cuda:4'), in_proj_covar=tensor([0.0440, 0.0504, 0.0481, 0.0446, 0.0528, 0.0509, 0.0585, 0.0411], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 22:53:48,214 INFO [train.py:904] (4/8) Epoch 30, batch 5900, loss[loss=0.1656, simple_loss=0.2618, pruned_loss=0.03464, over 16668.00 frames. ], tot_loss[loss=0.2014, simple_loss=0.2888, pruned_loss=0.05702, over 3044693.37 frames. ], batch size: 89, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:54:30,646 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.22 vs. limit=5.0 2023-05-02 22:55:05,751 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6564, 4.7969, 5.0009, 4.7329, 4.8373, 5.3643, 4.8676, 4.5700], device='cuda:4'), covar=tensor([0.1258, 0.2024, 0.2761, 0.2016, 0.2267, 0.0967, 0.1777, 0.2591], device='cuda:4'), in_proj_covar=tensor([0.0433, 0.0640, 0.0715, 0.0520, 0.0694, 0.0729, 0.0550, 0.0690], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:55:09,235 INFO [train.py:904] (4/8) Epoch 30, batch 5950, loss[loss=0.2058, simple_loss=0.2921, pruned_loss=0.0598, over 16653.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2896, pruned_loss=0.05612, over 3035084.80 frames. ], batch size: 62, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:55:12,894 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.875e+02 2.782e+02 3.249e+02 4.004e+02 6.648e+02, threshold=6.499e+02, percent-clipped=2.0 2023-05-02 22:56:10,744 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3966, 3.4282, 2.4232, 2.1741, 2.2195, 2.0884, 3.4751, 2.9784], device='cuda:4'), covar=tensor([0.3471, 0.0814, 0.2446, 0.2955, 0.3003, 0.2792, 0.0738, 0.1579], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0275, 0.0314, 0.0329, 0.0306, 0.0280, 0.0304, 0.0352], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003], device='cuda:4') 2023-05-02 22:56:25,387 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-05-02 22:56:29,059 INFO [train.py:904] (4/8) Epoch 30, batch 6000, loss[loss=0.1678, simple_loss=0.2651, pruned_loss=0.03522, over 16875.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2884, pruned_loss=0.05587, over 3035855.18 frames. ], batch size: 102, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:56:29,059 INFO [train.py:929] (4/8) Computing validation loss 2023-05-02 22:56:36,471 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6460, 3.6454, 4.3674, 2.4680, 3.6442, 2.8936, 4.0543, 3.8985], device='cuda:4'), covar=tensor([0.0173, 0.0935, 0.0415, 0.2094, 0.0642, 0.0961, 0.0472, 0.0947], device='cuda:4'), in_proj_covar=tensor([0.0161, 0.0171, 0.0171, 0.0156, 0.0148, 0.0133, 0.0145, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 22:56:39,790 INFO [train.py:938] (4/8) Epoch 30, validation: loss=0.1471, simple_loss=0.2591, pruned_loss=0.01755, over 944034.00 frames. 2023-05-02 22:56:39,791 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-02 22:57:57,808 INFO [train.py:904] (4/8) Epoch 30, batch 6050, loss[loss=0.1973, simple_loss=0.2903, pruned_loss=0.05212, over 16345.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2875, pruned_loss=0.05517, over 3048546.37 frames. ], batch size: 35, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:58:01,295 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.745e+02 2.541e+02 2.904e+02 3.584e+02 6.768e+02, threshold=5.808e+02, percent-clipped=1.0 2023-05-02 22:58:11,709 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.60 vs. limit=5.0 2023-05-02 22:58:29,299 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=300425.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:58:35,424 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=300429.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:59:06,572 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3873, 3.2466, 3.6559, 1.8969, 3.7873, 3.7898, 2.9822, 2.8685], device='cuda:4'), covar=tensor([0.0892, 0.0340, 0.0227, 0.1299, 0.0099, 0.0201, 0.0463, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0111, 0.0103, 0.0139, 0.0088, 0.0134, 0.0132, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 22:59:15,657 INFO [train.py:904] (4/8) Epoch 30, batch 6100, loss[loss=0.1946, simple_loss=0.2868, pruned_loss=0.05116, over 16933.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2869, pruned_loss=0.05429, over 3076313.52 frames. ], batch size: 109, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 22:59:16,343 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=300454.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 22:59:47,466 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=300473.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 22:59:48,967 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0911, 3.9837, 4.1593, 4.2920, 4.3867, 4.0066, 4.3593, 4.4119], device='cuda:4'), covar=tensor([0.1802, 0.1218, 0.1407, 0.0706, 0.0608, 0.1335, 0.0830, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0681, 0.0828, 0.0954, 0.0847, 0.0642, 0.0666, 0.0703, 0.0817], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:00:32,802 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=300502.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 23:00:34,966 INFO [train.py:904] (4/8) Epoch 30, batch 6150, loss[loss=0.171, simple_loss=0.2654, pruned_loss=0.03834, over 16892.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.285, pruned_loss=0.05399, over 3060271.50 frames. ], batch size: 90, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:00:38,226 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.879e+02 2.817e+02 3.397e+02 3.962e+02 7.100e+02, threshold=6.795e+02, percent-clipped=3.0 2023-05-02 23:01:53,681 INFO [train.py:904] (4/8) Epoch 30, batch 6200, loss[loss=0.1902, simple_loss=0.2789, pruned_loss=0.05072, over 16230.00 frames. ], tot_loss[loss=0.1952, simple_loss=0.2829, pruned_loss=0.05374, over 3067922.96 frames. ], batch size: 165, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:03:13,157 INFO [train.py:904] (4/8) Epoch 30, batch 6250, loss[loss=0.178, simple_loss=0.2811, pruned_loss=0.03745, over 16736.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2826, pruned_loss=0.05335, over 3075065.35 frames. ], batch size: 89, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:03:16,359 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.883e+02 2.845e+02 3.157e+02 4.016e+02 6.897e+02, threshold=6.314e+02, percent-clipped=1.0 2023-05-02 23:03:20,523 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4350, 3.5382, 3.6481, 3.6320, 3.6573, 3.4976, 3.5222, 3.5337], device='cuda:4'), covar=tensor([0.0411, 0.0640, 0.0507, 0.0479, 0.0530, 0.0562, 0.0786, 0.0604], device='cuda:4'), in_proj_covar=tensor([0.0440, 0.0503, 0.0481, 0.0446, 0.0526, 0.0509, 0.0584, 0.0411], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 23:03:34,832 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=300617.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:04:06,440 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0599, 3.1357, 3.3958, 2.0897, 2.8812, 2.2359, 3.5460, 3.5542], device='cuda:4'), covar=tensor([0.0246, 0.0988, 0.0733, 0.2266, 0.0987, 0.1153, 0.0603, 0.1003], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0172, 0.0172, 0.0158, 0.0149, 0.0134, 0.0146, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 23:04:27,624 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5483, 2.8049, 2.3767, 2.5726, 3.1207, 2.7813, 3.1279, 3.3037], device='cuda:4'), covar=tensor([0.0162, 0.0454, 0.0550, 0.0477, 0.0290, 0.0411, 0.0297, 0.0291], device='cuda:4'), in_proj_covar=tensor([0.0231, 0.0241, 0.0232, 0.0232, 0.0243, 0.0239, 0.0240, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:04:31,360 INFO [train.py:904] (4/8) Epoch 30, batch 6300, loss[loss=0.1972, simple_loss=0.2809, pruned_loss=0.05679, over 16217.00 frames. ], tot_loss[loss=0.1938, simple_loss=0.282, pruned_loss=0.05276, over 3078781.75 frames. ], batch size: 35, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:05:11,440 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=300678.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:05:23,166 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0175, 4.0797, 4.3271, 4.3051, 4.3162, 4.0783, 4.0655, 4.0614], device='cuda:4'), covar=tensor([0.0446, 0.0870, 0.0645, 0.0603, 0.0625, 0.0726, 0.0908, 0.0722], device='cuda:4'), in_proj_covar=tensor([0.0442, 0.0504, 0.0481, 0.0447, 0.0527, 0.0510, 0.0586, 0.0412], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 23:05:40,761 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9491, 4.1661, 4.0120, 4.0409, 3.7694, 3.7930, 3.8570, 4.1633], device='cuda:4'), covar=tensor([0.1137, 0.0906, 0.1060, 0.0902, 0.0824, 0.1858, 0.0970, 0.1051], device='cuda:4'), in_proj_covar=tensor([0.0723, 0.0874, 0.0718, 0.0679, 0.0558, 0.0555, 0.0731, 0.0685], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:05:50,440 INFO [train.py:904] (4/8) Epoch 30, batch 6350, loss[loss=0.1955, simple_loss=0.281, pruned_loss=0.05504, over 16628.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.2826, pruned_loss=0.05342, over 3093871.06 frames. ], batch size: 134, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:05:53,979 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.720e+02 2.820e+02 3.153e+02 4.098e+02 7.645e+02, threshold=6.307e+02, percent-clipped=4.0 2023-05-02 23:06:16,404 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9840, 4.1481, 4.3884, 4.3436, 4.3719, 4.1242, 4.0275, 4.1058], device='cuda:4'), covar=tensor([0.0591, 0.0772, 0.0517, 0.0590, 0.0667, 0.0653, 0.1300, 0.0660], device='cuda:4'), in_proj_covar=tensor([0.0442, 0.0504, 0.0482, 0.0447, 0.0528, 0.0511, 0.0586, 0.0412], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 23:06:29,183 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=300729.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:06:49,335 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-05-02 23:07:07,467 INFO [train.py:904] (4/8) Epoch 30, batch 6400, loss[loss=0.2794, simple_loss=0.337, pruned_loss=0.1109, over 11352.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.2841, pruned_loss=0.05524, over 3072427.13 frames. ], batch size: 248, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:07:41,793 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3316, 3.1192, 3.5062, 1.7213, 3.6436, 3.6742, 2.9664, 2.6804], device='cuda:4'), covar=tensor([0.0881, 0.0329, 0.0230, 0.1359, 0.0106, 0.0217, 0.0449, 0.0566], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0103, 0.0140, 0.0088, 0.0135, 0.0132, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 23:07:42,791 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=300777.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:08:19,451 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=300802.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:08:21,337 INFO [train.py:904] (4/8) Epoch 30, batch 6450, loss[loss=0.2164, simple_loss=0.2938, pruned_loss=0.06949, over 11843.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2846, pruned_loss=0.05508, over 3058539.16 frames. ], batch size: 246, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:08:24,279 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.818e+02 2.836e+02 3.339e+02 4.246e+02 8.060e+02, threshold=6.678e+02, percent-clipped=2.0 2023-05-02 23:09:38,120 INFO [train.py:904] (4/8) Epoch 30, batch 6500, loss[loss=0.1693, simple_loss=0.2648, pruned_loss=0.03694, over 16820.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2823, pruned_loss=0.05435, over 3048059.01 frames. ], batch size: 96, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:09:52,528 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=300863.0, num_to_drop=1, layers_to_drop={0} 2023-05-02 23:10:27,487 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9465, 2.1886, 2.1814, 3.5057, 2.0391, 2.4977, 2.2334, 2.3085], device='cuda:4'), covar=tensor([0.1634, 0.3622, 0.3190, 0.0676, 0.4394, 0.2549, 0.4021, 0.3410], device='cuda:4'), in_proj_covar=tensor([0.0426, 0.0478, 0.0390, 0.0340, 0.0447, 0.0551, 0.0452, 0.0561], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:10:58,874 INFO [train.py:904] (4/8) Epoch 30, batch 6550, loss[loss=0.2281, simple_loss=0.319, pruned_loss=0.06858, over 15400.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2854, pruned_loss=0.05487, over 3066191.10 frames. ], batch size: 191, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:11:01,652 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.987e+02 2.626e+02 3.208e+02 3.708e+02 1.015e+03, threshold=6.415e+02, percent-clipped=2.0 2023-05-02 23:12:11,521 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5296, 3.6083, 3.3849, 2.9791, 3.2373, 3.5093, 3.3444, 3.3532], device='cuda:4'), covar=tensor([0.0650, 0.0672, 0.0307, 0.0276, 0.0507, 0.0487, 0.1247, 0.0465], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0478, 0.0368, 0.0370, 0.0365, 0.0426, 0.0253, 0.0438], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:12:19,421 INFO [train.py:904] (4/8) Epoch 30, batch 6600, loss[loss=0.1861, simple_loss=0.2787, pruned_loss=0.04671, over 15181.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.287, pruned_loss=0.05473, over 3089790.99 frames. ], batch size: 190, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:12:48,780 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=300973.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:13:38,992 INFO [train.py:904] (4/8) Epoch 30, batch 6650, loss[loss=0.1844, simple_loss=0.2722, pruned_loss=0.04835, over 16868.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2871, pruned_loss=0.05531, over 3098998.09 frames. ], batch size: 109, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:13:43,931 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.884e+02 2.674e+02 3.483e+02 3.945e+02 7.489e+02, threshold=6.965e+02, percent-clipped=2.0 2023-05-02 23:13:49,702 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-02 23:14:44,971 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7196, 3.5640, 4.1132, 2.0879, 4.2871, 4.3118, 3.1452, 3.2169], device='cuda:4'), covar=tensor([0.0823, 0.0322, 0.0243, 0.1275, 0.0083, 0.0163, 0.0451, 0.0484], device='cuda:4'), in_proj_covar=tensor([0.0151, 0.0112, 0.0104, 0.0141, 0.0089, 0.0135, 0.0133, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 23:14:55,111 INFO [train.py:904] (4/8) Epoch 30, batch 6700, loss[loss=0.2444, simple_loss=0.3095, pruned_loss=0.08963, over 11521.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2861, pruned_loss=0.05582, over 3085318.55 frames. ], batch size: 246, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:15:39,449 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.0954, 2.4721, 2.5689, 1.9233, 2.7059, 2.7855, 2.4076, 2.3542], device='cuda:4'), covar=tensor([0.0726, 0.0298, 0.0272, 0.1010, 0.0150, 0.0302, 0.0495, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0152, 0.0113, 0.0105, 0.0141, 0.0089, 0.0136, 0.0133, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 23:16:11,407 INFO [train.py:904] (4/8) Epoch 30, batch 6750, loss[loss=0.1934, simple_loss=0.2837, pruned_loss=0.05158, over 16271.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2857, pruned_loss=0.05664, over 3068439.89 frames. ], batch size: 165, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:16:15,728 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.912e+02 2.866e+02 3.351e+02 4.005e+02 5.751e+02, threshold=6.702e+02, percent-clipped=0.0 2023-05-02 23:16:16,924 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7983, 1.9671, 2.3685, 2.7634, 2.7215, 3.1238, 2.0411, 3.0756], device='cuda:4'), covar=tensor([0.0253, 0.0611, 0.0381, 0.0402, 0.0380, 0.0213, 0.0622, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0199, 0.0188, 0.0194, 0.0211, 0.0168, 0.0204, 0.0168], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:17:07,594 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=301140.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:17:22,070 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.5031, 1.7979, 2.1295, 2.4668, 2.5072, 2.8073, 1.8708, 2.7098], device='cuda:4'), covar=tensor([0.0265, 0.0615, 0.0385, 0.0380, 0.0384, 0.0238, 0.0632, 0.0183], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0200, 0.0188, 0.0194, 0.0212, 0.0169, 0.0205, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 23:17:29,009 INFO [train.py:904] (4/8) Epoch 30, batch 6800, loss[loss=0.1874, simple_loss=0.2771, pruned_loss=0.04886, over 17018.00 frames. ], tot_loss[loss=0.1983, simple_loss=0.2848, pruned_loss=0.05594, over 3075050.75 frames. ], batch size: 55, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:17:35,471 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=301158.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 23:18:43,873 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=301201.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:18:47,466 INFO [train.py:904] (4/8) Epoch 30, batch 6850, loss[loss=0.2, simple_loss=0.2936, pruned_loss=0.05323, over 16722.00 frames. ], tot_loss[loss=0.1997, simple_loss=0.2867, pruned_loss=0.05636, over 3091424.16 frames. ], batch size: 62, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:18:51,707 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.931e+02 2.806e+02 3.426e+02 4.107e+02 8.215e+02, threshold=6.851e+02, percent-clipped=6.0 2023-05-02 23:19:20,112 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.97 vs. limit=5.0 2023-05-02 23:20:04,219 INFO [train.py:904] (4/8) Epoch 30, batch 6900, loss[loss=0.1975, simple_loss=0.2908, pruned_loss=0.05214, over 16451.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2886, pruned_loss=0.05554, over 3101695.66 frames. ], batch size: 75, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:20:35,488 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=301273.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:21:22,698 INFO [train.py:904] (4/8) Epoch 30, batch 6950, loss[loss=0.1951, simple_loss=0.2794, pruned_loss=0.05541, over 16544.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.2904, pruned_loss=0.05773, over 3064171.34 frames. ], batch size: 35, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:21:26,933 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.877e+02 3.278e+02 4.034e+02 6.348e+02, threshold=6.555e+02, percent-clipped=0.0 2023-05-02 23:21:50,651 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=301321.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:22:25,324 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-05-02 23:22:38,930 INFO [train.py:904] (4/8) Epoch 30, batch 7000, loss[loss=0.1959, simple_loss=0.297, pruned_loss=0.04735, over 16627.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2905, pruned_loss=0.05703, over 3073923.79 frames. ], batch size: 57, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:23:54,782 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8752, 5.1267, 4.8916, 4.9070, 4.6271, 4.6532, 4.5770, 5.2028], device='cuda:4'), covar=tensor([0.1192, 0.0822, 0.1096, 0.0910, 0.0866, 0.1056, 0.1217, 0.0847], device='cuda:4'), in_proj_covar=tensor([0.0724, 0.0875, 0.0720, 0.0682, 0.0558, 0.0557, 0.0733, 0.0687], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:23:55,569 INFO [train.py:904] (4/8) Epoch 30, batch 7050, loss[loss=0.2073, simple_loss=0.297, pruned_loss=0.05883, over 16752.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2911, pruned_loss=0.05719, over 3067168.32 frames. ], batch size: 124, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:24:00,750 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.808e+02 2.564e+02 3.175e+02 3.880e+02 7.852e+02, threshold=6.351e+02, percent-clipped=2.0 2023-05-02 23:24:38,123 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=301430.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:24:40,109 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8657, 2.7028, 2.8479, 2.1923, 2.6904, 2.1749, 2.6451, 2.8987], device='cuda:4'), covar=tensor([0.0277, 0.0794, 0.0522, 0.1824, 0.0822, 0.0965, 0.0569, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0173, 0.0172, 0.0158, 0.0150, 0.0135, 0.0146, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 23:25:14,075 INFO [train.py:904] (4/8) Epoch 30, batch 7100, loss[loss=0.1932, simple_loss=0.2843, pruned_loss=0.05105, over 16879.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2888, pruned_loss=0.05603, over 3081425.02 frames. ], batch size: 109, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:25:20,940 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=301458.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:25:25,165 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8770, 3.1370, 3.3224, 1.9971, 2.8971, 2.2083, 3.3193, 3.3870], device='cuda:4'), covar=tensor([0.0256, 0.0831, 0.0640, 0.2171, 0.0873, 0.1054, 0.0621, 0.0977], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0173, 0.0172, 0.0158, 0.0149, 0.0134, 0.0147, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 23:25:40,167 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6520, 3.4204, 3.9012, 1.9697, 4.0592, 4.1085, 3.0983, 3.0307], device='cuda:4'), covar=tensor([0.0782, 0.0309, 0.0214, 0.1232, 0.0087, 0.0164, 0.0459, 0.0480], device='cuda:4'), in_proj_covar=tensor([0.0150, 0.0112, 0.0104, 0.0139, 0.0088, 0.0134, 0.0132, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 23:26:14,375 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=301491.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:26:21,593 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=301496.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:26:33,760 INFO [train.py:904] (4/8) Epoch 30, batch 7150, loss[loss=0.1845, simple_loss=0.2786, pruned_loss=0.04524, over 16491.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2866, pruned_loss=0.05537, over 3083632.05 frames. ], batch size: 62, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:26:36,531 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=301506.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:26:37,393 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.858e+02 2.724e+02 3.109e+02 4.036e+02 9.536e+02, threshold=6.218e+02, percent-clipped=1.0 2023-05-02 23:26:52,197 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7661, 2.8354, 2.8025, 5.1412, 3.8569, 4.2368, 1.9058, 3.1458], device='cuda:4'), covar=tensor([0.1361, 0.0860, 0.1216, 0.0162, 0.0353, 0.0408, 0.1527, 0.0812], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0182, 0.0202, 0.0207, 0.0207, 0.0219, 0.0211, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 23:27:03,209 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.7961, 3.0176, 3.2564, 1.9743, 2.8729, 2.1307, 3.3129, 3.3186], device='cuda:4'), covar=tensor([0.0252, 0.0964, 0.0668, 0.2304, 0.0921, 0.1133, 0.0634, 0.1117], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0173, 0.0172, 0.0158, 0.0149, 0.0134, 0.0146, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 23:27:49,042 INFO [train.py:904] (4/8) Epoch 30, batch 7200, loss[loss=0.1749, simple_loss=0.2603, pruned_loss=0.04473, over 17249.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2837, pruned_loss=0.05317, over 3100969.03 frames. ], batch size: 45, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:28:15,684 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.1462, 5.7096, 5.8937, 5.5437, 5.6826, 6.1868, 5.6281, 5.4261], device='cuda:4'), covar=tensor([0.0828, 0.1593, 0.1954, 0.1699, 0.1868, 0.0753, 0.1441, 0.2063], device='cuda:4'), in_proj_covar=tensor([0.0433, 0.0643, 0.0716, 0.0525, 0.0694, 0.0733, 0.0554, 0.0698], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-02 23:28:35,610 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0519, 4.1492, 3.9681, 3.6783, 3.6956, 4.0850, 3.7089, 3.8702], device='cuda:4'), covar=tensor([0.0619, 0.0588, 0.0315, 0.0281, 0.0742, 0.0484, 0.1139, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0472, 0.0364, 0.0365, 0.0361, 0.0421, 0.0251, 0.0433], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:28:59,102 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.3266, 3.9886, 3.9327, 2.5609, 3.5872, 4.0313, 3.5836, 2.2843], device='cuda:4'), covar=tensor([0.0637, 0.0060, 0.0067, 0.0472, 0.0122, 0.0113, 0.0119, 0.0482], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0092, 0.0093, 0.0136, 0.0104, 0.0117, 0.0100, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 23:29:07,490 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.9363, 4.1592, 4.0004, 4.0135, 3.7273, 3.7926, 3.8250, 4.1663], device='cuda:4'), covar=tensor([0.1151, 0.0949, 0.1050, 0.0898, 0.0815, 0.1720, 0.0982, 0.0994], device='cuda:4'), in_proj_covar=tensor([0.0722, 0.0871, 0.0719, 0.0681, 0.0556, 0.0557, 0.0732, 0.0685], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:29:10,980 INFO [train.py:904] (4/8) Epoch 30, batch 7250, loss[loss=0.1786, simple_loss=0.2633, pruned_loss=0.04695, over 16386.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2817, pruned_loss=0.05252, over 3085195.13 frames. ], batch size: 146, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:29:15,148 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.563e+02 2.328e+02 2.684e+02 3.448e+02 5.554e+02, threshold=5.368e+02, percent-clipped=0.0 2023-05-02 23:30:27,540 INFO [train.py:904] (4/8) Epoch 30, batch 7300, loss[loss=0.2493, simple_loss=0.3114, pruned_loss=0.09359, over 11392.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2809, pruned_loss=0.05205, over 3097990.74 frames. ], batch size: 247, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:31:44,853 INFO [train.py:904] (4/8) Epoch 30, batch 7350, loss[loss=0.2019, simple_loss=0.289, pruned_loss=0.05742, over 16863.00 frames. ], tot_loss[loss=0.1949, simple_loss=0.2828, pruned_loss=0.05349, over 3096229.30 frames. ], batch size: 116, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:31:50,960 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.792e+02 2.739e+02 3.075e+02 3.816e+02 5.602e+02, threshold=6.150e+02, percent-clipped=1.0 2023-05-02 23:32:38,474 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2037, 1.6130, 1.9733, 2.1327, 2.2437, 2.4057, 1.7770, 2.3424], device='cuda:4'), covar=tensor([0.0273, 0.0570, 0.0323, 0.0375, 0.0381, 0.0245, 0.0580, 0.0187], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0200, 0.0189, 0.0195, 0.0213, 0.0169, 0.0206, 0.0170], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-02 23:32:44,239 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=301743.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:33:00,863 INFO [train.py:904] (4/8) Epoch 30, batch 7400, loss[loss=0.1937, simple_loss=0.2852, pruned_loss=0.05107, over 16889.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.2842, pruned_loss=0.0543, over 3099020.43 frames. ], batch size: 96, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:33:02,213 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7274, 2.9888, 2.6427, 5.2449, 3.9765, 4.3046, 1.7235, 3.0165], device='cuda:4'), covar=tensor([0.1478, 0.0840, 0.1345, 0.0184, 0.0360, 0.0443, 0.1721, 0.0886], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0183, 0.0203, 0.0208, 0.0208, 0.0220, 0.0212, 0.0202], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 23:33:54,209 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=301786.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:34:01,435 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.92 vs. limit=5.0 2023-05-02 23:34:10,817 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=301796.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:34:23,458 INFO [train.py:904] (4/8) Epoch 30, batch 7450, loss[loss=0.1933, simple_loss=0.2816, pruned_loss=0.05248, over 16722.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2855, pruned_loss=0.05539, over 3099422.33 frames. ], batch size: 89, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:34:24,258 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=301804.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:34:29,849 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.990e+02 2.877e+02 3.334e+02 4.200e+02 6.879e+02, threshold=6.668e+02, percent-clipped=3.0 2023-05-02 23:35:21,966 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4175, 4.5022, 4.3042, 3.9948, 3.9935, 4.4100, 4.1173, 4.1515], device='cuda:4'), covar=tensor([0.0647, 0.0599, 0.0320, 0.0327, 0.0806, 0.0540, 0.0724, 0.0623], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0472, 0.0363, 0.0365, 0.0360, 0.0420, 0.0251, 0.0433], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:35:30,252 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=301844.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:35:45,204 INFO [train.py:904] (4/8) Epoch 30, batch 7500, loss[loss=0.2109, simple_loss=0.2972, pruned_loss=0.06227, over 15286.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2858, pruned_loss=0.05488, over 3106631.36 frames. ], batch size: 190, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:36:56,668 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7668, 1.8535, 1.7008, 1.5229, 2.0082, 1.6179, 1.5972, 1.9434], device='cuda:4'), covar=tensor([0.0220, 0.0383, 0.0483, 0.0410, 0.0247, 0.0304, 0.0202, 0.0237], device='cuda:4'), in_proj_covar=tensor([0.0232, 0.0243, 0.0234, 0.0235, 0.0245, 0.0241, 0.0240, 0.0243], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:36:57,868 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=301899.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:37:05,267 INFO [train.py:904] (4/8) Epoch 30, batch 7550, loss[loss=0.1878, simple_loss=0.2722, pruned_loss=0.05171, over 17000.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2847, pruned_loss=0.05443, over 3119462.82 frames. ], batch size: 55, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:37:07,762 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=301905.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:37:11,511 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.823e+02 2.767e+02 3.333e+02 4.100e+02 7.380e+02, threshold=6.666e+02, percent-clipped=1.0 2023-05-02 23:38:21,597 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1342, 3.5800, 3.5334, 2.2511, 3.3467, 3.6411, 3.3622, 1.8503], device='cuda:4'), covar=tensor([0.0666, 0.0073, 0.0089, 0.0536, 0.0123, 0.0137, 0.0123, 0.0611], device='cuda:4'), in_proj_covar=tensor([0.0140, 0.0092, 0.0093, 0.0136, 0.0103, 0.0117, 0.0100, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0004], device='cuda:4') 2023-05-02 23:38:23,957 INFO [train.py:904] (4/8) Epoch 30, batch 7600, loss[loss=0.1916, simple_loss=0.2701, pruned_loss=0.0565, over 11815.00 frames. ], tot_loss[loss=0.196, simple_loss=0.2835, pruned_loss=0.05424, over 3113241.64 frames. ], batch size: 247, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:38:34,803 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=301960.0, num_to_drop=1, layers_to_drop={3} 2023-05-02 23:38:43,710 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=301966.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:39:19,328 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-05-02 23:39:20,865 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.9677, 5.2511, 5.0560, 5.0378, 4.7764, 4.7200, 4.6735, 5.3960], device='cuda:4'), covar=tensor([0.1306, 0.0944, 0.1030, 0.0958, 0.0829, 0.1137, 0.1252, 0.0926], device='cuda:4'), in_proj_covar=tensor([0.0723, 0.0871, 0.0718, 0.0680, 0.0553, 0.0557, 0.0730, 0.0683], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:39:47,150 INFO [train.py:904] (4/8) Epoch 30, batch 7650, loss[loss=0.2328, simple_loss=0.3017, pruned_loss=0.0819, over 11483.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.2844, pruned_loss=0.05503, over 3106384.31 frames. ], batch size: 250, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:39:53,136 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 2.050e+02 2.820e+02 3.410e+02 4.459e+02 8.175e+02, threshold=6.820e+02, percent-clipped=3.0 2023-05-02 23:40:34,700 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.3558, 3.4924, 3.6750, 2.2432, 3.1359, 2.4070, 3.6468, 3.8161], device='cuda:4'), covar=tensor([0.0255, 0.0892, 0.0591, 0.2178, 0.0922, 0.1054, 0.0624, 0.0994], device='cuda:4'), in_proj_covar=tensor([0.0162, 0.0173, 0.0172, 0.0159, 0.0150, 0.0135, 0.0147, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0005], device='cuda:4') 2023-05-02 23:41:06,720 INFO [train.py:904] (4/8) Epoch 30, batch 7700, loss[loss=0.1948, simple_loss=0.2828, pruned_loss=0.05343, over 16773.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.284, pruned_loss=0.05518, over 3115394.00 frames. ], batch size: 124, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:41:49,137 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.03 vs. limit=5.0 2023-05-02 23:41:57,590 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=302086.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:42:12,060 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.3692, 3.2594, 3.2906, 3.4810, 3.4813, 3.2581, 3.4734, 3.5455], device='cuda:4'), covar=tensor([0.1350, 0.1205, 0.1508, 0.0847, 0.0942, 0.3026, 0.1598, 0.1161], device='cuda:4'), in_proj_covar=tensor([0.0675, 0.0819, 0.0948, 0.0841, 0.0636, 0.0658, 0.0700, 0.0809], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:42:18,121 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=302099.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:42:26,380 INFO [train.py:904] (4/8) Epoch 30, batch 7750, loss[loss=0.1965, simple_loss=0.2977, pruned_loss=0.04763, over 16782.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2837, pruned_loss=0.05492, over 3110585.89 frames. ], batch size: 124, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:42:33,832 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.870e+02 2.636e+02 3.091e+02 3.789e+02 8.720e+02, threshold=6.183e+02, percent-clipped=2.0 2023-05-02 23:43:12,802 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=302134.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:43:41,534 INFO [train.py:904] (4/8) Epoch 30, batch 7800, loss[loss=0.2057, simple_loss=0.2906, pruned_loss=0.06037, over 17060.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2848, pruned_loss=0.0558, over 3101155.22 frames. ], batch size: 55, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:44:13,215 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.75 vs. limit=5.0 2023-05-02 23:44:21,646 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8529, 3.9288, 4.1506, 4.1128, 4.1263, 3.9216, 3.9143, 3.9382], device='cuda:4'), covar=tensor([0.0399, 0.0682, 0.0475, 0.0452, 0.0572, 0.0523, 0.0911, 0.0543], device='cuda:4'), in_proj_covar=tensor([0.0439, 0.0501, 0.0479, 0.0444, 0.0523, 0.0507, 0.0584, 0.0408], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 23:44:55,036 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=302200.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:45:00,017 INFO [train.py:904] (4/8) Epoch 30, batch 7850, loss[loss=0.1929, simple_loss=0.3031, pruned_loss=0.04133, over 16858.00 frames. ], tot_loss[loss=0.1979, simple_loss=0.2853, pruned_loss=0.05519, over 3097517.31 frames. ], batch size: 96, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:45:06,747 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.895e+02 2.702e+02 3.253e+02 3.935e+02 9.747e+02, threshold=6.506e+02, percent-clipped=3.0 2023-05-02 23:46:15,231 INFO [train.py:904] (4/8) Epoch 30, batch 7900, loss[loss=0.2055, simple_loss=0.2915, pruned_loss=0.05978, over 16646.00 frames. ], tot_loss[loss=0.1962, simple_loss=0.2839, pruned_loss=0.05423, over 3109205.57 frames. ], batch size: 57, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:46:16,926 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=302255.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 23:46:25,229 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=302261.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:46:25,354 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=302261.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:46:53,025 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.7090, 2.5366, 2.2110, 3.7854, 2.2688, 3.6769, 1.5047, 2.5807], device='cuda:4'), covar=tensor([0.1534, 0.0983, 0.1576, 0.0242, 0.0248, 0.0529, 0.1952, 0.1069], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0182, 0.0201, 0.0207, 0.0206, 0.0219, 0.0211, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-02 23:47:22,397 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=302296.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:47:32,970 INFO [train.py:904] (4/8) Epoch 30, batch 7950, loss[loss=0.199, simple_loss=0.2796, pruned_loss=0.05918, over 16547.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2851, pruned_loss=0.05511, over 3105531.98 frames. ], batch size: 57, lr: 2.23e-03, grad_scale: 4.0 2023-05-02 23:47:41,026 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.902e+02 2.712e+02 3.186e+02 3.569e+02 8.017e+02, threshold=6.372e+02, percent-clipped=2.0 2023-05-02 23:48:32,625 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4174, 3.3576, 3.4431, 3.5080, 3.5482, 3.3106, 3.5152, 3.6022], device='cuda:4'), covar=tensor([0.1247, 0.0930, 0.0988, 0.0621, 0.0706, 0.2668, 0.1125, 0.0948], device='cuda:4'), in_proj_covar=tensor([0.0672, 0.0816, 0.0944, 0.0837, 0.0636, 0.0655, 0.0696, 0.0806], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:48:47,237 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-05-02 23:48:48,632 INFO [train.py:904] (4/8) Epoch 30, batch 8000, loss[loss=0.2161, simple_loss=0.3118, pruned_loss=0.06025, over 16677.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2861, pruned_loss=0.05585, over 3097654.47 frames. ], batch size: 89, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:48:54,500 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=302357.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:49:41,588 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.22 vs. limit=5.0 2023-05-02 23:49:57,471 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=302399.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:50:03,889 INFO [train.py:904] (4/8) Epoch 30, batch 8050, loss[loss=0.2093, simple_loss=0.2966, pruned_loss=0.06097, over 16868.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2862, pruned_loss=0.05569, over 3098739.70 frames. ], batch size: 116, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:50:11,661 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.866e+02 2.982e+02 3.316e+02 3.990e+02 1.007e+03, threshold=6.633e+02, percent-clipped=4.0 2023-05-02 23:50:14,985 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2215, 4.2000, 4.1014, 3.2353, 4.1691, 1.6611, 3.9500, 3.6638], device='cuda:4'), covar=tensor([0.0161, 0.0137, 0.0225, 0.0367, 0.0119, 0.3159, 0.0155, 0.0339], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0175, 0.0216, 0.0186, 0.0192, 0.0219, 0.0203, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:51:10,641 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=302447.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:51:21,032 INFO [train.py:904] (4/8) Epoch 30, batch 8100, loss[loss=0.1819, simple_loss=0.2699, pruned_loss=0.04692, over 16689.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2851, pruned_loss=0.05454, over 3129721.23 frames. ], batch size: 57, lr: 2.23e-03, grad_scale: 8.0 2023-05-02 23:51:56,223 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=302477.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:52:35,994 INFO [train.py:904] (4/8) Epoch 30, batch 8150, loss[loss=0.1812, simple_loss=0.2713, pruned_loss=0.0455, over 16307.00 frames. ], tot_loss[loss=0.196, simple_loss=0.2834, pruned_loss=0.05428, over 3108551.03 frames. ], batch size: 35, lr: 2.22e-03, grad_scale: 8.0 2023-05-02 23:52:43,486 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.950e+02 2.749e+02 3.387e+02 4.131e+02 6.245e+02, threshold=6.774e+02, percent-clipped=0.0 2023-05-02 23:53:28,258 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=302538.0, num_to_drop=1, layers_to_drop={2} 2023-05-02 23:53:52,238 INFO [train.py:904] (4/8) Epoch 30, batch 8200, loss[loss=0.1801, simple_loss=0.2685, pruned_loss=0.04589, over 16898.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2805, pruned_loss=0.05344, over 3113429.38 frames. ], batch size: 109, lr: 2.22e-03, grad_scale: 8.0 2023-05-02 23:53:53,927 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=302555.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:53:55,104 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=302556.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:54:03,194 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=302561.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:55:13,022 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=302603.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:55:13,852 INFO [train.py:904] (4/8) Epoch 30, batch 8250, loss[loss=0.1844, simple_loss=0.2833, pruned_loss=0.04272, over 16209.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2791, pruned_loss=0.05113, over 3074899.65 frames. ], batch size: 165, lr: 2.22e-03, grad_scale: 8.0 2023-05-02 23:55:22,071 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.591e+02 2.554e+02 2.948e+02 3.684e+02 6.863e+02, threshold=5.896e+02, percent-clipped=1.0 2023-05-02 23:55:23,079 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=302609.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:56:02,637 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=302632.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:56:35,851 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=302652.0, num_to_drop=0, layers_to_drop=set() 2023-05-02 23:56:38,704 INFO [train.py:904] (4/8) Epoch 30, batch 8300, loss[loss=0.1601, simple_loss=0.2649, pruned_loss=0.02769, over 16765.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2765, pruned_loss=0.04834, over 3059811.24 frames. ], batch size: 124, lr: 2.22e-03, grad_scale: 8.0 2023-05-02 23:56:42,917 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8774, 2.3335, 2.0099, 2.1356, 2.5884, 2.3075, 2.2677, 2.7344], device='cuda:4'), covar=tensor([0.0246, 0.0484, 0.0586, 0.0504, 0.0338, 0.0437, 0.0273, 0.0325], device='cuda:4'), in_proj_covar=tensor([0.0230, 0.0242, 0.0232, 0.0232, 0.0244, 0.0240, 0.0240, 0.0242], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:57:43,517 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=302693.0, num_to_drop=1, layers_to_drop={1} 2023-05-02 23:58:01,687 INFO [train.py:904] (4/8) Epoch 30, batch 8350, loss[loss=0.1547, simple_loss=0.2555, pruned_loss=0.02695, over 16618.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2759, pruned_loss=0.04598, over 3084832.77 frames. ], batch size: 57, lr: 2.22e-03, grad_scale: 8.0 2023-05-02 23:58:09,435 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.510e+02 2.283e+02 2.623e+02 3.174e+02 4.916e+02, threshold=5.246e+02, percent-clipped=0.0 2023-05-02 23:58:31,333 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7406, 4.9972, 4.7959, 4.8334, 4.5763, 4.5364, 4.3682, 5.0797], device='cuda:4'), covar=tensor([0.1082, 0.0815, 0.0968, 0.0818, 0.0775, 0.1181, 0.1290, 0.0803], device='cuda:4'), in_proj_covar=tensor([0.0726, 0.0873, 0.0718, 0.0681, 0.0555, 0.0557, 0.0731, 0.0685], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:58:52,574 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5195, 4.6257, 4.3743, 4.0566, 4.0803, 4.5104, 4.2420, 4.2233], device='cuda:4'), covar=tensor([0.0648, 0.0688, 0.0366, 0.0366, 0.0870, 0.0552, 0.0611, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0473, 0.0364, 0.0365, 0.0359, 0.0421, 0.0252, 0.0434], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:59:15,980 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.8638, 3.7665, 3.9377, 4.0289, 4.1085, 3.6830, 4.0737, 4.1445], device='cuda:4'), covar=tensor([0.1567, 0.1098, 0.1230, 0.0676, 0.0618, 0.2021, 0.0781, 0.0824], device='cuda:4'), in_proj_covar=tensor([0.0670, 0.0812, 0.0940, 0.0836, 0.0632, 0.0653, 0.0692, 0.0803], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-02 23:59:22,769 INFO [train.py:904] (4/8) Epoch 30, batch 8400, loss[loss=0.1767, simple_loss=0.2705, pruned_loss=0.04146, over 15440.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2735, pruned_loss=0.04423, over 3084712.17 frames. ], batch size: 191, lr: 2.22e-03, grad_scale: 8.0 2023-05-02 23:59:28,132 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.0280, 4.0195, 4.3260, 4.2885, 4.3114, 4.0951, 4.0707, 4.0952], device='cuda:4'), covar=tensor([0.0423, 0.0791, 0.0493, 0.0529, 0.0543, 0.0532, 0.0941, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0439, 0.0500, 0.0478, 0.0444, 0.0523, 0.0507, 0.0583, 0.0409], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-02 23:59:37,969 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.28 vs. limit=5.0 2023-05-03 00:00:21,872 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.1090, 3.9730, 4.1754, 4.2911, 4.4163, 3.9740, 4.3786, 4.4475], device='cuda:4'), covar=tensor([0.1795, 0.1266, 0.1464, 0.0754, 0.0574, 0.1427, 0.0747, 0.0797], device='cuda:4'), in_proj_covar=tensor([0.0668, 0.0810, 0.0937, 0.0833, 0.0630, 0.0652, 0.0690, 0.0802], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:00:44,411 INFO [train.py:904] (4/8) Epoch 30, batch 8450, loss[loss=0.1659, simple_loss=0.2565, pruned_loss=0.0377, over 16716.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2717, pruned_loss=0.04273, over 3088398.27 frames. ], batch size: 57, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:00:52,080 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.493e+02 2.019e+02 2.324e+02 2.808e+02 4.179e+02, threshold=4.647e+02, percent-clipped=0.0 2023-05-03 00:01:31,131 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=302833.0, num_to_drop=1, layers_to_drop={2} 2023-05-03 00:01:42,062 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7228, 4.7023, 4.5080, 3.7625, 4.5963, 1.7639, 4.3763, 4.2009], device='cuda:4'), covar=tensor([0.0115, 0.0110, 0.0231, 0.0372, 0.0124, 0.2965, 0.0144, 0.0309], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0174, 0.0213, 0.0184, 0.0190, 0.0217, 0.0201, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:02:04,570 INFO [train.py:904] (4/8) Epoch 30, batch 8500, loss[loss=0.152, simple_loss=0.2383, pruned_loss=0.03287, over 12041.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2675, pruned_loss=0.04041, over 3055316.83 frames. ], batch size: 246, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:02:08,475 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=302856.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:02:48,217 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=302881.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:02:55,760 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-05-03 00:03:07,416 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.5874, 4.3709, 4.6291, 4.7559, 4.9262, 4.5020, 4.9341, 4.9408], device='cuda:4'), covar=tensor([0.1927, 0.1472, 0.1842, 0.0888, 0.0665, 0.1090, 0.0650, 0.0864], device='cuda:4'), in_proj_covar=tensor([0.0666, 0.0808, 0.0935, 0.0832, 0.0630, 0.0651, 0.0689, 0.0800], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:03:13,611 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6628, 2.6454, 1.9267, 2.7864, 2.1709, 2.8295, 2.1841, 2.4558], device='cuda:4'), covar=tensor([0.0290, 0.0330, 0.1192, 0.0315, 0.0621, 0.0404, 0.1190, 0.0501], device='cuda:4'), in_proj_covar=tensor([0.0173, 0.0179, 0.0193, 0.0170, 0.0176, 0.0216, 0.0200, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-03 00:03:26,607 INFO [train.py:904] (4/8) Epoch 30, batch 8550, loss[loss=0.1764, simple_loss=0.2752, pruned_loss=0.03886, over 16769.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2658, pruned_loss=0.03981, over 3041686.45 frames. ], batch size: 83, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:03:27,650 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=302904.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:03:38,834 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.520e+02 2.175e+02 2.493e+02 2.958e+02 5.835e+02, threshold=4.986e+02, percent-clipped=2.0 2023-05-03 00:03:40,182 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1729, 2.5038, 2.5761, 2.0289, 2.7498, 2.7409, 2.5335, 2.5302], device='cuda:4'), covar=tensor([0.0631, 0.0264, 0.0263, 0.0898, 0.0137, 0.0301, 0.0451, 0.0414], device='cuda:4'), in_proj_covar=tensor([0.0148, 0.0110, 0.0101, 0.0137, 0.0086, 0.0132, 0.0130, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-03 00:04:00,400 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.66 vs. limit=5.0 2023-05-03 00:04:11,659 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-05-03 00:04:45,055 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=302942.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:04:56,150 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8774, 2.2906, 2.3484, 3.1492, 1.7686, 3.2675, 1.7594, 2.7318], device='cuda:4'), covar=tensor([0.1293, 0.0726, 0.1131, 0.0212, 0.0101, 0.0414, 0.1593, 0.0771], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0181, 0.0201, 0.0206, 0.0206, 0.0218, 0.0211, 0.0200], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-03 00:05:04,976 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=302952.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:05:08,027 INFO [train.py:904] (4/8) Epoch 30, batch 8600, loss[loss=0.1583, simple_loss=0.2607, pruned_loss=0.0279, over 16870.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.266, pruned_loss=0.03879, over 3032415.49 frames. ], batch size: 90, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:06:16,016 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=302988.0, num_to_drop=1, layers_to_drop={0} 2023-05-03 00:06:21,392 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-05-03 00:06:34,012 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1665, 3.1747, 1.9705, 3.4570, 2.3675, 3.4670, 2.0858, 2.6338], device='cuda:4'), covar=tensor([0.0377, 0.0494, 0.1743, 0.0338, 0.1001, 0.0649, 0.1733, 0.0891], device='cuda:4'), in_proj_covar=tensor([0.0175, 0.0180, 0.0194, 0.0171, 0.0178, 0.0218, 0.0202, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-03 00:06:39,317 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=303000.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:06:46,603 INFO [train.py:904] (4/8) Epoch 30, batch 8650, loss[loss=0.1748, simple_loss=0.2731, pruned_loss=0.03825, over 15317.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2645, pruned_loss=0.03775, over 3022560.16 frames. ], batch size: 191, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:06:58,831 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.356e+02 2.161e+02 2.643e+02 3.389e+02 5.700e+02, threshold=5.286e+02, percent-clipped=3.0 2023-05-03 00:07:01,455 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-05-03 00:08:31,137 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=303052.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:08:34,579 INFO [train.py:904] (4/8) Epoch 30, batch 8700, loss[loss=0.1613, simple_loss=0.2585, pruned_loss=0.0321, over 16693.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2621, pruned_loss=0.03682, over 3031896.09 frames. ], batch size: 134, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:10:14,430 INFO [train.py:904] (4/8) Epoch 30, batch 8750, loss[loss=0.1858, simple_loss=0.2872, pruned_loss=0.04225, over 15247.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2624, pruned_loss=0.03656, over 3038480.09 frames. ], batch size: 191, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:10:25,189 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.363e+02 2.109e+02 2.507e+02 3.042e+02 7.094e+02, threshold=5.015e+02, percent-clipped=1.0 2023-05-03 00:10:38,662 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=303113.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:11:23,584 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=303133.0, num_to_drop=1, layers_to_drop={0} 2023-05-03 00:11:48,208 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1110, 1.9004, 1.7321, 1.5833, 2.0335, 1.6855, 1.5758, 2.0198], device='cuda:4'), covar=tensor([0.0228, 0.0327, 0.0471, 0.0414, 0.0270, 0.0310, 0.0194, 0.0254], device='cuda:4'), in_proj_covar=tensor([0.0227, 0.0239, 0.0229, 0.0230, 0.0241, 0.0237, 0.0236, 0.0238], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:12:06,341 INFO [train.py:904] (4/8) Epoch 30, batch 8800, loss[loss=0.1779, simple_loss=0.2656, pruned_loss=0.04507, over 12748.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.261, pruned_loss=0.03575, over 3035812.12 frames. ], batch size: 248, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:12:42,024 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.8010, 1.4460, 1.7898, 1.7639, 1.9318, 1.9474, 1.7574, 1.8764], device='cuda:4'), covar=tensor([0.0393, 0.0552, 0.0322, 0.0473, 0.0449, 0.0282, 0.0564, 0.0198], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0200, 0.0189, 0.0193, 0.0211, 0.0168, 0.0206, 0.0169], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0002], device='cuda:4') 2023-05-03 00:13:03,254 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=303181.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:13:50,921 INFO [train.py:904] (4/8) Epoch 30, batch 8850, loss[loss=0.1438, simple_loss=0.2405, pruned_loss=0.02353, over 12325.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2637, pruned_loss=0.03537, over 3024514.82 frames. ], batch size: 246, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:14:00,695 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 2.249e+02 2.609e+02 3.163e+02 6.486e+02, threshold=5.217e+02, percent-clipped=3.0 2023-05-03 00:15:03,333 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=303237.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:15:38,682 INFO [train.py:904] (4/8) Epoch 30, batch 8900, loss[loss=0.1735, simple_loss=0.2739, pruned_loss=0.03656, over 12839.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.265, pruned_loss=0.03501, over 3050551.61 frames. ], batch size: 247, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:15:49,472 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=303258.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:15:58,402 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-05-03 00:17:03,911 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=303288.0, num_to_drop=1, layers_to_drop={0} 2023-05-03 00:17:32,990 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-05-03 00:17:43,147 INFO [train.py:904] (4/8) Epoch 30, batch 8950, loss[loss=0.1589, simple_loss=0.2561, pruned_loss=0.03091, over 16670.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2636, pruned_loss=0.0347, over 3070691.54 frames. ], batch size: 134, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:17:53,311 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 2.074e+02 2.389e+02 2.805e+02 4.954e+02, threshold=4.779e+02, percent-clipped=0.0 2023-05-03 00:18:16,177 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=303319.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:18:53,962 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=303336.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:19:32,192 INFO [train.py:904] (4/8) Epoch 30, batch 9000, loss[loss=0.1465, simple_loss=0.2415, pruned_loss=0.02568, over 16815.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2603, pruned_loss=0.03354, over 3060935.13 frames. ], batch size: 124, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:19:32,192 INFO [train.py:929] (4/8) Computing validation loss 2023-05-03 00:19:42,082 INFO [train.py:938] (4/8) Epoch 30, validation: loss=0.1431, simple_loss=0.2465, pruned_loss=0.01984, over 944034.00 frames. 2023-05-03 00:19:42,083 INFO [train.py:939] (4/8) Maximum memory allocated so far is 17768MB 2023-05-03 00:21:28,490 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=303403.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:21:29,164 INFO [train.py:904] (4/8) Epoch 30, batch 9050, loss[loss=0.1646, simple_loss=0.2589, pruned_loss=0.03517, over 16739.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2608, pruned_loss=0.03382, over 3077647.92 frames. ], batch size: 76, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:21:38,011 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=303408.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:21:39,555 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.357e+02 2.159e+02 2.466e+02 3.038e+02 8.070e+02, threshold=4.932e+02, percent-clipped=3.0 2023-05-03 00:22:15,663 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1916, 2.5761, 2.6724, 1.9309, 2.7778, 2.8121, 2.5652, 2.5630], device='cuda:4'), covar=tensor([0.0621, 0.0246, 0.0220, 0.1024, 0.0124, 0.0252, 0.0418, 0.0418], device='cuda:4'), in_proj_covar=tensor([0.0146, 0.0108, 0.0100, 0.0136, 0.0085, 0.0129, 0.0128, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-03 00:23:13,932 INFO [train.py:904] (4/8) Epoch 30, batch 9100, loss[loss=0.1706, simple_loss=0.2666, pruned_loss=0.03731, over 16700.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2601, pruned_loss=0.03414, over 3071106.58 frames. ], batch size: 134, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:23:33,671 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=303464.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:24:43,221 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.5667, 2.6411, 2.3863, 2.3292, 2.9010, 2.5429, 2.9198, 3.1171], device='cuda:4'), covar=tensor([0.0144, 0.0470, 0.0559, 0.0558, 0.0353, 0.0500, 0.0280, 0.0320], device='cuda:4'), in_proj_covar=tensor([0.0225, 0.0239, 0.0229, 0.0230, 0.0241, 0.0237, 0.0235, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:24:54,224 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=3.33 vs. limit=5.0 2023-05-03 00:24:59,457 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.88 vs. limit=5.0 2023-05-03 00:25:12,849 INFO [train.py:904] (4/8) Epoch 30, batch 9150, loss[loss=0.1475, simple_loss=0.2433, pruned_loss=0.02586, over 16618.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.261, pruned_loss=0.03406, over 3062241.29 frames. ], batch size: 68, lr: 2.22e-03, grad_scale: 4.0 2023-05-03 00:25:25,337 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 2.268e+02 2.639e+02 3.152e+02 6.290e+02, threshold=5.277e+02, percent-clipped=4.0 2023-05-03 00:25:55,426 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.2385, 3.3640, 3.3972, 2.2351, 3.1360, 3.4347, 3.2360, 2.0528], device='cuda:4'), covar=tensor([0.0582, 0.0082, 0.0081, 0.0487, 0.0151, 0.0118, 0.0106, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0089, 0.0091, 0.0134, 0.0102, 0.0114, 0.0097, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-03 00:26:23,064 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=303537.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:26:54,943 INFO [train.py:904] (4/8) Epoch 30, batch 9200, loss[loss=0.1764, simple_loss=0.2778, pruned_loss=0.03749, over 15361.00 frames. ], tot_loss[loss=0.1617, simple_loss=0.2571, pruned_loss=0.03319, over 3065176.75 frames. ], batch size: 191, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:27:08,071 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6261, 3.7033, 2.3727, 4.2211, 2.9079, 4.1369, 2.5511, 3.0973], device='cuda:4'), covar=tensor([0.0304, 0.0399, 0.1540, 0.0401, 0.0825, 0.0560, 0.1483, 0.0777], device='cuda:4'), in_proj_covar=tensor([0.0174, 0.0178, 0.0192, 0.0168, 0.0176, 0.0215, 0.0201, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-03 00:27:31,974 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=303574.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:27:52,353 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=303585.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:28:27,174 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.9059, 2.1839, 2.4288, 3.1983, 2.2279, 2.3375, 2.3504, 2.2943], device='cuda:4'), covar=tensor([0.1427, 0.4184, 0.3137, 0.0805, 0.4665, 0.3094, 0.4034, 0.4055], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0471, 0.0386, 0.0332, 0.0441, 0.0540, 0.0445, 0.0551], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:28:29,157 INFO [train.py:904] (4/8) Epoch 30, batch 9250, loss[loss=0.1528, simple_loss=0.25, pruned_loss=0.02781, over 15301.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2567, pruned_loss=0.03312, over 3076084.56 frames. ], batch size: 191, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:28:41,899 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.600e+02 2.148e+02 2.583e+02 3.115e+02 7.022e+02, threshold=5.165e+02, percent-clipped=1.0 2023-05-03 00:28:49,554 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=303614.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:29:38,977 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9855, 2.8334, 2.6910, 2.0236, 2.6070, 2.8681, 2.7269, 1.9676], device='cuda:4'), covar=tensor([0.0451, 0.0094, 0.0090, 0.0384, 0.0161, 0.0116, 0.0111, 0.0456], device='cuda:4'), in_proj_covar=tensor([0.0138, 0.0090, 0.0091, 0.0134, 0.0103, 0.0114, 0.0097, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-03 00:29:39,001 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=303635.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:30:19,432 INFO [train.py:904] (4/8) Epoch 30, batch 9300, loss[loss=0.1498, simple_loss=0.2413, pruned_loss=0.02911, over 17034.00 frames. ], tot_loss[loss=0.1598, simple_loss=0.2548, pruned_loss=0.03243, over 3049708.93 frames. ], batch size: 50, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:30:25,038 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.7023, 3.8962, 3.9806, 2.8651, 3.5367, 3.9901, 3.6530, 2.3179], device='cuda:4'), covar=tensor([0.0472, 0.0061, 0.0050, 0.0363, 0.0133, 0.0093, 0.0087, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0089, 0.0091, 0.0134, 0.0102, 0.0114, 0.0097, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-03 00:31:05,525 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.6413, 4.6576, 4.9923, 4.9784, 4.9745, 4.7417, 4.6881, 4.6429], device='cuda:4'), covar=tensor([0.0398, 0.0891, 0.0529, 0.0553, 0.0720, 0.0517, 0.0948, 0.0493], device='cuda:4'), in_proj_covar=tensor([0.0432, 0.0491, 0.0471, 0.0437, 0.0516, 0.0498, 0.0570, 0.0402], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-03 00:31:53,185 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.8742, 4.7491, 4.9416, 5.0773, 5.2875, 4.7371, 5.2885, 5.3151], device='cuda:4'), covar=tensor([0.2073, 0.1328, 0.1792, 0.0833, 0.0541, 0.0886, 0.0622, 0.0693], device='cuda:4'), in_proj_covar=tensor([0.0657, 0.0796, 0.0919, 0.0824, 0.0621, 0.0645, 0.0682, 0.0789], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:32:05,583 INFO [train.py:904] (4/8) Epoch 30, batch 9350, loss[loss=0.1803, simple_loss=0.2693, pruned_loss=0.04563, over 16983.00 frames. ], tot_loss[loss=0.1601, simple_loss=0.2551, pruned_loss=0.03257, over 3078878.10 frames. ], batch size: 109, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:32:13,813 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=303708.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:32:16,690 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.429e+02 2.024e+02 2.345e+02 2.902e+02 6.854e+02, threshold=4.690e+02, percent-clipped=2.0 2023-05-03 00:33:46,246 INFO [train.py:904] (4/8) Epoch 30, batch 9400, loss[loss=0.1464, simple_loss=0.238, pruned_loss=0.02743, over 12388.00 frames. ], tot_loss[loss=0.1598, simple_loss=0.2548, pruned_loss=0.03235, over 3062727.34 frames. ], batch size: 247, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:33:50,774 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=303756.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:33:56,165 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=303759.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:35:14,451 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.4278, 4.5443, 4.7262, 4.4867, 4.6055, 5.0618, 4.6365, 4.3560], device='cuda:4'), covar=tensor([0.1448, 0.1877, 0.2045, 0.1951, 0.2347, 0.0988, 0.1466, 0.2287], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0623, 0.0699, 0.0507, 0.0673, 0.0719, 0.0540, 0.0677], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-05-03 00:35:25,718 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8268, 3.5113, 3.8712, 2.0230, 4.0004, 4.0670, 3.1788, 3.2006], device='cuda:4'), covar=tensor([0.0642, 0.0268, 0.0212, 0.1181, 0.0084, 0.0163, 0.0430, 0.0406], device='cuda:4'), in_proj_covar=tensor([0.0144, 0.0107, 0.0098, 0.0134, 0.0084, 0.0128, 0.0127, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-05-03 00:35:26,335 INFO [train.py:904] (4/8) Epoch 30, batch 9450, loss[loss=0.1413, simple_loss=0.2459, pruned_loss=0.01841, over 16925.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2567, pruned_loss=0.03242, over 3064534.37 frames. ], batch size: 96, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:35:29,343 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.8971, 2.1766, 2.4534, 3.2103, 2.2151, 2.3741, 2.3596, 2.3208], device='cuda:4'), covar=tensor([0.1521, 0.3992, 0.2947, 0.0801, 0.4702, 0.2941, 0.3992, 0.3950], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0471, 0.0386, 0.0332, 0.0441, 0.0540, 0.0445, 0.0550], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:35:36,997 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.278e+02 2.120e+02 2.504e+02 3.073e+02 7.133e+02, threshold=5.009e+02, percent-clipped=4.0 2023-05-03 00:35:38,598 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=303810.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:35:59,012 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-05-03 00:37:08,345 INFO [train.py:904] (4/8) Epoch 30, batch 9500, loss[loss=0.1553, simple_loss=0.2553, pruned_loss=0.02765, over 16202.00 frames. ], tot_loss[loss=0.1603, simple_loss=0.2562, pruned_loss=0.0322, over 3066986.91 frames. ], batch size: 165, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:37:43,990 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=303871.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:38:36,435 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=303896.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:38:54,494 INFO [train.py:904] (4/8) Epoch 30, batch 9550, loss[loss=0.1964, simple_loss=0.2954, pruned_loss=0.04875, over 15253.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2562, pruned_loss=0.03265, over 3070705.95 frames. ], batch size: 190, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:39:08,517 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 2.050e+02 2.496e+02 3.159e+02 5.329e+02, threshold=4.992e+02, percent-clipped=3.0 2023-05-03 00:39:18,191 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=303914.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:39:49,601 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=303930.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:40:36,097 INFO [train.py:904] (4/8) Epoch 30, batch 9600, loss[loss=0.1676, simple_loss=0.2761, pruned_loss=0.0296, over 15418.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2579, pruned_loss=0.03338, over 3063641.38 frames. ], batch size: 191, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:40:43,422 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=303957.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:40:53,326 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=303962.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:41:16,237 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.7319, 4.7286, 4.5001, 3.8046, 4.5929, 1.5912, 4.3870, 4.1988], device='cuda:4'), covar=tensor([0.0094, 0.0089, 0.0225, 0.0322, 0.0128, 0.3153, 0.0135, 0.0320], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0173, 0.0211, 0.0181, 0.0188, 0.0217, 0.0199, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:41:16,259 INFO [zipformer.py:625] (4/8) warmup_begin=2666.7, warmup_end=3333.3, batch_count=303974.0, num_to_drop=1, layers_to_drop={1} 2023-05-03 00:42:24,230 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([5.7328, 6.0604, 5.8116, 5.8706, 5.4955, 5.4753, 5.4125, 6.1731], device='cuda:4'), covar=tensor([0.1358, 0.1015, 0.1081, 0.0842, 0.0791, 0.0664, 0.1290, 0.0843], device='cuda:4'), in_proj_covar=tensor([0.0707, 0.0850, 0.0699, 0.0663, 0.0542, 0.0542, 0.0712, 0.0668], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:42:31,192 INFO [train.py:904] (4/8) Epoch 30, batch 9650, loss[loss=0.153, simple_loss=0.2452, pruned_loss=0.03043, over 12054.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2601, pruned_loss=0.03379, over 3061293.38 frames. ], batch size: 247, lr: 2.22e-03, grad_scale: 4.0 2023-05-03 00:42:48,096 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.484e+02 2.176e+02 2.625e+02 3.264e+02 9.247e+02, threshold=5.250e+02, percent-clipped=2.0 2023-05-03 00:43:41,195 INFO [zipformer.py:625] (4/8) warmup_begin=3333.3, warmup_end=4000.0, batch_count=304035.0, num_to_drop=1, layers_to_drop={2} 2023-05-03 00:44:20,836 INFO [train.py:904] (4/8) Epoch 30, batch 9700, loss[loss=0.1732, simple_loss=0.2673, pruned_loss=0.03956, over 16943.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2587, pruned_loss=0.03316, over 3083959.78 frames. ], batch size: 109, lr: 2.22e-03, grad_scale: 4.0 2023-05-03 00:44:30,046 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=304059.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:46:04,096 INFO [train.py:904] (4/8) Epoch 30, batch 9750, loss[loss=0.1599, simple_loss=0.2568, pruned_loss=0.03145, over 16502.00 frames. ], tot_loss[loss=0.1619, simple_loss=0.2574, pruned_loss=0.03321, over 3092231.84 frames. ], batch size: 68, lr: 2.22e-03, grad_scale: 4.0 2023-05-03 00:46:09,339 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=304107.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:46:17,048 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.422e+02 2.052e+02 2.422e+02 2.861e+02 5.308e+02, threshold=4.844e+02, percent-clipped=2.0 2023-05-03 00:47:41,835 INFO [train.py:904] (4/8) Epoch 30, batch 9800, loss[loss=0.1454, simple_loss=0.2368, pruned_loss=0.02697, over 12244.00 frames. ], tot_loss[loss=0.1612, simple_loss=0.2575, pruned_loss=0.03245, over 3104668.57 frames. ], batch size: 247, lr: 2.22e-03, grad_scale: 4.0 2023-05-03 00:48:04,642 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=304166.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:48:38,219 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.1533, 3.2372, 3.2543, 2.2310, 2.9352, 3.2753, 3.1436, 1.9812], device='cuda:4'), covar=tensor([0.0537, 0.0072, 0.0076, 0.0430, 0.0162, 0.0099, 0.0096, 0.0513], device='cuda:4'), in_proj_covar=tensor([0.0137, 0.0090, 0.0091, 0.0134, 0.0102, 0.0114, 0.0097, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0002, 0.0002, 0.0004, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-05-03 00:48:53,974 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([4.2762, 4.2885, 4.6175, 4.5700, 4.6148, 4.3672, 4.3121, 4.3198], device='cuda:4'), covar=tensor([0.0403, 0.0810, 0.0488, 0.0495, 0.0465, 0.0543, 0.1003, 0.0499], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0488, 0.0469, 0.0435, 0.0513, 0.0495, 0.0566, 0.0399], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0003], device='cuda:4') 2023-05-03 00:49:25,007 INFO [train.py:904] (4/8) Epoch 30, batch 9850, loss[loss=0.1613, simple_loss=0.2598, pruned_loss=0.03142, over 16439.00 frames. ], tot_loss[loss=0.1618, simple_loss=0.2587, pruned_loss=0.03247, over 3092961.37 frames. ], batch size: 146, lr: 2.22e-03, grad_scale: 4.0 2023-05-03 00:49:39,351 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.408e+02 2.072e+02 2.330e+02 2.794e+02 5.653e+02, threshold=4.660e+02, percent-clipped=2.0 2023-05-03 00:50:19,586 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=304230.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:51:13,301 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=304252.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:51:16,009 INFO [train.py:904] (4/8) Epoch 30, batch 9900, loss[loss=0.1668, simple_loss=0.2752, pruned_loss=0.0292, over 16764.00 frames. ], tot_loss[loss=0.1617, simple_loss=0.2589, pruned_loss=0.03222, over 3083757.82 frames. ], batch size: 83, lr: 2.22e-03, grad_scale: 4.0 2023-05-03 00:52:13,220 INFO [zipformer.py:625] (4/8) warmup_begin=666.7, warmup_end=1333.3, batch_count=304278.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:52:42,632 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.4679, 3.4042, 3.5182, 3.5915, 3.6408, 3.3492, 3.6250, 3.6963], device='cuda:4'), covar=tensor([0.1502, 0.1042, 0.1172, 0.0803, 0.0713, 0.2300, 0.1044, 0.0892], device='cuda:4'), in_proj_covar=tensor([0.0650, 0.0786, 0.0904, 0.0815, 0.0613, 0.0636, 0.0675, 0.0781], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:52:56,006 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.1440, 2.3156, 2.1358, 2.1538, 2.6978, 2.3787, 2.5600, 2.8502], device='cuda:4'), covar=tensor([0.0186, 0.0535, 0.0583, 0.0581, 0.0325, 0.0494, 0.0281, 0.0304], device='cuda:4'), in_proj_covar=tensor([0.0225, 0.0239, 0.0229, 0.0230, 0.0241, 0.0238, 0.0233, 0.0236], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:52:59,570 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([3.6560, 2.6059, 2.4622, 2.4612, 3.0418, 2.7561, 3.0559, 3.1970], device='cuda:4'), covar=tensor([0.0140, 0.0511, 0.0564, 0.0537, 0.0289, 0.0444, 0.0338, 0.0289], device='cuda:4'), in_proj_covar=tensor([0.0225, 0.0239, 0.0229, 0.0230, 0.0241, 0.0238, 0.0233, 0.0236], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-05-03 00:53:14,777 INFO [train.py:904] (4/8) Epoch 30, batch 9950, loss[loss=0.1759, simple_loss=0.2743, pruned_loss=0.03881, over 16271.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2614, pruned_loss=0.03246, over 3097551.11 frames. ], batch size: 165, lr: 2.22e-03, grad_scale: 4.0 2023-05-03 00:53:31,713 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.284e+02 2.067e+02 2.374e+02 2.801e+02 5.953e+02, threshold=4.748e+02, percent-clipped=1.0 2023-05-03 00:53:50,470 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=2.70 vs. limit=5.0 2023-05-03 00:54:21,392 INFO [zipformer.py:625] (4/8) warmup_begin=1333.3, warmup_end=2000.0, batch_count=304330.0, num_to_drop=1, layers_to_drop={2} 2023-05-03 00:54:37,534 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([1.9229, 2.2882, 2.2293, 3.1859, 1.7996, 3.2669, 1.7702, 2.6844], device='cuda:4'), covar=tensor([0.1455, 0.0795, 0.1251, 0.0180, 0.0088, 0.0331, 0.1796, 0.0773], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0178, 0.0197, 0.0200, 0.0199, 0.0213, 0.0208, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-03 00:55:15,632 INFO [train.py:904] (4/8) Epoch 30, batch 10000, loss[loss=0.1728, simple_loss=0.2686, pruned_loss=0.03854, over 16458.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2602, pruned_loss=0.0323, over 3100937.88 frames. ], batch size: 147, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:56:00,045 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=192, metric=1.48 vs. limit=2.0 2023-05-03 00:56:57,637 INFO [train.py:904] (4/8) Epoch 30, batch 10050, loss[loss=0.162, simple_loss=0.2658, pruned_loss=0.02917, over 16708.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2606, pruned_loss=0.03272, over 3102596.40 frames. ], batch size: 83, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:57:10,557 INFO [optim.py:368] (4/8) Clipping_scale=2.0, grad-norm quartiles 1.432e+02 2.064e+02 2.395e+02 2.910e+02 6.626e+02, threshold=4.790e+02, percent-clipped=3.0 2023-05-03 00:57:13,612 INFO [zipformer.py:1454] (4/8) attn_weights_entropy = tensor([2.6820, 2.6328, 1.9403, 2.7918, 2.0851, 2.8511, 2.1611, 2.4171], device='cuda:4'), covar=tensor([0.0374, 0.0415, 0.1359, 0.0303, 0.0793, 0.0554, 0.1278, 0.0681], device='cuda:4'), in_proj_covar=tensor([0.0172, 0.0176, 0.0190, 0.0167, 0.0175, 0.0212, 0.0199, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-05-03 00:57:17,970 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-05-03 00:57:54,758 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-05-03 00:58:05,156 INFO [scaling.py:679] (4/8) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-05-03 00:58:34,282 INFO [train.py:904] (4/8) Epoch 30, batch 10100, loss[loss=0.165, simple_loss=0.2419, pruned_loss=0.04409, over 12409.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2604, pruned_loss=0.03295, over 3091403.58 frames. ], batch size: 248, lr: 2.22e-03, grad_scale: 8.0 2023-05-03 00:58:57,144 INFO [zipformer.py:625] (4/8) warmup_begin=2000.0, warmup_end=2666.7, batch_count=304466.0, num_to_drop=0, layers_to_drop=set() 2023-05-03 00:59:20,377 INFO [scaling.py:679] (4/8) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-05-03 00:59:55,822 INFO [train.py:1169] (4/8) Done!